linux_dsm_epyc7002/drivers/gpu/drm/i915/intel_ringbuffer.c

2798 lines
75 KiB
C
Raw Normal View History

/*
* Copyright © 2008-2010 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Authors:
* Eric Anholt <eric@anholt.net>
* Zou Nan hai <nanhai.zou@intel.com>
* Xiang Hai hao<haihao.xiang@intel.com>
*
*/
#include <linux/log2.h>
#include <drm/drmP.h>
#include "i915_drv.h"
#include <drm/i915_drm.h>
#include "i915_trace.h"
#include "intel_drv.h"
/* Rough estimate of the typical request size, performing a flush,
* set-context and then emitting the batch.
*/
#define LEGACY_REQUEST_SIZE 200
int __intel_ring_space(int head, int tail, int size)
{
int space = head - tail;
if (space <= 0)
space += size;
return space - I915_RING_FREE_SPACE;
}
void intel_ring_update_space(struct intel_ring *ring)
{
if (ring->last_retired_head != -1) {
ring->head = ring->last_retired_head;
ring->last_retired_head = -1;
}
ring->space = __intel_ring_space(ring->head & HEAD_ADDR,
ring->tail, ring->size);
}
static int
gen2_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
u32 cmd;
int ret;
cmd = MI_FLUSH;
if (mode & EMIT_INVALIDATE)
cmd |= MI_READ_FLUSH;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring, cmd);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static int
gen4_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
u32 cmd;
int ret;
/*
* read/write caches:
*
* I915_GEM_DOMAIN_RENDER is always invalidated, but is
* only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
* also flushed at 2d versus 3d pipeline switches.
*
* read-only caches:
*
* I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
* MI_READ_FLUSH is set, and is always flushed on 965.
*
* I915_GEM_DOMAIN_COMMAND may not exist?
*
* I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
* invalidated when MI_EXE_FLUSH is set.
*
* I915_GEM_DOMAIN_VERTEX, which exists on 965, is
* invalidated with every MI_FLUSH.
*
* TLBs:
*
* On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
* and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
* I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
* are flushed at any MI_FLUSH.
*/
cmd = MI_FLUSH;
if (mode & EMIT_INVALIDATE) {
cmd |= MI_EXE_FLUSH;
if (IS_G4X(req->i915) || IS_GEN5(req->i915))
cmd |= MI_INVALIDATE_ISP;
}
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring, cmd);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
/**
* Emits a PIPE_CONTROL with a non-zero post-sync operation, for
* implementing two workarounds on gen6. From section 1.4.7.1
* "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
*
* [DevSNB-C+{W/A}] Before any depth stall flush (including those
* produced by non-pipelined state commands), software needs to first
* send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
* 0.
*
* [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
* =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
*
* And the workaround for these two requires this workaround first:
*
* [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
* BEFORE the pipe-control with a post-sync op and no write-cache
* flushes.
*
* And this last workaround is tricky because of the requirements on
* that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
* volume 2 part 1:
*
* "1 of the following must also be set:
* - Render Target Cache Flush Enable ([12] of DW1)
* - Depth Cache Flush Enable ([0] of DW1)
* - Stall at Pixel Scoreboard ([1] of DW1)
* - Depth Stall ([13] of DW1)
* - Post-Sync Operation ([13] of DW1)
* - Notify Enable ([8] of DW1)"
*
* The cache flushes require the workaround flush that triggered this
* one, so we can't use it. Depth stall would trigger the same.
* Post-sync nonzero is what triggered this second workaround, so we
* can't use that one either. Notify enable is IRQs, which aren't
* really our business. That leaves only stall at scoreboard.
*/
static int
intel_emit_post_sync_nonzero_flush(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
u32 scratch_addr =
i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
int ret;
ret = intel_ring_begin(req, 6);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
PIPE_CONTROL_STALL_AT_SCOREBOARD);
intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
intel_ring_emit(ring, 0); /* low dword */
intel_ring_emit(ring, 0); /* high dword */
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
ret = intel_ring_begin(req, 6);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE);
intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static int
gen6_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
u32 scratch_addr =
i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
u32 flags = 0;
int ret;
/* Force SNB workarounds for PIPE_CONTROL flushes */
ret = intel_emit_post_sync_nonzero_flush(req);
if (ret)
return ret;
/* Just flush everything. Experiments have shown that reducing the
* number of bits based on the write domains has little performance
* impact.
*/
if (mode & EMIT_FLUSH) {
flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
/*
* Ensure that any following seqno writes only happen
* when the render cache is indeed flushed.
*/
flags |= PIPE_CONTROL_CS_STALL;
}
if (mode & EMIT_INVALIDATE) {
flags |= PIPE_CONTROL_TLB_INVALIDATE;
flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
/*
* TLB invalidate requires a post-sync write.
*/
flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
}
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
intel_ring_emit(ring, flags);
intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
intel_ring_emit(ring, 0);
intel_ring_advance(ring);
return 0;
}
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
static int
gen7_render_ring_cs_stall_wa(struct drm_i915_gem_request *req)
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
{
struct intel_ring *ring = req->ring;
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
int ret;
ret = intel_ring_begin(req, 4);
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
intel_ring_emit(ring,
PIPE_CONTROL_CS_STALL |
PIPE_CONTROL_STALL_AT_SCOREBOARD);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, 0);
intel_ring_advance(ring);
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
return 0;
}
static int
gen7_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
u32 scratch_addr =
i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
u32 flags = 0;
int ret;
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
/*
* Ensure that any following seqno writes only happen when the render
* cache is indeed flushed.
*
* Workaround: 4th PIPE_CONTROL command (except the ones with only
* read-cache invalidate bits set) must have the CS_STALL bit set. We
* don't try to be clever and just set it unconditionally.
*/
flags |= PIPE_CONTROL_CS_STALL;
/* Just flush everything. Experiments have shown that reducing the
* number of bits based on the write domains has little performance
* impact.
*/
if (mode & EMIT_FLUSH) {
flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
flags |= PIPE_CONTROL_FLUSH_ENABLE;
}
if (mode & EMIT_INVALIDATE) {
flags |= PIPE_CONTROL_TLB_INVALIDATE;
flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
/*
* TLB invalidate requires a post-sync write.
*/
flags |= PIPE_CONTROL_QW_WRITE;
flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
drm/i915: add workarounds to gen7_render_ring_flush From Bspec, Vol 2a, Section 1.9.3.4 "PIPE_CONTROL", intro section detailing the various workarounds: "[DevIVB {W/A}, DevHSW {W/A}]: Pipe_control with CS-stall bit set must be issued before a pipe-control command that has the State Cache Invalidate bit set." Note that public Bspec has different numbering, it's Vol2Part1, Section 1.10.4.1 "PIPE_CONTROL" there. There's also a second workaround for the PIPE_CONTROL command itself: "[DevIVB, DevVLV, DevHSW] {WA}: Every 4th PIPE_CONTROL command, not counting the PIPE_CONTROL with only read-cache-invalidate bit(s) set, must have a CS_STALL bit set" For simplicity we simply set the CS_STALL bit on every pipe_control on gen7+ Note that this massively helps on some hsw machines, together with the following patch to unconditionally set the CS_STALL bit on every pipe_control it prevents a gpu hang every few seconds. This is a regression that has been introduced in the pipe_control cleanup: commit 6c6cf5aa9c583478b19e23149feaa92d01fb8c2d Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Jul 20 18:02:28 2012 +0100 drm/i915: Only apply the SNB pipe control w/a to gen6 It looks like the massive snb pipe_control workaround also papered over any issues on ivb and hsw. Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> [danvet: squashed both workarounds together, pimped commit message with Bsepc citations, regression commit citation and changed the comment in the code a bit to clarify that we unconditionally set CS_STALL to avoid being hurt by trying to be clever.] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-08-18 04:35:43 +07:00
/* Workaround: we must issue a pipe_control with CS-stall bit
* set before a pipe_control command that has the state cache
* invalidate bit set. */
gen7_render_ring_cs_stall_wa(req);
}
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
intel_ring_emit(ring, flags);
intel_ring_emit(ring, scratch_addr);
intel_ring_emit(ring, 0);
intel_ring_advance(ring);
return 0;
}
static int
gen8_emit_pipe_control(struct drm_i915_gem_request *req,
u32 flags, u32 scratch_addr)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 6);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
intel_ring_emit(ring, flags);
intel_ring_emit(ring, scratch_addr);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, 0);
intel_ring_advance(ring);
return 0;
}
static int
gen8_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
u32 scratch_addr =
i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
u32 flags = 0;
int ret;
flags |= PIPE_CONTROL_CS_STALL;
if (mode & EMIT_FLUSH) {
flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
flags |= PIPE_CONTROL_FLUSH_ENABLE;
}
if (mode & EMIT_INVALIDATE) {
flags |= PIPE_CONTROL_TLB_INVALIDATE;
flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
flags |= PIPE_CONTROL_QW_WRITE;
flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
/* WaCsStallBeforeStateCacheInvalidate:bdw,chv */
ret = gen8_emit_pipe_control(req,
PIPE_CONTROL_CS_STALL |
PIPE_CONTROL_STALL_AT_SCOREBOARD,
0);
if (ret)
return ret;
}
return gen8_emit_pipe_control(req, flags, scratch_addr);
}
u64 intel_engine_get_active_head(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
u64 acthd;
if (INTEL_GEN(dev_priv) >= 8)
acthd = I915_READ64_2x32(RING_ACTHD(engine->mmio_base),
RING_ACTHD_UDW(engine->mmio_base));
else if (INTEL_GEN(dev_priv) >= 4)
acthd = I915_READ(RING_ACTHD(engine->mmio_base));
else
acthd = I915_READ(ACTHD);
return acthd;
}
static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
u32 addr;
addr = dev_priv->status_page_dmah->busaddr;
if (INTEL_GEN(dev_priv) >= 4)
addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
I915_WRITE(HWS_PGA, addr);
}
static void intel_ring_setup_status_page(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
drm/i915: Type safe register read/write Make I915_READ and I915_WRITE more type safe by wrapping the register offset in a struct. This should eliminate most of the fumbles we've had with misplaced parens. This only takes care of normal mmio registers. We could extend the idea to other register types and define each with its own struct. That way you wouldn't be able to accidentally pass the wrong thing to a specific register access function. The gpio_reg setup is probably the ugliest thing left. But I figure I'd just leave it for now, and wait for some divine inspiration to strike before making it nice. As for the generated code, it's actually a bit better sometimes. Eg. looking at i915_irq_handler(), we can see the following change: lea 0x70024(%rdx,%rax,1),%r9d mov $0x1,%edx - movslq %r9d,%r9 - mov %r9,%rsi - mov %r9,-0x58(%rbp) - callq *0xd8(%rbx) + mov %r9d,%esi + mov %r9d,-0x48(%rbp) callq *0xd8(%rbx) So previously gcc thought the register offset might be signed and decided to sign extend it, just in case. The rest appears to be mostly just minor shuffling of instructions. v2: i915_mmio_reg_{offset,equal,valid}() helpers added s/_REG/_MMIO/ in the register defines mo more switch statements left to worry about ring_emit stuff got sorted in a prep patch cmd parser, lrc context and w/a batch buildup also in prep patch vgpu stuff cleaned up and moved to a prep patch all other unrelated changes split out v3: Rebased due to BXT DSI/BLC, MOCS, etc. v4: Rebased due to churn, s/i915_mmio_reg_t/i915_reg_t/ Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Link: http://patchwork.freedesktop.org/patch/msgid/1447853606-2751-1-git-send-email-ville.syrjala@linux.intel.com
2015-11-18 20:33:26 +07:00
i915_reg_t mmio;
/* The ring status page addresses are no longer next to the rest of
* the ring registers as of gen7.
*/
if (IS_GEN7(dev_priv)) {
switch (engine->id) {
case RCS:
mmio = RENDER_HWS_PGA_GEN7;
break;
case BCS:
mmio = BLT_HWS_PGA_GEN7;
break;
/*
* VCS2 actually doesn't exist on Gen7. Only shut up
* gcc switch check warning
*/
case VCS2:
case VCS:
mmio = BSD_HWS_PGA_GEN7;
break;
case VECS:
mmio = VEBOX_HWS_PGA_GEN7;
break;
}
} else if (IS_GEN6(dev_priv)) {
mmio = RING_HWS_PGA_GEN6(engine->mmio_base);
} else {
/* XXX: gen8 returns to sanity */
mmio = RING_HWS_PGA(engine->mmio_base);
}
I915_WRITE(mmio, engine->status_page.ggtt_offset);
POSTING_READ(mmio);
/*
* Flush the TLB for this page
*
* FIXME: These two bits have disappeared on gen8, so a question
* arises: do we still need this and if so how should we go about
* invalidating the TLB?
*/
if (IS_GEN(dev_priv, 6, 7)) {
i915_reg_t reg = RING_INSTPM(engine->mmio_base);
/* ring should be idle before issuing a sync flush*/
WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
I915_WRITE(reg,
_MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
INSTPM_SYNC_FLUSH));
if (intel_wait_for_register(dev_priv,
reg, INSTPM_SYNC_FLUSH, 0,
1000))
DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
engine->name);
}
}
static bool stop_ring(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
if (INTEL_GEN(dev_priv) > 2) {
I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
if (intel_wait_for_register(dev_priv,
RING_MI_MODE(engine->mmio_base),
MODE_IDLE,
MODE_IDLE,
1000)) {
DRM_ERROR("%s : timed out trying to stop ring\n",
engine->name);
/* Sometimes we observe that the idle flag is not
* set even though the ring is empty. So double
* check before giving up.
