AuxXxilium/linux_dsm_epyc7002
1
0
Fork 0
mirror of https://github.com/AuxXxilium/linux_dsm_epyc7002.git synced 2024-12-15 11:16:40 +07:00
Code Issues Actions Packages Projects Releases Wiki Activity
f72180cc93
linux_dsm_epyc7002/arch/powerpc/kernel/tm.S
Cyril Bur eb5c3f1c86 powerpc: Always save/restore checkpointed regs during treclaim/trecheckpoint 
Lazy save and restore of FP/Altivec means that a userspace process can
be sent to userspace with FP or Altivec disabled and loaded only as
required (by way of an FP/Altivec unavailable exception). Transactional
Memory complicates this situation as a transaction could be started
without FP/Altivec being loaded up. This causes the hardware to
checkpoint incorrect registers. Handling FP/Altivec unavailable
exceptions while a thread is transactional requires a reclaim and
recheckpoint to ensure the CPU has correct state for both sets of
registers.

tm_reclaim() has optimisations to not always save the FP/Altivec
registers to the checkpointed save area. This was originally done
because the caller might have information that the checkpointed
registers aren't valid due to lazy save and restore. We've also been a
little vague as to how tm_reclaim() leaves the FP/Altivec state since it
doesn't necessarily always save it to the thread struct. This has lead
to an (incorrect) assumption that it leaves the checkpointed state on
the CPU.

tm_recheckpoint() has similar optimisations in reverse. It may not
always reload the checkpointed FP/Altivec registers from the thread
struct before the trecheckpoint. It is therefore quite unclear where it
expects to get the state from. This didn't help with the assumption
made about tm_reclaim().

These optimisations sit in what is by definition a slow path. If a
process has to go through a reclaim/recheckpoint then its transaction
will be doomed on returning to userspace. This mean that the process
will be unable to complete its transaction and be forced to its failure
handler. This is already an out if line case for userspace. Furthermore,
the cost of copying 64 times 128 bits from registers isn't very long[0]
(at all) on modern processors. As such it appears these optimisations
have only served to increase code complexity and are unlikely to have
had a measurable performance impact.

Our transactional memory handling has been riddled with bugs. A cause
of this has been difficulty in following the code flow, code complexity
has not been our friend here. It makes sense to remove these
optimisations in favour of a (hopefully) more stable implementation.

This patch does mean that some times the assembly will needlessly save
'junk' registers which will subsequently get overwritten with the
correct value by the C code which calls the assembly function. This
small inefficiency is far outweighed by the reduction in complexity for
general TM code, context switching paths, and transactional facility
unavailable exception handler.

0: I tried to measure it once for other work and found that it was
hiding in the noise of everything else I was working with. I find it
exceedingly likely this will be the case here.

Signed-off-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2017-11-06 20:39:33 +11:00

