Deal with TM on kernel entry and exit
Subsystems (3)
cover Deal with TM on kernel entry and exit
This is very much a proof of concept and if it isn't clear from the commit names, still a work in progress.
I believe I have something that works - all the powerpc selftests pass. I would like to get some eyes on it to a) see if I've missed anything big and b) some opinions on if it is looking like a net improvement.
Obviously it is still a bit rough around the edges, I'll have to convince myself that the SPR code is correct. I don't think the TM_KERNEL_ENTRY macro needs to check that we came from userspace, if TM is on then we can probably assume. Maybe a check not in the fastpath. Some of the BUG_ON()s will probably go.
Background:
Currently TM is dealt with when we need to. That is, when we switch processes, we'll (if nessesary) reclaim the outgoing process and (if nessesary) recheckpoint the incoming process. Same with signals, if we need to deliver a signal, we'll ensure we've reclaimed in order to have all the information and go from there.
I, along with some others got curious to see what it would look like if we did the 'opposite'.
At all kernel entry points that won't simply just zoom straight to an RFID we now check if the thread was transactional and do the reclaim. Correspondingly do the recheckpoint quite late on exception exit. It turns out we already had a lot of the code pathes set up on the exit path as there were things that TM had special cased on exit already. I wasn't sure it it would lead to more or less complexity and though I'd have to try it to see. I feel like it was almost a win but SPRs did add some annoying caveats.
In order to get this past Michael I'm going to prove it performs, or rather, doesn't slow anything down - workload suggestions welcome.
Thanks,
Cyril Bur (12):
powerpc/tm: Remove struct thread_info param from tm_reclaim_thread() selftests/powerpc: Fix tm.h helpers
selftests/powerpc: Add tm-signal-drop-transaction TM test selftests/powerpc: Use less common thread names
[WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
[WIP] powerpc/tm: Remove dead code from __switch_to_tm() [WIP] powerpc/tm: Add TM_KERNEL_ENTRY in more delicate exception
pathes
[WIP] powerpc/tm: Fix *unavailable_tm exceptions
[WIP] powerpc/tm: Tweak signal code to handle new reclaim/recheckpoint
times
[WIP] powerpc/tm: Correctly save/restore checkpointed sprs [WIP] powerpc/tm: Afterthoughts
[WIP] selftests/powerpc: Remove incorrect tm-syscall selftest
arch/powerpc/include/asm/exception-64s.h | 25 ++++ arch/powerpc/kernel/entry_64.S | 20 ++- arch/powerpc/kernel/exceptions-64s.S | 31 ++++- arch/powerpc/kernel/process.c | 145 ++++++++++++++++++--- arch/powerpc/kernel/ptrace.c | 9 +- arch/powerpc/kernel/signal.c | 11 +- arch/powerpc/kernel/signal_32.c | 16 +-- arch/powerpc/kernel/signal_64.c | 41 ++++-- arch/powerpc/kernel/traps.c | 3 - tools/testing/selftests/powerpc/tm/Makefile | 5 +- .../powerpc/tm/tm-signal-drop-transaction.c | 74 +++++++++++ .../testing/selftests/powerpc/tm/tm-syscall-asm.S | 28 ---- tools/testing/selftests/powerpc/tm/tm-syscall.c | 106 --------------- .../testing/selftests/powerpc/tm/tm-unavailable.c | 4 +- tools/testing/selftests/powerpc/tm/tm.h | 10 +- 15 files changed, 319 insertions(+), 209 deletions(-) create mode 100644 tools/testing/selftests/powerpc/tm/tm-signal-drop-transaction.c delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall-asm.S delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall.c
Signature · 1 line
2.16.2
RFC PATCH 1/12 powerpc/tm: Remove struct thread_info param from tm_reclaim_thread()
tm_reclaim_thread() doesn't use the parameter anymore, both callers have to bother getting it as they have no need for a struct thread_info either.
It was previously used but became unused in dc3106690b20 ("powerpc: tm:
Always use fp_state and vr_state to store live registers")
Just remove it and adjust the callers.
--- arch/powerpc/kernel/process.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
arch/powerpc/kernel/process.c +3 -4
@@ -850,8 +850,7 @@ static inline bool tm_enabled(struct task_struct *tsk)return tsk && tsk->thread.regs && (tsk->thread.regs->msr &MSR_TM ); } -static void tm_reclaim_thread (struct thread_struct *thr, -struct thread_info *ti,uint8_t cause) +static void tm_reclaim_thread (struct thread_struct *thr,uint8_t cause) {/* * Use the current MSR TM suspended bit to track if we have
@@ -898,7 +897,7 @@ static void tm_reclaim_thread(struct thread_struct *thr,void tm_reclaim_current (uint8_t cause) {tm_enable (); -tm_reclaim_thread (¤t->thread,current_thread_info (), cause); +tm_reclaim_thread (¤t->thread, cause); }static inline void tm_reclaim_task (struct task_struct *tsk)
@@ -929,7 +928,7 @@ static inline void tm_reclaim_task(struct task_struct *tsk) thr->regs->ccr, thr->regs->msr, thr->regs->trap); -tm_reclaim_thread (thr,task_thread_info (tsk),TM_CAUSE_RESCHED ); +tm_reclaim_thread (thr,TM_CAUSE_RESCHED );TM_DEBUG ("--- tm_reclaim on pid %d complete\n" , tsk->pid);
Signature · 1 line
2.16.2
RFC PATCH 2/12 selftests/powerpc: Fix tm.h helpers
In reply to Cyril BurTurns out the tcheck() helpers were subtly wrong
--- tools/testing/selftests/powerpc/tm/tm.h | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
tools/testing/selftests/powerpc/tm/tm.h +5 -5
@@ -57,11 +57,11 @@ static inline bool failure_is_nesting(void)return (__builtin_get_texasru () &0x400000 ); } -static inline int tcheck (void ) +static inline uint8_t tcheck (void ) { -long cr; - asmvolatile ("tcheck 0" :"=r" (cr) : :"cr0" ); -return (cr >>28 ) &4 ; +unsigned long cr; + asmvolatile ("tcheck 0; mfcr %0;" :"=r" (cr) : :"cr0" ); +return (cr >>28 ) &0xF ; }static inline bool tcheck_doomed (void )
@@ -81,7 +81,7 @@ static inline bool tcheck_suspended(void)static inline bool tcheck_transactional (void ) { -return tcheck () &6 ; +return (tcheck_active ()) || (tcheck_suspended ()); }#endif /* _SELFTESTS_POWERPC_TM_TM_H */
Signature · 1 line
2.16.2
RFC PATCH 3/12 selftests/powerpc: Add tm-signal-drop-transaction TM test
In reply to Cyril BurThis test uses a signal to 'discard' a transaction. That is, it will take a signal of a thread in a suspended transaction and just remove the suspended MSR bit. Because this will send the userspace thread back to the tebgin + 4 address, we should also set CR0 to be nice.
--- tools/testing/selftests/powerpc/tm/Makefile | 1 + .../powerpc/tm/tm-signal-drop-transaction.c | 74 ++++++++++++++++++++++ 2 files changed, 75 insertions(+) create mode 100644 tools/testing/selftests/powerpc/tm/tm-signal-drop-transaction.c
tools/testing/selftests/powerpc/tm/Makefile +1 -0
@@ -4,6 +4,7 @@ SIGNAL_CONTEXT_CHK_TESTS := tm-signal-context-chk-gpr tm-signal-context-chk-fpu TEST_GEN_PROGS := tm-resched-dscr tm-syscall tm-signal-msr-resv tm-signal-stack \ tm-vmxcopy tm-fork tm-tar tm-tmspr tm-vmx-unavail tm-unavailable tm-trap \ + tm-signal-drop-transaction \ $(SIGNAL_CONTEXT_CHK_TESTS) include ../../lib.mk
tools/testing/selftests/powerpc/tm/tm-signal-drop-transaction.c new file +74 -0
74 lines not on this page · raw message
Signature · 1 line
2.16.2
RFC PATCH 4/12 selftests/powerpc: Use less common thread names
In reply to Cyril Bur"ping" and "pong" (in particular "ping") are common names. If a selftests causes a kernel BUG_ON or any kind of backtrace the process name is displayed. Setting a more unique name avoids confusion as to which process caused the problem.