*/
if (I915_READ_HEAD(engine) != I915_READ_TAIL(engine))
return false;
}
}
I915_WRITE_CTL(engine, 0);
I915_WRITE_HEAD(engine, 0);
I915_WRITE_TAIL(engine, 0);
if (INTEL_GEN(dev_priv) > 2) {
(void)I915_READ_CTL(engine);
I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
}
return (I915_READ_HEAD(engine) & HEAD_ADDR) == 0;
}
static int init_ring_common(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
struct intel_ring *ring = engine->buffer;
int ret = 0;
intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
if (!stop_ring(engine)) {
/* G45 ring initialization often fails to reset head to zero */
DRM_DEBUG_KMS("%s head not reset to zero "
"ctl %08x head %08x tail %08x start %08x\n",
engine->name,
I915_READ_CTL(engine),
I915_READ_HEAD(engine),
I915_READ_TAIL(engine),
I915_READ_START(engine));
if (!stop_ring(engine)) {
DRM_ERROR("failed to set %s head to zero "
"ctl %08x head %08x tail %08x start %08x\n",
engine->name,
I915_READ_CTL(engine),
I915_READ_HEAD(engine),
I915_READ_TAIL(engine),
I915_READ_START(engine));
ret = -EIO;
goto out;
}
}
if (HWS_NEEDS_PHYSICAL(dev_priv))
ring_setup_phys_status_page(engine);
else
intel_ring_setup_status_page(engine);
/* Enforce ordering by reading HEAD register back */
I915_READ_HEAD(engine);
/* Initialize the ring. This must happen _after_ we've cleared the ring
* registers with the above sequence (the readback of the HEAD registers
* also enforces ordering), otherwise the hw might lose the new ring
* register values. */
I915_WRITE_START(engine, i915_ggtt_offset(ring->vma));
/* WaClearRingBufHeadRegAtInit:ctg,elk */
if (I915_READ_HEAD(engine))
DRM_DEBUG("%s initialization failed [head=%08x], fudging\n",
engine->name, I915_READ_HEAD(engine));
I915_WRITE_HEAD(engine, 0);
(void)I915_READ_HEAD(engine);
I915_WRITE_CTL(engine,
((ring->size - PAGE_SIZE) & RING_NR_PAGES)
drm/i915: Remove use of the autoreported ringbuffer HEAD position This is a revert of 6aa56062eaba67adfb247cded244fd877329588d. This was originally introduced to workaround reads of the ringbuffer registers returning 0 on SandyBridge causing hangs due to ringbuffer overflow. The root cause here was reads through the GT powerwell require the forcewake dance, something we only learnt of later. Now it appears that reading the reported head position from the HWS is returning garbage, leading once again to hangs. For example, on q35 the autoreported head reports: [ 217.975608] head now 00010000, actual 00010000 [ 436.725613] head now 00200000, actual 00200000 [ 462.956033] head now 00210000, actual 00210010 [ 485.501409] head now 00400000, actual 00400020 [ 508.064280] head now 00410000, actual 00410000 [ 530.576078] head now 00600000, actual 00600020 [ 553.273489] head now 00610000, actual 00610018 which appears reasonably sane. In contrast, if we look at snb: [ 141.970680] head now 00e10000, actual 00008238 [ 141.974062] head now 02734000, actual 000083c8 [ 141.974425] head now 00e10000, actual 00008488 [ 141.980374] head now 032b5000, actual 000088b8 [ 141.980885] head now 03271000, actual 00008950 [ 142.040628] head now 02101000, actual 00008b40 [ 142.180173] head now 02734000, actual 00009050 [ 142.181090] head now 00000000, actual 00000ae0 [ 142.183737] head now 02734000, actual 00009050 In addition, the automatic reporting of the head position is scheduled to be defeatured in the future. It has no more utility, remove it. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=45492 Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch> Tested-by: Eric Anholt <eric@anholt.net> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
2012-02-08 20:34:13 +07:00
| RING_VALID);
/* If the head is still not zero, the ring is dead */
if (wait_for((I915_READ_CTL(engine) & RING_VALID) != 0 &&
I915_READ_START(engine) == i915_ggtt_offset(ring->vma) &&
(I915_READ_HEAD(engine) & HEAD_ADDR) == 0, 50)) {
DRM_ERROR("%s initialization failed "
"ctl %08x (valid? %d) head %08x tail %08x start %08x [expected %08x]\n",
engine->name,
I915_READ_CTL(engine),
I915_READ_CTL(engine) & RING_VALID,
I915_READ_HEAD(engine), I915_READ_TAIL(engine),
I915_READ_START(engine),
i915_ggtt_offset(ring->vma));
ret = -EIO;
goto out;
}
ring->last_retired_head = -1;
ring->head = I915_READ_HEAD(engine);
ring->tail = I915_READ_TAIL(engine) & TAIL_ADDR;
intel_ring_update_space(ring);
intel_engine_init_hangcheck(engine);
out:
intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
return ret;
}
static int intel_ring_workarounds_emit(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
struct i915_workarounds *w = &req->i915->workarounds;
int ret, i;
if (w->count == 0)
return 0;
ret = req->engine->emit_flush(req, EMIT_BARRIER);
if (ret)
return ret;
ret = intel_ring_begin(req, (w->count * 2 + 2));
if (ret)
return ret;
intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(w->count));
for (i = 0; i < w->count; i++) {
intel_ring_emit_reg(ring, w->reg[i].addr);
intel_ring_emit(ring, w->reg[i].value);
}
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
ret = req->engine->emit_flush(req, EMIT_BARRIER);
if (ret)
return ret;
DRM_DEBUG_DRIVER("Number of Workarounds emitted: %d\n", w->count);
return 0;
}
static int intel_rcs_ctx_init(struct drm_i915_gem_request *req)
{
int ret;
ret = intel_ring_workarounds_emit(req);
if (ret != 0)
return ret;
ret = i915_gem_render_state_init(req);
if (ret)
return ret;
return 0;
}
static int wa_add(struct drm_i915_private *dev_priv,
drm/i915: Type safe register read/write Make I915_READ and I915_WRITE more type safe by wrapping the register offset in a struct. This should eliminate most of the fumbles we've had with misplaced parens. This only takes care of normal mmio registers. We could extend the idea to other register types and define each with its own struct. That way you wouldn't be able to accidentally pass the wrong thing to a specific register access function. The gpio_reg setup is probably the ugliest thing left. But I figure I'd just leave it for now, and wait for some divine inspiration to strike before making it nice. As for the generated code, it's actually a bit better sometimes. Eg. looking at i915_irq_handler(), we can see the following change: lea 0x70024(%rdx,%rax,1),%r9d mov $0x1,%edx - movslq %r9d,%r9 - mov %r9,%rsi - mov %r9,-0x58(%rbp) - callq *0xd8(%rbx) + mov %r9d,%esi + mov %r9d,-0x48(%rbp) callq *0xd8(%rbx) So previously gcc thought the register offset might be signed and decided to sign extend it, just in case. The rest appears to be mostly just minor shuffling of instructions. v2: i915_mmio_reg_{offset,equal,valid}() helpers added s/_REG/_MMIO/ in the register defines mo more switch statements left to worry about ring_emit stuff got sorted in a prep patch cmd parser, lrc context and w/a batch buildup also in prep patch vgpu stuff cleaned up and moved to a prep patch all other unrelated changes split out v3: Rebased due to BXT DSI/BLC, MOCS, etc. v4: Rebased due to churn, s/i915_mmio_reg_t/i915_reg_t/ Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Link: http://patchwork.freedesktop.org/patch/msgid/1447853606-2751-1-git-send-email-ville.syrjala@linux.intel.com
2015-11-18 20:33:26 +07:00
i915_reg_t addr,
const u32 mask, const u32 val)
{
const u32 idx = dev_priv->workarounds.count;
if (WARN_ON(idx >= I915_MAX_WA_REGS))
return -ENOSPC;
dev_priv->workarounds.reg[idx].addr = addr;
dev_priv->workarounds.reg[idx].value = val;
dev_priv->workarounds.reg[idx].mask = mask;
dev_priv->workarounds.count++;
return 0;
}
#define WA_REG(addr, mask, val) do { \
const int r = wa_add(dev_priv, (addr), (mask), (val)); \
if (r) \
return r; \
} while (0)
#define WA_SET_BIT_MASKED(addr, mask) \
WA_REG(addr, (mask), _MASKED_BIT_ENABLE(mask))
#define WA_CLR_BIT_MASKED(addr, mask) \
WA_REG(addr, (mask), _MASKED_BIT_DISABLE(mask))
drm/i915/bdw: Fix the write setting up the WIZ hashing mode I was playing with clang and oh surprise! a warning trigerred by -Wshift-overflow (gcc doesn't have this one): WA_SET_BIT_MASKED(GEN7_GT_MODE, GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4); drivers/gpu/drm/i915/intel_ringbuffer.c:786:2: warning: signed shift result (0x28002000000) requires 43 bits to represent, but 'int' only has 32 bits [-Wshift-overflow] WA_SET_BIT_MASKED(GEN7_GT_MODE, ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ drivers/gpu/drm/i915/intel_ringbuffer.c:737:15: note: expanded from macro 'WA_SET_BIT_MASKED' WA_REG(addr, _MASKED_BIT_ENABLE(mask), (mask) & 0xffff) Turned out GEN6_WIZ_HASHING_MASK was already shifted by 16, and we were trying to shift it a bit more. The other thing is that it's not the usual case of setting WA bits here, we need to have separate mask and value. To fix this, I've introduced a new _MASKED_FIELD() macro that takes both the (unshifted) mask and the desired value and the rest of the patch ripples through from it. This bug was introduced when reworking the WA emission in: Commit 7225342ab501befdb64bcec76ded41f5897c0855 Author: Mika Kuoppala <mika.kuoppala@linux.intel.com> Date: Tue Oct 7 17:21:26 2014 +0300 drm/i915: Build workaround list in ring initialization v2: Invert the order of the mask and value arguments (Daniel Vetter) Rewrite _MASKED_BIT_ENABLE() and _MASKED_BIT_DISABLE() with _MASKED_FIELD() (Jani Nikula) Make sure we only evaluate 'a' once in _MASKED_BIT_ENABLE() (Dave Gordon) Add check to ensure the value is within the mask boundaries (Chris Wilson) v3: Ensure the the value and mask are 16 bits (Dave Gordon) Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Arun Siluvery <arun.siluvery@linux.intel.com> Signed-off-by: Damien Lespiau <damien.lespiau@intel.com> Signed-off-by: Jani Nikula <jani.nikula@intel.com>
2014-12-09 00:33:51 +07:00
#define WA_SET_FIELD_MASKED(addr, mask, value) \
WA_REG(addr, mask, _MASKED_FIELD(mask, value))
#define WA_SET_BIT(addr, mask) WA_REG(addr, mask, I915_READ(addr) | (mask))
#define WA_CLR_BIT(addr, mask) WA_REG(addr, mask, I915_READ(addr) & ~(mask))
#define WA_WRITE(addr, val) WA_REG(addr, 0xffffffff, val)
static int wa_ring_whitelist_reg(struct intel_engine_cs *engine,
i915_reg_t reg)
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
{
struct drm_i915_private *dev_priv = engine->i915;
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
struct i915_workarounds *wa = &dev_priv->workarounds;
const uint32_t index = wa->hw_whitelist_count[engine->id];
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
if (WARN_ON(index >= RING_MAX_NONPRIV_SLOTS))
return -EINVAL;
WA_WRITE(RING_FORCE_TO_NONPRIV(engine->mmio_base, index),
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
i915_mmio_reg_offset(reg));
wa->hw_whitelist_count[engine->id]++;
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
return 0;
}
static int gen8_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
WA_SET_BIT_MASKED(INSTPM, INSTPM_FORCE_ORDERING);
/* WaDisableAsyncFlipPerfMode:bdw,chv */
WA_SET_BIT_MASKED(MI_MODE, ASYNC_FLIP_PERF_DISABLE);
/* WaDisablePartialInstShootdown:bdw,chv */
WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN,
PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
/* Use Force Non-Coherent whenever executing a 3D context. This is a
* workaround for for a possible hang in the unlikely event a TLB
* invalidation occurs during a PSD flush.
*/
/* WaForceEnableNonCoherent:bdw,chv */
/* WaHdcDisableFetchWhenMasked:bdw,chv */
WA_SET_BIT_MASKED(HDC_CHICKEN0,
HDC_DONOT_FETCH_MEM_WHEN_MASKED |
HDC_FORCE_NON_COHERENT);
/* From the Haswell PRM, Command Reference: Registers, CACHE_MODE_0:
* "The Hierarchical Z RAW Stall Optimization allows non-overlapping
* polygons in the same 8x4 pixel/sample area to be processed without
* stalling waiting for the earlier ones to write to Hierarchical Z
* buffer."
*
* This optimization is off by default for BDW and CHV; turn it on.
*/
WA_CLR_BIT_MASKED(CACHE_MODE_0_GEN7, HIZ_RAW_STALL_OPT_DISABLE);
/* Wa4x4STCOptimizationDisable:bdw,chv */
WA_SET_BIT_MASKED(CACHE_MODE_1, GEN8_4x4_STC_OPTIMIZATION_DISABLE);
/*
* BSpec recommends 8x4 when MSAA is used,
* however in practice 16x4 seems fastest.
*
* Note that PS/WM thread counts depend on the WIZ hashing
* disable bit, which we don't touch here, but it's good
* to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
*/
WA_SET_FIELD_MASKED(GEN7_GT_MODE,
GEN6_WIZ_HASHING_MASK,
GEN6_WIZ_HASHING_16x4);
return 0;
}
static int bdw_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
ret = gen8_init_workarounds(engine);
if (ret)
return ret;
/* WaDisableThreadStallDopClockGating:bdw (pre-production) */
WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
/* WaDisableDopClockGating:bdw */
WA_SET_BIT_MASKED(GEN7_ROW_CHICKEN2,
DOP_CLOCK_GATING_DISABLE);
WA_SET_BIT_MASKED(HALF_SLICE_CHICKEN3,
GEN8_SAMPLER_POWER_BYPASS_DIS);
WA_SET_BIT_MASKED(HDC_CHICKEN0,
/* WaForceContextSaveRestoreNonCoherent:bdw */
HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
/* WaDisableFenceDestinationToSLM:bdw (pre-prod) */
(IS_BDW_GT3(dev_priv) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
return 0;
}
static int chv_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
ret = gen8_init_workarounds(engine);
if (ret)
return ret;
/* WaDisableThreadStallDopClockGating:chv */
WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN, STALL_DOP_GATING_DISABLE);
/* Improve HiZ throughput on CHV. */
WA_SET_BIT_MASKED(HIZ_CHICKEN, CHV_HZ_8X8_MODE_IN_1X);
return 0;
}
static int gen9_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
/* WaConextSwitchWithConcurrentTLBInvalidate:skl,bxt,kbl */
I915_WRITE(GEN9_CSFE_CHICKEN1_RCS, _MASKED_BIT_ENABLE(GEN9_PREEMPT_GPGPU_SYNC_SWITCH_DISABLE));
/* WaEnableLbsSlaRetryTimerDecrement:skl,bxt,kbl */
I915_WRITE(BDW_SCRATCH1, I915_READ(BDW_SCRATCH1) |
GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
/* WaDisableKillLogic:bxt,skl,kbl */
I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) |
ECOCHK_DIS_TLB);
/* WaClearFlowControlGpgpuContextSave:skl,bxt,kbl */
/* WaDisablePartialInstShootdown:skl,bxt,kbl */
WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN,
FLOW_CONTROL_ENABLE |
PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE);
/* Syncing dependencies between camera and graphics:skl,bxt,kbl */
WA_SET_BIT_MASKED(HALF_SLICE_CHICKEN3,
GEN9_DISABLE_OCL_OOB_SUPPRESS_LOGIC);
/* WaDisableDgMirrorFixInHalfSliceChicken5:skl,bxt */
if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_B0) ||
IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
WA_CLR_BIT_MASKED(GEN9_HALF_SLICE_CHICKEN5,
GEN9_DG_MIRROR_FIX_ENABLE);
/* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_B0) ||
IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
WA_SET_BIT_MASKED(GEN7_COMMON_SLICE_CHICKEN1,
GEN9_RHWO_OPTIMIZATION_DISABLE);
/*
* WA also requires GEN9_SLICE_COMMON_ECO_CHICKEN0[14:14] to be set
* but we do that in per ctx batchbuffer as there is an issue
* with this register not getting restored on ctx restore
*/
}
/* WaEnableYV12BugFixInHalfSliceChicken7:skl,bxt,kbl */
/* WaEnableSamplerGPGPUPreemptionSupport:skl,bxt,kbl */
WA_SET_BIT_MASKED(GEN9_HALF_SLICE_CHICKEN7,
GEN9_ENABLE_YV12_BUGFIX |
GEN9_ENABLE_GPGPU_PREEMPTION);
/* Wa4x4STCOptimizationDisable:skl,bxt,kbl */
/* WaDisablePartialResolveInVc:skl,bxt,kbl */
WA_SET_BIT_MASKED(CACHE_MODE_1, (GEN8_4x4_STC_OPTIMIZATION_DISABLE |
GEN9_PARTIAL_RESOLVE_IN_VC_DISABLE));
/* WaCcsTlbPrefetchDisable:skl,bxt,kbl */
WA_CLR_BIT_MASKED(GEN9_HALF_SLICE_CHICKEN5,
GEN9_CCS_TLB_PREFETCH_ENABLE);
/* WaDisableMaskBasedCammingInRCC:skl,bxt */
if (IS_SKL_REVID(dev_priv, SKL_REVID_C0, SKL_REVID_C0) ||
IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
WA_SET_BIT_MASKED(SLICE_ECO_CHICKEN0,
PIXEL_MASK_CAMMING_DISABLE);
/* WaForceContextSaveRestoreNonCoherent:skl,bxt,kbl */
WA_SET_BIT_MASKED(HDC_CHICKEN0,
HDC_FORCE_CONTEXT_SAVE_RESTORE_NON_COHERENT |
HDC_FORCE_CSR_NON_COHERENT_OVR_DISABLE);
/* WaForceEnableNonCoherent and WaDisableHDCInvalidation are
* both tied to WaForceContextSaveRestoreNonCoherent
* in some hsds for skl. We keep the tie for all gen9. The
* documentation is a bit hazy and so we want to get common behaviour,
* even though there is no clear evidence we would need both on kbl/bxt.
