printk, tracing: fix console tracepoint
PATCH 1/1 printk, tracing: fix console tracepoint
The original intent of the 'console' tracepoint per 95100358491a ("printk/tracing: Add console output tracing") had been to "[...] record any printk messages into the trace, regardless of the current console loglevel. This can help correlate (existing) printk debugging with other tracing."
Petr points out [1] that calling trace_console_rcuidle() in call_console_driver() had been the wrong thing for a while, because "printk() always used console_trylock() and the message was flushed to the console only when the trylock succeeded. And it was always deferred in NMI or when printed via printk_deferred()."
With 09c5ba0aa2fc ("printk: add kthread console printers"), things only got worse, and calls to call_console_driver() no longer happen with typical printk() calls but always appear deferred [2].
As such, the tracepoint can no longer serve its purpose to clearly correlate printk() calls and other tracing, as well as breaks usecases that expect every printk() call to result in a callback of the console tracepoint. Notably, the KFENCE and KCSAN test suites, which want to capture console output and assume a printk() immediately gives us a callback to the console tracepoint.
Fix the console tracepoint by moving it into printk_sprint() [3].
One notable difference is that by moving tracing into printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on the printk() call and isn't delayed.
--- kernel/printk/printk.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
kernel/printk/printk.c +2 -2
@@ -2064,8 +2064,6 @@ static void call_console_driver(struct console *con, const char *text, size_t le {size_t dropped_len; -trace_console_rcuidle (text, len); -if (con->dropped && dropped_text) { dropped_len = snprintf(dropped_text, DROPPED_TEXT_MAX, "** %lu printk messages dropped **\n",
@@ -2240,6 +2238,8 @@ static u16 printk_sprint(char *text, u16 size, int facility, } } +trace_console_rcuidle (text, text_len); +return text_len; }
Signature · 1 line
2.36.0.464.gb9c8b46e94-goog
Quoted 23 lines · On Tue, 3 May 2022 09:38:44 +0200 Marco Elver <[email protected]> wrote:
The original intent of the 'console' tracepoint per
95100358491a("printk/tracing: Add console output tracing") had been to "[...] record any printk messages into the trace, regardless of the current console loglevel. This can help correlate (existing) printk debugging with other tracing."Petr points out [1] that calling
trace_console_rcuidle()incall_console_driver()had been the wrong thing for a while, because "printk()always usedconsole_trylock()and the message was flushed to the console only when the trylock succeeded. And it was always deferred in NMI or when printed viaprintk_deferred()."With
09c5ba0aa2fc("printk: add kthread console printers"), things only got worse, and calls tocall_console_driver()no longer happen with typicalprintk()calls but always appear deferred [2].As such, the tracepoint can no longer serve its purpose to clearly correlate
printk()calls and other tracing, as well as breaks usecases that expect everyprintk()call to result in a callback of the console tracepoint. Notably, the KFENCE and KCSAN test suites, which want to capture console output and assume aprintk()immediately gives us a callback to the console tracepoint.
Quoted 7 lines
Fix the console tracepoint by moving it into
printk_sprint()[3].One notable difference is that by moving tracing into
printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on theprintk()call and isn't delayed.
I'm OK with this change, but I don't know everyone that uses the trace printk feature. I am worried that this could cause regressions in people's workloads.
I'd like to hear more feedback from others, but for me:
-- Steve
Quoted 33 lines
Link https://lore.kernel.org/all/CA+G9fYu2kS0wR4WqMRsj2rePKV9XLgOU1PiXnMvpT+Z=c2ucHA@mail.gmail.com/ [2] open hereSigned-off-by Marco Elver <[email protected]>Cc John Ogness <[email protected]>Cc Petr Mladek <[email protected]>--- kernel/printk/printk.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)kernel/printk/printk.c +2
-2@@ -2064,8 +2064,6 @@ static void call_console_driver(struct console *con, const char *text, size_t le {size_t dropped_len; -trace_console_rcuidle (text, len); -if (con->dropped && dropped_text) { dropped_len = snprintf(dropped_text, DROPPED_TEXT_MAX, "** %lu printk messages dropped **\n",@@ -2240,6 +2238,8 @@ static u16 printk_sprint(char *text, u16 size, int facility, } } +trace_console_rcuidle (text, text_len); +return text_len; }
Quoted 4 lines · Marco Elver
One notable difference is that by moving tracing into
printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on theprintk()call and isn't delayed.
Another slight difference is that messages composed of LOG_CONT pieces will trigger the tracepoint for each individual piece and _never_ as a complete line.
It was never guaranteed that all LOG_CONT pieces make it into the final printed line anyway, but with this change it will be guaranteed that they are always handled separately.
I am OK with this change, but like Steven, I agree the the users of that tracepoint need to chime in.
The ongoing printbuf/seq_buf work [0] will hopefully someday do away with LOG_CONT altogether.
John
Quoted 10 lines · John Ogness
Quoted 4 lines · Marco Elver
One notable difference is that by moving tracing into
printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on theprintk()call and isn't delayed.Another slight difference is that messages composed of
LOG_CONTpieces will trigger the tracepoint for each individual piece and _never_ as a complete line.