--- tools/testing/selftests/powerpc/tm/tm-unavailable.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
tools/testing/selftests/powerpc/tm/tm-unavailable.c +2 -2
@@ -315,7 +315,7 @@ void test_fp_vec(int fp, int vec, pthread_attr_t *attr) rc =pthread_create (&t0, attr, ping, (void *) &flags);if (rc)pr_err (rc,"pthread_create()" ); - rc =pthread_setname_np (t0,"ping" ); + rc =pthread_setname_np (t0,"tm-unavailable-ping" );if (rc)pr_warn (rc,"pthread_setname_np" ); rc =pthread_join (t0, &ret_value);
@@ -359,7 +359,7 @@ int main(int argc, char **argv)pr_err (rc,"pthread_create()" );/* Name it for systemtap convenience */ - rc =pthread_setname_np (t1,"pong" ); + rc =pthread_setname_np (t1,"tm-unavailable-pong" );if (rc)pr_warn (rc,"pthread_create()" );
Signature · 1 line
2.16.2
RFC PATCH 5/12 powerpc/tm: Reclaim/recheckpoint on entry/exit
In reply to Cyril Bur--- arch/powerpc/include/asm/exception-64s.h | 25 +++++++++++++++++++++ arch/powerpc/kernel/entry_64.S | 5 +++++ arch/powerpc/kernel/process.c | 37 ++++++++++++++++++++++++++++---- 3 files changed, 63 insertions(+), 4 deletions(-)
arch/powerpc/include/asm/exception-64s.h +25 -0
@@ -35,6 +35,7 @@* implementations as possible. */ #include <asm/head-64.h> +#include <asm/tm.h> /* PACA save area offsets (exgen, exmc, etc) */ #define EX_R9 0
@@ -127,6 +128,26 @@hrfid ; \ b hrfi_flush_fallback +#ifdef CONFIG_PPC_TRANSACTIONAL_MEM +#define TM_KERNEL_ENTRY \ + ld r3,_MSR(r1); \ + /* Probably don't need to check if coming from user/kernel */ \ +/* If TM is suspended or active then we must have come from*/ \ +/* userspace */ \ + andi. r0,r3,MSR_PR; \ + beq 1f; \ + rldicl. r3,r3,(64-MSR_TS_LG),(64-2);/* SUSPENDED or ACTIVE*/ \ + beql+1f ;/* Not SUSPENDED or ACTIVE */ \ +bl save_nvgprs; \ +RECONCILE_IRQ_STATE (r10,r11); \ +li r3,TM_CAUSE_MISC ; \ +bl tm_reclaim_current;/* uint8 cause */ \ +1: + +#else /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#define TM_KERNEL_ENTRY +#endif /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#ifdef CONFIG_RELOCATABLE #define __EXCEPTION_RELON_PROLOG_PSERIES_1 (label, h) \ mfspr r11,SPRN_##h##SRR0;/* save SRR0 */ \
@@ -675,6 +696,9 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON (trap, area); \/* Volatile regs are potentially clobbered here */ \ additions; \ +/* This is going to need to go somewhere else as well */ \ +/* See comment in tm_recheckpoint() */ \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr; \b ret
@@ -689,6 +713,7 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON_3 (trap); \/* Volatile regs are potentially clobbered here */ \ additions; \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr
arch/powerpc/kernel/entry_64.S +5 -0
@@ -126,6 +126,11 @@ BEGIN_FW_FTR_SECTION 33: END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR) #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE && CONFIG_PPC_SPLPAR */ + TM_KERNEL_ENTRY + REST_GPR(0,r1) + REST_4GPRS(3,r1) + REST_2GPRS(7,r1) + addi r9,r1,STACK_FRAME_OVERHEAD /* * A syscall should always be called with interrupts enabled
arch/powerpc/kernel/process.c +33 -4
@@ -951,6 +951,23 @@ void tm_recheckpoint(struct thread_struct *thread)if (!(thread->regs->msr &MSR_TM ))return ; +/* +* This is 'that' comment. +* +* If we get where with tm suspended or active then something +* has gone wrong. I've added this now as a proof of concept. +* +* The problem I'm seeing without it is an attempt to +* recheckpoint a CPU without a previous reclaim. +* +* I'm probably missed an exception entry with the +* TM_KERNEL_ENTRY macro. Should be easy enough to find. +*/ +if (MSR_TM_ACTIVE (mfmsr ())) +return ; + +tm_enable (); +/* We really can't be interrupted here as the TEXASR registers can't * change and later in the trecheckpoint code, we have a userspace R1. * So let's hard disable over this region.
@@ -1009,6 +1026,13 @@ static inline void tm_recheckpoint_new_task(struct task_struct *new)static inline void __switch_to_tm (struct task_struct *prev,struct task_struct *new) { +/* +* So, with the rework none of this code should not be needed. +* I've left in the reclaim for now. This *should* save us +* from any mistake in the new code. Also the +* enabling/disabling logic of MSR_TM really should be +* refactored into a common way with MSR_{FP,VEC,VSX} +*/ if (cpu_has_feature (CPU_FTR_TM )) {if (tm_enabled (prev) ||tm_enabled (new))tm_enable ();
@@ -1016,11 +1040,14 @@ static inline void __switch_to_tm(struct task_struct *prev,if (tm_enabled (prev)) { prev->thread.load_tm++;tm_reclaim_task (prev); -if (!MSR_TM_ACTIVE (prev->thread.regs->msr) && prev->thread.load_tm ==0 ) - prev->thread.regs->msr &= ~MSR_TM ; +/* +* The disabling logic may be confused don't +* disable for now +* +* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0) +* prev->thread.regs->msr &= ~MSR_TM; +*/ } - -tm_recheckpoint_new_task (new); } }
@@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs) msr_diff = current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=MSR_FP |MSR_VEC |MSR_VSX ; +tm_recheckpoint (¤t->thread); +/* Ensure that restore_math() will restore */ if (msr_diff &MSR_FP ) current->thread.load_fp =1 ;
Signature · 1 line
2.16.2
Re: [RFC PATCH 05/12] [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
On Tue, 2018-02-20 at 11:22 +1100, Cyril Bur wrote:
The comment from the cover sheet should be here
Quoted 4 lines
--- arch/powerpc/include/asm/exception-64s.h | 25 +++++++++++++++++++++ arch/powerpc/kernel/entry_64.S | 5 +++++
arch/powerpc/kernel/process.c| 37 ++++++++++++++++++++++++++=
++----
arch/powerpc/incl=
Quoted 3 lines
3 files changed, 63 insertions(+), 4 deletions(-)=20
arch/powerpc/incl= renamed from arch/powerpc/include/asm/exception-64s.h +0
-0
ude/asm/exception-64s.h
arch/powerpc/kernel/entry_6=
Quoted 57 lines
index
471b2274fbeb..f904f19a9ec2100644arch/powerpc/include/asm/exception-64s.h +1
-0@@ -35,6 +35,7 @@* implementations as possible. */ #include <asm/head-64.h> +#include <asm/tm.h> =20
/* PACA save area offsets (exgen, exmc, etc) */
#defineEX_R90arch/powerpc/include/asm/exception-64s.h +0
-0@@ -127,6 +128,26 @@ hrfid; \b hrfi_flush_fallback=20
arch/powerpc/include/asm/exception-64s.h +24
-0+#ifdef CONFIG_PPC_TRANSACTIONAL_MEM +#define TM_KERNEL_ENTRY \ + ld r3,_MSR(r1); \ + /* Probably don't need to check if coming from user/kernel */ \ +/* If TM is suspended or active then we must have come from*/ \ +/* userspace */ \ + andi. r0,r3,MSR_PR; \ + beq 1f; \ + rldicl. r3,r3,(64-MSR_TS_LG),(64-2);/* SUSPENDED or ACTIVE*/ \ + beql+1f ;/* Not SUSPENDED or ACTIVE */ \ +bl save_nvgprs; \ +RECONCILE_IRQ_STATE (r10,r11); \ +li r3,TM_CAUSE_MISC ; \ +bl tm_reclaim_current;/* uint8 cause */ \ +1: + +#else /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#define TM_KERNEL_ENTRY +#endif /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#ifdef CONFIG_RELOCATABLE #define __EXCEPTION_RELON_PROLOG_PSERIES_1 (label, h) \ mfspr r11,SPRN_##h##SRR0;/* save SRR0 */ \@@ -675,6 +696,9 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON (trap, area); \/* Volatile regs are potentially clobbered here */ \ additions; \ +/* This is going to need to go somewhere else as well */ \ +/* See comment in tm_recheckpoint() */ \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr; \b ret@@ -689,6 +713,7 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON_3 (trap); \/* Volatile regs are potentially clobbered here */ \ additions; \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr=20
arch/powerpc/kernel/entry_6= renamed from arch/powerpc/kernel/entry_64.S +0
-0
4.S
arch/powerpc/kernel/entry_64.S BEGIN_FW_FTR_SECTION · line 133
Quoted 12 lines
index
2cb5109a7ea3..107c15c6f48b100644arch/powerpc/kernel/entry_64.S +5
-0@@ -126,6 +126,11 @@ BEGIN_FW_FTR_SECTION 33: END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR) #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE && CONFIG_PPC_SPLPAR */ + TM_KERNEL_ENTRY + REST_GPR(0,r1) + REST_4GPRS(3,r1) + REST_2GPRS(7,r1) + addi r9,r1,STACK_FRAME_OVERHEAD
Why are we doing these restores here now?
arch/powerpc/kernel/process.=
Quoted 4 lines
=20
arch/powerpc/kernel/entry_64.S +0
-0/* * A syscall should always be called with interrupts enabledarch/powerpc/kernel/process.= renamed from arch/powerpc/kernel/process.c +0
-0
c
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 955
Quoted 9 lines
index
77dc6d8288eb..ea75da0fd506100644arch/powerpc/kernel/process.c +0
-0@@ -951,6 +951,23 @@ void tm_recheckpoint(struct thread_struct *thread)if (!(thread->regs->msr &MSR_TM ))return ;=20
arch/powerpc/kernel/process.c +2
-0+/* +* This is 'that' comment.
I think I'm in the loop here but I don't actually know what this means.=20
Senior Mikey moment or Crazy Cyril comments? I'll let the peanut gallery de= cide.
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 957
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+* +* If we get where with tm suspended or active then something
s/where/here/
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 967
Quoted 10 lines
arch/powerpc/kernel/process.c +10
-0+* has gone wrong. I've added this now as a proof of concept. +* +* The problem I'm seeing without it is an attempt to +* recheckpoint a CPU without a previous reclaim. +* +* I'm probably missed an exception entry with the +* TM_KERNEL_ENTRY macro. Should be easy enough to find. +*/ +if (MSR_TM_ACTIVE (mfmsr ())) +return ;
I don't really get this. Wouldn't this test apply now?