* This area has been source of system hangs so we play it safe
* and mimic the skl regardless of what bspec says.
*
* Use Force Non-Coherent whenever executing a 3D context. This
* is a workaround for a possible hang in the unlikely event
* a TLB invalidation occurs during a PSD flush.
*/
/* WaForceEnableNonCoherent:skl,bxt,kbl */
WA_SET_BIT_MASKED(HDC_CHICKEN0,
HDC_FORCE_NON_COHERENT);
/* WaDisableHDCInvalidation:skl,bxt,kbl */
I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) |
BDW_DISABLE_HDC_INVALIDATION);
/* WaDisableSamplerPowerBypassForSOPingPong:skl,bxt,kbl */
if (IS_SKYLAKE(dev_priv) ||
IS_KABYLAKE(dev_priv) ||
IS_BXT_REVID(dev_priv, 0, BXT_REVID_B0))
WA_SET_BIT_MASKED(HALF_SLICE_CHICKEN3,
GEN8_SAMPLER_POWER_BYPASS_DIS);
/* WaDisableSTUnitPowerOptimization:skl,bxt,kbl */
WA_SET_BIT_MASKED(HALF_SLICE_CHICKEN2, GEN8_ST_PO_DISABLE);
/* WaOCLCoherentLineFlush:skl,bxt,kbl */
I915_WRITE(GEN8_L3SQCREG4, (I915_READ(GEN8_L3SQCREG4) |
GEN8_LQSC_FLUSH_COHERENT_LINES));
/* WaVFEStateAfterPipeControlwithMediaStateClear:skl,bxt */
ret = wa_ring_whitelist_reg(engine, GEN9_CTX_PREEMPT_REG);
if (ret)
return ret;
/* WaEnablePreemptionGranularityControlByUMD:skl,bxt,kbl */
ret= wa_ring_whitelist_reg(engine, GEN8_CS_CHICKEN1);
if (ret)
return ret;
/* WaAllowUMDToModifyHDCChicken1:skl,bxt,kbl */
ret = wa_ring_whitelist_reg(engine, GEN8_HDC_CHICKEN1);
if (ret)
return ret;
return 0;
}
static int skl_tune_iz_hashing(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
u8 vals[3] = { 0, 0, 0 };
unsigned int i;
for (i = 0; i < 3; i++) {
u8 ss;
/*
* Only consider slices where one, and only one, subslice has 7
* EUs
*/
if (!is_power_of_2(INTEL_INFO(dev_priv)->sseu.subslice_7eu[i]))
continue;
/*
* subslice_7eu[i] != 0 (because of the check above) and
* ss_max == 4 (maximum number of subslices possible per slice)
*
* -> 0 <= ss <= 3;
*/
ss = ffs(INTEL_INFO(dev_priv)->sseu.subslice_7eu[i]) - 1;
vals[i] = 3 - ss;
}
if (vals[0] == 0 && vals[1] == 0 && vals[2] == 0)
return 0;
/* Tune IZ hashing. See intel_device_info_runtime_init() */
WA_SET_FIELD_MASKED(GEN7_GT_MODE,
GEN9_IZ_HASHING_MASK(2) |
GEN9_IZ_HASHING_MASK(1) |
GEN9_IZ_HASHING_MASK(0),
GEN9_IZ_HASHING(2, vals[2]) |
GEN9_IZ_HASHING(1, vals[1]) |
GEN9_IZ_HASHING(0, vals[0]));
return 0;
}
static int skl_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
ret = gen9_init_workarounds(engine);
if (ret)
return ret;
/*
* Actual WA is to disable percontext preemption granularity control
* until D0 which is the default case so this is equivalent to
* !WaDisablePerCtxtPreemptionGranularityControl:skl
*/
if (IS_SKL_REVID(dev_priv, SKL_REVID_E0, REVID_FOREVER)) {
I915_WRITE(GEN7_FF_SLICE_CS_CHICKEN1,
_MASKED_BIT_ENABLE(GEN9_FFSC_PERCTX_PREEMPT_CTRL));
}
if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0)) {
/* WaDisableChickenBitTSGBarrierAckForFFSliceCS:skl */
I915_WRITE(FF_SLICE_CS_CHICKEN2,
_MASKED_BIT_ENABLE(GEN9_TSG_BARRIER_ACK_DISABLE));
}
/* GEN8_L3SQCREG4 has a dependency with WA batch so any new changes
* involving this register should also be added to WA batch as required.
*/
if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0))
/* WaDisableLSQCROPERFforOCL:skl */
I915_WRITE(GEN8_L3SQCREG4, I915_READ(GEN8_L3SQCREG4) |
GEN8_LQSC_RO_PERF_DIS);
/* WaEnableGapsTsvCreditFix:skl */
if (IS_SKL_REVID(dev_priv, SKL_REVID_C0, REVID_FOREVER)) {
I915_WRITE(GEN8_GARBCNTL, (I915_READ(GEN8_GARBCNTL) |
GEN9_GAPS_TSV_CREDIT_DISABLE));
}
/* WaDisablePowerCompilerClockGating:skl */
if (IS_SKL_REVID(dev_priv, SKL_REVID_B0, SKL_REVID_B0))
WA_SET_BIT_MASKED(HIZ_CHICKEN,
BDW_HIZ_POWER_COMPILER_CLOCK_GATING_DISABLE);
/* WaBarrierPerformanceFixDisable:skl */
if (IS_SKL_REVID(dev_priv, SKL_REVID_C0, SKL_REVID_D0))
WA_SET_BIT_MASKED(HDC_CHICKEN0,
HDC_FENCE_DEST_SLM_DISABLE |
HDC_BARRIER_PERFORMANCE_DISABLE);
/* WaDisableSbeCacheDispatchPortSharing:skl */
if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_F0))
WA_SET_BIT_MASKED(
GEN7_HALF_SLICE_CHICKEN1,
GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
/* WaDisableGafsUnitClkGating:skl */
WA_SET_BIT(GEN7_UCGCTL4, GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
/* WaInPlaceDecompressionHang:skl */
if (IS_SKL_REVID(dev_priv, SKL_REVID_H0, REVID_FOREVER))
WA_SET_BIT(GEN9_GAMT_ECO_REG_RW_IA,
GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
/* WaDisableLSQCROPERFforOCL:skl */
ret = wa_ring_whitelist_reg(engine, GEN8_L3SQCREG4);
if (ret)
return ret;
return skl_tune_iz_hashing(engine);
}
static int bxt_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
ret = gen9_init_workarounds(engine);
if (ret)
return ret;
/* WaStoreMultiplePTEenable:bxt */
/* This is a requirement according to Hardware specification */
if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_TLBPF);
/* WaSetClckGatingDisableMedia:bxt */
if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
I915_WRITE(GEN7_MISCCPCTL, (I915_READ(GEN7_MISCCPCTL) &
~GEN8_DOP_CLOCK_GATE_MEDIA_ENABLE));
}
/* WaDisableThreadStallDopClockGating:bxt */
WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN,
STALL_DOP_GATING_DISABLE);
/* WaDisablePooledEuLoadBalancingFix:bxt */
if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER)) {
WA_SET_BIT_MASKED(FF_SLICE_CS_CHICKEN2,
GEN9_POOLED_EU_LOAD_BALANCING_FIX_DISABLE);
}
/* WaDisableSbeCacheDispatchPortSharing:bxt */
if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_B0)) {
WA_SET_BIT_MASKED(
GEN7_HALF_SLICE_CHICKEN1,
GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
}
/* WaDisableObjectLevelPreemptionForTrifanOrPolygon:bxt */
/* WaDisableObjectLevelPreemptionForInstancedDraw:bxt */
/* WaDisableObjectLevelPreemtionForInstanceId:bxt */
/* WaDisableLSQCROPERFforOCL:bxt */
if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
ret = wa_ring_whitelist_reg(engine, GEN9_CS_DEBUG_MODE1);
if (ret)
return ret;
ret = wa_ring_whitelist_reg(engine, GEN8_L3SQCREG4);
if (ret)
return ret;
}
/* WaProgramL3SqcReg1DefaultForPerf:bxt */
if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
I915_WRITE(GEN8_L3SQCREG1, L3_GENERAL_PRIO_CREDITS(62) |
L3_HIGH_PRIO_CREDITS(2));
/* WaToEnableHwFixForPushConstHWBug:bxt */
if (IS_BXT_REVID(dev_priv, BXT_REVID_C0, REVID_FOREVER))
WA_SET_BIT_MASKED(COMMON_SLICE_CHICKEN2,
GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
/* WaInPlaceDecompressionHang:bxt */
if (IS_BXT_REVID(dev_priv, BXT_REVID_C0, REVID_FOREVER))
WA_SET_BIT(GEN9_GAMT_ECO_REG_RW_IA,
GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
return 0;
}
static int kbl_init_workarounds(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
ret = gen9_init_workarounds(engine);
if (ret)
return ret;
/* WaEnableGapsTsvCreditFix:kbl */
I915_WRITE(GEN8_GARBCNTL, (I915_READ(GEN8_GARBCNTL) |
GEN9_GAPS_TSV_CREDIT_DISABLE));
/* WaDisableDynamicCreditSharing:kbl */
if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
WA_SET_BIT(GAMT_CHKN_BIT_REG,
GAMT_CHKN_DISABLE_DYNAMIC_CREDIT_SHARING);
/* WaDisableFenceDestinationToSLM:kbl (pre-prod) */
if (IS_KBL_REVID(dev_priv, KBL_REVID_A0, KBL_REVID_A0))
WA_SET_BIT_MASKED(HDC_CHICKEN0,
HDC_FENCE_DEST_SLM_DISABLE);
/* GEN8_L3SQCREG4 has a dependency with WA batch so any new changes
* involving this register should also be added to WA batch as required.
*/
if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_E0))
/* WaDisableLSQCROPERFforOCL:kbl */
I915_WRITE(GEN8_L3SQCREG4, I915_READ(GEN8_L3SQCREG4) |
GEN8_LQSC_RO_PERF_DIS);
/* WaToEnableHwFixForPushConstHWBug:kbl */
if (IS_KBL_REVID(dev_priv, KBL_REVID_C0, REVID_FOREVER))
WA_SET_BIT_MASKED(COMMON_SLICE_CHICKEN2,
GEN8_SBE_DISABLE_REPLAY_BUF_OPTIMIZATION);
/* WaDisableGafsUnitClkGating:kbl */
WA_SET_BIT(GEN7_UCGCTL4, GEN8_EU_GAUNIT_CLOCK_GATE_DISABLE);
/* WaDisableSbeCacheDispatchPortSharing:kbl */
WA_SET_BIT_MASKED(
GEN7_HALF_SLICE_CHICKEN1,
GEN7_SBE_SS_CACHE_DISPATCH_PORT_SHARING_DISABLE);
/* WaInPlaceDecompressionHang:kbl */
WA_SET_BIT(GEN9_GAMT_ECO_REG_RW_IA,
GAMT_ECO_ENABLE_IN_PLACE_DECOMPRESS);
/* WaDisableLSQCROPERFforOCL:kbl */
ret = wa_ring_whitelist_reg(engine, GEN8_L3SQCREG4);
if (ret)
return ret;
return 0;
}
int init_workarounds_ring(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
WARN_ON(engine->id != RCS);
dev_priv->workarounds.count = 0;
drm/i915/gen9: Add framework to whitelist specific GPU registers Some of the HW registers are privileged and cannot be written to from non-privileged batch buffers coming from userspace unless they are added to the HW whitelist. This whitelist is maintained by HW and it is different from SW whitelist. Userspace need write access to them to implement preemption related WA. The reason for using this approach is, the register bits that control preemption granularity at the HW level are not context save/restored; so even if we set these bits always in kernel they are going to change once the context is switched out. We can consider making them non-privileged by default but these registers also contain other chicken bits which should not be allowed to be modified. In the later revisions controlling bits are save/restored at context level but in the existing revisions these are exported via other debug registers and should be on the whitelist. This patch adds changes to provide HW with a list of registers to be whitelisted. HW checks this list during execution and provides access accordingly. HW imposes a limit on the number of registers on whitelist and it is per-engine. At this point we are only enabling whitelist for RCS and we don't foresee any requirement for other engines. The registers to be whitelisted are added using generic workaround list mechanism, even these are only enablers for userspace workarounds. But by sharing this mechanism we get some test assets without additional cost (Mika). v2: rebase v3: parameterize RING_FORCE_TO_NONPRIV() as _MMIO() should be limited to i915_reg.h (Ville), drop inline for wa_ring_whitelist_reg (Mika). v4: improvements suggested by Chris Wilson. Clarify that this is HW whitelist and different from the one maintained in driver. This list is engine specific but it gets initialized along with other WA which is RCS specific thing, so make it clear that we are not doing any cross engine setup during initialization. Make HW whitelist count of each engine available in debugfs. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Arun Siluvery <arun.siluvery@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453412634-29238-2-git-send-email-arun.siluvery@linux.intel.com Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-22 04:43:47 +07:00
dev_priv->workarounds.hw_whitelist_count[RCS] = 0;
if (IS_BROADWELL(dev_priv))
return bdw_init_workarounds(engine);
if (IS_CHERRYVIEW(dev_priv))
return chv_init_workarounds(engine);
if (IS_SKYLAKE(dev_priv))
return skl_init_workarounds(engine);
if (IS_BROXTON(dev_priv))
return bxt_init_workarounds(engine);
if (IS_KABYLAKE(dev_priv))
return kbl_init_workarounds(engine);
return 0;
}
static int init_render_ring(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret = init_ring_common(engine);
if (ret)
return ret;
/* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
if (IS_GEN(dev_priv, 4, 6))
I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
/* We need to disable the AsyncFlip performance optimisations in order
* to use MI_WAIT_FOR_EVENT within the CS. It should already be
* programmed to '1' on all products.
*
* WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
*/
if (IS_GEN(dev_priv, 6, 7))
I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
/* Required for the hardware to program scanline values for waiting */
/* WaEnableFlushTlbInvalidationMode:snb */
if (IS_GEN6(dev_priv))
I915_WRITE(GFX_MODE,
_MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
/* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
if (IS_GEN7(dev_priv))
I915_WRITE(GFX_MODE_GEN7,
_MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
_MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
if (IS_GEN6(dev_priv)) {
/* From the Sandybridge PRM, volume 1 part 3, page 24:
* "If this bit is set, STCunit will have LRA as replacement
* policy. [...] This bit must be reset. LRA replacement
* policy is not supported."