Quoted 6 lines
It was never guaranteed that all
LOG_CONTpieces make it into the final printed line anyway, but with this change it will be guaranteed that they are always handled separately.I am OK with this change, but like Steven, I agree the the users of that tracepoint need to chime in.
My feeling is that the feature is not used much. Otherwise people would complain that it was asynchronous and hard to use.
I mean that the printk() messages appeared in the trace log asynchronously. So it required some post processing to correctly sort them against other tracing messages. The same result can be achieved by processing printk log buffer, dmesg.log, journalctl.
I guess that we will only find the answer when we push the change into linux-next and mainline. I am going to do so.
Best Regards,
Petr
Petr Mladek
maintainerQuoted 21 lines · Petr Mladek
Quoted 10 lines · John Ogness
Quoted 4 lines · Marco Elver
One notable difference is that by moving tracing into
printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on theprintk()call and isn't delayed.Another slight difference is that messages composed of
LOG_CONTpieces will trigger the tracepoint for each individual piece and _never_ as a complete line.Quoted 6 lines
It was never guaranteed that all
LOG_CONTpieces make it into the final printed line anyway, but with this change it will be guaranteed that they are always handled separately.I am OK with this change, but like Steven, I agree the the users of that tracepoint need to chime in.
My feeling is that the feature is not used much. Otherwise people would complain that it was asynchronous and hard to use.
Quoted 7 lines
I mean that the
printk()messages appeared in the trace log asynchronously. So it required some post processing to correctly sort them against other tracing messages. The same result can be achieved by processing printk log buffer, dmesg.log, journalctl.I guess that we will only find the answer when we push the change into linux-next and mainline. I am going to do so.
JFYI, the patch has been committed into printk/linux.git, branch rework/kthreads.
Best Regards,
Petr
Quoted 28 lines · Petr Mladek
Quoted 26 lines · Petr Mladek
Quoted 10 lines · John Ogness
Quoted 4 lines · Marco Elver
One notable difference is that by moving tracing into
printk_sprint(), the 'text' will no longer include the "header" (loglevel and timestamp), but only the raw message. Arguably this is less of a problem now that the console tracepoint happens on theprintk()call and isn't delayed.Another slight difference is that messages composed of
LOG_CONTpieces will trigger the tracepoint for each individual piece and _never_ as a complete line.Quoted 6 lines
It was never guaranteed that all
LOG_CONTpieces make it into the final printed line anyway, but with this change it will be guaranteed that they are always handled separately.I am OK with this change, but like Steven, I agree the the users of that tracepoint need to chime in.
My feeling is that the feature is not used much. Otherwise people would complain that it was asynchronous and hard to use.
I mean that the
printk()messages appeared in the trace log asynchronously. So it required some post processing to correctly sort them against other tracing messages. The same result can be achieved by processing printk log buffer, dmesg.log, journalctl.
Quoted 5 lines
Quoted 2 lines
I guess that we will only find the answer when we push the change into linux-next and mainline. I am going to do so.
JFYI, the patch has been committed into printk/linux.git, branch rework/kthreads.
Thank you, sounds good.
Thanks,
-- Marco
Steven Rostedt
maintainerI know I acked this, but I finally got a tree where it is included in my testing, and I hit this:
INFO: NMI handler (perf_event_nmi_handler) took too long to run: 9.860 msecs
------------[ cut here ]------------ WARNING: CPU: 1 PID: 16462 at include/trace/events/printk.h:10 printk_sprint+0x81/0xda Modules linked in: ppdev parport_pc parport CPU: 1 PID: 16462 Comm: event_benchmark Not tainted 5.19.0-rc5-test+ #5 Hardware name: MSI MS-7823/CSM-H87M-G43 (MS-7823), BIOS V1.6 02/22/2014 EIP: printk_sprint+0x81/0xda Code: 89 d8 e8 88 fc 33 00 e9 02 00 00 00 eb 6b 64 a1 a4 b8 91 c1 e8 fd d6 ff ff 84 c0 74 5c 64 a1 14 08 92 c1 a9 00 00 f0 00 74 02 <0f> 0b 64 ff 05 14 08 92 c1 b8 e0 c4 6b c1 e8 a5 dc 00 00 89 c7 e8 EAX: 80110001 EBX: c20a52f8 ECX: 0000000c EDX: 6d203036 ESI: 3df6004c EDI: 00000000 EBP: c61fbd7c ESP: c61fbd70 DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068 EFLAGS: 00010006 CR0: 80050033 CR2: b7efc000 CR3: 05b80000 CR4: 001506f0 Call Trace: vprintk_store+0x24b/0x2ff perf: interrupt took too long (7980 > 7977), lowering kernel.perf_event_max_sample_rate to 25000 vprintk+0x37/0x4d _printk+0x14/0x16 nmi_handle+0x1ef/0x24e ? find_next_bit.part.0+0x13/0x13 ? find_next_bit.part.0+0x13/0x13 ? function_trace_call+0xd8/0xd9 default_do_nmi+0x57/0x1af ? trace_hardirqs_off_finish+0x2a/0xd9 ? to_kthread+0xf/0xf exc_nmi+0x9b/0xf4 asm_exc_nmi+0xae/0x29c
Quoted 5 lines · On Tue, 3 May 2022 09:38:44 +0200 Marco Elver <[email protected]> wrote:
Petr points out [1] that calling
trace_console_rcuidle()incall_console_driver()had been the wrong thing for a while, because "printk()always usedconsole_trylock()and the message was flushed to the console only when the trylock succeeded. And it was always deferred in NMI or when printed viaprintk_deferred()."
