| Message ID | 20241203-arm64-sme-reenable-v1-5-d853479d1b77@kernel.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | arm64/sme: Collected SME fixes | expand |
On Tue, Dec 03, 2024 at 12:45:57PM +0000, Mark Brown wrote: > We intend that signal handlers are entered with PSTATE.{SM,ZA}={0,0}. > The logic for this in setup_return() manipulates the saved state and > live CPU state in an unsafe manner, and consequently, when a task enters > a signal handler: > > * The task entering the signal handler might not have its PSTATE.{SM,ZA} > bits cleared, and other register state that is affected by changes to > PSTATE.{SM,ZA} might not be zeroed as expected. > > * An unrelated task might have its PSTATE.{SM,ZA} bits cleared > unexpectedly, potentially zeroing other register state that is > affected by changes to PSTATE.{SM,ZA}. > > Tasks which do not set PSTATE.{SM,ZA} (i.e. those only using plain > FPSIMD or non-streaming SVE) are not affected, as there is no > resulting change to PSTATE.{SM,ZA}. > > Consider for example two tasks on one CPU: > > A: Begins signal entry in kernel mode, is preempted prior to SMSTOP. > B: Using SM and/or ZA in userspace with register state current on the > CPU, is preempted. > A: Scheduled in, no register state changes made as in kernel mode. > A: Executes SMSTOP, modifying live register state. > A: Scheduled out. > B: Scheduled in, fpsimd_thread_switch() sees the register state on the > CPU is tracked as being that for task B so the state is not reloaded > prior to returning to userspace. > > Task B is now running with SM and ZA incorrectly cleared. > > Fix this by: > > * Checking TIF_FOREIGN_FPSTATE, and only updating the saved or live > state as appropriate. > > * Using {get,put}_cpu_fpsimd_context() to ensure mutual exclusion > against other code which manipulates this state. To allow their use, > the logic is moved into a new fpsimd_enter_sighandler() helper in > fpsimd.c. > > This race has been observed intermittently with fp-stress, especially > with preempt disabled, commonly but not exclusively reporting "Bad SVCR: 0". > > While we're at it also fix a discrepancy between in register and in memory > entries. When operating on the register state we issue a SMSTOP, exiting > streaming mode if we were in it. This clears the V/Z and P register and > FPMR but nothing else. The in memory version clears all the user FPSIMD > state including FPCR and FPSR but does not clear FPMR. Add the clear of > FPMR and limit the existing memset() to only cover the vregs, preserving > the state of FPCR and FPSR like SMSTOP does. > > Fixes: 40a8e87bb3285 ("arm64/sme: Disable ZA and streaming mode when handling signals") > Signed-off-by: Mark Brown <broonie@kernel.org> > Cc: stable@vger.kernel.org > --- > arch/arm64/include/asm/fpsimd.h | 1 + > arch/arm64/kernel/fpsimd.c | 39 +++++++++++++++++++++++++++++++++++++++ > arch/arm64/kernel/signal.c | 19 +------------------ > 3 files changed, 41 insertions(+), 18 deletions(-) > > diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h > index f2a84efc361858d4deda99faf1967cc7cac386c1..09af7cfd9f6c2cec26332caa4c254976e117b1bf 100644 > --- a/arch/arm64/include/asm/fpsimd.h > +++ b/arch/arm64/include/asm/fpsimd.h > @@ -76,6 +76,7 @@ extern void fpsimd_load_state(struct user_fpsimd_state *state); > extern void fpsimd_thread_switch(struct task_struct *next); > extern void fpsimd_flush_thread(void); > > +extern void fpsimd_enter_sighandler(void); > extern void fpsimd_signal_preserve_current_state(void); > extern