*/
I915_WRITE(CACHE_MODE_0,
_MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
}
if (IS_GEN(dev_priv, 6, 7))
I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
if (INTEL_INFO(dev_priv)->gen >= 6)
I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
return init_workarounds_ring(engine);
}
static void render_ring_cleanup(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
i915_vma_unpin_and_release(&dev_priv->semaphore);
}
static int gen8_rcs_signal(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
struct drm_i915_private *dev_priv = req->i915;
struct intel_engine_cs *waiter;
enum intel_engine_id id;
int ret, num_rings;
num_rings = INTEL_INFO(dev_priv)->num_rings;
ret = intel_ring_begin(req, (num_rings-1) * 8);
if (ret)
return ret;
for_each_engine_id(waiter, dev_priv, id) {
u64 gtt_offset = req->engine->semaphore.signal_ggtt[id];
if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
continue;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
intel_ring_emit(ring,
PIPE_CONTROL_GLOBAL_GTT_IVB |
PIPE_CONTROL_QW_WRITE |
PIPE_CONTROL_CS_STALL);
intel_ring_emit(ring, lower_32_bits(gtt_offset));
intel_ring_emit(ring, upper_32_bits(gtt_offset));
intel_ring_emit(ring, req->fence.seqno);
intel_ring_emit(ring, 0);
intel_ring_emit(ring,
MI_SEMAPHORE_SIGNAL |
MI_SEMAPHORE_TARGET(waiter->hw_id));
intel_ring_emit(ring, 0);
}
intel_ring_advance(ring);
return 0;
}
static int gen8_xcs_signal(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
struct drm_i915_private *dev_priv = req->i915;
struct intel_engine_cs *waiter;
enum intel_engine_id id;
int ret, num_rings;
num_rings = INTEL_INFO(dev_priv)->num_rings;
ret = intel_ring_begin(req, (num_rings-1) * 6);
if (ret)
return ret;
for_each_engine_id(waiter, dev_priv, id) {
u64 gtt_offset = req->engine->semaphore.signal_ggtt[id];
if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
continue;
intel_ring_emit(ring,
(MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
intel_ring_emit(ring,
lower_32_bits(gtt_offset) |
MI_FLUSH_DW_USE_GTT);
intel_ring_emit(ring, upper_32_bits(gtt_offset));
intel_ring_emit(ring, req->fence.seqno);
intel_ring_emit(ring,
MI_SEMAPHORE_SIGNAL |
MI_SEMAPHORE_TARGET(waiter->hw_id));
intel_ring_emit(ring, 0);
}
intel_ring_advance(ring);
return 0;
}
static int gen6_signal(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
struct drm_i915_private *dev_priv = req->i915;
struct intel_engine_cs *engine;
int ret, num_rings;
num_rings = INTEL_INFO(dev_priv)->num_rings;
ret = intel_ring_begin(req, round_up((num_rings-1) * 3, 2));
if (ret)
return ret;
for_each_engine(engine, dev_priv) {
i915_reg_t mbox_reg;
if (!(BIT(engine->hw_id) & GEN6_SEMAPHORES_MASK))
continue;
drm/i915: Type safe register read/write Make I915_READ and I915_WRITE more type safe by wrapping the register offset in a struct. This should eliminate most of the fumbles we've had with misplaced parens. This only takes care of normal mmio registers. We could extend the idea to other register types and define each with its own struct. That way you wouldn't be able to accidentally pass the wrong thing to a specific register access function. The gpio_reg setup is probably the ugliest thing left. But I figure I'd just leave it for now, and wait for some divine inspiration to strike before making it nice. As for the generated code, it's actually a bit better sometimes. Eg. looking at i915_irq_handler(), we can see the following change: lea 0x70024(%rdx,%rax,1),%r9d mov $0x1,%edx - movslq %r9d,%r9 - mov %r9,%rsi - mov %r9,-0x58(%rbp) - callq *0xd8(%rbx) + mov %r9d,%esi + mov %r9d,-0x48(%rbp) callq *0xd8(%rbx) So previously gcc thought the register offset might be signed and decided to sign extend it, just in case. The rest appears to be mostly just minor shuffling of instructions. v2: i915_mmio_reg_{offset,equal,valid}() helpers added s/_REG/_MMIO/ in the register defines mo more switch statements left to worry about ring_emit stuff got sorted in a prep patch cmd parser, lrc context and w/a batch buildup also in prep patch vgpu stuff cleaned up and moved to a prep patch all other unrelated changes split out v3: Rebased due to BXT DSI/BLC, MOCS, etc. v4: Rebased due to churn, s/i915_mmio_reg_t/i915_reg_t/ Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Link: http://patchwork.freedesktop.org/patch/msgid/1447853606-2751-1-git-send-email-ville.syrjala@linux.intel.com
2015-11-18 20:33:26 +07:00
mbox_reg = req->engine->semaphore.mbox.signal[engine->hw_id];
drm/i915: Type safe register read/write Make I915_READ and I915_WRITE more type safe by wrapping the register offset in a struct. This should eliminate most of the fumbles we've had with misplaced parens. This only takes care of normal mmio registers. We could extend the idea to other register types and define each with its own struct. That way you wouldn't be able to accidentally pass the wrong thing to a specific register access function. The gpio_reg setup is probably the ugliest thing left. But I figure I'd just leave it for now, and wait for some divine inspiration to strike before making it nice. As for the generated code, it's actually a bit better sometimes. Eg. looking at i915_irq_handler(), we can see the following change: lea 0x70024(%rdx,%rax,1),%r9d mov $0x1,%edx - movslq %r9d,%r9 - mov %r9,%rsi - mov %r9,-0x58(%rbp) - callq *0xd8(%rbx) + mov %r9d,%esi + mov %r9d,-0x48(%rbp) callq *0xd8(%rbx) So previously gcc thought the register offset might be signed and decided to sign extend it, just in case. The rest appears to be mostly just minor shuffling of instructions. v2: i915_mmio_reg_{offset,equal,valid}() helpers added s/_REG/_MMIO/ in the register defines mo more switch statements left to worry about ring_emit stuff got sorted in a prep patch cmd parser, lrc context and w/a batch buildup also in prep patch vgpu stuff cleaned up and moved to a prep patch all other unrelated changes split out v3: Rebased due to BXT DSI/BLC, MOCS, etc. v4: Rebased due to churn, s/i915_mmio_reg_t/i915_reg_t/ Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Link: http://patchwork.freedesktop.org/patch/msgid/1447853606-2751-1-git-send-email-ville.syrjala@linux.intel.com
2015-11-18 20:33:26 +07:00
if (i915_mmio_reg_valid(mbox_reg)) {
intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
intel_ring_emit_reg(ring, mbox_reg);
intel_ring_emit(ring, req->fence.seqno);
}
}
/* If num_dwords was rounded, make sure the tail pointer is correct */
if (num_rings % 2 == 0)
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static void i9xx_submit_request(struct drm_i915_gem_request *request)
{
struct drm_i915_private *dev_priv = request->i915;
I915_WRITE_TAIL(request->engine,
intel_ring_offset(request->ring, request->tail));
}
static int i9xx_emit_request(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
intel_ring_emit(ring, req->fence.seqno);
intel_ring_emit(ring, MI_USER_INTERRUPT);
intel_ring_advance(ring);
req->tail = ring->tail;
return 0;
}
/**
* gen6_sema_emit_request - Update the semaphore mailbox registers
*
* @request - request to write to the ring
*
* Update the mailbox registers in the *other* rings with the current seqno.
* This acts like a signal in the canonical semaphore.
*/
static int gen6_sema_emit_request(struct drm_i915_gem_request *req)
{
int ret;
ret = req->engine->semaphore.signal(req);
if (ret)
return ret;
return i9xx_emit_request(req);
}
static int gen8_render_emit_request(struct drm_i915_gem_request *req)
{
struct intel_engine_cs *engine = req->engine;
struct intel_ring *ring = req->ring;
int ret;
if (engine->semaphore.signal) {
ret = engine->semaphore.signal(req);
if (ret)
return ret;
}
ret = intel_ring_begin(req, 8);
if (ret)
return ret;
intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
intel_ring_emit(ring, (PIPE_CONTROL_GLOBAL_GTT_IVB |
PIPE_CONTROL_CS_STALL |
PIPE_CONTROL_QW_WRITE));
intel_ring_emit(ring, intel_hws_seqno_address(engine));
intel_ring_emit(ring, 0);
intel_ring_emit(ring, i915_gem_request_get_seqno(req));
/* We're thrashing one dword of HWS. */
intel_ring_emit(ring, 0);
intel_ring_emit(ring, MI_USER_INTERRUPT);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
req->tail = ring->tail;
return 0;
}
/**
* intel_ring_sync - sync the waiter to the signaller on seqno
*
* @waiter - ring that is waiting
* @signaller - ring which has, or will signal
* @seqno - seqno which the waiter will block on
*/
static int
gen8_ring_sync_to(struct drm_i915_gem_request *req,
struct drm_i915_gem_request *signal)
{
struct intel_ring *ring = req->ring;
struct drm_i915_private *dev_priv = req->i915;
u64 offset = GEN8_WAIT_OFFSET(req->engine, signal->engine->id);
struct i915_hw_ppgtt *ppgtt;
int ret;
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
intel_ring_emit(ring,
MI_SEMAPHORE_WAIT |
MI_SEMAPHORE_GLOBAL_GTT |
MI_SEMAPHORE_SAD_GTE_SDD);
intel_ring_emit(ring, signal->fence.seqno);
intel_ring_emit(ring, lower_32_bits(offset));
intel_ring_emit(ring, upper_32_bits(offset));
intel_ring_advance(ring);
/* When the !RCS engines idle waiting upon a semaphore, they lose their
* pagetables and we must reload them before executing the batch.
* We do this on the i915_switch_context() following the wait and
* before the dispatch.
*/
ppgtt = req->ctx->ppgtt;
if (ppgtt && req->engine->id != RCS)
ppgtt->pd_dirty_rings |= intel_engine_flag(req->engine);
return 0;
}
static int
gen6_ring_sync_to(struct drm_i915_gem_request *req,
struct drm_i915_gem_request *signal)
{
struct intel_ring *ring = req->ring;
u32 dw1 = MI_SEMAPHORE_MBOX |
MI_SEMAPHORE_COMPARE |
MI_SEMAPHORE_REGISTER;
u32 wait_mbox = signal->engine->semaphore.mbox.wait[req->engine->hw_id];
int ret;
WARN_ON(wait_mbox == MI_SEMAPHORE_SYNC_INVALID);
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
intel_ring_emit(ring, dw1 | wait_mbox);
/* Throughout all of the GEM code, seqno passed implies our current
* seqno is >= the last seqno executed. However for hardware the
* comparison is strictly greater than.
*/
intel_ring_emit(ring, signal->fence.seqno - 1);
intel_ring_emit(ring, 0);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
static void
gen5_seqno_barrier(struct intel_engine_cs *engine)
{
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
/* MI_STORE are internally buffered by the GPU and not flushed
* either by MI_FLUSH or SyncFlush or any other combination of
* MI commands.
*
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
* "Only the submission of the store operation is guaranteed.
* The write result will be complete (coherent) some time later
* (this is practically a finite period but there is no guaranteed
* latency)."
*
* Empirically, we observe that we need a delay of at least 75us to
* be sure that the seqno write is visible by the CPU.
*/
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
usleep_range(125, 250);
}
static void
gen6_seqno_barrier(struct intel_engine_cs *engine)
2012-12-14 22:01:25 +07:00
{
struct drm_i915_private *dev_priv = engine->i915;
2012-12-14 22:01:25 +07:00
/* Workaround to force correct ordering between irq and seqno writes on
* ivb (and maybe also on snb) by reading from a CS register (like
* ACTHD) before reading the status page.
*
* Note that this effectively stalls the read by the time it takes to
* do a memory transaction, which more or less ensures that the write
* from the GPU has sufficient time to invalidate the CPU cacheline.
* Alternatively we could delay the interrupt from the CS ring to give
* the write time to land, but that would incur a delay after every
* batch i.e. much more frequent than a delay when waiting for the
* interrupt (with the same net latency).
*
* Also note that to prevent whole machine hangs on gen7, we have to
* take the spinlock to guard against concurrent cacheline access.
*/
spin_lock_irq(&dev_priv->uncore.lock);
POSTING_READ_FW(RING_ACTHD(engine->mmio_base));
spin_unlock_irq(&dev_priv->uncore.lock);
2012-12-14 22:01:25 +07:00
}
static void
gen5_irq_enable(struct intel_engine_cs *engine)
{
gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
}
static void
gen5_irq_disable(struct intel_engine_cs *engine)
{
gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
}
static void
i9xx_irq_enable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
dev_priv->irq_mask &= ~engine->irq_enable_mask;
I915_WRITE(IMR, dev_priv->irq_mask);
POSTING_READ_FW(RING_IMR(engine->mmio_base));
}
static void
i9xx_irq_disable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
dev_priv->irq_mask |= engine->irq_enable_mask;
I915_WRITE(IMR, dev_priv->irq_mask);
}
static void
i8xx_irq_enable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
dev_priv->irq_mask &= ~engine->irq_enable_mask;
I915_WRITE16(IMR, dev_priv->irq_mask);
POSTING_READ16(RING_IMR(engine->mmio_base));
}
static void
i8xx_irq_disable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
dev_priv->irq_mask |= engine->irq_enable_mask;
I915_WRITE16(IMR, dev_priv->irq_mask);
}
static int
bsd_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring, MI_FLUSH);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static void
gen6_irq_enable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
I915_WRITE_IMR(engine,
~(engine->irq_enable_mask |
engine->irq_keep_mask));
gen5_enable_gt_irq(dev_priv, engine->irq_enable_mask);
}
static void
gen6_irq_disable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
gen5_disable_gt_irq(dev_priv, engine->irq_enable_mask);
}
static void
hsw_vebox_irq_enable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
I915_WRITE_IMR(engine, ~engine->irq_enable_mask);
gen6_enable_pm_irq(dev_priv, engine->irq_enable_mask);
}
static void
hsw_vebox_irq_disable(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
I915_WRITE_IMR(engine, ~0);
gen6_disable_pm_irq(dev_priv, engine->irq_enable_mask);
}
static void
gen8_irq_enable(struct intel_engine_cs *engine)
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
{
struct drm_i915_private *dev_priv = engine->i915;
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
I915_WRITE_IMR(engine,
~(engine->irq_enable_mask |
engine->irq_keep_mask));
POSTING_READ_FW(RING_IMR(engine->mmio_base));
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
}
static void
gen8_irq_disable(struct intel_engine_cs *engine)
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
{
struct drm_i915_private *dev_priv = engine->i915;
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
}
static int
i965_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 length,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring,
MI_BATCH_BUFFER_START |
MI_BATCH_GTT |
(dispatch_flags & I915_DISPATCH_SECURE ?
0 : MI_BATCH_NON_SECURE_I965));
intel_ring_emit(ring, offset);
intel_ring_advance(ring);
return 0;
}
/* Just userspace ABI convention to limit the wa batch bo to a resonable size */
#define I830_BATCH_LIMIT (256*1024)
#define I830_TLB_ENTRIES (2)
#define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
static int
i830_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 len,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
u32 cs_offset = i915_ggtt_offset(req->engine->scratch);
int ret;
ret = intel_ring_begin(req, 6);
if (ret)
return ret;
/* Evict the invalid PTE TLBs */
intel_ring_emit(ring, COLOR_BLT_CMD | BLT_WRITE_RGBA);
intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096);
intel_ring_emit(ring, I830_TLB_ENTRIES << 16 | 4); /* load each page */
intel_ring_emit(ring, cs_offset);
intel_ring_emit(ring, 0xdeadbeef);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
if (len > I830_BATCH_LIMIT)
return -ENOSPC;
ret = intel_ring_begin(req, 6 + 2);
if (ret)
return ret;
/* Blit the batch (which has now all relocs applied) to the
* stable batch scratch bo area (so that the CS never
* stumbles over its tlb invalidation bug) ...
*/
intel_ring_emit(ring, SRC_COPY_BLT_CMD | BLT_WRITE_RGBA);
intel_ring_emit(ring,
BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096);
intel_ring_emit(ring, DIV_ROUND_UP(len, 4096) << 16 | 4096);
intel_ring_emit(ring, cs_offset);
intel_ring_emit(ring, 4096);
intel_ring_emit(ring, offset);
intel_ring_emit(ring, MI_FLUSH);
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
/* ... and execute it. */
offset = cs_offset;
}
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_GTT);
intel_ring_emit(ring, offset | (dispatch_flags & I915_DISPATCH_SECURE ?
0 : MI_BATCH_NON_SECURE));
intel_ring_advance(ring);
return 0;
}
static int
i915_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 len,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_GTT);
intel_ring_emit(ring, offset | (dispatch_flags & I915_DISPATCH_SECURE ?