489 lines
12 KiB
ArmAsm
Raw Blame History

/*
* Transactional memory support routines to reclaim and recheckpoint
* transactional process state.
*
* Copyright 2012 Matt Evans & Michael Neuling, IBM Corporation.
*/
#include <asm/asm-offsets.h>
#include <asm/ppc_asm.h>
#include <asm/ppc-opcode.h>
#include <asm/ptrace.h>
#include <asm/reg.h>
#include <asm/bug.h>
#ifdef CONFIG_VSX
/* See fpu.S, this is borrowed from there */
#define __SAVE_32FPRS_VSRS(n,c,base) \
BEGIN_FTR_SECTION \
b 2f; \
END_FTR_SECTION_IFSET(CPU_FTR_VSX); \
SAVE_32FPRS(n,base); \
b 3f; \
2: SAVE_32VSRS(n,c,base); \
3:
#define __REST_32FPRS_VSRS(n,c,base) \
BEGIN_FTR_SECTION \
b 2f; \
END_FTR_SECTION_IFSET(CPU_FTR_VSX); \
REST_32FPRS(n,base); \
b 3f; \
2: REST_32VSRS(n,c,base); \
3:
#else
#define __SAVE_32FPRS_VSRS(n,c,base) SAVE_32FPRS(n, base)
#define __REST_32FPRS_VSRS(n,c,base) REST_32FPRS(n, base)
#endif
#define SAVE_32FPRS_VSRS(n,c,base) \
__SAVE_32FPRS_VSRS(n,__REG_##c,__REG_##base)
#define REST_32FPRS_VSRS(n,c,base) \
__REST_32FPRS_VSRS(n,__REG_##c,__REG_##base)
/* Stack frame offsets for local variables. */
#define TM_FRAME_L0 TM_FRAME_SIZE-16
#define TM_FRAME_L1 TM_FRAME_SIZE-8
/* In order to access the TM SPRs, TM must be enabled. So, do so: */
_GLOBAL(tm_enable)
mfmsr r4
li r3, MSR_TM >> 32
sldi r3, r3, 32
and. r0, r4, r3
bne 1f
or r4, r4, r3
mtmsrd r4
1: blr
_GLOBAL(tm_save_sprs)
mfspr r0, SPRN_TFHAR
std r0, THREAD_TM_TFHAR(r3)
mfspr r0, SPRN_TEXASR
std r0, THREAD_TM_TEXASR(r3)
mfspr r0, SPRN_TFIAR
std r0, THREAD_TM_TFIAR(r3)
blr
_GLOBAL(tm_restore_sprs)
ld r0, THREAD_TM_TFHAR(r3)
mtspr SPRN_TFHAR, r0
ld r0, THREAD_TM_TEXASR(r3)
mtspr SPRN_TEXASR, r0
ld r0, THREAD_TM_TFIAR(r3)
mtspr SPRN_TFIAR, r0
blr
/* Passed an 8-bit failure cause as first argument. */
_GLOBAL(tm_abort)
TABORT(R3)
blr
/* void tm_reclaim(struct thread_struct *thread,
* uint8_t cause)
*
* - Performs a full reclaim. This destroys outstanding
* transactions and updates thread->regs.tm_ckpt_* with the
* original checkpointed state. Note that thread->regs is
* unchanged.
*
* Purpose is to both abort transactions of, and preserve the state of,
* a transactions at a context switch. We preserve/restore both sets of process
* state to restore them when the thread's scheduled again. We continue in
* userland as though nothing happened, but when the transaction is resumed
* they will abort back to the checkpointed state we save out here.
*
* Call with IRQs off, stacks get all out of sync for some periods in here!
*/
_GLOBAL(tm_reclaim)
mfcr r5
mflr r0
stw r5, 8(r1)
std r0, 16(r1)
std r2, STK_GOT(r1)
stdu r1, -TM_FRAME_SIZE(r1)
/* We've a struct pt_regs at [r1+STACK_FRAME_OVERHEAD]. */
std r3, STK_PARAM(R3)(r1)
SAVE_NVGPRS(r1)
/* We need to setup MSR for VSX register save instructions. */
mfmsr r14
mr r15, r14