void fpsimd_preserve_current_state(void); > extern void fpsimd_restore_current_state(void); > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c > index f02762762dbcf954e9add6dfd3575ae7055b6b0e..c5465c8ec467cb1ab8bd211dc5370f91aa2bcf35 100644 > --- a/arch/arm64/kernel/fpsimd.c > +++ b/arch/arm64/kernel/fpsimd.c > @@ -1696,6 +1696,45 @@ void fpsimd_signal_preserve_current_state(void) > sve_to_fpsimd(current); > } > > +/* > + * Called by the signal handling code when preparing current to enter > + * a signal handler. Currently this only needs to take care of exiting > + * streaming mode and clearing ZA on SME systems. > + */ > +void fpsimd_enter_sighandler(void) > +{ > + if (!system_supports_sme()) > + return; > + > + get_cpu_fpsimd_context(); > + > + if (test_thread_flag(TIF_FOREIGN_FPSTATE)) { > + /* > + * Exiting streaming mode zeros the V/Z and P > + * registers and FPMR. Zero FPMR and the V registers, > + * marking the state as FPSIMD only to force a clear > + * of the remaining bits during reload if needed. > + */ > + if (current->thread.svcr & SVCR_SM_MASK) { > + memset(¤t->thread.uw.fpsimd_state.vregs, 0, > + sizeof(current->thread.uw.fpsimd_state.vregs)); Do we need to hold the CPU fpsimd context across this memset? IIRC, TIF_FOREIGN_FPSTATE can be spontaneously cleared along with dumping of the regs into thread_struct (from current's PoV), but never spontaneously set again. So ... -> [*] > + current->thread.uw.fpmr = 0; > + current->thread.fp_type = FP_STATE_FPSIMD; > + } > + > + current->thread.svcr &= ~(SVCR_ZA_MASK | > + SVCR_SM_MASK); > + > + /* Ensure any copies on other CPUs aren't reused */ > + fpsimd_flush_task_state(current); (This is very similar to fpsimd_flush_thread(); can they be unified?) > + } else { > + /* The register state is current, just update it. */ > + sme_smstop(); ... [*] the critical thing seems to be that the CPU fpsimd context is held from the test on TIF_FOREIGN_FPSTATE, across this else clause. (Whether or not this is a worthwhile optimisation is another matter. But if the behaviour of TIF_FOREIGN_FPSTATE is still the same, then it may be a good idea to avoid sending mixed messages about this in the code.) (A similar argument applies in fpsimd_flush_thread().) [...] Cheers ---Dave
On Tue, Dec 03, 2024 at 03:33:18PM +0000, Dave Martin wrote: > On Tue, Dec 03, 2024 at 12:45:57PM +0000, Mark Brown wrote: > > + get_cpu_fpsimd_context(); > > + if (current->thread.svcr & SVCR_SM_MASK) { > > + memset(¤t->thread.uw.fpsimd_state.vregs, 0, > > + sizeof(current->thread.uw.fpsimd_state.vregs)); > Do we need to hold the CPU fpsimd context across this memset? > IIRC, TIF_FOREIGN_FPSTATE can be spontaneously cleared along with > dumping of the regs into thread_struct (from current's PoV), but never > spontaneously set again. So ... -> [*] Yes, we could drop the lock here. OTOH this is very simple and easy to understand. > > + /* Ensure any copies on other CPUs aren't reused */ > > + fpsimd_flush_task_state(current); > (This is very similar to fpsimd_flush_thread(); can they be unified?) I have a half finished series to replace the whole setup around accessing the state with get/put operations for working on the state which should remove all these functions. The pile of similarly and confusingly named operations we have for working on the state is one of the major sources of issues with this code, even when actively working on the code it's hard to remember exactly which operation does what never mind the rules for which is needed.
