linux_dsm_epyc7002/drivers/gpu/drm/i915/i915_gem_userptr.c
Daniel Vetter 7741b547b6 drm/i915: Preallocate our mmu notifier workequeu to unbreak cpu hotplug deadlock
4.14-rc1 gained the fancy new cross-release support in lockdep, which
seems to have uncovered a few more rules about what is allowed and
isn't.

This one here seems to indicate that allocating a work-queue while
holding mmap_sem is a no-go, so let's try to preallocate it.

Of course another way to break this chain would be somewhere in the
cpu hotplug code, since this isn't the only trace we're finding now
which goes through msr_create_device.

Full lockdep splat:

======================================================
WARNING: possible circular locking dependency detected
4.14.0-rc1-CI-CI_DRM_3118+  Tainted: G     U
------------------------------------------------------
prime_mmap/1551 is trying to acquire lock:
 (cpu_hotplug_lock.rw_sem){++++}, at: [<ffffffff8109dbb7>] apply_workqueue_attrs+0x17/0x50

but task is already holding lock:
 (&dev_priv->mm_lock){+.+.}, at: [<ffffffffa01a7b2a>] i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

->  (&dev_priv->mm_lock){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __mutex_lock+0x86/0x9b0
       mutex_lock_nested+0x1b/0x20
       i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]
       i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
       drm_ioctl_kernel+0x69/0xb0
       drm_ioctl+0x2f9/0x3d0
       do_vfs_ioctl+0x94/0x670
       SyS_ioctl+0x41/0x70
       entry_SYSCALL_64_fastpath+0x1c/0xb1

->  (&mm->mmap_sem){++++}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __might_fault+0x68/0x90
       _copy_to_user+0x23/0x70
       filldir+0xa5/0x120
       dcache_readdir+0xf9/0x170
       iterate_dir+0x69/0x1a0
       SyS_getdents+0xa5/0x140
       entry_SYSCALL_64_fastpath+0x1c/0xb1

->  (&sb->s_type->i_mutex_key#5){++++}:
       down_write+0x3b/0x70
       handle_create+0xcb/0x1e0
       devtmpfsd+0x139/0x180
       kthread+0x152/0x190
       ret_from_fork+0x27/0x40

->  ((complete)&req.done){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       wait_for_common+0x58/0x210
       wait_for_completion+0x1d/0x20
       devtmpfs_create_node+0x13d/0x160
       device_add+0x5eb/0x620
       device_create_groups_vargs+0xe0/0xf0
       device_create+0x3a/0x40
       msr_device_create+0x2b/0x40
       cpuhp_invoke_callback+0xa3/0x840
       cpuhp_thread_fun+0x7a/0x150
       smpboot_thread_fn+0x18a/0x280
       kthread+0x152/0x190
       ret_from_fork+0x27/0x40

->  (cpuhp_state){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       cpuhp_issue_call+0x10b/0x170
       __cpuhp_setup_state_cpuslocked+0x134/0x2a0
       __cpuhp_setup_state+0x46/0x60
       page_writeback_init+0x43/0x67
       pagecache_init+0x3d/0x42
       start_kernel+0x3a8/0x3fc
       x86_64_start_reservations+0x2a/0x2c
       x86_64_start_kernel+0x6d/0x70
       verify_cpu+0x0/0xfb

->  (cpuhp_state_mutex){+.+.}:
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       __mutex_lock+0x86/0x9b0
       mutex_lock_nested+0x1b/0x20
       __cpuhp_setup_state_cpuslocked+0x52/0x2a0
       __cpuhp_setup_state+0x46/0x60
       page_alloc_init+0x28/0x30
       start_kernel+0x145/0x3fc
       x86_64_start_reservations+0x2a/0x2c
       x86_64_start_kernel+0x6d/0x70
       verify_cpu+0x0/0xfb