ori r15, r15, MSR_FP
li r16, 0
ori r16, r16, MSR_EE /* IRQs hard off */
andc r15, r15, r16
oris r15, r15, MSR_VEC@h
#ifdef CONFIG_VSX
BEGIN_FTR_SECTION
oris r15,r15, MSR_VSX@h
END_FTR_SECTION_IFSET(CPU_FTR_VSX)
#endif
mtmsrd r15
std r14, TM_FRAME_L0(r1)
/* Do sanity check on MSR to make sure we are suspended */
li r7, (MSR_TS_S)@higher
srdi r6, r14, 32
and r6, r6, r7
1: tdeqi r6, 0
EMIT_BUG_ENTRY 1b,__FILE__,__LINE__,0
/* Stash the stack pointer away for use after reclaim */
std r1, PACAR1(r13)
/* Clear MSR RI since we are about to change r1, EE is already off. */
li r5, 0
mtmsrd r5, 1
/*
* BE CAREFUL HERE:
* At this point we can't take an SLB miss since we have MSR_RI
* off. Load only to/from the stack/paca which are in SLB bolted regions
* until we turn MSR RI back on.
*
* The moment we treclaim, ALL of our GPRs will switch
* to user register state. (FPRs, CCR etc. also!)
* Use an sprg and a tm_scratch in the PACA to shuffle.
*/
TRECLAIM(R4) /* Cause in r4 */
/* ******************** GPRs ******************** */
/* Stash the checkpointed r13 away in the scratch SPR and get the real
* paca
*/
SET_SCRATCH0(r13)
GET_PACA(r13)
/* Stash the checkpointed r1 away in paca tm_scratch and get the real
* stack pointer back
*/
std r1, PACATMSCRATCH(r13)
ld r1, PACAR1(r13)
/* Store the PPR in r11 and reset to decent value */
std r11, GPR11(r1) /* Temporary stash */
/* Reset MSR RI so we can take SLB faults again */
li r11, MSR_RI
mtmsrd r11, 1
mfspr r11, SPRN_PPR
HMT_MEDIUM
/* Now get some more GPRS free */
std r7, GPR7(r1) /* Temporary stash */
std r12, GPR12(r1) /* '' '' '' */
ld r12, STK_PARAM(R3)(r1) /* Param 0, thread_struct * */
std r11, THREAD_TM_PPR(r12) /* Store PPR and free r11 */
addi r7, r12, PT_CKPT_REGS /* Thread's ckpt_regs */
/* Make r7 look like an exception frame so that we
* can use the neat GPRx(n) macros. r7 is NOT a pt_regs ptr!
*/
subi r7, r7, STACK_FRAME_OVERHEAD
/* Sync the userland GPRs 2-12, 14-31 to thread->regs: */
SAVE_GPR(0, r7) /* user r0 */
SAVE_GPR(2, r7) /* user r2 */
SAVE_4GPRS(3, r7) /* user r3-r6 */
SAVE_GPR(8, r7) /* user r8 */
SAVE_GPR(9, r7) /* user r9 */
SAVE_GPR(10, r7) /* user r10 */
ld r3, PACATMSCRATCH(r13) /* user r1 */
ld r4, GPR7(r1) /* user r7 */
ld r5, GPR11(r1) /* user r11 */
ld r6, GPR12(r1) /* user r12 */
GET_SCRATCH0(8) /* user r13 */
std r3, GPR1(r7)
std r4, GPR7(r7)
std r5, GPR11(r7)
std r6, GPR12(r7)
std r8, GPR13(r7)
SAVE_NVGPRS(r7) /* user r14-r31 */
/* ******************** NIP ******************** */
mfspr r3, SPRN_TFHAR
std r3, _NIP(r7) /* Returns to failhandler */
/* The checkpointed NIP is ignored when rescheduling/rechkpting,
* but is used in signal return to 'wind back' to the abort handler.
*/
/* ******************** CR,LR,CCR,MSR ********** */
mfctr r3
mflr r4
mfcr r5
mfxer r6