0 : MI_BATCH_NON_SECURE));
intel_ring_advance(ring);
return 0;
}
static void cleanup_phys_status_page(struct intel_engine_cs *engine)
drm/i915: Cleanup phys status page too Restore the lost phys status page cleanup. Fixes the following splat with DMA_API_DEBUG=y: WARNING: CPU: 0 PID: 21615 at ../lib/dma-debug.c:974 dma_debug_device_change+0x190/0x1f0() pci 0000:00:02.0: DMA-API: device driver has pending DMA allocations while released from device [count=1] One of leaked entries details: [device address=0x0000000023163000] [size=4096 bytes] [mapped with DMA_BIDIRECTIONAL] [mapped as coherent] Modules linked in: i915(-) i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops drm sha256_generic hmac drbg ctr ccm sch_fq_codel binfmt_misc joydev mousedev arc4 ath5k iTCO_wdt mac80211 smsc_ircc2 ath snd_intel8x0m snd_intel8x0 snd_ac97_codec ac97_bus psmouse snd_pcm input_leds i2c_i801 pcspkr snd_timer cfg80211 snd soundcore i2c_core ehci_pci firewire_ohci ehci_hcd firewire_core lpc_ich 8139too rfkill crc_itu_t mfd_core mii usbcore rng_core intel_agp intel_gtt usb_common agpgart irda crc_ccitt fujitsu_laptop led_class parport_pc video parport evdev backlight CPU: 0 PID: 21615 Comm: rmmod Tainted: G U 4.4.0-rc4-mgm-ovl+ #4 Hardware name: FUJITSU SIEMENS LIFEBOOK S6120/FJNB16C, BIOS Version 1.26 05/10/2004 e31a3de0 e31a3de0 e31a3d9c c128d4bd e31a3dd0 c1045a0c c15e00c4 e31a3dfc 0000546f c15dfad2 000003ce c12b3740 000003ce c12b3740 00000000 00000001 f61fb8a0 e31a3de8 c1045a83 00000009 e31a3de0 c15e00c4 e31a3dfc e31a3e4c Call Trace: [<c128d4bd>] dump_stack+0x16/0x19 [<c1045a0c>] warn_slowpath_common+0x8c/0xd0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c1045a83>] warn_slowpath_fmt+0x33/0x40 [<c12b3740>] dma_debug_device_change+0x190/0x1f0 [<c1065499>] notifier_call_chain+0x59/0x70 [<c10655af>] __blocking_notifier_call_chain+0x3f/0x80 [<c106560f>] blocking_notifier_call_chain+0x1f/0x30 [<c134cfb3>] __device_release_driver+0xc3/0xf0 [<c134d0d7>] driver_detach+0x97/0xa0 [<c134c440>] bus_remove_driver+0x40/0x90 [<c134db18>] driver_unregister+0x28/0x60 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c12c0618>] pci_unregister_driver+0x18/0x80 [<f83e96e7>] drm_pci_exit+0x87/0xb0 [drm] [<f8b3be2d>] i915_exit+0x1b/0x1ee [i915] [<c10b999c>] SyS_delete_module+0x14c/0x210 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c115a9bd>] ? ____fput+0xd/0x10 [<c1002014>] do_fast_syscall_32+0xa4/0x450 [<c149f6fa>] sysenter_past_esp+0x3b/0x5d ---[ end trace c2ecbc77760f10a0 ]--- Mapped at: [<c12b3183>] debug_dma_alloc_coherent+0x33/0x90 [<f83e989c>] drm_pci_alloc+0x18c/0x1e0 [drm] [<f8acd59f>] intel_init_ring_buffer+0x2af/0x490 [i915] [<f8acd8b0>] intel_init_render_ring_buffer+0x130/0x750 [i915] [<f8aaea4e>] i915_gem_init_rings+0x1e/0x110 [i915] v2: s/BUG_ON/WARN_ON/ since dim doens't like the former anymore Cc: Chris Wilson <chris@chris-wilson.co.uk> Fixes: 5c6c600 ("drm/i915: Remove DRI1 ring accessors and API") Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Link: http://patchwork.freedesktop.org/patch/msgid/1452538112-5331-1-git-send-email-ville.syrjala@linux.intel.com Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-12 01:48:32 +07:00
{
struct drm_i915_private *dev_priv = engine->i915;
drm/i915: Cleanup phys status page too Restore the lost phys status page cleanup. Fixes the following splat with DMA_API_DEBUG=y: WARNING: CPU: 0 PID: 21615 at ../lib/dma-debug.c:974 dma_debug_device_change+0x190/0x1f0() pci 0000:00:02.0: DMA-API: device driver has pending DMA allocations while released from device [count=1] One of leaked entries details: [device address=0x0000000023163000] [size=4096 bytes] [mapped with DMA_BIDIRECTIONAL] [mapped as coherent] Modules linked in: i915(-) i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops drm sha256_generic hmac drbg ctr ccm sch_fq_codel binfmt_misc joydev mousedev arc4 ath5k iTCO_wdt mac80211 smsc_ircc2 ath snd_intel8x0m snd_intel8x0 snd_ac97_codec ac97_bus psmouse snd_pcm input_leds i2c_i801 pcspkr snd_timer cfg80211 snd soundcore i2c_core ehci_pci firewire_ohci ehci_hcd firewire_core lpc_ich 8139too rfkill crc_itu_t mfd_core mii usbcore rng_core intel_agp intel_gtt usb_common agpgart irda crc_ccitt fujitsu_laptop led_class parport_pc video parport evdev backlight CPU: 0 PID: 21615 Comm: rmmod Tainted: G U 4.4.0-rc4-mgm-ovl+ #4 Hardware name: FUJITSU SIEMENS LIFEBOOK S6120/FJNB16C, BIOS Version 1.26 05/10/2004 e31a3de0 e31a3de0 e31a3d9c c128d4bd e31a3dd0 c1045a0c c15e00c4 e31a3dfc 0000546f c15dfad2 000003ce c12b3740 000003ce c12b3740 00000000 00000001 f61fb8a0 e31a3de8 c1045a83 00000009 e31a3de0 c15e00c4 e31a3dfc e31a3e4c Call Trace: [<c128d4bd>] dump_stack+0x16/0x19 [<c1045a0c>] warn_slowpath_common+0x8c/0xd0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c1045a83>] warn_slowpath_fmt+0x33/0x40 [<c12b3740>] dma_debug_device_change+0x190/0x1f0 [<c1065499>] notifier_call_chain+0x59/0x70 [<c10655af>] __blocking_notifier_call_chain+0x3f/0x80 [<c106560f>] blocking_notifier_call_chain+0x1f/0x30 [<c134cfb3>] __device_release_driver+0xc3/0xf0 [<c134d0d7>] driver_detach+0x97/0xa0 [<c134c440>] bus_remove_driver+0x40/0x90 [<c134db18>] driver_unregister+0x28/0x60 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c12c0618>] pci_unregister_driver+0x18/0x80 [<f83e96e7>] drm_pci_exit+0x87/0xb0 [drm] [<f8b3be2d>] i915_exit+0x1b/0x1ee [i915] [<c10b999c>] SyS_delete_module+0x14c/0x210 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c115a9bd>] ? ____fput+0xd/0x10 [<c1002014>] do_fast_syscall_32+0xa4/0x450 [<c149f6fa>] sysenter_past_esp+0x3b/0x5d ---[ end trace c2ecbc77760f10a0 ]--- Mapped at: [<c12b3183>] debug_dma_alloc_coherent+0x33/0x90 [<f83e989c>] drm_pci_alloc+0x18c/0x1e0 [drm] [<f8acd59f>] intel_init_ring_buffer+0x2af/0x490 [i915] [<f8acd8b0>] intel_init_render_ring_buffer+0x130/0x750 [i915] [<f8aaea4e>] i915_gem_init_rings+0x1e/0x110 [i915] v2: s/BUG_ON/WARN_ON/ since dim doens't like the former anymore Cc: Chris Wilson <chris@chris-wilson.co.uk> Fixes: 5c6c600 ("drm/i915: Remove DRI1 ring accessors and API") Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Link: http://patchwork.freedesktop.org/patch/msgid/1452538112-5331-1-git-send-email-ville.syrjala@linux.intel.com Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-12 01:48:32 +07:00
if (!dev_priv->status_page_dmah)
return;
drm_pci_free(&dev_priv->drm, dev_priv->status_page_dmah);
engine->status_page.page_addr = NULL;
drm/i915: Cleanup phys status page too Restore the lost phys status page cleanup. Fixes the following splat with DMA_API_DEBUG=y: WARNING: CPU: 0 PID: 21615 at ../lib/dma-debug.c:974 dma_debug_device_change+0x190/0x1f0() pci 0000:00:02.0: DMA-API: device driver has pending DMA allocations while released from device [count=1] One of leaked entries details: [device address=0x0000000023163000] [size=4096 bytes] [mapped with DMA_BIDIRECTIONAL] [mapped as coherent] Modules linked in: i915(-) i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops drm sha256_generic hmac drbg ctr ccm sch_fq_codel binfmt_misc joydev mousedev arc4 ath5k iTCO_wdt mac80211 smsc_ircc2 ath snd_intel8x0m snd_intel8x0 snd_ac97_codec ac97_bus psmouse snd_pcm input_leds i2c_i801 pcspkr snd_timer cfg80211 snd soundcore i2c_core ehci_pci firewire_ohci ehci_hcd firewire_core lpc_ich 8139too rfkill crc_itu_t mfd_core mii usbcore rng_core intel_agp intel_gtt usb_common agpgart irda crc_ccitt fujitsu_laptop led_class parport_pc video parport evdev backlight CPU: 0 PID: 21615 Comm: rmmod Tainted: G U 4.4.0-rc4-mgm-ovl+ #4 Hardware name: FUJITSU SIEMENS LIFEBOOK S6120/FJNB16C, BIOS Version 1.26 05/10/2004 e31a3de0 e31a3de0 e31a3d9c c128d4bd e31a3dd0 c1045a0c c15e00c4 e31a3dfc 0000546f c15dfad2 000003ce c12b3740 000003ce c12b3740 00000000 00000001 f61fb8a0 e31a3de8 c1045a83 00000009 e31a3de0 c15e00c4 e31a3dfc e31a3e4c Call Trace: [<c128d4bd>] dump_stack+0x16/0x19 [<c1045a0c>] warn_slowpath_common+0x8c/0xd0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c1045a83>] warn_slowpath_fmt+0x33/0x40 [<c12b3740>] dma_debug_device_change+0x190/0x1f0 [<c1065499>] notifier_call_chain+0x59/0x70 [<c10655af>] __blocking_notifier_call_chain+0x3f/0x80 [<c106560f>] blocking_notifier_call_chain+0x1f/0x30 [<c134cfb3>] __device_release_driver+0xc3/0xf0 [<c134d0d7>] driver_detach+0x97/0xa0 [<c134c440>] bus_remove_driver+0x40/0x90 [<c134db18>] driver_unregister+0x28/0x60 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c12c0618>] pci_unregister_driver+0x18/0x80 [<f83e96e7>] drm_pci_exit+0x87/0xb0 [drm] [<f8b3be2d>] i915_exit+0x1b/0x1ee [i915] [<c10b999c>] SyS_delete_module+0x14c/0x210 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c115a9bd>] ? ____fput+0xd/0x10 [<c1002014>] do_fast_syscall_32+0xa4/0x450 [<c149f6fa>] sysenter_past_esp+0x3b/0x5d ---[ end trace c2ecbc77760f10a0 ]--- Mapped at: [<c12b3183>] debug_dma_alloc_coherent+0x33/0x90 [<f83e989c>] drm_pci_alloc+0x18c/0x1e0 [drm] [<f8acd59f>] intel_init_ring_buffer+0x2af/0x490 [i915] [<f8acd8b0>] intel_init_render_ring_buffer+0x130/0x750 [i915] [<f8aaea4e>] i915_gem_init_rings+0x1e/0x110 [i915] v2: s/BUG_ON/WARN_ON/ since dim doens't like the former anymore Cc: Chris Wilson <chris@chris-wilson.co.uk> Fixes: 5c6c600 ("drm/i915: Remove DRI1 ring accessors and API") Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Link: http://patchwork.freedesktop.org/patch/msgid/1452538112-5331-1-git-send-email-ville.syrjala@linux.intel.com Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-12 01:48:32 +07:00
}
static void cleanup_status_page(struct intel_engine_cs *engine)
{
struct i915_vma *vma;
vma = fetch_and_zero(&engine->status_page.vma);
if (!vma)
return;
i915_vma_unpin(vma);
i915_gem_object_unpin_map(vma->obj);
i915_vma_put(vma);
}
static int init_status_page(struct intel_engine_cs *engine)
{
struct drm_i915_gem_object *obj;
struct i915_vma *vma;
unsigned int flags;
int ret;
obj = i915_gem_object_create(&engine->i915->drm, 4096);
if (IS_ERR(obj)) {
DRM_ERROR("Failed to allocate status page\n");
return PTR_ERR(obj);
}
ret = i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
if (ret)
goto err;
vma = i915_vma_create(obj, &engine->i915->ggtt.base, NULL);
if (IS_ERR(vma)) {
ret = PTR_ERR(vma);
goto err;
}
flags = PIN_GLOBAL;
if (!HAS_LLC(engine->i915))
/* On g33, we cannot place HWS above 256MiB, so
* restrict its pinning to the low mappable arena.
* Though this restriction is not documented for
* gen4, gen5, or byt, they also behave similarly
* and hang if the HWS is placed at the top of the
* GTT. To generalise, it appears that all !llc
* platforms have issues with us placing the HWS
* above the mappable region (even though we never
* actualy map it).