->  (cpu_hotplug_lock.rw_sem){++++}:
       check_prev_add+0x430/0x840
       __lock_acquire+0x1420/0x15e0
       lock_acquire+0xb0/0x200
       cpus_read_lock+0x3d/0xb0
       apply_workqueue_attrs+0x17/0x50
       __alloc_workqueue_key+0x1d8/0x4d9
       i915_gem_userptr_init__mmu_notifier+0x1fb/0x270 [i915]
       i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
       drm_ioctl_kernel+0x69/0xb0
       drm_ioctl+0x2f9/0x3d0
       do_vfs_ioctl+0x94/0x670
       SyS_ioctl+0x41/0x70
       entry_SYSCALL_64_fastpath+0x1c/0xb1

other info that might help us debug this:

Chain exists of:
  cpu_hotplug_lock.rw_sem --> &mm->mmap_sem --> &dev_priv->mm_lock

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&dev_priv->mm_lock);
                               lock(&mm->mmap_sem);
                               lock(&dev_priv->mm_lock);
  lock(cpu_hotplug_lock.rw_sem);

 *** DEADLOCK ***

2 locks held by prime_mmap/1551:
 :  (&mm->mmap_sem){++++}, at: [<ffffffffa01a7b18>] i915_gem_userptr_init__mmu_notifier+0x138/0x270 [i915]
 :  (&dev_priv->mm_lock){+.+.}, at: [<ffffffffa01a7b2a>] i915_gem_userptr_init__mmu_notifier+0x14a/0x270 [i915]

stack backtrace:
CPU: 4 PID: 1551 Comm: prime_mmap Tainted: G     U          4.14.0-rc1-CI-CI_DRM_3118+ 
Hardware name: Dell Inc. XPS 8300  /0Y2MRG, BIOS A06 10/17/2011
Call Trace:
 dump_stack+0x68/0x9f
 print_circular_bug+0x235/0x3c0
 ? lockdep_init_map_crosslock+0x20/0x20
 check_prev_add+0x430/0x840
 __lock_acquire+0x1420/0x15e0
 ? __lock_acquire+0x1420/0x15e0
 ? lockdep_init_map_crosslock+0x20/0x20
 lock_acquire+0xb0/0x200
 ? apply_workqueue_attrs+0x17/0x50
 cpus_read_lock+0x3d/0xb0
 ? apply_workqueue_attrs+0x17/0x50
 apply_workqueue_attrs+0x17/0x50
 __alloc_workqueue_key+0x1d8/0x4d9
 ? __lockdep_init_map+0x57/0x1c0
 i915_gem_userptr_init__mmu_notifier+0x1fb/0x270 [i915]
 i915_gem_userptr_ioctl+0x222/0x2c0 [i915]
 ? i915_gem_userptr_release+0x140/0x140 [i915]
 drm_ioctl_kernel+0x69/0xb0
 drm_ioctl+0x2f9/0x3d0
 ? i915_gem_userptr_release+0x140/0x140 [i915]
 ? __do_page_fault+0x2a4/0x570
 do_vfs_ioctl+0x94/0x670
 ? entry_SYSCALL_64_fastpath+0x5/0xb1
 ? __this_cpu_preempt_check+0x13/0x20
 ? trace_hardirqs_on_caller+0xe3/0x1b0
 SyS_ioctl+0x41/0x70
 entry_SYSCALL_64_fastpath+0x1c/0xb1
RIP: 0033:0x7fbb83c39587
RSP: 002b:00007fff188dc228 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
RAX: ffffffffffffffda RBX: ffffffff81492963 RCX: 00007fbb83c39587
RDX: 00007fff188dc260 RSI: 00000000c0186473 RDI: 0000000000000003
RBP: ffffc90001487f88 R08: 0000000000000000 R09: 00007fff188dc2ac
R10: 00007fbb83efcb58 R11: 0000000000000246 R12: 0000000000000000
R13: 0000000000000003 R14: 00000000c0186473 R15: 00007fff188dc2ac
 ? __this_cpu_preempt_check+0x13/0x20

Note that this also has the minor benefit of slightly reducing the
critical section where we hold mmap_sem.

v2: Set ret correctly when we raced with another thread.

v3: Use Chris' diff. Attach the right lockdep splat.

v4: Repaint in Tvrtko's colors (aka don't report ENOMEM if we race and
some other thread managed to not also get an ENOMEM and successfully
install the mmu notifier. Note that the kernel guarantees that small
allocations succeed, so this never actually happens).