std r3, _CTR(r7)
std r4, _LINK(r7)
std r5, _CCR(r7)
std r6, _XER(r7)
/* ******************** TAR, DSCR ********** */
mfspr r3, SPRN_TAR
mfspr r4, SPRN_DSCR
std r3, THREAD_TM_TAR(r12)
std r4, THREAD_TM_DSCR(r12)
/* MSR and flags: We don't change CRs, and we don't need to alter
* MSR.
*/
/* ******************** FPR/VR/VSRs ************
* After reclaiming, capture the checkpointed FPRs/VRs.
*
* We enabled VEC/FP/VSX in the msr above, so we can execute these
* instructions!
*/
mr r3, r12
/* Altivec (VEC/VMX/VR)*/
addi r7, r3, THREAD_CKVRSTATE
SAVE_32VRS(0, r6, r7) /* r6 scratch, r7 transact vr state */
mfvscr v0
li r6, VRSTATE_VSCR
stvx v0, r7, r6
/* VRSAVE */
mfspr r0, SPRN_VRSAVE
std r0, THREAD_CKVRSAVE(r3)
/* Floating Point (FP) */
addi r7, r3, THREAD_CKFPSTATE
SAVE_32FPRS_VSRS(0, R6, R7) /* r6 scratch, r7 transact fp state */
mffs fr0
stfd fr0,FPSTATE_FPSCR(r7)
/* TM regs, incl TEXASR -- these live in thread_struct. Note they've
* been updated by the treclaim, to explain to userland the failure
* cause (aborted).
*/
mfspr r0, SPRN_TEXASR
mfspr r3, SPRN_TFHAR
mfspr r4, SPRN_TFIAR
std r0, THREAD_TM_TEXASR(r12)
std r3, THREAD_TM_TFHAR(r12)
std r4, THREAD_TM_TFIAR(r12)
/* AMR is checkpointed too, but is unsupported by Linux. */
/* Restore original MSR/IRQ state & clear TM mode */
ld r14, TM_FRAME_L0(r1) /* Orig MSR */
li r15, 0
rldimi r14, r15, MSR_TS_LG, (63-MSR_TS_LG)-1
mtmsrd r14
REST_NVGPRS(r1)
addi r1, r1, TM_FRAME_SIZE
lwz r4, 8(r1)
ld r0, 16(r1)
mtcr r4
mtlr r0
ld r2, STK_GOT(r1)
/* Load CPU's default DSCR */
ld r0, PACA_DSCR_DEFAULT(r13)
mtspr SPRN_DSCR, r0
blr
/* void __tm_recheckpoint(struct thread_struct *thread,
* unsigned long orig_msr)
* - Restore the checkpointed register state saved by tm_reclaim
* when we switch_to a process.
*
* Call with IRQs off, stacks get all out of sync for
* some periods in here!
*/
_GLOBAL(__tm_recheckpoint)
mfcr r5
mflr r0
stw r5, 8(r1)
std r0, 16(r1)
std r2, STK_GOT(r1)
stdu r1, -TM_FRAME_SIZE(r1)
/* We've a struct pt_regs at [r1+STACK_FRAME_OVERHEAD].
* This is used for backing up the NVGPRs:
*/
SAVE_NVGPRS(r1)
/* Load complete register state from ts_ckpt* registers */
addi r7, r3, PT_CKPT_REGS /* Thread's ckpt_regs */
/* Make r7 look like an exception frame so that we
* can use the neat GPRx(n) macros. r7 is now NOT a pt_regs ptr!
*/
subi r7, r7, STACK_FRAME_OVERHEAD
/* We need to setup MSR for FP/VMX/VSX register save instructions. */
mfmsr r6
mr r5, r6
ori r5, r5, MSR_FP
#ifdef CONFIG_ALTIVEC
oris r5, r5, MSR_VEC@h
#endif
#ifdef CONFIG_VSX
BEGIN_FTR_SECTION
oris r5,r5, MSR_VSX@h
END_FTR_SECTION_IFSET(CPU_FTR_VSX)
#endif
mtmsrd r5
#ifdef CONFIG_ALTIVEC
/*
* FP and VEC registers: These are recheckpointed from
* thread.ckfp_state and thread.ckvr_state respectively. The
* thread.fp_state[] version holds the 'live' (transactional)