*/
flags |= PIN_MAPPABLE;
ret = i915_vma_pin(vma, 0, 4096, flags);
if (ret)
goto err;
engine->status_page.vma = vma;
engine->status_page.ggtt_offset = i915_ggtt_offset(vma);
engine->status_page.page_addr =
i915_gem_object_pin_map(obj, I915_MAP_WB);
DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
engine->name, i915_ggtt_offset(vma));
return 0;
err:
i915_gem_object_put(obj);
return ret;
}
static int init_phys_status_page(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
dev_priv->status_page_dmah =
drm_pci_alloc(&dev_priv->drm, PAGE_SIZE, PAGE_SIZE);
if (!dev_priv->status_page_dmah)
return -ENOMEM;
engine->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
memset(engine->status_page.page_addr, 0, PAGE_SIZE);
return 0;
}
int intel_ring_pin(struct intel_ring *ring)
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
{
/* Ring wraparound at offset 0 sometimes hangs. No idea why. */
unsigned int flags = PIN_GLOBAL | PIN_OFFSET_BIAS | 4096;
enum i915_map_type map;
struct i915_vma *vma = ring->vma;
void *addr;
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
int ret;
GEM_BUG_ON(ring->vaddr);
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
map = HAS_LLC(ring->engine->i915) ? I915_MAP_WB : I915_MAP_WC;
if (vma->obj->stolen)
flags |= PIN_MAPPABLE;
if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
if (flags & PIN_MAPPABLE || map == I915_MAP_WC)
ret = i915_gem_object_set_to_gtt_domain(vma->obj, true);
else
ret = i915_gem_object_set_to_cpu_domain(vma->obj, true);
if (unlikely(ret))
return ret;
}
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
ret = i915_vma_pin(vma, 0, PAGE_SIZE, flags);
if (unlikely(ret))
return ret;
if (i915_vma_is_map_and_fenceable(vma))
addr = (void __force *)i915_vma_pin_iomap(vma);
else
addr = i915_gem_object_pin_map(vma->obj, map);
if (IS_ERR(addr))
goto err;
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
ring->vaddr = addr;
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
return 0;
err:
i915_vma_unpin(vma);
return PTR_ERR(addr);
drm/i915/bdw: Pin the ringbuffer backing object to GGTT on-demand Same as with the context, pinning to GGTT regardless is harmful (it badly fragments the GGTT and can even exhaust it). Unfortunately, this case is also more complex than the previous one because we need to map and access the ringbuffer in several places along the execbuffer path (and we cannot make do by leaving the default ringbuffer pinned, as before). Also, the context object itself contains a pointer to the ringbuffer address that we have to keep updated if we are going to allow the ringbuffer to move around. v2: Same as with the context pinning, we cannot really do it during an interrupt. Also, pin the default ringbuffers objects regardless (makes error capture a lot easier). v3: Rebased. Take a pin reference of the ringbuffer for each item in the execlist request queue because the hardware may still be using the ringbuffer after the MI_USER_INTERRUPT to notify the seqno update is executed. The ringbuffer must remain pinned until the context save is complete. No longer pin and unpin ringbuffer in populate_lr_context() - this transient address is meaningless and the pinning can cause a sleep while atomic. v4: Moved ringbuffer pin and unpin into the lr_context_pin functions. Downgraded pinning check BUG_ONs to WARN_ONs. v5: Reinstated WARN_ONs for unexpected execlist states. Removed unused variable. Issue: VIZ-4277 Signed-off-by: Oscar Mateo <oscar.mateo@intel.com> Signed-off-by: Thomas Daniel <thomas.daniel@intel.com> Reviewed-by: Akash Goel <akash.goels@gmail.com> Reviewed-by: Deepak S<deepak.s@linux.intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2014-11-13 17:28:56 +07:00
}
void intel_ring_unpin(struct intel_ring *ring)
{
GEM_BUG_ON(!ring->vma);
GEM_BUG_ON(!ring->vaddr);
if (i915_vma_is_map_and_fenceable(ring->vma))
i915_vma_unpin_iomap(ring->vma);
else
i915_gem_object_unpin_map(ring->vma->obj);
ring->vaddr = NULL;
i915_vma_unpin(ring->vma);
}
static struct i915_vma *
intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
{
struct drm_i915_gem_object *obj;
struct i915_vma *vma;
obj = i915_gem_object_create_stolen(&dev_priv->drm, size);
if (!obj)
obj = i915_gem_object_create(&dev_priv->drm, size);
if (IS_ERR(obj))
return ERR_CAST(obj);
/* mark ring buffers as read-only from GPU side by default */
obj->gt_ro = 1;
vma = i915_vma_create(obj, &dev_priv->ggtt.base, NULL);
if (IS_ERR(vma))
goto err;
return vma;
err:
i915_gem_object_put(obj);
return vma;
}
struct intel_ring *
intel_engine_create_ring(struct intel_engine_cs *engine, int size)
{
struct intel_ring *ring;
struct i915_vma *vma;
GEM_BUG_ON(!is_power_of_2(size));
ring = kzalloc(sizeof(*ring), GFP_KERNEL);
if (!ring)
return ERR_PTR(-ENOMEM);
ring->engine = engine;
INIT_LIST_HEAD(&ring->request_list);
ring->size = size;
/* Workaround an erratum on the i830 which causes a hang if
* the TAIL pointer points to within the last 2 cachelines
* of the buffer.
*/
ring->effective_size = size;
if (IS_I830(engine->i915) || IS_845G(engine->i915))
ring->effective_size -= 2 * CACHELINE_BYTES;
ring->last_retired_head = -1;
intel_ring_update_space(ring);
vma = intel_ring_create_vma(engine->i915, size);
if (IS_ERR(vma)) {
kfree(ring);
return ERR_CAST(vma);
}
ring->vma = vma;
list_add(&ring->link, &engine->buffers);
return ring;
}
void
intel_ring_free(struct intel_ring *ring)
{
i915_vma_put(ring->vma);
list_del(&ring->link);
kfree(ring);
}
static int intel_ring_context_pin(struct i915_gem_context *ctx,
struct intel_engine_cs *engine)
{
struct intel_context *ce = &ctx->engine[engine->id];
int ret;
lockdep_assert_held(&ctx->i915->drm.struct_mutex);
if (ce->pin_count++)
return 0;
if (ce->state) {
ret = i915_gem_object_set_to_gtt_domain(ce->state->obj, false);
if (ret)
goto error;
ret = i915_vma_pin(ce->state, 0, ctx->ggtt_alignment,
PIN_GLOBAL | PIN_HIGH);
if (ret)
goto error;
}
/* The kernel context is only used as a placeholder for flushing the
* active context. It is never used for submitting user rendering and
* as such never requires the golden render context, and so we can skip
* emitting it when we switch to the kernel context. This is required
* as during eviction we cannot allocate and pin the renderstate in
* order to initialise the context.
*/
if (ctx == ctx->i915->kernel_context)
ce->initialised = true;
i915_gem_context_get(ctx);
return 0;
error:
ce->pin_count = 0;
return ret;
}
static void intel_ring_context_unpin(struct i915_gem_context *ctx,
struct intel_engine_cs *engine)
{
struct intel_context *ce = &ctx->engine[engine->id];
lockdep_assert_held(&ctx->i915->drm.struct_mutex);
if (--ce->pin_count)
return;
if (ce->state)
i915_vma_unpin(ce->state);
i915_gem_context_put(ctx);
}
static int intel_init_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
struct intel_ring *ring;
int ret;
WARN_ON(engine->buffer);
intel_engine_setup_common(engine);
memset(engine->semaphore.sync_seqno, 0,
sizeof(engine->semaphore.sync_seqno));
ret = intel_engine_init_common(engine);
drm/i915: Slaughter the thundering i915_wait_request herd One particularly stressful scenario consists of many independent tasks all competing for GPU time and waiting upon the results (e.g. realtime transcoding of many, many streams). One bottleneck in particular is that each client waits on its own results, but every client is woken up after every batchbuffer - hence the thunder of hooves as then every client must do its heavyweight dance to read a coherent seqno to see if it is the lucky one. Ideally, we only want one client to wake up after the interrupt and check its request for completion. Since the requests must retire in order, we can select the first client on the oldest request to be woken. Once that client has completed his wait, we can then wake up the next client and so on. However, all clients then incur latency as every process in the chain may be delayed for scheduling - this may also then cause some priority inversion. To reduce the latency, when a client is added or removed from the list, we scan the tree for completed seqno and wake up all the completed waiters in parallel. Using igt/benchmarks/gem_latency, we can demonstrate this effect. The benchmark measures the number of GPU cycles between completion of a batch and the client waking up from a call to wait-ioctl. With many concurrent waiters, with each on a different request, we observe that the wakeup latency before the patch scales nearly linearly with the number of waiters (before external factors kick in making the scaling much worse). After applying the patch, we can see that only the single waiter for the request is being woken up, providing a constant wakeup latency for every operation. However, the situation is not quite as rosy for many waiters on the same request, though to the best of my knowledge this is much less likely in practice. Here, we can observe that the concurrent waiters incur extra latency from being woken up by the solitary bottom-half, rather than directly by the interrupt. This appears to be scheduler induced (having discounted adverse effects from having a rbtree walk/erase in the wakeup path), each additional wake_up_process() costs approximately 1us on big core. Another effect of performing the secondary wakeups from the first bottom-half is the incurred delay this imposes on high priority threads - rather than immediately returning to userspace and leaving the interrupt handler to wake the others. To offset the delay incurred with additional waiters on a request, we could use a hybrid scheme that did a quick read in the interrupt handler and dequeued all the completed waiters (incurring the overhead in the interrupt handler, not the best plan either as we then incur GPU submission latency) but we would still have to wake up the bottom-half every time to do the heavyweight slow read. Or we could only kick the waiters on the seqno with the same priority as the current task (i.e. in the realtime waiter scenario, only it is woken up immediately by the interrupt and simply queues the next waiter before returning to userspace, minimising its delay at the expense of the chain, and also reducing contention on its scheduler runqueue). This is effective at avoid long pauses in the interrupt handler and at avoiding the extra latency in realtime/high-priority waiters. v2: Convert from a kworker per engine into a dedicated kthread for the bottom-half. v3: Rename request members and tweak comments. v4: Use a per-engine spinlock in the breadcrumbs bottom-half. v5: Fix race in locklessly checking waiter status and kicking the task on adding a new waiter. v6: Fix deciding when to force the timer to hide missing interrupts. v7: Move the bottom-half from the kthread to the first client process. v8: Reword a few comments v9: Break the busy loop when the interrupt is unmasked or has fired. v10: Comments, unnecessary churn, better debugging from Tvrtko v11: Wake all completed waiters on removing the current bottom-half to reduce the latency of waking up a herd of clients all waiting on the same request. v12: Rearrange missed-interrupt fault injection so that it works with igt/drv_missed_irq_hang v13: Rename intel_breadcrumb and friends to intel_wait in preparation for signal handling. v14: RCU commentary, assert_spin_locked v15: Hide BUG_ON behind the compiler; report on gem_latency findings. v16: Sort seqno-groups by priority so that first-waiter has the highest task priority (and so avoid priority inversion). v17: Add waiters to post-mortem GPU hang state. v18: Return early for a completed wait after acquiring the spinlock. Avoids adding ourselves to the tree if the is already complete, and skips the awkward question of why we don't do completion wakeups for waits earlier than or equal to ourselves. v19: Prepare for init_breadcrumbs to fail. Later patches may want to allocate during init, so be prepared to propagate back the error code. Testcase: igt/gem_concurrent_blit Testcase: igt/benchmarks/gem_latency Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: "Rogozhkin, Dmitry V" <dmitry.v.rogozhkin@intel.com> Cc: "Gong, Zhipeng" <zhipeng.gong@intel.com> Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> Cc: Dave Gordon <david.s.gordon@intel.com> Cc: "Goel, Akash" <akash.goel@intel.com> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com> #v18 Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-6-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:15 +07:00
if (ret)
goto error;
/* We may need to do things with the shrinker which
* require us to immediately switch back to the default
* context. This can cause a problem as pinning the
* default context also requires GTT space which may not
* be available. To avoid this we always pin the default
* context.
*/
ret = intel_ring_context_pin(dev_priv->kernel_context, engine);
if (ret)
goto error;
ring = intel_engine_create_ring(engine, 32 * PAGE_SIZE);
if (IS_ERR(ring)) {
ret = PTR_ERR(ring);
goto error;
}
if (HWS_NEEDS_PHYSICAL(dev_priv)) {
WARN_ON(engine->id != RCS);
ret = init_phys_status_page(engine);
if (ret)
goto error;
} else {
ret = init_status_page(engine);
if (ret)
goto error;
}
ret = intel_ring_pin(ring);
if (ret) {
intel_ring_free(ring);
goto error;
}
engine->buffer = ring;
return 0;
error:
intel_engine_cleanup(engine);
return ret;
}
void intel_engine_cleanup(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv;
if (!intel_engine_initialized(engine))
return;
dev_priv = engine->i915;
if (engine->buffer) {
WARN_ON(INTEL_GEN(dev_priv) > 2 &&
(I915_READ_MODE(engine) & MODE_IDLE) == 0);
intel_ring_unpin(engine->buffer);
intel_ring_free(engine->buffer);
engine->buffer = NULL;
}
if (engine->cleanup)
engine->cleanup(engine);
if (HWS_NEEDS_PHYSICAL(dev_priv)) {
WARN_ON(engine->id != RCS);
cleanup_phys_status_page(engine);
} else {
cleanup_status_page(engine);
drm/i915: Cleanup phys status page too Restore the lost phys status page cleanup. Fixes the following splat with DMA_API_DEBUG=y: WARNING: CPU: 0 PID: 21615 at ../lib/dma-debug.c:974 dma_debug_device_change+0x190/0x1f0() pci 0000:00:02.0: DMA-API: device driver has pending DMA allocations while released from device [count=1] One of leaked entries details: [device address=0x0000000023163000] [size=4096 bytes] [mapped with DMA_BIDIRECTIONAL] [mapped as coherent] Modules linked in: i915(-) i2c_algo_bit drm_kms_helper syscopyarea sysfillrect sysimgblt fb_sys_fops drm sha256_generic hmac drbg ctr ccm sch_fq_codel binfmt_misc joydev mousedev arc4 ath5k iTCO_wdt mac80211 smsc_ircc2 ath snd_intel8x0m snd_intel8x0 snd_ac97_codec ac97_bus psmouse snd_pcm input_leds i2c_i801 pcspkr snd_timer cfg80211 snd soundcore i2c_core ehci_pci firewire_ohci ehci_hcd firewire_core lpc_ich 8139too rfkill crc_itu_t mfd_core mii usbcore rng_core intel_agp intel_gtt usb_common agpgart irda crc_ccitt fujitsu_laptop led_class parport_pc video parport evdev backlight CPU: 0 PID: 21615 Comm: rmmod Tainted: G U 4.4.0-rc4-mgm-ovl+ #4 Hardware name: FUJITSU SIEMENS LIFEBOOK S6120/FJNB16C, BIOS Version 1.26 05/10/2004 e31a3de0 e31a3de0 e31a3d9c c128d4bd e31a3dd0 c1045a0c c15e00c4 e31a3dfc 0000546f c15dfad2 000003ce c12b3740 000003ce c12b3740 00000000 00000001 f61fb8a0 e31a3de8 c1045a83 00000009 e31a3de0 c15e00c4 e31a3dfc e31a3e4c Call Trace: [<c128d4bd>] dump_stack+0x16/0x19 [<c1045a0c>] warn_slowpath_common+0x8c/0xd0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c12b3740>] ? dma_debug_device_change+0x190/0x1f0 [<c1045a83>] warn_slowpath_fmt+0x33/0x40 [<c12b3740>] dma_debug_device_change+0x190/0x1f0 [<c1065499>] notifier_call_chain+0x59/0x70 [<c10655af>] __blocking_notifier_call_chain+0x3f/0x80 [<c106560f>] blocking_notifier_call_chain+0x1f/0x30 [<c134cfb3>] __device_release_driver+0xc3/0xf0 [<c134d0d7>] driver_detach+0x97/0xa0 [<c134c440>] bus_remove_driver+0x40/0x90 [<c134db18>] driver_unregister+0x28/0x60 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c12c0618>] pci_unregister_driver+0x18/0x80 [<f83e96e7>] drm_pci_exit+0x87/0xb0 [drm] [<f8b3be2d>] i915_exit+0x1b/0x1ee [i915] [<c10b999c>] SyS_delete_module+0x14c/0x210 [<c1079e8c>] ? trace_hardirqs_on_caller+0x12c/0x1d0 [<c115a9bd>] ? ____fput+0xd/0x10 [<c1002014>] do_fast_syscall_32+0xa4/0x450 [<c149f6fa>] sysenter_past_esp+0x3b/0x5d ---[ end trace c2ecbc77760f10a0 ]--- Mapped at: [<c12b3183>] debug_dma_alloc_coherent+0x33/0x90 [<f83e989c>] drm_pci_alloc+0x18c/0x1e0 [drm] [<f8acd59f>] intel_init_ring_buffer+0x2af/0x490 [i915] [<f8acd8b0>] intel_init_render_ring_buffer+0x130/0x750 [i915] [<f8aaea4e>] i915_gem_init_rings+0x1e/0x110 [i915] v2: s/BUG_ON/WARN_ON/ since dim doens't like the former anymore Cc: Chris Wilson <chris@chris-wilson.co.uk> Fixes: 5c6c600 ("drm/i915: Remove DRI1 ring accessors and API") Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Link: http://patchwork.freedesktop.org/patch/msgid/1452538112-5331-1-git-send-email-ville.syrjala@linux.intel.com Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2016-01-12 01:48:32 +07:00
}
intel_engine_cleanup_common(engine);
intel_ring_context_unpin(dev_priv->kernel_context, engine);
engine->i915 = NULL;
}
int intel_ring_alloc_request_extras(struct drm_i915_gem_request *request)
drm/i915: Preallocate next seqno before touching the ring Based on the work by Mika Kuoppala, we realised that we need to handle seqno wraparound prior to committing our changes to the ring. The most obvious point then is to grab the seqno inside intel_ring_begin(), and then to reuse that seqno for all ring operations until the next request. As intel_ring_begin() can fail, the callers must already be prepared to handle such failure and so we can safely add further checks. This patch looks like it should be split up into the interface changes and the tweaks to move seqno wrapping from the execbuffer into the core seqno increment. However, I found no easy way to break it into incremental steps without introducing further broken behaviour. v2: Mika found a silly mistake and a subtle error in the existing code; inside i915_gem_retire_requests() we were resetting the sync_seqno of the target ring based on the seqno from this ring - which are only related by the order of their allocation, not retirement. Hence we were applying the optimisation that the rings were synchronised too early, fortunately the only real casualty there is the handling of seqno wrapping. v3: Do not forget to reset the sync_seqno upon module reinitialisation, ala resume. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=863861 Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> [v2] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-11-27 23:22:52 +07:00
{
int ret;
/* Flush enough space to reduce the likelihood of waiting after
* we start building the request - in which case we will just
* have to repeat work.