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Sasha Levin <alexander.levin@verizon.com>
Cc: Marta Lofstedt <marta.lofstedt@intel.com>
Cc: Tejun Heo <tj@kernel.org>
References: https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_3180/shard-hsw3/igt@prime_mmap@test_userptr.html
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=102939
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171009164401.16035-1-daniel.vetter@ffwll.ch
2017-10-10 12:57:03 +02:00

840 lines
22 KiB
C

/*
* Copyright © 2012-2014 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.
*
*/
#include <drm/drmP.h>
#include <drm/i915_drm.h>
#include "i915_drv.h"
#include "i915_trace.h"
#include "intel_drv.h"
#include <linux/mmu_context.h>
#include <linux/mmu_notifier.h>
#include <linux/mempolicy.h>
#include <linux/swap.h>
#include <linux/sched/mm.h>
struct i915_mm_struct {
struct mm_struct *mm;
struct drm_i915_private *i915;
struct i915_mmu_notifier *mn;
struct hlist_node node;
struct kref kref;
struct work_struct work;
};
#if defined(CONFIG_MMU_NOTIFIER)
#include <linux/interval_tree.h>
struct i915_mmu_notifier {
spinlock_t lock;
struct hlist_node node;
struct mmu_notifier mn;
struct rb_root_cached objects;
struct workqueue_struct *wq;
};
struct i915_mmu_object {
struct i915_mmu_notifier *mn;
struct drm_i915_gem_object *obj;
struct interval_tree_node it;
struct list_head link;
struct work_struct work;
bool attached;
};
static void cancel_userptr(struct work_struct *work)
{
struct i915_mmu_object *mo = container_of(work, typeof(*mo), work);
struct drm_i915_gem_object *obj = mo->obj;
struct work_struct *active;
/* Cancel any active worker and force us to re-evaluate gup */
mutex_lock(&obj->mm.lock);
active = fetch_and_zero(&obj->userptr.work);
mutex_unlock(&obj->mm.lock);
if (active)
goto out;
i915_gem_object_wait(obj, I915_WAIT_ALL, MAX_SCHEDULE_TIMEOUT, NULL);
mutex_lock(&obj->base.dev->struct_mutex);
/* We are inside a kthread context and can't be interrupted */
if (i915_gem_object_unbind(obj) == 0)
__i915_gem_object_put_pages(obj, I915_MM_NORMAL);
WARN_ONCE(obj->mm.pages,
"Failed to release pages: bind_count=%d, pages_pin_count=%d, pin_display=%d\n",
obj->bind_count,
atomic_read(&obj->mm.pages_pin_count),
obj->pin_display);
mutex_unlock(&obj->base.dev->struct_mutex);
out:
i915_gem_object_put(obj);
}
static void add_object(struct i915_mmu_object *mo)
{
if (mo->attached)
return;
interval_tree_insert(&mo->it, &mo->mn->objects);
mo->attached = true;
}
static void del_object(struct i915_mmu_object *mo)
{
if (!mo->attached)
return;
interval_tree_remove(&mo->it, &mo->mn->objects);
mo->attached = false;
}
static void i915_gem_userptr_mn_invalidate_range_start(struct mmu_notifier *_mn,
struct mm_struct *mm,
unsigned long start,
unsigned long end)
{
struct i915_mmu_notifier *mn =
container_of(_mn, struct i915_mmu_notifier, mn);
struct i915_mmu_object *mo;
struct interval_tree_node *it;
LIST_HEAD(cancelled);
if (RB_EMPTY_ROOT(&mn->objects.rb_root))
return;
/* interval ranges are inclusive, but invalidate range is exclusive */
end--;
spin_lock(&mn->lock);
it = interval_tree_iter_first(&mn->objects, start, end);
while (it) {
/* The mmu_object is released late when destroying the
* GEM object so it is entirely possible to gain a
* reference on an object in the process of being freed
* since our serialisation is via the spinlock and not
* the struct_mutex - and consequently use it after it
* is freed and then double free it. To prevent that