Hi, On Tue, Dec 03, 2024 at 04:12:33PM +0000, Mark Brown wrote: > On Tue, Dec 03, 2024 at 03:33:18PM +0000, Dave Martin wrote: > > On Tue, Dec 03, 2024 at 12:45:57PM +0000, Mark Brown wrote: > > > > + get_cpu_fpsimd_context(); > > > > + if (current->thread.svcr & SVCR_SM_MASK) { > > > + memset(¤t->thread.uw.fpsimd_state.vregs, 0, > > > + sizeof(current->thread.uw.fpsimd_state.vregs)); > > > Do we need to hold the CPU fpsimd context across this memset? > > > IIRC, TIF_FOREIGN_FPSTATE can be spontaneously cleared along with > > dumping of the regs into thread_struct (from current's PoV), but never > > spontaneously set again. So ... -> [*] > > Yes, we could drop the lock here. OTOH this is very simple and easy to > understand. Ack; it works either way. Since this is a Fixes: patch, it may be better to keep it simple. > > > > + /* Ensure any copies on other CPUs aren't reused */ > > > + fpsimd_flush_task_state(current); > > > (This is very similar to fpsimd_flush_thread(); can they be unified?) > > I have a half finished series to replace the whole setup around > accessing the state with get/put operations for working on the state > which should remove all these functions. The pile of similarly and > confusingly named operations we have for working on the state is one of > the major sources of issues with this code, even when actively working > on the code it's hard to remember exactly which operation does what > never mind the rules for which is needed. Sure, something like that would definitely help. Cheers ---Dave
diff --git a/arch/arm64/include/asm/fpsimd.h b/arch/arm64/include/asm/fpsimd.h index f2a84efc361858d4deda99faf1967cc7cac386c1..09af7cfd9f6c2cec26332caa4c254976e117b1bf 100644 --- a/arch/arm64/include/asm/fpsimd.h +++ b/arch/arm64/include/asm/fpsimd.h @@ -76,6 +76,7 @@ extern void fpsimd_load_state(struct user_fpsimd_state *state); extern void fpsimd_thread_switch(struct task_struct *next); extern void fpsimd_flush_thread(void); +extern void fpsimd_enter_sighandler(void); extern void fpsimd_signal_preserve_current_state(void); extern void fpsimd_preserve_current_state(void); extern void fpsimd_restore_current_state(void); diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c index f02762762dbcf954e9add6dfd3575ae7055b6b0e..c5465c8ec467cb1ab8bd211dc5370f91aa2bcf35 100644 --- a/arch/arm64/kernel/fpsimd.c +++ b/arch/arm64/kernel/fpsimd.c @@ -1696,6 +1696,45 @@ void fpsimd_signal_preserve_current_state(void) sve_to_fpsimd(current); } +/* + * Called by the signal handling code when preparing current to enter + * a signal handler. Currently this only needs to take care of exiting + * streaming mode and clearing ZA on SME systems. + */ +void fpsimd_enter_sighandler(void) +{ + if (!system_supports_sme()) + return; + + get_cpu_fpsimd_context(); + + if (test_thread_flag(TIF_FOREIGN_FPSTATE)) { + /* + * Exiting streaming mode zeros the V/Z and P + * registers and FPMR. Zero FPMR and the V registers, + * marking the state as FPSIMD only to force a clear + * of the remaining bits during reload if needed. + */ + if (current->thread.svcr & SVCR_SM_MASK) { + memset(¤t->thread.uw.fpsimd_state.vregs, 0, + sizeof(current->thread.uw.fpsimd_state.vregs)); + current->thread.uw.fpmr = 0; + current->thread.fp_type = FP_STATE_FPSIMD; + } + + current->thread.svcr &= ~(SVCR_ZA_MASK | + SVCR_SM_MASK); + + /* Ensure any copies on other CPUs aren't reused */ + fpsimd_flush_task_state(current); + } else { + /* The register state is current, just update it. */ + sme_smstop(); + } + + put_cpu_fpsimd_context(); +} + /* * Called by KVM when entering the guest. */ diff --git a/arch/arm64/kernel/signal.c b/arch/arm64/kernel/signal.c