*/
request->reserved_space += LEGACY_REQUEST_SIZE;
request->ring = request->engine->buffer;
ret = intel_ring_begin(request, 0);
if (ret)
return ret;
request->reserved_space -= LEGACY_REQUEST_SIZE;
return 0;
drm/i915: Preallocate next seqno before touching the ring Based on the work by Mika Kuoppala, we realised that we need to handle seqno wraparound prior to committing our changes to the ring. The most obvious point then is to grab the seqno inside intel_ring_begin(), and then to reuse that seqno for all ring operations until the next request. As intel_ring_begin() can fail, the callers must already be prepared to handle such failure and so we can safely add further checks. This patch looks like it should be split up into the interface changes and the tweaks to move seqno wrapping from the execbuffer into the core seqno increment. However, I found no easy way to break it into incremental steps without introducing further broken behaviour. v2: Mika found a silly mistake and a subtle error in the existing code; inside i915_gem_retire_requests() we were resetting the sync_seqno of the target ring based on the seqno from this ring - which are only related by the order of their allocation, not retirement. Hence we were applying the optimisation that the rings were synchronised too early, fortunately the only real casualty there is the handling of seqno wrapping. v3: Do not forget to reset the sync_seqno upon module reinitialisation, ala resume. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@intel.com> Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=863861 Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com> [v2] Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2012-11-27 23:22:52 +07:00
}
static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
{
struct intel_ring *ring = req->ring;
struct drm_i915_gem_request *target;
int ret;
intel_ring_update_space(ring);
if (ring->space >= bytes)
return 0;
/*
* Space is reserved in the ringbuffer for finalising the request,
* as that cannot be allowed to fail. During request finalisation,
* reserved_space is set to 0 to stop the overallocation and the
* assumption is that then we never need to wait (which has the
* risk of failing with EINTR).
*
* See also i915_gem_request_alloc() and i915_add_request().
*/
GEM_BUG_ON(!req->reserved_space);
list_for_each_entry(target, &ring->request_list, ring_link) {
unsigned space;
/* Would completion of this request free enough space? */
space = __intel_ring_space(target->postfix, ring->tail,
ring->size);
if (space >= bytes)
break;
}
drm/i915: Reserve ring buffer space for i915_add_request() commands It is a bad idea for i915_add_request() to fail. The work will already have been send to the ring and will be processed, but there will not be any tracking or management of that work. The only way the add request call can fail is if it can't write its epilogue commands to the ring (cache flushing, seqno updates, interrupt signalling). The reasons for that are mostly down to running out of ring buffer space and the problems associated with trying to get some more. This patch prevents that situation from happening in the first place. When a request is created, it marks sufficient space as reserved for the epilogue commands. Thus guaranteeing that by the time the epilogue is written, there will be plenty of space for it. Note that a ring_begin() call is required to actually reserve the space (and do any potential waiting). However, that is not currently done at request creation time. This is because the ring_begin() code can allocate a request. Hence calling begin() from the request allocation code would lead to infinite recursion! Later patches in this series remove the need for begin() to do the allocate. At that point, it becomes safe for the allocate to call begin() and really reserve the space. Until then, there is a potential for insufficient space to be available at the point of calling i915_add_request(). However, that would only be in the case where the request was created and immediately submitted without ever calling ring_begin() and adding any work to that request. Which should never happen. And even if it does, and if that request happens to fall down the tiny window of opportunity for failing due to being out of ring space then does it really matter because the request wasn't doing anything in the first place? v2: Updated the 'reserved space too small' warning to include the offending sizes. Added a 'cancel' operation to clean up when a request is abandoned. Added re-initialisation of tracking state after a buffer wrap to keep the sanity checks accurate. v3: Incremented the reserved size to accommodate Ironlake (after finally managing to run on an ILK system). Also fixed missing wrap code in LRC mode. v4: Added extra comment and removed duplicate WARN (feedback from Tomas). For: VIZ-5115 CC: Tomas Elf <tomas.elf@intel.com> Signed-off-by: John Harrison <John.C.Harrison@Intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-06-18 19:10:09 +07:00
if (WARN_ON(&target->ring_link == &ring->request_list))
return -ENOSPC;
ret = i915_wait_request(target, true, NULL, NO_WAITBOOST);
if (ret)
return ret;
if (i915_reset_in_progress(&target->i915->gpu_error))
return -EAGAIN;
i915_gem_request_retire_upto(target);
intel_ring_update_space(ring);
GEM_BUG_ON(ring->space < bytes);
return 0;
drm/i915: Reserve ring buffer space for i915_add_request() commands It is a bad idea for i915_add_request() to fail. The work will already have been send to the ring and will be processed, but there will not be any tracking or management of that work. The only way the add request call can fail is if it can't write its epilogue commands to the ring (cache flushing, seqno updates, interrupt signalling). The reasons for that are mostly down to running out of ring buffer space and the problems associated with trying to get some more. This patch prevents that situation from happening in the first place. When a request is created, it marks sufficient space as reserved for the epilogue commands. Thus guaranteeing that by the time the epilogue is written, there will be plenty of space for it. Note that a ring_begin() call is required to actually reserve the space (and do any potential waiting). However, that is not currently done at request creation time. This is because the ring_begin() code can allocate a request. Hence calling begin() from the request allocation code would lead to infinite recursion! Later patches in this series remove the need for begin() to do the allocate. At that point, it becomes safe for the allocate to call begin() and really reserve the space. Until then, there is a potential for insufficient space to be available at the point of calling i915_add_request(). However, that would only be in the case where the request was created and immediately submitted without ever calling ring_begin() and adding any work to that request. Which should never happen. And even if it does, and if that request happens to fall down the tiny window of opportunity for failing due to being out of ring space then does it really matter because the request wasn't doing anything in the first place? v2: Updated the 'reserved space too small' warning to include the offending sizes. Added a 'cancel' operation to clean up when a request is abandoned. Added re-initialisation of tracking state after a buffer wrap to keep the sanity checks accurate. v3: Incremented the reserved size to accommodate Ironlake (after finally managing to run on an ILK system). Also fixed missing wrap code in LRC mode. v4: Added extra comment and removed duplicate WARN (feedback from Tomas). For: VIZ-5115 CC: Tomas Elf <tomas.elf@intel.com> Signed-off-by: John Harrison <John.C.Harrison@Intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-06-18 19:10:09 +07:00
}
int intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
{
struct intel_ring *ring = req->ring;
int remain_actual = ring->size - ring->tail;
int remain_usable = ring->effective_size - ring->tail;
int bytes = num_dwords * sizeof(u32);
int total_bytes, wait_bytes;
bool need_wrap = false;
drm/i915: Reserve ring buffer space for i915_add_request() commands It is a bad idea for i915_add_request() to fail. The work will already have been send to the ring and will be processed, but there will not be any tracking or management of that work. The only way the add request call can fail is if it can't write its epilogue commands to the ring (cache flushing, seqno updates, interrupt signalling). The reasons for that are mostly down to running out of ring buffer space and the problems associated with trying to get some more. This patch prevents that situation from happening in the first place. When a request is created, it marks sufficient space as reserved for the epilogue commands. Thus guaranteeing that by the time the epilogue is written, there will be plenty of space for it. Note that a ring_begin() call is required to actually reserve the space (and do any potential waiting). However, that is not currently done at request creation time. This is because the ring_begin() code can allocate a request. Hence calling begin() from the request allocation code would lead to infinite recursion! Later patches in this series remove the need for begin() to do the allocate. At that point, it becomes safe for the allocate to call begin() and really reserve the space. Until then, there is a potential for insufficient space to be available at the point of calling i915_add_request(). However, that would only be in the case where the request was created and immediately submitted without ever calling ring_begin() and adding any work to that request. Which should never happen. And even if it does, and if that request happens to fall down the tiny window of opportunity for failing due to being out of ring space then does it really matter because the request wasn't doing anything in the first place? v2: Updated the 'reserved space too small' warning to include the offending sizes. Added a 'cancel' operation to clean up when a request is abandoned. Added re-initialisation of tracking state after a buffer wrap to keep the sanity checks accurate. v3: Incremented the reserved size to accommodate Ironlake (after finally managing to run on an ILK system). Also fixed missing wrap code in LRC mode. v4: Added extra comment and removed duplicate WARN (feedback from Tomas). For: VIZ-5115 CC: Tomas Elf <tomas.elf@intel.com> Signed-off-by: John Harrison <John.C.Harrison@Intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-06-18 19:10:09 +07:00
total_bytes = bytes + req->reserved_space;
drm/i915: Reserve ring buffer space for i915_add_request() commands It is a bad idea for i915_add_request() to fail. The work will already have been send to the ring and will be processed, but there will not be any tracking or management of that work. The only way the add request call can fail is if it can't write its epilogue commands to the ring (cache flushing, seqno updates, interrupt signalling). The reasons for that are mostly down to running out of ring buffer space and the problems associated with trying to get some more. This patch prevents that situation from happening in the first place. When a request is created, it marks sufficient space as reserved for the epilogue commands. Thus guaranteeing that by the time the epilogue is written, there will be plenty of space for it. Note that a ring_begin() call is required to actually reserve the space (and do any potential waiting). However, that is not currently done at request creation time. This is because the ring_begin() code can allocate a request. Hence calling begin() from the request allocation code would lead to infinite recursion! Later patches in this series remove the need for begin() to do the allocate. At that point, it becomes safe for the allocate to call begin() and really reserve the space. Until then, there is a potential for insufficient space to be available at the point of calling i915_add_request(). However, that would only be in the case where the request was created and immediately submitted without ever calling ring_begin() and adding any work to that request. Which should never happen. And even if it does, and if that request happens to fall down the tiny window of opportunity for failing due to being out of ring space then does it really matter because the request wasn't doing anything in the first place? v2: Updated the 'reserved space too small' warning to include the offending sizes. Added a 'cancel' operation to clean up when a request is abandoned. Added re-initialisation of tracking state after a buffer wrap to keep the sanity checks accurate. v3: Incremented the reserved size to accommodate Ironlake (after finally managing to run on an ILK system). Also fixed missing wrap code in LRC mode. v4: Added extra comment and removed duplicate WARN (feedback from Tomas). For: VIZ-5115 CC: Tomas Elf <tomas.elf@intel.com> Signed-off-by: John Harrison <John.C.Harrison@Intel.com> Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
2015-06-18 19:10:09 +07:00
if (unlikely(bytes > remain_usable)) {
/*
* Not enough space for the basic request. So need to flush
* out the remainder and then wait for base + reserved.
*/
wait_bytes = remain_actual + total_bytes;
need_wrap = true;
} else if (unlikely(total_bytes > remain_usable)) {
/*
* The base request will fit but the reserved space
* falls off the end. So we don't need an immediate wrap
* and only need to effectively wait for the reserved
* size space from the start of ringbuffer.
*/
wait_bytes = remain_actual + req->reserved_space;
} else {
/* No wrapping required, just waiting. */
wait_bytes = total_bytes;
}
if (wait_bytes > ring->space) {
int ret = wait_for_space(req, wait_bytes);
if (unlikely(ret))
return ret;
}
if (unlikely(need_wrap)) {
GEM_BUG_ON(remain_actual > ring->space);
GEM_BUG_ON(ring->tail + remain_actual > ring->size);
/* Fill the tail with MI_NOOP */
memset(ring->vaddr + ring->tail, 0, remain_actual);
ring->tail = 0;
ring->space -= remain_actual;
}
ring->space -= bytes;
GEM_BUG_ON(ring->space < 0);
return 0;
}
/* Align the ring tail to a cacheline boundary */
int intel_ring_cacheline_align(struct drm_i915_gem_request *req)
{
struct intel_ring *ring = req->ring;
int num_dwords =
(ring->tail & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
int ret;
if (num_dwords == 0)
return 0;
num_dwords = CACHELINE_BYTES / sizeof(uint32_t) - num_dwords;
ret = intel_ring_begin(req, num_dwords);
if (ret)
return ret;
while (num_dwords--)
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static void gen6_bsd_submit_request(struct drm_i915_gem_request *request)
{
struct drm_i915_private *dev_priv = request->i915;
intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
/* Every tail move must follow the sequence below */
/* Disable notification that the ring is IDLE. The GT
* will then assume that it is busy and bring it out of rc6.
*/
I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
_MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
/* Clear the context id. Here be magic! */
I915_WRITE64_FW(GEN6_BSD_RNCID, 0x0);
/* Wait for the ring not to be idle, i.e. for it to wake up. */
if (intel_wait_for_register_fw(dev_priv,
GEN6_BSD_SLEEP_PSMI_CONTROL,
GEN6_BSD_SLEEP_INDICATOR,
0,
50))
DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
/* Now that the ring is fully powered up, update the tail */
i9xx_submit_request(request);
/* Let the ring send IDLE messages to the GT again,
* and so let it sleep to conserve power when idle.
*/
I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
_MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
}
static int gen6_bsd_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
uint32_t cmd;
int ret;
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
cmd = MI_FLUSH_DW;
if (INTEL_GEN(req->i915) >= 8)
cmd += 1;
/* We always require a command barrier so that subsequent
* commands, such as breadcrumb interrupts, are strictly ordered
* wrt the contents of the write cache being flushed to memory
* (and thus being coherent from the CPU).
*/
cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
/*
* Bspec vol 1c.5 - video engine command streamer:
* "If ENABLED, all TLBs will be invalidated once the flush
* operation is complete. This bit is only valid when the
* Post-Sync Operation field is a value of 1h or 3h."
*/
if (mode & EMIT_INVALIDATE)
cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD;
intel_ring_emit(ring, cmd);
intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
if (INTEL_GEN(req->i915) >= 8) {
intel_ring_emit(ring, 0); /* upper addr */
intel_ring_emit(ring, 0); /* value */
} else {
intel_ring_emit(ring, 0);
intel_ring_emit(ring, MI_NOOP);
}
intel_ring_advance(ring);
return 0;
}
static int
gen8_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 len,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
bool ppgtt = USES_PPGTT(req->i915) &&
!(dispatch_flags & I915_DISPATCH_SECURE);
int ret;
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
/* FIXME(BDW): Address space and security selectors. */
intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8) |
(dispatch_flags & I915_DISPATCH_RS ?
MI_BATCH_RESOURCE_STREAMER : 0));
intel_ring_emit(ring, lower_32_bits(offset));
intel_ring_emit(ring, upper_32_bits(offset));
intel_ring_emit(ring, MI_NOOP);
intel_ring_advance(ring);
return 0;
}
static int
hsw_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 len,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring,
MI_BATCH_BUFFER_START |
(dispatch_flags & I915_DISPATCH_SECURE ?
0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW) |
(dispatch_flags & I915_DISPATCH_RS ?