* use-after-free we only acquire a reference on the
* object if it is not in the process of being destroyed.
*/
mo = container_of(it, struct i915_mmu_object, it);
if (kref_get_unless_zero(&mo->obj->base.refcount))
queue_work(mn->wq, &mo->work);
list_add(&mo->link, &cancelled);
it = interval_tree_iter_next(it, start, end);
}
list_for_each_entry(mo, &cancelled, link)
del_object(mo);
spin_unlock(&mn->lock);
if (!list_empty(&cancelled))
flush_workqueue(mn->wq);
}
static const struct mmu_notifier_ops i915_gem_userptr_notifier = {
.invalidate_range_start = i915_gem_userptr_mn_invalidate_range_start,
};
static struct i915_mmu_notifier *
i915_mmu_notifier_create(struct mm_struct *mm)
{
struct i915_mmu_notifier *mn;
mn = kmalloc(sizeof(*mn), GFP_KERNEL);
if (mn == NULL)
return ERR_PTR(-ENOMEM);
spin_lock_init(&mn->lock);
mn->mn.ops = &i915_gem_userptr_notifier;
mn->objects = RB_ROOT_CACHED;
mn->wq = alloc_workqueue("i915-userptr-release", WQ_UNBOUND, 0);
if (mn->wq == NULL) {
kfree(mn);
return ERR_PTR(-ENOMEM);
}
return mn;
}
static void
i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
{
struct i915_mmu_object *mo;
mo = obj->userptr.mmu_object;
if (mo == NULL)
return;
spin_lock(&mo->mn->lock);
del_object(mo);
spin_unlock(&mo->mn->lock);
kfree(mo);
obj->userptr.mmu_object = NULL;
}
static struct i915_mmu_notifier *
i915_mmu_notifier_find(struct i915_mm_struct *mm)
{
struct i915_mmu_notifier *mn;
int err = 0;
mn = mm->mn;
if (mn)
return mn;
mn = i915_mmu_notifier_create(mm->mm);
if (IS_ERR(mn))
err = PTR_ERR(mn);
down_write(&mm->mm->mmap_sem);
mutex_lock(&mm->i915->mm_lock);
if (mm->mn == NULL && !err) {
/* Protected by mmap_sem (write-lock) */
err = __mmu_notifier_register(&mn->mn, mm->mm);
if (!err) {
/* Protected by mm_lock */
mm->mn = fetch_and_zero(&mn);
}
} else {
/* someone else raced and successfully installed the mmu
* notifier, we can cancel our own errors */
err = 0;
}
mutex_unlock(&mm->i915->mm_lock);
up_write(&mm->mm->mmap_sem);
if (mn) {
destroy_workqueue(mn->wq);
kfree(mn);
}
return err ? ERR_PTR(err) : mm->mn;
}
static int
i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
unsigned flags)
{
struct i915_mmu_notifier *mn;
struct i915_mmu_object *mo;
if (flags & I915_USERPTR_UNSYNCHRONIZED)
return capable(CAP_SYS_ADMIN) ? 0 : -EPERM;
if (WARN_ON(obj->userptr.mm == NULL))
return -EINVAL;
mn = i915_mmu_notifier_find(obj->userptr.mm);
if (IS_ERR(mn))
return PTR_ERR(mn);
mo = kzalloc(sizeof(*mo), GFP_KERNEL);
if (mo == NULL)
return -ENOMEM;
mo->mn = mn;
mo->obj = obj;
mo->it.start = obj->userptr.ptr;
mo->it.last = obj->userptr.ptr + obj->base.size - 1;
INIT_WORK(&mo->work, cancel_userptr);
obj->userptr.mmu_object = mo;
return 0;
}
static void
i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
struct mm_struct *mm)
{
if (mn == NULL)
return;
mmu_notifier_unregister(&mn->mn, mm);
destroy_workqueue(mn->wq);
kfree(mn);
}
#else
static void
i915_gem_userptr_release__mmu_notifier(struct drm_i915_gem_object *obj)
{
}
static int
i915_gem_userptr_init__mmu_notifier(struct drm_i915_gem_object *obj,
unsigned flags)
{
if ((flags & I915_USERPTR_UNSYNCHRONIZED) == 0)
return -ENODEV;
if (!capable(CAP_SYS_ADMIN))
return -EPERM;
return 0;
}
static void
i915_mmu_notifier_free(struct i915_mmu_notifier *mn,
struct mm_struct *mm)
{
}
#endif
static struct i915_mm_struct *
__i915_mm_struct_find(struct drm_i915_private *dev_priv, struct mm_struct *real)
{
struct i915_mm_struct *mm;