index abd0907061fe664bf22d1995319f9559c4bbed91..335c2327baf74eac9634cf594855dbf26a7d6b01 100644 --- a/arch/arm64/kernel/signal.c +++ b/arch/arm64/kernel/signal.c @@ -1461,24 +1461,7 @@ static int setup_return(struct pt_regs *regs, struct ksignal *ksig, /* TCO (Tag Check Override) always cleared for signal handlers */ regs->pstate &= ~PSR_TCO_BIT; - /* Signal handlers are invoked with ZA and streaming mode disabled */ - if (system_supports_sme()) { - /* - * If we were in streaming mode the saved register - * state was SVE but we will exit SM and use the - * FPSIMD register state - flush the saved FPSIMD - * register state in case it gets loaded. - */ - if (current->thread.svcr & SVCR_SM_MASK) { - memset(¤t->thread.uw.fpsimd_state, 0, - sizeof(current->thread.uw.fpsimd_state)); - current->thread.fp_type = FP_STATE_FPSIMD; - } - - current->thread.svcr &= ~(SVCR_ZA_MASK | - SVCR_SM_MASK); - sme_smstop(); - } + fpsimd_enter_sighandler(); if (ksig->ka.sa.sa_flags & SA_RESTORER) sigtramp = ksig->ka.sa.sa_restorer;
We intend that signal handlers are entered with PSTATE.{SM,ZA}={0,0}. The logic for this in setup_return() manipulates the saved state and live CPU state in an unsafe manner, and consequently, when a task enters a signal handler: * The task entering the signal handler might not have its PSTATE.{SM,ZA} bits cleared, and other register state that is affected by changes to PSTATE.{SM,ZA} might not be zeroed as expected. * An unrelated task might have its PSTATE.{SM,ZA} bits cleared unexpectedly, potentially zeroing other register state that is affected by changes to PSTATE.{SM,ZA}. Tasks which do not set PSTATE.{SM,ZA} (i.e. those only using plain FPSIMD or non-streaming SVE) are not affected, as there is no resulting change to PSTATE.{SM,ZA}. Consider for example two tasks on one CPU: A: Begins signal entry in kernel mode, is preempted prior to SMSTOP. B: Using SM and/or ZA in userspace with register state current on the CPU, is preempted. A: Scheduled in, no register state changes made as in kernel mode. A: Executes SMSTOP, modifying live register state. A: Scheduled out. B: Scheduled in, fpsimd_thread_switch() sees the register state on the CPU is tracked as being that for task B so the state is not reloaded prior to returning to userspace. Task B is now running with SM and ZA incorrectly cleared. Fix this by: * Checking TIF_FOREIGN_FPSTATE, and only updating the saved or live state as appropriate. * Using {get,put}_cpu_fpsimd_context() to ensure mutual exclusion against other code which manipulates this state. To allow their use, the logic is moved into a new fpsimd_enter_sighandler() helper in fpsimd.c. This race has been observed intermittently with fp-stress, especially with preempt disabled, commonly but not exclusively reporting "Bad SVCR: 0". While we're at it also fix a discrepancy between in register and in memory entries. When operating on the register state we issue a SMSTOP, exiting streaming mode if we were in it. This clears the V/Z and P register and FPMR but nothing else. The in memory version clears all the user FPSIMD state including FPCR and FPSR but does not clear FPMR. Add the clear of FPMR and limit the existing memset() to only cover the vregs, preserving the state of FPCR and FPSR like SMSTOP does. Fixes: 40a8e87bb3285 ("arm64/sme: Disable ZA and streaming mode when handling signals") Signed-off-by: Mark Brown <broonie@kernel.org> Cc: stable@vger.kernel.org --- arch/arm64/include/asm/fpsimd.h | 1 + arch/arm64/kernel/fpsimd.c | 39 +++++++++++++++++++++++++++++++++++++++ arch/arm64/kernel/signal.c | 19 +------------------ 3 files changed, 41 insertions(+), 18 deletions(-)