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 969
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+ +tm_enable ();
Why did we add this?
arch/powerpc/kernel/process.c static inline void tm_recheckpoint_new_task(struct=
Quoted 5 lines
arch/powerpc/kernel/process.c +1
-0+/* We really can't be interrupted here as the TEXASR registers can't * change and later in the trecheckpoint code, we have a userspace R1. * So let's hard disable over this region. @@ -1009,6 +1026,13 @@ static inline void tm_recheckpoint_new_task(struct=
task_struct *new)
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_str=
Quoted 14 lines
arch/powerpc/kernel/process.c +7
-0static inline void __switch_to_tm (struct task_struct *prev,struct task_struct *new) { +/* +* So, with the rework none of this code should not be needed. +* I've left in the reclaim for now. This *should* save us +* from any mistake in the new code. Also the +* enabling/disabling logic of MSR_TM really should be +* refactored into a common way with MSR_{FP,VEC,VSX} +*/ if (cpu_has_feature (CPU_FTR_TM )) {if (tm_enabled (prev) ||tm_enabled (new))tm_enable ();@@ -1016,11 +1040,14 @@ static inline void __switch_to_tm(struct task_str=
uct *prev,
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_struct *prev,
Quoted 4 lines
arch/powerpc/kernel/process.c +0
-1if (tm_enabled (prev)) { prev->thread.load_tm++;tm_reclaim_task (prev); - if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm =
=3D=3D 0)
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_struct *prev,
Quoted 6 lines
arch/powerpc/kernel/process.c +5
-1- prev->thread.regs->msr &=3D ~MSR_TM; +/* +* The disabling logic may be confused don't +* disable for now +* +* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm=
=3D=3D 0)
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_str= · line 1049
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+* prev->thread.regs->msr &=3D ~MSR_TM; +*/
Why are you doing this when you just remove all this code in the next patch= ?
arch/powerpc/kernel/process.c void restore_tm_state(struct pt_regs *regs) · line 1086
Quoted 12 lines
arch/powerpc/kernel/process.c +0
-2} - -tm_recheckpoint_new_task (new); } }=20
arch/powerpc/kernel/process.c +0
-0@@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs) msr_diff =3D current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=3D MSR_FP | MSR_VEC | MSR_VSX;=20
arch/powerpc/kernel/process.c +2
-0+tm_recheckpoint (¤t->thread); +
So why do we do tm_recheckpoint at all? Shouldn't most of the tm_blah code = go
away in process.c after all this?
Quoted 3 lines
arch/powerpc/kernel/process.c +0
-0/* Ensure that restore_math() will restore */ if (msr_diff &MSR_FP ) current->thread.load_fp =3D 1;
Re: [RFC PATCH 05/12] [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
Quoted 68 lines · Michael Neuling
On Tue, 2018-02-20 at 11:22 +1100, Cyril Bur wrote:
The comment from the cover sheet should be here
Quoted 63 lines
--- arch/powerpc/include/asm/exception-64s.h | 25 +++++++++++++++++++++ arch/powerpc/kernel/entry_64.S | 5 +++++ arch/powerpc/kernel/process.c | 37 ++++++++++++++++++++++++++++---- 3 files changed, 63 insertions(+), 4 deletions(-)arch/powerpc/include/asm/exception-64s.h +25
-0@@ -35,6 +35,7 @@* implementations as possible. */ #include <asm/head-64.h> +#include <asm/tm.h> /* PACA save area offsets (exgen, exmc, etc) */ #define EX_R9 0@@ -127,6 +128,26 @@hrfid ; \ b hrfi_flush_fallback +#ifdef CONFIG_PPC_TRANSACTIONAL_MEM +#define TM_KERNEL_ENTRY \ + ld r3,_MSR(r1); \ + /* Probably don't need to check if coming from user/kernel */ \ +/* If TM is suspended or active then we must have come from*/ \ +/* userspace */ \ + andi. r0,r3,MSR_PR; \ + beq 1f; \ + rldicl. r3,r3,(64-MSR_TS_LG),(64-2);/* SUSPENDED or ACTIVE*/ \ + beql+1f ;/* Not SUSPENDED or ACTIVE */ \ +bl save_nvgprs; \ +RECONCILE_IRQ_STATE (r10,r11); \ +li r3,TM_CAUSE_MISC ; \ +bl tm_reclaim_current;/* uint8 cause */ \ +1: + +#else /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#define TM_KERNEL_ENTRY +#endif /* CONFIG_PPC_TRANSACTIONAL_MEM */ +#ifdef CONFIG_RELOCATABLE #define __EXCEPTION_RELON_PROLOG_PSERIES_1 (label, h) \ mfspr r11,SPRN_##h##SRR0;/* save SRR0 */ \@@ -675,6 +696,9 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON (trap, area); \/* Volatile regs are potentially clobbered here */ \ additions; \ +/* This is going to need to go somewhere else as well */ \ +/* See comment in tm_recheckpoint() */ \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr; \b ret@@ -689,6 +713,7 @@ END_FTR_SECTION_IFSET(CPU_FTR_CTRL)EXCEPTION_PROLOG_COMMON_3 (trap); \/* Volatile regs are potentially clobbered here */ \ additions; \ +TM_KERNEL_ENTRY ; \addi r3,r1,STACK_FRAME_OVERHEAD ; \bl hdlr
Quoted 15 lines
arch/powerpc/kernel/entry_64.S BEGIN_FW_FTR_SECTION · line 133
Quoted 13 lines
arch/powerpc/kernel/entry_64.S +5
-0@@ -126,6 +126,11 @@ BEGIN_FW_FTR_SECTION 33: END_FW_FTR_SECTION_IFSET(FW_FEATURE_SPLPAR) #endif /* CONFIG_VIRT_CPU_ACCOUNTING_NATIVE && CONFIG_PPC_SPLPAR */ + TM_KERNEL_ENTRY + REST_GPR(0,r1) + REST_4GPRS(3,r1) + REST_2GPRS(7,r1) + addi r9,r1,STACK_FRAME_OVERHEADWhy are we doing these restores here now?
The syscall handler expects the syscall params to still be in their respective regs.
Quoted 12 lines
Quoted 12 lines
arch/powerpc/kernel/entry_64.S +0
-0/* * A syscall should always be called with interrupts enabledarch/powerpc/kernel/process.c +0
-0@@ -951,6 +951,23 @@ void tm_recheckpoint(struct thread_struct *thread)if (!(thread->regs->msr &MSR_TM ))return ;
Quoted 7 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 955
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+/* +* This is 'that' comment. I think I'm in the loop here but I don't actually know what this means.
Senior Mikey moment or Crazy Cyril comments? I'll let the peanut gallery decide.
Oh quite possibly crazy Cyril comment that will have to be... normalised. I should actually delete this and see if that's still the case.
Quoted 18 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 957
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+* +* If we get where with tm suspended or active then something s/where/here/
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 967
Quoted 10 lines
arch/powerpc/kernel/process.c +10
-0+* has gone wrong. I've added this now as a proof of concept. +* +* The problem I'm seeing without it is an attempt to +* recheckpoint a CPU without a previous reclaim. +* +* I'm probably missed an exception entry with the +* TM_KERNEL_ENTRY macro. Should be easy enough to find. +*/ +if (MSR_TM_ACTIVE (mfmsr ())) +return ;I don't really get this. Wouldn't this test apply now?
Quoted 5 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 969
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+ +tm_enable ();Why did we add this?
Ah yes that was a cleanup I noticed along the way and clearly forgot to finish.
At the moment there's a bunch of tm_enable()s either before calling functions like tm_recheckpoint() or tm_reclaim_current() or inside helpers (tm_reclaim_current() for example again). I feel like callers shouldn't have to worry, it should be up to the function actually doing the TM work to enable it.
Quoted 25 lines
Quoted 25 lines
arch/powerpc/kernel/process.c +9
-2+/* We really can't be interrupted here as the TEXASR registers can't * change and later in the trecheckpoint code, we have a userspace R1. * So let's hard disable over this region. @@ -1009,6 +1026,13 @@ static inline void tm_recheckpoint_new_task(struct task_struct *new)static inline void __switch_to_tm (struct task_struct *prev,struct task_struct *new) { +/* +* So, with the rework none of this code should not be needed. +* I've left in the reclaim for now. This *should* save us +* from any mistake in the new code. Also the +* enabling/disabling logic of MSR_TM really should be +* refactored into a common way with MSR_{FP,VEC,VSX} +*/ if (cpu_has_feature (CPU_FTR_TM )) {if (tm_enabled (prev) ||tm_enabled (new))tm_enable ();@@ -1016,11 +1040,14 @@ static inline void __switch_to_tm(struct task_struct *prev,if (tm_enabled (prev)) { prev->thread.load_tm++;tm_reclaim_task (prev); -if (!MSR_TM_ACTIVE (prev->thread.regs->msr) && prev->thread.load_tm ==0 ) - prev->thread.regs->msr &= ~MSR_TM ; +/*
Quoted 8 lines
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_struct *prev, · line 1049
Quoted 6 lines
arch/powerpc/kernel/process.c +6
-0+* The disabling logic may be confused don't +* disable for now +* +* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0) +* prev->thread.regs->msr &= ~MSR_TM; +*/ Why are you doing this when you just remove all this code in the next patch?
The next 3 or so patches will need squashing into this one before merging.
Quoted 10 lines
Quoted 10 lines
arch/powerpc/kernel/process.c +0
-2} - -tm_recheckpoint_new_task (new); } }@@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs) msr_diff = current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=MSR_FP |MSR_VEC |MSR_VSX ;
Quoted 6 lines
arch/powerpc/kernel/process.c void restore_tm_state(struct pt_regs *regs) · line 1086
Quoted 2 lines
arch/powerpc/kernel/process.c +2
-0+tm_recheckpoint (¤t->thread); +So why do we do
tm_recheckpointat all? Shouldn't most of thetm_blahcode go away inprocess.cafter all this?
I'm not sure I follow, we need to recheckpoint because we're going back to userspace? Or would you rather calling the tm.S code directly from the exception return path?
Yes, I hope we'll be able to have a fairly big cleanup commit of tm_ code in process.c at the end of this series.