/* Protected by dev_priv->mm_lock */
hash_for_each_possible(dev_priv->mm_structs, mm, node, (unsigned long)real)
if (mm->mm == real)
return mm;
return NULL;
}
static int
i915_gem_userptr_init__mm_struct(struct drm_i915_gem_object *obj)
{
struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
struct i915_mm_struct *mm;
int ret = 0;
/* During release of the GEM object we hold the struct_mutex. This
* precludes us from calling mmput() at that time as that may be
* the last reference and so call exit_mmap(). exit_mmap() will
* attempt to reap the vma, and if we were holding a GTT mmap
* would then call drm_gem_vm_close() and attempt to reacquire
* the struct mutex. So in order to avoid that recursion, we have
* to defer releasing the mm reference until after we drop the
* struct_mutex, i.e. we need to schedule a worker to do the clean
* up.
*/
mutex_lock(&dev_priv->mm_lock);
mm = __i915_mm_struct_find(dev_priv, current->mm);
if (mm == NULL) {
mm = kmalloc(sizeof(*mm), GFP_KERNEL);
if (mm == NULL) {
ret = -ENOMEM;
goto out;
}
kref_init(&mm->kref);
mm->i915 = to_i915(obj->base.dev);
mm->mm = current->mm;
mmgrab(current->mm);
mm->mn = NULL;
/* Protected by dev_priv->mm_lock */
hash_add(dev_priv->mm_structs,
&mm->node, (unsigned long)mm->mm);
} else
kref_get(&mm->kref);
obj->userptr.mm = mm;
out:
mutex_unlock(&dev_priv->mm_lock);
return ret;
}
static void
__i915_mm_struct_free__worker(struct work_struct *work)
{
struct i915_mm_struct *mm = container_of(work, typeof(*mm), work);
i915_mmu_notifier_free(mm->mn, mm->mm);
mmdrop(mm->mm);
kfree(mm);
}
static void
__i915_mm_struct_free(struct kref *kref)
{
struct i915_mm_struct *mm = container_of(kref, typeof(*mm), kref);
/* Protected by dev_priv->mm_lock */
hash_del(&mm->node);
mutex_unlock(&mm->i915->mm_lock);
INIT_WORK(&mm->work, __i915_mm_struct_free__worker);
queue_work(mm->i915->mm.userptr_wq, &mm->work);
}
static void
i915_gem_userptr_release__mm_struct(struct drm_i915_gem_object *obj)
{
if (obj->userptr.mm == NULL)
return;
kref_put_mutex(&obj->userptr.mm->kref,
__i915_mm_struct_free,
&to_i915(obj->base.dev)->mm_lock);
obj->userptr.mm = NULL;
}
struct get_pages_work {
struct work_struct work;
struct drm_i915_gem_object *obj;
struct task_struct *task;
};
static struct sg_table *
__i915_gem_userptr_alloc_pages(struct drm_i915_gem_object *obj,
struct page **pvec, int num_pages)
{
unsigned int max_segment = i915_sg_segment_size();
struct sg_table *st;
unsigned int sg_page_sizes;
int ret;
st = kmalloc(sizeof(*st), GFP_KERNEL);
if (!st)
return ERR_PTR(-ENOMEM);
alloc_table:
ret = __sg_alloc_table_from_pages(st, pvec, num_pages,
0, num_pages << PAGE_SHIFT,
max_segment,
GFP_KERNEL);
if (ret) {
kfree(st);
return ERR_PTR(ret);
}
ret = i915_gem_gtt_prepare_pages(obj, st);
if (ret) {
sg_free_table(st);
if (max_segment > PAGE_SIZE) {
max_segment = PAGE_SIZE;
goto alloc_table;
}
kfree(st);
return ERR_PTR(ret);
}
sg_page_sizes = i915_sg_page_sizes(st->sgl);
__i915_gem_object_set_pages(obj, st, sg_page_sizes);
return st;
}
static int
__i915_gem_userptr_set_active(struct drm_i915_gem_object *obj,
bool value)
{
int ret = 0;
/* During mm_invalidate_range we need to cancel any userptr that
* overlaps the range being invalidated. Doing so requires the
* struct_mutex, and that risks recursion. In order to cause
* recursion, the user must alias the userptr address space with
* a GTT mmapping (possible with a MAP_FIXED) - then when we have