MI_BATCH_RESOURCE_STREAMER : 0));
/* bit0-7 is the length on GEN6+ */
intel_ring_emit(ring, offset);
intel_ring_advance(ring);
return 0;
}
static int
gen6_emit_bb_start(struct drm_i915_gem_request *req,
u64 offset, u32 len,
unsigned int dispatch_flags)
{
struct intel_ring *ring = req->ring;
int ret;
ret = intel_ring_begin(req, 2);
if (ret)
return ret;
intel_ring_emit(ring,
MI_BATCH_BUFFER_START |
(dispatch_flags & I915_DISPATCH_SECURE ?
0 : MI_BATCH_NON_SECURE_I965));
/* bit0-7 is the length on GEN6+ */
intel_ring_emit(ring, offset);
intel_ring_advance(ring);
return 0;
}
/* Blitter support (SandyBridge+) */
static int gen6_ring_flush(struct drm_i915_gem_request *req, u32 mode)
{
struct intel_ring *ring = req->ring;
uint32_t cmd;
int ret;
ret = intel_ring_begin(req, 4);
if (ret)
return ret;
cmd = MI_FLUSH_DW;
if (INTEL_GEN(req->i915) >= 8)
cmd += 1;
/* We always require a command barrier so that subsequent
* commands, such as breadcrumb interrupts, are strictly ordered
* wrt the contents of the write cache being flushed to memory
* (and thus being coherent from the CPU).
*/
cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
/*
* Bspec vol 1c.3 - blitter engine command streamer:
* "If ENABLED, all TLBs will be invalidated once the flush
* operation is complete. This bit is only valid when the
* Post-Sync Operation field is a value of 1h or 3h."
*/
if (mode & EMIT_INVALIDATE)
cmd |= MI_INVALIDATE_TLB;
intel_ring_emit(ring, cmd);
intel_ring_emit(ring,
I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
if (INTEL_GEN(req->i915) >= 8) {
intel_ring_emit(ring, 0); /* upper addr */
intel_ring_emit(ring, 0); /* value */
} else {
intel_ring_emit(ring, 0);
intel_ring_emit(ring, MI_NOOP);
}
intel_ring_advance(ring);
return 0;
}
static void intel_ring_init_semaphores(struct drm_i915_private *dev_priv,
struct intel_engine_cs *engine)
{
struct drm_i915_gem_object *obj;
int ret, i;
if (!i915.semaphores)
return;
if (INTEL_GEN(dev_priv) >= 8 && !dev_priv->semaphore) {
struct i915_vma *vma;
obj = i915_gem_object_create(&dev_priv->drm, 4096);
if (IS_ERR(obj))
goto err;
vma = i915_vma_create(obj, &dev_priv->ggtt.base, NULL);
if (IS_ERR(vma))
goto err_obj;
ret = i915_gem_object_set_to_gtt_domain(obj, false);
if (ret)
goto err_obj;
ret = i915_vma_pin(vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
if (ret)
goto err_obj;
dev_priv->semaphore = vma;
}
if (INTEL_GEN(dev_priv) >= 8) {
u32 offset = i915_ggtt_offset(dev_priv->semaphore);
engine->semaphore.sync_to = gen8_ring_sync_to;
engine->semaphore.signal = gen8_xcs_signal;
for (i = 0; i < I915_NUM_ENGINES; i++) {
u32 ring_offset;
if (i != engine->id)
ring_offset = offset + GEN8_SEMAPHORE_OFFSET(engine->id, i);
else
ring_offset = MI_SEMAPHORE_SYNC_INVALID;
engine->semaphore.signal_ggtt[i] = ring_offset;
}
} else if (INTEL_GEN(dev_priv) >= 6) {
engine->semaphore.sync_to = gen6_ring_sync_to;
engine->semaphore.signal = gen6_signal;
/*
* The current semaphore is only applied on pre-gen8
* platform. And there is no VCS2 ring on the pre-gen8
* platform. So the semaphore between RCS and VCS2 is
* initialized as INVALID. Gen8 will initialize the
* sema between VCS2 and RCS later.
*/
for (i = 0; i < GEN6_NUM_SEMAPHORES; i++) {
static const struct {
u32 wait_mbox;
i915_reg_t mbox_reg;
} sem_data[GEN6_NUM_SEMAPHORES][GEN6_NUM_SEMAPHORES] = {
[RCS_HW] = {
[VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RV, .mbox_reg = GEN6_VRSYNC },
[BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RB, .mbox_reg = GEN6_BRSYNC },
[VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RVE, .mbox_reg = GEN6_VERSYNC },
},
[VCS_HW] = {
[RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VR, .mbox_reg = GEN6_RVSYNC },
[BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VB, .mbox_reg = GEN6_BVSYNC },
[VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VVE, .mbox_reg = GEN6_VEVSYNC },
},
[BCS_HW] = {
[RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BR, .mbox_reg = GEN6_RBSYNC },
[VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BV, .mbox_reg = GEN6_VBSYNC },
[VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BVE, .mbox_reg = GEN6_VEBSYNC },
},
[VECS_HW] = {
[RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VER, .mbox_reg = GEN6_RVESYNC },
[VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VEV, .mbox_reg = GEN6_VVESYNC },
[BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VEB, .mbox_reg = GEN6_BVESYNC },
},
};
u32 wait_mbox;
i915_reg_t mbox_reg;
if (i == engine->hw_id) {
wait_mbox = MI_SEMAPHORE_SYNC_INVALID;
mbox_reg = GEN6_NOSYNC;
} else {
wait_mbox = sem_data[engine->hw_id][i].wait_mbox;
mbox_reg = sem_data[engine->hw_id][i].mbox_reg;
}
engine->semaphore.mbox.wait[i] = wait_mbox;
engine->semaphore.mbox.signal[i] = mbox_reg;
}
}
return;
err_obj:
i915_gem_object_put(obj);
err:
DRM_DEBUG_DRIVER("Failed to allocate space for semaphores, disabling\n");
i915.semaphores = 0;
}
static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
struct intel_engine_cs *engine)
{
engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << engine->irq_shift;
if (INTEL_GEN(dev_priv) >= 8) {
engine->irq_enable = gen8_irq_enable;
engine->irq_disable = gen8_irq_disable;
engine->irq_seqno_barrier = gen6_seqno_barrier;
} else if (INTEL_GEN(dev_priv) >= 6) {
engine->irq_enable = gen6_irq_enable;
engine->irq_disable = gen6_irq_disable;
engine->irq_seqno_barrier = gen6_seqno_barrier;
} else if (INTEL_GEN(dev_priv) >= 5) {
engine->irq_enable = gen5_irq_enable;
engine->irq_disable = gen5_irq_disable;
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
engine->irq_seqno_barrier = gen5_seqno_barrier;
} else if (INTEL_GEN(dev_priv) >= 3) {
engine->irq_enable = i9xx_irq_enable;
engine->irq_disable = i9xx_irq_disable;
} else {
engine->irq_enable = i8xx_irq_enable;
engine->irq_disable = i8xx_irq_disable;
}
}
static void intel_ring_default_vfuncs(struct drm_i915_private *dev_priv,
struct intel_engine_cs *engine)
{
intel_ring_init_irq(dev_priv, engine);
intel_ring_init_semaphores(dev_priv, engine);
engine->init_hw = init_ring_common;
engine->emit_request = i9xx_emit_request;
if (i915.semaphores)
engine->emit_request = gen6_sema_emit_request;
engine->submit_request = i9xx_submit_request;
if (INTEL_GEN(dev_priv) >= 8)
engine->emit_bb_start = gen8_emit_bb_start;
else if (INTEL_GEN(dev_priv) >= 6)
engine->emit_bb_start = gen6_emit_bb_start;
else if (INTEL_GEN(dev_priv) >= 4)
engine->emit_bb_start = i965_emit_bb_start;
else if (IS_I830(dev_priv) || IS_845G(dev_priv))
engine->emit_bb_start = i830_emit_bb_start;
else
engine->emit_bb_start = i915_emit_bb_start;
}
int intel_init_render_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
int ret;
intel_ring_default_vfuncs(dev_priv, engine);
if (HAS_L3_DPF(dev_priv))
engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
if (INTEL_GEN(dev_priv) >= 8) {
engine->init_context = intel_rcs_ctx_init;
engine->emit_request = gen8_render_emit_request;
engine->emit_flush = gen8_render_ring_flush;
if (i915.semaphores)
engine->semaphore.signal = gen8_rcs_signal;
} else if (INTEL_GEN(dev_priv) >= 6) {
engine->init_context = intel_rcs_ctx_init;
engine->emit_flush = gen7_render_ring_flush;
if (IS_GEN6(dev_priv))
engine->emit_flush = gen6_render_ring_flush;
} else if (IS_GEN5(dev_priv)) {
engine->emit_flush = gen4_render_ring_flush;
} else {
if (INTEL_GEN(dev_priv) < 4)
engine->emit_flush = gen2_render_ring_flush;
else
engine->emit_flush = gen4_render_ring_flush;
engine->irq_enable_mask = I915_USER_INTERRUPT;
}
if (IS_HASWELL(dev_priv))
engine->emit_bb_start = hsw_emit_bb_start;
engine->init_hw = init_render_ring;
engine->cleanup = render_ring_cleanup;
ret = intel_init_ring_buffer(engine);
if (ret)
return ret;
drm/i915: Add a delay between interrupt and inspecting the final seqno (ilk) On Ironlake, there is no command nor register to ensure that the write from a MI_STORE command is completed (and coherent on the CPU) before the command parser continues. This means that the ordering between the seqno write and the subsequent user interrupt is undefined (like gen6+). So to ensure that the seqno write is completed after the final user interrupt we need to delay the read sufficiently to allow the write to complete. This delay is undefined by the bspec, and empirically requires 75us even though a register read combined with a clflush is less than 500ns. Hence, the delay is due to an on-chip buffer rather than the latency of the write to memory. Note that the render ring controls this by filling the PIPE_CONTROL fifo with stalling commands that force the earliest pipe-control with the seqno to be completed before the command parser continues. Given that we need a barrier operation for BSD, we may as well forgo the extra per-batch latency by using a common per-interrupt barrier. Studying the impact of adding the usleep shows that in both sequences of and individual synchronous no-op batches is negligible for the media engine (where the write now is unordered with the interrupt). Converting the render engine over from the current glutton of pie-controls over to the per-interrupt delays speeds up both the sequential and individual synchronous no-ops by 20% and 60%, respectively. This speed up holds even when looking at the throughput of small copies (4KiB->4MiB), both serial and synchronous, by about 20%. This is because despite adding a significant delay to the interrupt, in all likelihood we will see the seqno write without having to apply the barrier (only in the rare corner cases where the write is delayed on the last required is the delay necessary). Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=94307 Testcase: igt/gem_sync #ilk Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1467390209-3576-12-git-send-email-chris@chris-wilson.co.uk
2016-07-01 23:23:21 +07:00
if (INTEL_GEN(dev_priv) >= 6) {
ret = intel_engine_create_scratch(engine, 4096);
if (ret)
return ret;
} else if (HAS_BROKEN_CS_TLB(dev_priv)) {
ret = intel_engine_create_scratch(engine, I830_WA_SIZE);
if (ret)
return ret;
}
return 0;
}
int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
intel_ring_default_vfuncs(dev_priv, engine);
if (INTEL_GEN(dev_priv) >= 6) {
/* gen6 bsd needs a special wa for tail updates */
if (IS_GEN6(dev_priv))
engine->submit_request = gen6_bsd_submit_request;
engine->emit_flush = gen6_bsd_ring_flush;
if (INTEL_GEN(dev_priv) < 8)
engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
} else {
engine->mmio_base = BSD_RING_BASE;
engine->emit_flush = bsd_ring_flush;
if (IS_GEN5(dev_priv))
engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
else
engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
}
return intel_init_ring_buffer(engine);
}
/**
* Initialize the second BSD ring (eg. Broadwell GT3, Skylake GT3)
*/
int intel_init_bsd2_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
intel_ring_default_vfuncs(dev_priv, engine);
engine->emit_flush = gen6_bsd_ring_flush;
return intel_init_ring_buffer(engine);
}
int intel_init_blt_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
intel_ring_default_vfuncs(dev_priv, engine);
engine->emit_flush = gen6_ring_flush;
if (INTEL_GEN(dev_priv) < 8)
engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
return intel_init_ring_buffer(engine);
}
int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine)
{
struct drm_i915_private *dev_priv = engine->i915;
intel_ring_default_vfuncs(dev_priv, engine);
engine->emit_flush = gen6_ring_flush;
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
if (INTEL_GEN(dev_priv) < 8) {
engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
engine->irq_enable = hsw_vebox_irq_enable;
engine->irq_disable = hsw_vebox_irq_disable;
drm/i915/bdw: Implement interrupt changes The interrupt handling implementation remains the same as previous generations with the 4 types of registers, status, identity, mask, and enable. However the layout of where the bits go have changed entirely. To address these changes, all of the interrupt vfuncs needed special gen8 code. The way it works is there is a top level status register now which informs the interrupt service routine which unit caused the interrupt, and therefore which interrupt registers to read to process the interrupt. For display the division is quite logical, a set of interrupt registers for each pipe, and in addition to those, a set each for "misc" and port. For GT the things get a bit hairy, as seen by the code. Each of the GT units has it's own bits defined. They all look *very similar* and resides in 16 bits of a GT register. As an example, RCS and BCS share register 0. To compact the code a bit, at a slight expense to complexity, this is exactly how the code works as well. 2 structures are added to the ring buffer so that our ring buffer interrupt handling code knows which ring shares the interrupt registers, and a shift value (ie. the top or bottom 16 bits of the register). The above allows us to kept the interrupt register caching scheme, the per interrupt enables, and the code to mask and unmask interrupts relatively clean (again at the cost of some more complexity). Most of the GT units mentioned above are command streamers, and so the symmetry should work quite well for even the yet to be implemented rings which Broadwell adds. v2: Fixes up a couple of bugs, and is more verbose about errors in the Broadwell interrupt handler. v3: fix DE_MISC IER offset v4: Simplify interrupts: I totally misread the docs the first time I implemented interrupts, and so this should greatly simplify the mess. Unlike GEN6, we never touch the regular mask registers in irq_get/put. v5: Rebased on to of recent pch hotplug setup changes. v6: Fixup on top of moving num_pipes to intel_info. v7: Rebased on top of Egbert Eich's hpd irq handling rework. Also wired up ibx_hpd_irq_setup for gen8. v8: Rebase on top of Jani's asle handling rework. v9: Rebase on top of Ben's VECS enabling for Haswell, where he unfortunately went OCD on the gt irq #defines. Not that they're still not yet fully consistent: - Used the GT_RENDER_ #defines + bdw shifts. - Dropped the shift from the L3_PARITY stuff, seemed clearer. - s/irq_refcount/irq_refcount.gt/ v10: Squash in VECS enabling patches and the gen8_gt_irq_handler refactoring from Zhao Yakui <yakui.zhao@intel.com> v11: Rebase on top of the interrupt cleanups in upstream. v12: Rebase on top of Ben's DPF changes in upstream. v13: Drop bdw from the HAS_L3_DPF feature flag for now, it's unclear what exactly needs to be done. Requested by Ben. v14: Fix the patch. - Drop the mask of reserved bits and assorted logic, it doesn't match the spec. - Do the posting read inconditionally instead of commenting it out. - Add a GEN8_MASTER_IRQ_CONTROL definition and use it. - Fix up the GEN8_PIPE interrupt defines and give the GEN8_ prefixes - we actually will need to use them. - Enclose macros in do {} while (0) (checkpatch). - Clear DE_MISC interrupt bits only after having processed them. - Fix whitespace fail (checkpatch). - Fix overtly long lines where appropriate (checkpatch). - Don't use typedef'ed private_t (maintainer-scripts). - Align the function parameter list correctly. Signed-off-by: Ben Widawsky <ben@bwidawsk.net> (v4) Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch> bikeshed
2013-11-03 11:07:09 +07:00
}
return intel_init_ring_buffer(engine);
}