Quoted 3 lines
Quoted 3 lines
arch/powerpc/kernel/process.c +0
-0/* Ensure that restore_math() will restore */ if (msr_diff &MSR_FP ) current->thread.load_fp =1 ;
Re: [RFC PATCH 05/12] [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
Quoted 9 lines
Quoted 9 lines
arch/powerpc/kernel/process.c void restore_tm_state(struct pt_regs *regs) · line 1086
Quoted 7 lines
(quoted diff) +0
-0@@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs) msr_diff =3D current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=3D MSR_FP | MSR_VEC | MSR_VSX;=20
arch/powerpc/kernel/process.c +2
-0+tm_recheckpoint (¤t->thread); +=20
So why do we dotm_recheckpointat all? Shouldn't most of thetm_blahc=
ode go
Quoted 6 lines
Quoted 2 lines
away in
process.cafter all this?
=20=20
I'm not sure I follow, we need to recheckpoint because we're going back to userspace? Or would you rather calling thetm.Scode directly from the exception return path?
Yeah, I was thinking the point of this series was. We do tm_reclaim right = on
entry and tm_recheckpoint right on exit. =20
The bits in between (ie. the tm_blah() calls process.c) would mostly go awa= y.
Quoted 2 lines
Yes, I hope we'll be able to have a fairly big cleanup commit of
tm_code inprocess.cat the end of this series.
Yep, agreed.
Mikey
Re: [RFC PATCH 05/12] [WIP] powerpc/tm: Reclaim/recheckpoint on entry/exit
Quoted 11 lines · Michael Neuling
Quoted 11 lines
Quoted 10 lines
arch/powerpc/kernel/process.c void restore_tm_state(struct pt_regs *regs) · line 1086
Quoted 6 lines
arch/powerpc/kernel/process.c +2
-0@@ -1055,6 +1082,8 @@ void restore_tm_state(struct pt_regs *regs) msr_diff = current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=MSR_FP |MSR_VEC |MSR_VSX ; +tm_recheckpoint (¤t->thread); +So why do we do
tm_recheckpointat all? Shouldn't most of thetm_blahcode go away inprocess.cafter all this?
Quoted 7 lines
Quoted 3 lines
I'm not sure I follow, we need to recheckpoint because we're going back to userspace? Or would you rather calling the
tm.Scode directly from the exception return path?Yeah, I was thinking the point of this series was. We do
tm_reclaimright on entry andtm_recheckpointright on exit.
Yeah that's the ultimate goal, considering I haven't been attacked or offered more drugs I feel like what I've done isn't crazy. Your feedback is great, thanks.
Quoted 9 lines
The bits in between (ie. the
tm_blah()callsprocess.c) would mostly go away.Quoted 2 lines
Yes, I hope we'll be able to have a fairly big cleanup commit of
tm_code inprocess.cat the end of this series.Yep, agreed.
Mikey
RFC PATCH 6/12 powerpc/tm: Remove dead code from __switch_to_tm()
In reply to Cyril Bur--- arch/powerpc/kernel/process.c | 24 +++++------------------- 1 file changed, 5 insertions(+), 19 deletions(-)
arch/powerpc/kernel/process.c +5 -19
@@ -1027,27 +1027,13 @@ static inline void __switch_to_tm(struct task_struct *prev, struct task_struct *new) {/* -* So, with the rework none of this code should not be needed. -* I've left in the reclaim for now. This *should* save us -* from any mistake in the new code. Also the -* enabling/disabling logic of MSR_TM really should be +* The enabling/disabling logic of MSR_TM really should be * refactored into a common way with MSR_{FP,VEC,VSX} */ -if (cpu_has_feature (CPU_FTR_TM )) { -if (tm_enabled (prev) ||tm_enabled (new)) -tm_enable (); - -if (tm_enabled (prev)) { - prev->thread.load_tm++; -tm_reclaim_task (prev); -/* -* The disabling logic may be confused don't -* disable for now -* -* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0) -* prev->thread.regs->msr &= ~MSR_TM; -*/ - } +if (cpu_has_feature (CPU_FTR_TM ) &&tm_enabled (prev)) { + prev->thread.load_tm++; +if (!MSR_TM_ACTIVE (prev->thread.regs->msr) && prev->thread.load_tm ==0 ) + prev->thread.regs->msr &= ~MSR_TM ; } }
Signature · 1 line
2.16.2
Re: [RFC PATCH 06/12] [WIP] powerpc/tm: Remove dead code from __switch_to_tm()
Not sure I understand this.. should it be merged with the last patch?
Needs a comment here.
arch/powerpc/kernel/process.=
Quoted 5 lines · Cyril Bur
--- arch/powerpc/kernel/process.c | 24 +++++------------------- 1 file changed, 5 insertions(+), 19 deletions(-)=20
arch/powerpc/kernel/process.= renamed from arch/powerpc/kernel/process.c +0
-0
c
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_str=
Quoted 4 lines
index
ea75da0fd506..574b05fe7d66100644arch/powerpc/kernel/process.c +0
-0@@ -1027,27 +1027,13 @@ static inline void __switch_to_tm(struct task_str=
uct *prev,
Quoted 12 lines
arch/powerpc/kernel/process.c +1
-6struct task_struct *new) {/* -* So, with the rework none of this code should not be needed. -* I've left in the reclaim for now. This *should* save us -* from any mistake in the new code. Also the -* enabling/disabling logic of MSR_TM really should be +* The enabling/disabling logic of MSR_TM really should be * refactored into a common way with MSR_{FP,VEC,VSX} */ -if (cpu_has_feature (CPU_FTR_TM )) { -if (tm_enabled (prev) ||tm_enabled (new))
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_struct *prev,
Quoted 10 lines
arch/powerpc/kernel/process.c +0
-10-tm_enable (); - -if (tm_enabled (prev)) { - prev->thread.load_tm++; -tm_reclaim_task (prev); -/* -* The disabling logic may be confused don't -* disable for now -* -* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm=
=3D=3D 0)
arch/powerpc/kernel/process.c static inline void __switch_to_tm(struct task_struct *prev,
Quoted 6 lines
arch/powerpc/kernel/process.c +3
-3-* prev->thread.regs->msr &=3D ~MSR_TM; -*/ - } + if (cpu_has_feature(CPU_FTR_TM) && tm_enabled(prev)) { + prev->thread.load_tm++; + if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm =3D=
=3D 0)
Quoted 4 lines
arch/powerpc/kernel/process.c +1
-0+ prev->thread.regs->msr &=3D ~MSR_TM; } }=20
Re: [RFC PATCH 06/12] [WIP] powerpc/tm: Remove dead code from __switch_to_tm()
Quoted 2 lines · Michael Neuling
Not sure I understand this.. should it be merged with the last patch?
Its all going to have to be one patch - I've left it split out to make it more obvious which bits have had to mess with, this series absolutely doesn't bisect.
Quoted 45 lines
Needs a comment here.
Quoted 41 lines · Cyril Bur
--- arch/powerpc/kernel/process.c | 24 +++++------------------- 1 file changed, 5 insertions(+), 19 deletions(-)arch/powerpc/kernel/process.c +5
-19@@ -1027,27 +1027,13 @@ static inline void __switch_to_tm(struct task_struct *prev, struct task_struct *new) {/* -* So, with the rework none of this code should not be needed. -* I've left in the reclaim for now. This *should* save us -* from any mistake in the new code. Also the -* enabling/disabling logic of MSR_TM really should be +* The enabling/disabling logic of MSR_TM really should be * refactored into a common way with MSR_{FP,VEC,VSX} */ -if (cpu_has_feature (CPU_FTR_TM )) { -if (tm_enabled (prev) ||tm_enabled (new)) -tm_enable (); - -if (tm_enabled (prev)) { - prev->thread.load_tm++; -tm_reclaim_task (prev); -/* -* The disabling logic may be confused don't -* disable for now -* -* if (!MSR_TM_ACTIVE(prev->thread.regs->msr) && prev->thread.load_tm == 0) -* prev->thread.regs->msr &= ~MSR_TM; -*/ - } +if (cpu_has_feature (CPU_FTR_TM ) &&tm_enabled (prev)) { + prev->thread.load_tm++; +if (!MSR_TM_ACTIVE (prev->thread.regs->msr) && prev->thread.load_tm ==0 ) + prev->thread.regs->msr &= ~MSR_TM ; } }
RFC PATCH 7/12 powerpc/tm: Add TM_KERNEL_ENTRY in more delicate exception pathes
In reply to Cyril Bur--- arch/powerpc/kernel/entry_64.S | 15 ++++++++++++++- arch/powerpc/kernel/exceptions-64s.S | 31 ++++++++++++++++++++++++++++--- 2 files changed, 42 insertions(+), 4 deletions(-)
arch/powerpc/kernel/entry_64.S +14 -1
@@ -967,7 +967,20 @@ END_FTR_SECTION_IFSET(CPU_FTR_HAS_PPR) bl __check_irq_replay cmpwi cr0,r3,0 beq .Lrestore_no_replay - + + /* + * We decide VERY late if we need to replay interrupts, theres + * not much which can be done about that so this will have to + * do + */ + TM_KERNEL_ENTRY + /* + * This will restore r3 that TM_KERNEL_ENTRY clobbered. + * Clearly not ideal! I wonder if we could change the trap + * value beforehand... + */ + bl __check_irq_replay + /* * We need to re-emit an interrupt. We do so by re-using our * existing exception frame. We first change the trap value,