* to invalidate that mmaping, mm_invalidate_range is called with
* the userptr address *and* the struct_mutex held. To prevent that
* we set a flag under the i915_mmu_notifier spinlock to indicate
* whether this object is valid.
*/
#if defined(CONFIG_MMU_NOTIFIER)
if (obj->userptr.mmu_object == NULL)
return 0;
spin_lock(&obj->userptr.mmu_object->mn->lock);
/* In order to serialise get_pages with an outstanding
* cancel_userptr, we must drop the struct_mutex and try again.
*/
if (!value)
del_object(obj->userptr.mmu_object);
else if (!work_pending(&obj->userptr.mmu_object->work))
add_object(obj->userptr.mmu_object);
else
ret = -EAGAIN;
spin_unlock(&obj->userptr.mmu_object->mn->lock);
#endif
return ret;
}
static void
__i915_gem_userptr_get_pages_worker(struct work_struct *_work)
{
struct get_pages_work *work = container_of(_work, typeof(*work), work);
struct drm_i915_gem_object *obj = work->obj;
const int npages = obj->base.size >> PAGE_SHIFT;
struct page **pvec;
int pinned, ret;
ret = -ENOMEM;
pinned = 0;
pvec = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
if (pvec != NULL) {
struct mm_struct *mm = obj->userptr.mm->mm;
unsigned int flags = 0;
if (!obj->userptr.read_only)
flags |= FOLL_WRITE;
ret = -EFAULT;
if (mmget_not_zero(mm)) {
down_read(&mm->mmap_sem);
while (pinned < npages) {
ret = get_user_pages_remote
(work->task, mm,
obj->userptr.ptr + pinned * PAGE_SIZE,
npages - pinned,
flags,
pvec + pinned, NULL, NULL);
if (ret < 0)
break;
pinned += ret;
}
up_read(&mm->mmap_sem);
mmput(mm);
}
}
mutex_lock(&obj->mm.lock);
if (obj->userptr.work == &work->work) {
struct sg_table *pages = ERR_PTR(ret);
if (pinned == npages) {
pages = __i915_gem_userptr_alloc_pages(obj, pvec,
npages);
if (!IS_ERR(pages)) {
pinned = 0;
pages = NULL;
}
}
obj->userptr.work = ERR_CAST(pages);
if (IS_ERR(pages))
__i915_gem_userptr_set_active(obj, false);
}
mutex_unlock(&obj->mm.lock);
release_pages(pvec, pinned, 0);
kvfree(pvec);
i915_gem_object_put(obj);
put_task_struct(work->task);
kfree(work);
}
static struct sg_table *
__i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
{
struct get_pages_work *work;
/* Spawn a worker so that we can acquire the
* user pages without holding our mutex. Access
* to the user pages requires mmap_sem, and we have
* a strict lock ordering of mmap_sem, struct_mutex -
* we already hold struct_mutex here and so cannot
* call gup without encountering a lock inversion.
*
* Userspace will keep on repeating the operation
* (thanks to EAGAIN) until either we hit the fast
* path or the worker completes. If the worker is
* cancelled or superseded, the task is still run
* but the results ignored. (This leads to
* complications that we may have a stray object
* refcount that we need to be wary of when
* checking for existing objects during creation.)
* If the worker encounters an error, it reports
* that error back to this function through
* obj->userptr.work = ERR_PTR.
*/
work = kmalloc(sizeof(*work), GFP_KERNEL);
if (work == NULL)
return ERR_PTR(-ENOMEM);
obj->userptr.work = &work->work;
work->obj = i915_gem_object_get(obj);
work->task = current;
get_task_struct(work->task);
INIT_WORK(&work->work, __i915_gem_userptr_get_pages_worker);
queue_work(to_i915(obj->base.dev)->mm.userptr_wq, &work->work);
return ERR_PTR(-EAGAIN);
}
static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
{
const int num_pages = obj->base.size >> PAGE_SHIFT;
struct mm_struct *mm = obj->userptr.mm->mm;
struct page **pvec;
struct sg_table *pages;