* and will be loaded subsequently by any FPUnavailable trap.
*/
addi r8, r3, THREAD_CKVRSTATE
li r5, VRSTATE_VSCR
lvx v0, r8, r5
mtvscr v0
REST_32VRS(0, r5, r8) /* r5 scratch, r8 ptr */
ld r5, THREAD_CKVRSAVE(r3)
mtspr SPRN_VRSAVE, r5
#endif
addi r8, r3, THREAD_CKFPSTATE
lfd fr0, FPSTATE_FPSCR(r8)
MTFSF_L(fr0)
REST_32FPRS_VSRS(0, R4, R8)
mtmsr r6 /* FP/Vec off again! */
restore_gprs:
/* ******************** CR,LR,CCR,MSR ********** */
ld r4, _CTR(r7)
ld r5, _LINK(r7)
ld r8, _XER(r7)
mtctr r4
mtlr r5
mtxer r8
/* ******************** TAR ******************** */
ld r4, THREAD_TM_TAR(r3)
mtspr SPRN_TAR, r4
/* Load up the PPR and DSCR in GPRs only at this stage */
ld r5, THREAD_TM_DSCR(r3)
ld r6, THREAD_TM_PPR(r3)
REST_GPR(0, r7) /* GPR0 */
REST_2GPRS(2, r7) /* GPR2-3 */
REST_GPR(4, r7) /* GPR4 */
REST_4GPRS(8, r7) /* GPR8-11 */
REST_2GPRS(12, r7) /* GPR12-13 */
REST_NVGPRS(r7) /* GPR14-31 */
/* Load up PPR and DSCR here so we don't run with user values for long
*/
mtspr SPRN_DSCR, r5
mtspr SPRN_PPR, r6
/* Do final sanity check on TEXASR to make sure FS is set. Do this
* here before we load up the userspace r1 so any bugs we hit will get
* a call chain */
mfspr r5, SPRN_TEXASR
srdi r5, r5, 16
li r6, (TEXASR_FS)@h
and r6, r6, r5
1: tdeqi r6, 0
EMIT_BUG_ENTRY 1b,__FILE__,__LINE__,0
/* Do final sanity check on MSR to make sure we are not transactional
* or suspended
*/
mfmsr r6
li r5, (MSR_TS_MASK)@higher
srdi r6, r6, 32
and r6, r6, r5
1: tdnei r6, 0
EMIT_BUG_ENTRY 1b,__FILE__,__LINE__,0
/* Restore CR */
ld r6, _CCR(r7)
mtcr r6
REST_GPR(6, r7)
/*
* Store r1 and r5 on the stack so that we can access them
* after we clear MSR RI.
*/
REST_GPR(5, r7)
std r5, -8(r1)
ld r5, GPR1(r7)
std r5, -16(r1)
REST_GPR(7, r7)
/* Clear MSR RI since we are about to change r1. EE is already off */
li r5, 0
mtmsrd r5, 1
/*
* BE CAREFUL HERE:
* At this point we can't take an SLB miss since we have MSR_RI
* off. Load only to/from the stack/paca which are in SLB bolted regions
* until we turn MSR RI back on.
*/
SET_SCRATCH0(r1)
ld r5, -8(r1)
ld r1, -16(r1)
/* Commit register state as checkpointed state: */
TRECHKPT
HMT_MEDIUM
/* Our transactional state has now changed.
*
* Now just get out of here. Transactional (current) state will be
* updated once restore is called on the return path in the _switch-ed
* -to process.
*/
GET_PACA(r13)
GET_SCRATCH0(r1)
/* R1 is restored, so we are recoverable again. EE is still off */
li r4, MSR_RI
mtmsrd r4, 1
REST_NVGPRS(r1)
addi r1, r1, TM_FRAME_SIZE
lwz r4, 8(r1)
ld r0, 16(r1)
mtcr r4
mtlr r0
ld r2, STK_GOT(r1)
/* Load CPU's default DSCR */
ld r0, PACA_DSCR_DEFAULT(r13)
mtspr SPRN_DSCR, r0
blr
/* ****************************************************************** */
Reference in New Issue View Git Blame Copy Permalink