arch/powerpc/kernel/exceptions-64s.S +28 -3
@@ -504,6 +504,11 @@ EXC_COMMON_BEGIN(data_access_common) li r5,0x300 std r3,_DAR(r1) std r4,_DSISR(r1) + /* + * Can't do TM_KERNEL_ENTRY here as do_hash_page might jump to + * very late in the expection exit code, well after any + * possiblity of doing a recheckpoint + */ BEGIN_MMU_FTR_SECTION b do_hash_page /* Try to handle as hpte fault */ MMU_FTR_SECTION_ELSE
@@ -548,6 +553,11 @@ EXC_COMMON_BEGIN(instruction_access_common) li r5,0x400 std r3,_DAR(r1) std r4,_DSISR(r1) + /* + * Can't do TM_KERNEL_ENTRY here as do_hash_page might jump to + * very late in the expection exit code, well after any + * possiblity of doing a recheckpoint + */ BEGIN_MMU_FTR_SECTION b do_hash_page /* Try to handle as hpte fault */ MMU_FTR_SECTION_ELSE
@@ -761,6 +771,7 @@ EXC_COMMON_BEGIN(alignment_common) std r4,_DSISR(r1) bl save_nvgprs RECONCILE_IRQ_STATE(r10, r11) + TM_KERNEL_ENTRY addi r3,r1,STACK_FRAME_OVERHEAD bl alignment_exception b ret_from_except
@@ -1668,7 +1679,9 @@ do_hash_page: /* Here we have a page fault that hash_page can't handle. */ handle_page_fault: -11: andis. r0,r4,DSISR_DABRMATCH@h +11: TM_KERNEL_ENTRY + ld r4,_DSISR(r1) + andis. r0,r4,DSISR_DABRMATCH@h bne- handle_dabr_fault ld r4,_DAR(r1) ld r5,_DSISR(r1)
@@ -1685,6 +1698,10 @@ handle_page_fault: /* We have a data breakpoint exception - handle it */ handle_dabr_fault: + /* + * Don't need to do TM_KERNEL_ENTRY here as we'll + * come from handle_page_fault: which has done it already + */ bl save_nvgprs ld r4,_DAR(r1) ld r5,_DSISR(r1)
@@ -1698,7 +1715,14 @@ handle_dabr_fault: * the PTE insertion */ 13: bl save_nvgprs - mr r5,r3 + /* + * Use a non-volatile as the TM code will call, r3 is the + * return value from __hash_page() so not exactly easy to get + * again. + */ + mr r31,r3 + TM_KERNEL_ENTRY + mr r5, r31 addi r3,r1,STACK_FRAME_OVERHEAD ld r4,_DAR(r1) bl low_hash_fault
@@ -1713,7 +1737,8 @@ handle_dabr_fault: * the access, or panic if there isn't a handler. */ 77: bl save_nvgprs - mr r4,r3 + TM_KERNEL_ENTRY + ld r4,_DAR(r1) addi r3,r1,STACK_FRAME_OVERHEAD li r5,SIGSEGV bl bad_page_fault
Signature · 1 line
2.16.2
RFC PATCH 8/12 powerpc/tm: Fix *unavailable_tm exceptions
In reply to Cyril Bur--- arch/powerpc/kernel/process.c | 11 ++++++++++- arch/powerpc/kernel/traps.c | 3 --- 2 files changed, 10 insertions(+), 4 deletions(-)
arch/powerpc/kernel/process.c +10 -1
@@ -500,11 +500,20 @@ void giveup_all(struct task_struct *tsk) usermsr = tsk->thread.regs->msr; +/* +* The *_unavailable_tm() functions might call this in a +* transaction but with not FP or VEC or VSX meaning that the +* if condition below will be true, this is bad since we will +* have preformed a reclaim but not set the TIF flag which +* must be set in order to trigger the recheckpoint. +* +* possibleTODO: Move setting the TIF flag into reclaim code +*/ +check_if_tm_restore_required (tsk);if ((usermsr & msr_all_available) ==0 )return ;msr_check_and_set (msr_all_available); -check_if_tm_restore_required (tsk);WARN_ON ((usermsr &MSR_VSX ) && !((usermsr &MSR_FP ) && (usermsr &MSR_VEC )));
arch/powerpc/kernel/traps.c +0 -3
@@ -1728,7 +1728,6 @@ void fp_unavailable_tm(struct pt_regs *regs)* If VMX is in use, the VRs now hold checkpointed values, * so we don't want to load the VRs from the thread_struct. */ -tm_recheckpoint (¤t->thread); }void altivec_unavailable_tm (struct pt_regs *regs)
@@ -1742,7 +1741,6 @@ void altivec_unavailable_tm(struct pt_regs *regs) regs->nip, regs->msr);tm_reclaim_current (TM_CAUSE_FAC_UNAV ); current->thread.load_vec =1 ; -tm_recheckpoint (¤t->thread); current->thread.used_vr =1 ; }
@@ -1767,7 +1765,6 @@ void vsx_unavailable_tm(struct pt_regs *regs) current->thread.load_vec =1 ; current->thread.load_fp = 1; - tm_recheckpoint(¤t->thread); }#endif /* CONFIG_PPC_TRANSACTIONAL_MEM */
Signature · 1 line
2.16.2
RFC PATCH 9/12 powerpc/tm: Tweak signal code to handle new reclaim/recheckpoint times
In reply to Cyril Bur--- arch/powerpc/kernel/process.c | 13 ++++++++++++- arch/powerpc/kernel/signal.c | 11 ++++++----- arch/powerpc/kernel/signal_32.c | 16 ++-------------- arch/powerpc/kernel/signal_64.c | 41 +++++++++++++++++++++++++++++------------ 4 files changed, 49 insertions(+), 32 deletions(-)
arch/powerpc/kernel/process.c +12 -1
@@ -1070,9 +1070,20 @@ void restore_tm_state(struct pt_regs *regs)* again, anything else could lead to an incorrect ckpt_msr being * saved and therefore incorrect signal contexts. */ -clear_thread_flag (TIF_RESTORE_TM ); + +/* +* So, on signals we're going to have cleared the TM bits from +* the MSR, meaning that heading to userspace signal handler +* this will be true. +* I'm not convinced clearing the TIF_RESTORE_TM flag is a +* good idea however, we should do it only if we actually +* recheckpoint, which we'll need to do once the signal +* hanlder is done and we're returning to the main thread of +* execution. +*/ if (!MSR_TM_ACTIVE (regs->msr))return ; +clear_thread_flag (TIF_RESTORE_TM ); msr_diff = current->thread.ckpt_regs.msr & ~regs->msr; msr_diff &=MSR_FP |MSR_VEC |MSR_VSX ;
arch/powerpc/kernel/signal.c +6 -5
@@ -191,16 +191,17 @@ unsigned long get_tm_stackpointer(struct task_struct *tsk)* * For signals taken in non-TM or suspended mode, we use the * normal/non-checkpointed stack pointer. +* +* We now do reclaims on kernel entry, we should absolutely +* never need to reclaim here. +* TODO Update the comment above if needed. */ #ifdef CONFIG_PPC_TRANSACTIONAL_MEM BUG_ON (tsk != current); -if (MSR_TM_ACTIVE (tsk->thread.regs->msr)) { -tm_reclaim_current (TM_CAUSE_SIGNAL ); -if (MSR_TM_TRANSACTIONAL (tsk->thread.regs->msr)) -return tsk->thread.ckpt_regs.gpr[1 ]; - } +if (MSR_TM_TRANSACTIONAL (tsk->thread.regs->msr)) +return tsk->thread.ckpt_regs.gpr[1 ];#endif return tsk->thread.regs->gpr[1 ]; }
arch/powerpc/kernel/signal_32.c +2 -14
@@ -860,21 +860,9 @@ static long restore_tm_user_regs(struct pt_regs *regs,tm_enable ();/* Make sure the transaction is marked as failed */ current->thread.tm_texasr |=TEXASR_FS ; -/* This loads the checkpointed FP/VEC state, if used */ -tm_recheckpoint (¤t->thread); -/* This loads the speculative FP/VEC state, if used */ -msr_check_and_set (msr & (MSR_FP |MSR_VEC )); -if (msr &MSR_FP ) { -load_fp_state (¤t->thread.fp_state); - regs->msr |= (MSR_FP | current->thread.fpexc_mode); - } -#ifdef CONFIG_ALTIVEC -if (msr &MSR_VEC ) { -load_vr_state (¤t->thread.vr_state); - regs->msr |=MSR_VEC ; - } -#endif +/* See comment in signal_64.c */ +set_thread_flag (TIF_RESTORE_TM );return 0 ; }
arch/powerpc/kernel/signal_64.c +29 -12
62 lines not on this page · raw message
Signature · 1 line
2.16.2
RFC PATCH 10/12 powerpc/tm: Correctly save/restore checkpointed sprs
In reply to Cyril Bur--- arch/powerpc/kernel/process.c | 57 +++++++++++++++++++++++++++++++++++++++++-- arch/powerpc/kernel/ptrace.c | 9 +++---- 2 files changed, 58 insertions(+), 8 deletions(-)
arch/powerpc/kernel/process.c +55 -2
@@ -859,6 +859,8 @@ static inline bool tm_enabled(struct task_struct *tsk)return tsk && tsk->thread.regs && (tsk->thread.regs->msr &MSR_TM ); } +static inline void save_sprs (struct thread_struct *t); +static void tm_reclaim_thread (struct thread_struct *thr,uint8_t cause) {/*
@@ -879,6 +881,8 @@ static void tm_reclaim_thread(struct thread_struct *thr, uint8_t cause)if (!MSR_TM_SUSPENDED (mfmsr ()))return ; +save_sprs (thr); +giveup_all (container_of(thr ,struct task_struct ,thread ));tm_reclaim (thr, cause);