bool active;
int pinned;
/* If userspace should engineer that these pages are replaced in
* the vma between us binding this page into the GTT and completion
* of rendering... Their loss. If they change the mapping of their
* pages they need to create a new bo to point to the new vma.
*
* However, that still leaves open the possibility of the vma
* being copied upon fork. Which falls under the same userspace
* synchronisation issue as a regular bo, except that this time
* the process may not be expecting that a particular piece of
* memory is tied to the GPU.
*
* Fortunately, we can hook into the mmu_notifier in order to
* discard the page references prior to anything nasty happening
* to the vma (discard or cloning) which should prevent the more
* egregious cases from causing harm.
*/
if (obj->userptr.work) {
/* active flag should still be held for the pending work */
if (IS_ERR(obj->userptr.work))
return PTR_ERR(obj->userptr.work);
else
return -EAGAIN;
}
pvec = NULL;
pinned = 0;
if (mm == current->mm) {
pvec = kvmalloc_array(num_pages, sizeof(struct page *),
GFP_KERNEL |
__GFP_NORETRY |
__GFP_NOWARN);
if (pvec) /* defer to worker if malloc fails */
pinned = __get_user_pages_fast(obj->userptr.ptr,
num_pages,
!obj->userptr.read_only,
pvec);
}
active = false;
if (pinned < 0) {
pages = ERR_PTR(pinned);
pinned = 0;
} else if (pinned < num_pages) {
pages = __i915_gem_userptr_get_pages_schedule(obj);
active = pages == ERR_PTR(-EAGAIN);
} else {
pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
active = !IS_ERR(pages);
}
if (active)
__i915_gem_userptr_set_active(obj, true);
if (IS_ERR(pages))
release_pages(pvec, pinned, 0);
kvfree(pvec);
return PTR_ERR_OR_ZERO(pages);
}
static void
i915_gem_userptr_put_pages(struct drm_i915_gem_object *obj,
struct sg_table *pages)
{
struct sgt_iter sgt_iter;
struct page *page;
BUG_ON(obj->userptr.work != NULL);
__i915_gem_userptr_set_active(obj, false);
if (obj->mm.madv != I915_MADV_WILLNEED)
obj->mm.dirty = false;
i915_gem_gtt_finish_pages(obj, pages);
for_each_sgt_page(page, sgt_iter, pages) {
if (obj->mm.dirty)
set_page_dirty(page);
mark_page_accessed(page);
put_page(page);
}
obj->mm.dirty = false;
sg_free_table(pages);
kfree(pages);
}
static void
i915_gem_userptr_release(struct drm_i915_gem_object *obj)
{
i915_gem_userptr_release__mmu_notifier(obj);
i915_gem_userptr_release__mm_struct(obj);
}
static int
i915_gem_userptr_dmabuf_export(struct drm_i915_gem_object *obj)
{
if (obj->userptr.mmu_object)
return 0;
return i915_gem_userptr_init__mmu_notifier(obj, 0);
}
static const struct drm_i915_gem_object_ops i915_gem_userptr_ops = {
.flags = I915_GEM_OBJECT_HAS_STRUCT_PAGE |
I915_GEM_OBJECT_IS_SHRINKABLE,
.get_pages = i915_gem_userptr_get_pages,
.put_pages = i915_gem_userptr_put_pages,
.dmabuf_export = i915_gem_userptr_dmabuf_export,
.release = i915_gem_userptr_release,
};
/**
* Creates a new mm object that wraps some normal memory from the process
* context - user memory.
*
* We impose several restrictions upon the memory being mapped
* into the GPU.
* 1. It must be page aligned (both start/end addresses, i.e ptr and size).
* 2. It must be normal system memory, not a pointer into another map of IO
* space (e.g. it must not be a GTT mmapping of another object).
* 3. We only allow a bo as large as we could in theory map into the GTT,
* that is we limit the size to the total size of the GTT.
* 4. The bo is marked as being snoopable. The backing pages are left