@@ -991,6 +995,37 @@ void tm_recheckpoint(struct thread_struct *thread)__tm_recheckpoint (thread); +/* +* This is a stripped down restore_sprs(), we need to do this +* now as we might go straight out to userspace and currently +* the checkpointed values are on the CPU. +* +* TODO: Improve +*/ +#ifdef CONFIG_ALTIVEC +if (cpu_has_feature (CPU_FTR_ALTIVEC )) +mtspr (SPRN_VRSAVE , thread->vrsave); +#endif +#ifdef CONFIG_PPC_BOOK3S_64 +if (cpu_has_feature (CPU_FTR_DSCR )) { +u64 dscr =get_paca ()->dscr_default; +if (thread->dscr_inherit) + dscr = thread->dscr; + +mtspr (SPRN_DSCR , dscr); + } + +if (cpu_has_feature (CPU_FTR_ARCH_207S )) { +/* The EBB regs aren't checkpointed */ +mtspr (SPRN_FSCR , thread->fscr); + +mtspr (SPRN_TAR , thread->tar); + } + +/* I think we don't need to */ +if (cpu_has_feature (CPU_FTR_ARCH_300 )) +mtspr (SPRN_TIDR , thread->tidr); +#endif local_irq_restore (flags); }
@@ -1193,6 +1228,11 @@ struct task_struct *__switch_to(struct task_struct *prev,#endif new_thread = &new->thread; +/* +* Why not &prev->thread; ? +* What is the difference between &prev->thread and +* ¤t->thread ? +*/ old_thread = ¤t->thread;WARN_ON (!irqs_disabled ());
@@ -1237,8 +1277,16 @@ struct task_struct *__switch_to(struct task_struct *prev,/* * We need to save SPRs before treclaim/trecheckpoint as these will * change a number of them. +* +* Because we're now reclaiming on kernel entry, we've had to +* already save them. Don't do it again. +* Note: To deliver a signal in the signal context, we'll have +* turned off TM because we don't want the signal context to +* have the transactional state of the main thread - what if +* we go through switch to at that point? Can we? */ -save_sprs (&prev->thread); +if (!prev->thread.regs || !MSR_TM_ACTIVE (prev->thread.regs->msr)) +save_sprs (&prev->thread);/* Save FPU, Altivec, VSX and SPE state */ giveup_all (prev);
@@ -1260,8 +1308,13 @@ struct task_struct *__switch_to(struct task_struct *prev,* for this is we manually create a stack frame for new tasks that * directly returns through ret_from_fork() or * ret_from_kernel_thread(). See copy_thread() for details. +* +* It isn't stricly nessesary that we avoid the restore here +* because we'll simply restore again after the recheckpoint, +* but we can avoid it for performance reasons. */ -restore_sprs (old_thread, new_thread); +if (!new_thread->regs || !MSR_TM_ACTIVE (new_thread->regs->msr)) +restore_sprs (old_thread, new_thread); last =_switch (old_thread, new_thread);
arch/powerpc/kernel/ptrace.c +3 -6
@@ -135,12 +135,9 @@ static void flush_tmregs_to_thread(struct task_struct *tsk)if ((!cpu_has_feature (CPU_FTR_TM )) || (tsk != current))return ; -if (MSR_TM_SUSPENDED (mfmsr ())) { -tm_reclaim_current (TM_CAUSE_SIGNAL ); - }else { -tm_enable (); -tm_save_sprs (&(tsk->thread)); - } +BUG_ON (MSR_TM_SUSPENDED (mfmsr ())); +tm_enable (); + tm_save_sprs(&(tsk->thread)); }#else static inline void flush_tmregs_to_thread (struct task_struct *tsk) { }
Signature · 1 line
2.16.2
Re: [RFC PATCH 10/12] [WIP] powerpc/tm: Correctly save/restore checkpointed sprs
This needs a description of what you're trying to do. "Correctly" doesn't really mean anything.
Quoted 2 lines · Cyril Bur
---
arch/powerpc/kernel/process.c| 57 +++++++++++++++++++++++++++++++++++++=
++++-
arch/powerpc/kernel/process.=
Quoted 5 lines
-
arch/powerpc/kernel/ptrace.c | 9 +++---- 2 files changed, 58 insertions(+), 8 deletions(-)=20
arch/powerpc/kernel/process.= renamed from arch/powerpc/kernel/process.c +0
-0
c
arch/powerpc/kernel/process.c static inline bool tm_enabled(struct task_struct *tsk=
Quoted 4 lines
index
cd3ae80a6878..674f75c56172100644arch/powerpc/kernel/process.c +0
-0@@ -859,6 +859,8 @@ static inline bool tm_enabled(struct task_struct *tsk=
)
arch/powerpc/kernel/process.c static void tm_reclaim_thread(struct thread_struct *t=
Quoted 9 lines
arch/powerpc/kernel/process.c +0
-0return tsk && tsk->thread.regs && (tsk->thread.regs->msr &MSR_TM ); }=20
arch/powerpc/kernel/process.c +2
-0+static inline void save_sprs (struct thread_struct *t); +static void tm_reclaim_thread (struct thread_struct *thr,uint8_t cause) {/* @@ -879,6 +881,8 @@ static void tm_reclaim_thread(struct thread_struct *t=
hr,
Quoted 34 lines
arch/powerpc/kernel/process.c +0
-0uint8_t cause)if (!MSR_TM_SUSPENDED (mfmsr ()))return ;=20
arch/powerpc/kernel/process.c +2
-0+save_sprs (thr); +giveup_all (container_of(thr ,struct task_struct ,thread ));=20
arch/powerpc/kernel/process.c +0
-0tm_reclaim (thr, cause);@@ -991,6 +995,37 @@ void tm_recheckpoint(struct thread_struct *thread)=20
__tm_recheckpoint(thread);
=20arch/powerpc/kernel/process.c +21
-0+/* +* This is a stripped down restore_sprs(), we need to do this +* now as we might go straight out to userspace and currently +* the checkpointed values are on the CPU. +* +* TODO: Improve +*/ +#ifdef CONFIG_ALTIVEC +if (cpu_has_feature (CPU_FTR_ALTIVEC )) +mtspr (SPRN_VRSAVE , thread->vrsave); +#endif +#ifdef CONFIG_PPC_BOOK3S_64 +if (cpu_has_feature (CPU_FTR_DSCR )) { + u64 dscr =3D get_paca()->dscr_default; +if (thread->dscr_inherit) + dscr =3D thread->dscr; + +mtspr (SPRN_DSCR , dscr); + } + +if (cpu_has_feature (CPU_FTR_ARCH_207S )) {
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 1028
Quoted 10 lines
arch/powerpc/kernel/process.c +10
-0+/* The EBB regs aren't checkpointed */ +mtspr (SPRN_FSCR , thread->fscr); + + mtspr(SPRN_TAR, thread->tar); + } + +/* I think we don't need to */ +if (cpu_has_feature (CPU_FTR_ARCH_300 )) +mtspr (SPRN_TIDR , thread->tidr); +#endif
Why are you touching all the above hunk?
arch/powerpc/kernel/process.c struct task_struct *__switch_to(struct task_struct *prev, · line 1235
Quoted 13 lines
local_irq_restore(flags);
}
=20arch/powerpc/kernel/process.c +0
-0@@ -1193,6 +1228,11 @@ struct task_struct *__switch_to(struct task_struct *prev,#endif =20
arch/powerpc/kernel/process.c +5
-0new_thread =3D &new->thread; +/* +* Why not &prev->thread; ? +* What is the difference between &prev->thread and +* ¤t->thread ? +*/
Why not just work it out and FIX THE CODE, rather than just rabbiting on ab= out
it! :-P
arch/powerpc/kernel/ptrace.c static void flush_tmregs_to_thread(struct task_struc=
Quoted 43 lines
old_thread=3D¤t->thread;
=20arch/powerpc/kernel/process.c +9
-1WARN_ON (!irqs_disabled ());@@ -1237,8 +1277,16 @@ struct task_struct *__switch_to(struct task_struct *prev,/* * We need to save SPRs before treclaim/trecheckpoint as these will * change a number of them. +* +* Because we're now reclaiming on kernel entry, we've had to +* already save them. Don't do it again. +* Note: To deliver a signal in the signal context, we'll have +* turned off TM because we don't want the signal context to +* have the transactional state of the main thread - what if +* we go through switch to at that point? Can we? */ -save_sprs (&prev->thread); + if (!prev->thread.regs || !MSR_TM_ACTIVE(prev->thread.regs->msr)) +save_sprs (&prev->thread);=20
arch/powerpc/kernel/process.c +6
-1/* Save FPU, Altivec, VSX and SPE state */ giveup_all (prev);@@ -1260,8 +1308,13 @@ struct task_struct *__switch_to(struct task_struct*prev, * for this is we manually create a stack frame for new tasks that * directly returns through ret_from_fork() or * ret_from_kernel_thread(). See copy_thread() for details. +* +* It isn't stricly nessesary that we avoid the restore here +* because we'll simply restore again after the recheckpoint, +* but we can avoid it for performance reasons. */ -restore_sprs (old_thread, new_thread); +if (!new_thread->regs || !MSR_TM_ACTIVE (new_thread->regs->msr)) +restore_sprs (old_thread, new_thread);=20
last =3D_switch(old_thread, new_thread);
=20arch/powerpc/kernel/ptrace.c +0
-0@@ -135,12 +135,9 @@ static void flush_tmregs_to_thread(struct task_struc=
t
Quoted 16 lines
arch/powerpc/kernel/ptrace.c +0
-0*tsk) if ((!cpu_has_feature(CPU_FTR_TM)) || (tsk !=3D current))return ;=20
arch/powerpc/kernel/ptrace.c +3
-6-if (MSR_TM_SUSPENDED (mfmsr ())) { -tm_reclaim_current (TM_CAUSE_SIGNAL ); - }else { -tm_enable (); -tm_save_sprs (&(tsk->thread)); - } +BUG_ON (MSR_TM_SUSPENDED (mfmsr ())); +tm_enable (); + tm_save_sprs(&(tsk->thread)); }#else static inline void flush_tmregs_to_thread (struct task_struct *tsk) { }
Re: [RFC PATCH 10/12] [WIP] powerpc/tm: Correctly save/restore checkpointed sprs
Quoted 63 lines · Michael Neuling
This needs a description of what you're trying to do. "Correctly" doesn't really mean anything.
Quoted 58 lines · Cyril Bur
--- arch/powerpc/kernel/process.c | 57 +++++++++++++++++++++++++++++++++++++++++-(quoted diff) +0
-1- arch/powerpc/kernel/ptrace.c | 9 +++---- 2 files changed, 58 insertions(+), 8 deletions(-)arch/powerpc/kernel/process.c +4
-0@@ -859,6 +859,8 @@ static inline bool tm_enabled(struct task_struct *tsk)return tsk && tsk->thread.regs && (tsk->thread.regs->msr &MSR_TM ); } +static inline void save_sprs (struct thread_struct *t); +static void tm_reclaim_thread (struct thread_struct *thr,uint8_t cause) {/* @@ -879,6 +881,8 @@ static void tm_reclaim_thread(struct thread_struct *thr, uint8_t cause)if (!MSR_TM_SUSPENDED (mfmsr ()))return ; +save_sprs (thr); +giveup_all (container_of(thr ,struct task_struct ,thread ));arch/powerpc/kernel/process.c +25
-0tm_reclaim (thr, cause);@@ -991,6 +995,37 @@ void tm_recheckpoint(struct thread_struct *thread)__tm_recheckpoint (thread); +/* +* This is a stripped down restore_sprs(), we need to do this +* now as we might go straight out to userspace and currently +* the checkpointed values are on the CPU. +* +* TODO: Improve +*/ +#ifdef CONFIG_ALTIVEC +if (cpu_has_feature (CPU_FTR_ALTIVEC )) +mtspr (SPRN_VRSAVE , thread->vrsave); +#endif +#ifdef CONFIG_PPC_BOOK3S_64 +if (cpu_has_feature (CPU_FTR_DSCR )) { +u64 dscr =get_paca ()->dscr_default; +if (thread->dscr_inherit) + dscr = thread->dscr; + +mtspr (SPRN_DSCR , dscr); + } + +if (cpu_has_feature (CPU_FTR_ARCH_207S )) { +/* The EBB regs aren't checkpointed */ +mtspr (SPRN_FSCR , thread->fscr); + +mtspr (SPRN_TAR , thread->tar);
Quoted 8 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 1028
Quoted 6 lines
arch/powerpc/kernel/process.c +6
-0+ } + +/* I think we don't need to */ +if (cpu_has_feature (CPU_FTR_ARCH_300 )) +mtspr (SPRN_TIDR , thread->tidr); +#endif Why are you touching all the above hunk?
I copied restore_sprs. I'm tidying that up now - we can't call restore_sprs because we don't have a prev and next thread.
Quoted 8 lines
Quoted 8 lines
arch/powerpc/kernel/process.c +0
-0local_irq_restore (flags); }@@ -1193,6 +1228,11 @@ struct task_struct *__switch_to(struct task_struct *prev,#endif
Quoted 9 lines
arch/powerpc/kernel/process.c struct task_struct *__switch_to(struct task_struct *prev, · line 1235
Quoted 6 lines
arch/powerpc/kernel/process.c +5
-0new_thread = &new->thread; +/* +* Why not &prev->thread; ? +* What is the difference between &prev->thread and +* ¤t->thread ? +*/ Why not just work it out and FIX THE CODE, rather than just rabbiting on about it! :-P
Agreed - I started to and then had a mini freakout that things would end really badly if they're not the same. So I left that comment as a reminder to investigate.
They should be the same though right?
Quoted 60 lines
Quoted 60 lines
arch/powerpc/kernel/process.c +0
-0old_thread = ¤t->thread;arch/powerpc/kernel/process.c +9
-1WARN_ON (!irqs_disabled ());@@ -1237,8 +1277,16 @@ struct task_struct *__switch_to(struct task_struct *prev,/* * We need to save SPRs before treclaim/trecheckpoint as these will * change a number of them. +* +* Because we're now reclaiming on kernel entry, we've had to +* already save them. Don't do it again. +* Note: To deliver a signal in the signal context, we'll have +* turned off TM because we don't want the signal context to +* have the transactional state of the main thread - what if +* we go through switch to at that point? Can we? */ -save_sprs (&prev->thread); + if (!prev->thread.regs || !MSR_TM_ACTIVE(prev->thread.regs->msr)) +save_sprs (&prev->thread);/* Save FPU, Altivec, VSX and SPE state */ arch/powerpc/kernel/process.c +6
-1giveup_all (prev);@@ -1260,8 +1308,13 @@ struct task_struct *__switch_to(struct task_struct*prev, * for this is we manually create a stack frame for new tasks that * directly returns through ret_from_fork() or * ret_from_kernel_thread(). See copy_thread() for details. +* +* It isn't stricly nessesary that we avoid the restore here +* because we'll simply restore again after the recheckpoint, +* but we can avoid it for performance reasons. */ -restore_sprs (old_thread, new_thread); +if (!new_thread->regs || !MSR_TM_ACTIVE (new_thread->regs->msr)) +restore_sprs (old_thread, new_thread); last =_switch (old_thread, new_thread);arch/powerpc/kernel/ptrace.c +3
-6@@ -135,12 +135,9 @@ static void flush_tmregs_to_thread(struct task_struct *tsk)if ((!cpu_has_feature (CPU_FTR_TM )) || (tsk != current))return ; -if (MSR_TM_SUSPENDED (mfmsr ())) { -tm_reclaim_current (TM_CAUSE_SIGNAL ); - }else { -tm_enable (); -tm_save_sprs (&(tsk->thread)); - } +BUG_ON (MSR_TM_SUSPENDED (mfmsr ())); +tm_enable (); + tm_save_sprs(&(tsk->thread)); }#else static inline void
flush_tmregs_to_thread(struct task_struct *tsk){ }
Re: [RFC PATCH 10/12] [WIP] powerpc/tm: Correctly save/restore checkpointed sprs
Quoted 2 lines · Cyril Bur
Quoted 1 line · Michael Neuling
This needs a description of what you're trying to do. "Correctly" does=
n't
Quoted 6 lines
Quoted 6 lines
really mean anything.
=20
=20Quoted 2 lines · Cyril Bur
---
arch/powerpc/kernel/process.c| 57 +++++++++++++++++++++++++++++++++=
++++++++-
Quoted 5 lines
Quoted 5 lines
arch/powerpc/kernel/proc=
Quoted 5 lines
-
arch/powerpc/kernel/ptrace.c | 9 +++---- 2 files changed, 58 insertions(+), 8 deletions(-)=20
arch/powerpc/kernel/proc= renamed from arch/powerpc/kernel/process.c +0
-0
ess.c
Quoted 4 lines
Quoted 4 lines
arch/powerpc/kernel/process.c static inline bool tm_enabled(struct task_struct =
Quoted 4 lines
index
cd3ae80a6878..674f75c56172100644arch/powerpc/kernel/process.c +0
-0@@ -859,6 +859,8 @@ static inline bool tm_enabled(struct task_struct =
*tsk)
Quoted 6 lines
Quoted 6 lines
arch/powerpc/kernel/process.c static inline bool tm_enabled(struct task_struct *tsk)
Quoted 6 lines
arch/powerpc/kernel/process.c +0
-0return tsk && tsk->thread.regs && (tsk->thread.regs->msr &MSR_TM ); }=20
arch/powerpc/kernel/process.c +2
-0+static inline void save_sprs (struct thread_struct *t); + static void tm_reclaim_thread(struct thread_struct *thr, uint8_t cau=
se)
Quoted 3 lines
Quoted 3 lines
arch/powerpc/kernel/process.c static void tm_reclaim_thread(struct thread_struc=
Quoted 3 lines
{
arch/powerpc/kernel/process.c +0
-0/* @@ -879,6 +881,8 @@ static void tm_reclaim_thread(struct thread_struc=
t *thr,
Quoted 10 lines
Quoted 10 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *threa=
Quoted 10 lines
arch/powerpc/kernel/process.c +0
-0uint8_t cause)if (!MSR_TM_SUSPENDED (mfmsr ()))return ;=20
arch/powerpc/kernel/process.c +2
-0+save_sprs (thr); +giveup_all (container_of(thr ,struct task_struct ,thread ));=20
arch/powerpc/kernel/process.c +0
-0tm_reclaim (thr, cause);@@ -991,6 +995,37 @@ void tm_recheckpoint(struct thread_struct *threa=
d)
Quoted 36 lines
Quoted 28 lines
Quoted 28 lines
=20
__tm_recheckpoint(thread);
=20arch/powerpc/kernel/process.c +25
-0+/* +* This is a stripped down restore_sprs(), we need to do this +* now as we might go straight out to userspace and currently +* the checkpointed values are on the CPU. +* +* TODO: Improve +*/ +#ifdef CONFIG_ALTIVEC +if (cpu_has_feature (CPU_FTR_ALTIVEC )) +mtspr (SPRN_VRSAVE , thread->vrsave); +#endif +#ifdef CONFIG_PPC_BOOK3S_64 +if (cpu_has_feature (CPU_FTR_DSCR )) { + u64 dscr =3D get_paca()->dscr_default; +if (thread->dscr_inherit) + dscr =3D thread->dscr; + +mtspr (SPRN_DSCR , dscr); + } + +if (cpu_has_feature (CPU_FTR_ARCH_207S )) { +/* The EBB regs aren't checkpointed */ +mtspr (SPRN_FSCR , thread->fscr); + +mtspr (SPRN_TAR , thread->tar);Quoted 8 lines
arch/powerpc/kernel/process.c void tm_recheckpoint(struct thread_struct *thread) · line 1028
Quoted 6 lines
arch/powerpc/kernel/process.c +6
-0+ } + +/* I think we don't need to */ +if (cpu_has_feature (CPU_FTR_ARCH_300 )) +mtspr (SPRN_TIDR , thread->tidr); +#endif =20
Why are you touching all the above hunk?