* accessible directly by the CPU, but reads and writes by the GPU may
* incur the cost of a snoop (unless you have an LLC architecture).
*
* Synchronisation between multiple users and the GPU is left to userspace
* through the normal set-domain-ioctl. The kernel will enforce that the
* GPU relinquishes the VMA before it is returned back to the system
* i.e. upon free(), munmap() or process termination. However, the userspace
* malloc() library may not immediately relinquish the VMA after free() and
* instead reuse it whilst the GPU is still reading and writing to the VMA.
* Caveat emptor.
*
* Also note, that the object created here is not currently a "first class"
* object, in that several ioctls are banned. These are the CPU access
* ioctls: mmap(), pwrite and pread. In practice, you are expected to use
* direct access via your pointer rather than use those ioctls. Another
* restriction is that we do not allow userptr surfaces to be pinned to the
* hardware and so we reject any attempt to create a framebuffer out of a
* userptr.
*
* If you think this is a good interface to use to pass GPU memory between
* drivers, please use dma-buf instead. In fact, wherever possible use
* dma-buf instead.
*/
int
i915_gem_userptr_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
{
struct drm_i915_private *dev_priv = to_i915(dev);
struct drm_i915_gem_userptr *args = data;
struct drm_i915_gem_object *obj;
int ret;
u32 handle;
if (!HAS_LLC(dev_priv) && !HAS_SNOOP(dev_priv)) {
/* We cannot support coherent userptr objects on hw without
* LLC and broken snooping.
*/
return -ENODEV;
}
if (args->flags & ~(I915_USERPTR_READ_ONLY |
I915_USERPTR_UNSYNCHRONIZED))
return -EINVAL;
if (offset_in_page(args->user_ptr | args->user_size))
return -EINVAL;
if (!access_ok(args->flags & I915_USERPTR_READ_ONLY ? VERIFY_READ : VERIFY_WRITE,
(char __user *)(unsigned long)args->user_ptr, args->user_size))
return -EFAULT;
if (args->flags & I915_USERPTR_READ_ONLY) {
/* On almost all of the current hw, we cannot tell the GPU that a
* page is readonly, so this is just a placeholder in the uAPI.
*/
return -ENODEV;
}
obj = i915_gem_object_alloc(dev_priv);
if (obj == NULL)
return -ENOMEM;
drm_gem_private_object_init(dev, &obj->base, args->user_size);
i915_gem_object_init(obj, &i915_gem_userptr_ops);
obj->base.read_domains = I915_GEM_DOMAIN_CPU;
obj->base.write_domain = I915_GEM_DOMAIN_CPU;
i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
obj->userptr.ptr = args->user_ptr;
obj->userptr.read_only = !!(args->flags & I915_USERPTR_READ_ONLY);
/* And keep a pointer to the current->mm for resolving the user pages
* at binding. This means that we need to hook into the mmu_notifier
* in order to detect if the mmu is destroyed.
*/
ret = i915_gem_userptr_init__mm_struct(obj);
if (ret == 0)
ret = i915_gem_userptr_init__mmu_notifier(obj, args->flags);
if (ret == 0)
ret = drm_gem_handle_create(file, &obj->base, &handle);
/* drop reference from allocate - handle holds it now */
i915_gem_object_put(obj);
if (ret)
return ret;
args->handle = handle;
return 0;
}
int i915_gem_init_userptr(struct drm_i915_private *dev_priv)
{
mutex_init(&dev_priv->mm_lock);
hash_init(dev_priv->mm_structs);
dev_priv->mm.userptr_wq =
alloc_workqueue("i915-userptr-acquire",
WQ_HIGHPRI | WQ_MEM_RECLAIM,
0);
if (!dev_priv->mm.userptr_wq)
return -ENOMEM;
return 0;
}
void i915_gem_cleanup_userptr(struct drm_i915_private *dev_priv)
{
destroy_workqueue(dev_priv->mm.userptr_wq);
}