Quoted 3 lines
=20
I copiedrestore_sprs. I'm tidying that up now - we can't callrestore_sprsbecause we don't have a prev and next thread.
Yeah needs to be tided up... we can't have another copy of the code.. obvio= usly.
Quoted 6 lines
=20
Quoted 5 lines
=20
arch/powerpc/kernel/process.c struct task_struct *__switch_to(struct task_st=
Quoted 4 lines
local_irq_restore(flags);
}
=20arch/powerpc/kernel/process.c +0
-0@@ -1193,6 +1228,11 @@ struct task_struct *__switch_to(struct task_st=
ruct
Quoted 11 lines
Quoted 11 lines
arch/powerpc/kernel/process.c struct task_struct *__switch_to(struct task_struct *prev, · line 1235
Quoted 9 lines
arch/powerpc/kernel/process.c +0
-0*prev,#endif =20
arch/powerpc/kernel/process.c +5
-0new_thread =3D &new->thread; +/* +* Why not &prev->thread; ? +* What is the difference between &prev->thread and +* ¤t->thread ? +*/ =20
Why not just work it out and FIX THE CODE, rather than just rabbiting o=
n about
Quoted 7 lines
Quoted 1 line
it! :-P
=20
Agreed - I started to and then had a mini freakout that things would end really badly if they're not the same. So I left that comment as a reminder to investigate.
=20
They should be the same though right?
Should be if prev =3D=3D current.
Mikey
RFC PATCH 11/12 powerpc/tm: Afterthoughts
In reply to Cyril Bur--- arch/powerpc/kernel/process.c | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-)
arch/powerpc/kernel/process.c +17 -1
@@ -1079,6 +1079,12 @@ static inline void __switch_to_tm(struct task_struct *prev,if (!MSR_TM_ACTIVE (prev->thread.regs->msr) && prev->thread.load_tm ==0 ) prev->thread.regs->msr &= ~MSR_TM ; } + +/* +* Now that we're reclaiming on kernel entry, we should never +* get here still with user checkpointed state on the CPU +*/ +BUG_ON (MSR_TM_ACTIVE (mfmsr ())); }/*
@@ -1326,7 +1332,17 @@ struct task_struct *__switch_to(struct task_struct *prev, }if (current_thread_info ()->task->thread.regs) { -restore_math (current_thread_info ()->task->thread.regs); +/* +* Calling this now has reloaded the live state, which +* gets overwritten with the checkpointed state right +* before the trecheckpoint. BUT the MSR still has +* that the live state is on the CPU, which it isn't. +* +* restore_math(current_thread_info()->task->thread.regs); +* Therefore: +*/ +if (!MSR_TM_ACTIVE (current_thread_info ()->task->thread.regs->msr)) +restore_math (current_thread_info ()->task->thread.regs);/* * The copy-paste buffer can only store into foreign real
Signature · 1 line
2.16.2
RFC PATCH 12/12 selftests/powerpc: Remove incorrect tm-syscall selftest
In reply to Cyril BurCurrently we perform transactional memory work at late as possible. That is we run in the kernel with the userspace checkpointed state on the CPU untill we absolultely must remove it and store it away. Likely a process switch, but possibly also signals or ptrace.
What this means is that if userspace does a system call in suspended mode, it is possible that we will handle the system call and return them without the need to to a reclaim/recheckpoint and so they can expect to resume their transaction.
This is what tm-syscall tests for - the ability to perform a system call in suspended state and still resume it afterwards.
TM reworks have meant that we now deal with any transactional state on entry to the kernel, no matter the reason for entry (some expections apply). We will categorically doom any suspended transaction that makes a system call, making that transaction unresumeable.
This test will now always fail no matter what. I would like to note here that this new behaviour does not break userspace at all. Hardware Transactional Memory gives zero guarantee of forward progress and any correct userspace has already had and will always have to implement a non HTM fallback. Relying on this specific kernel behaviour also meant relying on the stars aligning in the hardware such that there was no cache overlaps and that it had a large enough footprint to handle any system call without dooming a transaction.
--- tools/testing/selftests/powerpc/tm/Makefile | 4 +- .../testing/selftests/powerpc/tm/tm-syscall-asm.S | 28 ------ tools/testing/selftests/powerpc/tm/tm-syscall.c | 106 --------------------- 3 files changed, 1 insertion(+), 137 deletions(-) delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall-asm.S delete mode 100644 tools/testing/selftests/powerpc/tm/tm-syscall.c
tools/testing/selftests/powerpc/tm/Makefile +1 -3
@@ -2,7 +2,7 @@ SIGNAL_CONTEXT_CHK_TESTS := tm-signal-context-chk-gpr tm-signal-context-chk-fpu \ tm-signal-context-chk-vmx tm-signal-context-chk-vsx -TEST_GEN_PROGS := tm-resched-dscr tm-syscall tm-signal-msr-resv tm-signal-stack \ +TEST_GEN_PROGS := tm-resched-dscr tm-signal-msr-resv tm-signal-stack \ tm-vmxcopy tm-fork tm-tar tm-tmspr tm-vmx-unavail tm-unavailable tm-trap \ tm-signal-drop-transaction \ $(SIGNAL_CONTEXT_CHK_TESTS)
@@ -13,8 +13,6 @@ $(TEST_GEN_PROGS): ../harness.c ../utils.c CFLAGS += -mhtm -$(OUTPUT)/tm-syscall: tm-syscall-asm.S -$(OUTPUT)/tm-syscall: CFLAGS += -I../../../../../usr/include $(OUTPUT)/tm-tmspr: CFLAGS += -pthread $(OUTPUT)/tm-vmx-unavail: CFLAGS += -pthread -m64 $(OUTPUT)/tm-resched-dscr: ../pmu/lib.o
tools/testing/selftests/powerpc/tm/tm-syscall-asm.S deleted +0 -28
@@ -1,28 +0,0 @@ -/* SPDX-License-Identifier: GPL-2.0 */ -#include <ppc-asm.h> -#include <asm/unistd.h> - - .text -FUNC_START(getppid_tm_active) - tbegin. - beq 1f - li r0, __NR_getppid - sc - tend. - blr -1: - li r3, -1 - blr - -FUNC_START(getppid_tm_suspended) - tbegin. - beq 1f - li r0, __NR_getppid - tsuspend. - sc - tresume. - tend. - blr -1: - li r3, -1 - blr
tools/testing/selftests/powerpc/tm/tm-syscall.c deleted +0 -106
106 lines not on this page · raw message
Signature · 1 line
2.16.2
Re: [RFC PATCH 12/12] [WIP] selftests/powerpc: Remove incorrect tm-syscall selftest
tools/testing/selftests/powerpc/tm/tm-syscall.c line 7
Quoted 10 lines
tools/testing/selftests/powerpc/tm/tm-syscall.c deleted +0
-7@@ -1,106 +0,0 @@ -/* -* Copyright 2015, Sam Bobroff, IBM Corp. -* Licensed under GPLv2. -* -* Test the kernel's system call code to ensure that a system call -* made from within an active HTM transaction is aborted with the -* correct failure code.
The above is still true
tools/testing/selftests/powerpc/tm/tm-syscall.c line 9
Quoted 2 lines
tools/testing/selftests/powerpc/tm/tm-syscall.c deleted +0
-2-* Conversely, ensure that a system call made from within a -* suspended transaction can succeed.
This is true anymore....
So can we just modify the test to remove the second part?
Mikey
Re: [RFC PATCH 12/12] [WIP] selftests/powerpc: Remove incorrect tm-syscall selftest
Quoted 13 lines · Michael Neuling
tools/testing/selftests/powerpc/tm/tm-syscall.c line 7
Quoted 10 lines
tools/testing/selftests/powerpc/tm/tm-syscall.c deleted +0
-7@@ -1,106 +0,0 @@ -/* -* Copyright 2015, Sam Bobroff, IBM Corp. -* Licensed under GPLv2. -* -* Test the kernel's system call code to ensure that a system call -* made from within an active HTM transaction is aborted with the -* correct failure code. The above is still true
Quoted 7 lines
tools/testing/selftests/powerpc/tm/tm-syscall.c line 9
Quoted 2 lines
tools/testing/selftests/powerpc/tm/tm-syscall.c deleted +0
-2-* Conversely, ensure that a system call made from within a -* suspended transaction can succeed. This is true anymore....
So can we just modify the test to remove the second part?
Oh true I overlooked that
Thanks
Quoted 1 line
Mikey
Hi Cyril,
Quoted 7 lines · Cyril Bur
This is very much a proof of concept and if it isn't clear from the commit names, still a work in progress.
I believe I have something that works - all the powerpc selftests pass. I would like to get some eyes on it to a) see if I've missed anything big and b) some opinions on if it is looking like a net improvement.
I started to look at this patchset. The patchset apply cleanly on top of the current kernel and I started to run some tests.
It seems there is a different behavior when there is a trap (as a illegal instruction or a 'trap' instruction) inside the transaction.
In this case, the signal handler does not seem to be called, and and the task segfaults. On current upstream, the signal handler is called and the program can continue.
4.17 pristine
-------------
$ ./illegal Failure
4.17 plus your patches
----------------------
$ ./illegal [1] 2504 segmentation fault ./illegal
Here is a minimal example that is able to recreate this behaviour:
#include <stdio.h>
#include <signal.h>
int htm(){
asm goto ("tbegin. \n\t"
"beq %l[failure] \n\t"
"li 3, 3 \n\t"
"trap \n\t"
"tend. \n\t"
: : : : failure);
return 0;
failure:
printf("Failure\n");
return 1;
}
void signal_handler(int signo, siginfo_t *si, void *data) {
// Do nothing
}
int main(){
struct sigaction sa;
sa.sa_flags = SA_SIGINFO;
sa.sa_sigaction = signal_handler;
sigaction(SIGTRAP, &sa, NULL);
sigaction(SIGILL, &sa, NULL);
return htm();
}