The issue is that we use "trace_console_rcuidle()" where the "_rcuidle()" version uses srcu, which the last I knew is not safe in NMI context.
Paul, has that changed?
Thus, we need to make sure that printk() is always called when "rcu is watching" and remove the _rcuidle() part, or we do not call it from nmi context. Or make srcu nmi safe.
For now, I'm reverting this in my local tree.
-- Steve
Quoted 41 lines · Steven Rostedt
I know I acked this, but I finally got a tree where it is included in my testing, and I hit this:
INFO: NMI handler (
perf_event_nmi_handler) took too long to run: 9.860 msecs------------[ cut here ]------------ WARNING: CPU: 1 PID: 16462 at include/trace/events/printk.h:10 printk_sprint+0x81/0xda Modules linked in: ppdev parport_pc parport CPU: 1 PID: 16462 Comm: event_benchmark Not tainted 5.19.0-rc5-test+ #5 Hardware name: MSI MS-7823/CSM-H87M-G43 (MS-7823), BIOS V1.6 02/22/2014 EIP: printk_sprint+0x81/0xda Code: 89 d8 e8 88 fc 33 00 e9 02 00 00 00 eb 6b 64 a1 a4 b8 91 c1 e8 fd d6 ff ff 84 c0 74 5c 64 a1 14 08 92 c1 a9 00 00 f0 00 74 02 <0f> 0b 64 ff 05 14 08 92 c1 b8 e0 c4 6b c1 e8 a5 dc 00 00 89 c7 e8 EAX: 80110001 EBX: c20a52f8 ECX: 0000000c EDX: 6d203036 ESI: 3df6004c EDI: 00000000 EBP: c61fbd7c ESP: c61fbd70 DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068 EFLAGS: 00010006 CR0: 80050033 CR2: b7efc000 CR3: 05b80000 CR4: 001506f0 Call Trace: vprintk_store+0x24b/0x2ff perf: interrupt took too long (7980 > 7977), lowering kernel.perf_event_max_sample_rate to 25000 vprintk+0x37/0x4d _printk+0x14/0x16 nmi_handle+0x1ef/0x24e ? find_next_bit.part.0+0x13/0x13 ? find_next_bit.part.0+0x13/0x13 ? function_trace_call+0xd8/0xd9 default_do_nmi+0x57/0x1af ? trace_hardirqs_off_finish+0x2a/0xd9 ? to_kthread+0xf/0xf exc_nmi+0x9b/0xf4 asm_exc_nmi+0xae/0x29cQuoted 5 lines · On Tue, 3 May 2022 09:38:44 +0200 Marco Elver <[email protected]> wrote:
Petr points out [1] that calling
trace_console_rcuidle()incall_console_driver()had been the wrong thing for a while, because "printk()always usedconsole_trylock()and the message was flushed to the console only when the trylock succeeded. And it was always deferred in NMI or when printed viaprintk_deferred()."
Quoted 4 lines
The issue is that we use "
trace_console_rcuidle()" where the "_rcuidle()" version uses srcu, which the last I knew is not safe in NMI context.Paul, has that changed?
On x86, both srcu_read_lock() and srcu_read_unlock() should be OK from NMI context, give or take their use of lockdep. Which is why we have srcu_read_lock_notrace() and srcu_read_unlock_notrace(), which do not use lockdep. Which __DO_TRACE() does in fact invoke. Ah, but you have
this: "WARN_ON_ONCE(rcuidle && in_nmi())".
Because all the world is not an x86.
For srcu_read_lock() and srcu_read_unlock() to be NMI-safe more generally, this_cpu_inc() would need to be NMI-safe on all platforms. This requires it be implemented with read-modify-write atomic instructions, which appears to be the case on arm64, s390 (for at least some configs, maybe all depending on what the "laa" instruction does), and again x86. The generic code relies on disabling interrrupts, which does nothing against NMIs.
So srcu_read_lock_notrace() and srcu_read_unlock_notrace() are NMI-safe only on arm64, s390, and x86. So that WARN_ON_ONCE() is necessary.
Quoted 3 lines
Thus, we need to make sure that
printk()is always called when "rcu is watching" and remove the_rcuidle()part, or we do not call it from nmi context. Or make srcu nmi safe.
Which means making this_cpu_inc() NMI-safe. Which could be done by making the generic this_cpu_inc() implementation use something like arch_atomic_fetch_add(). Or a cmpxchg() loop. Or making __srcu_read_lock() and srcu_read_unlock() check the architecture and Do The Right Thing. Maybe a CONFIG_THIS_CPU_NMI_SAFE that is set by x86, s390, and arm64?
Is either approach something that those other architectures would be willing to countenance? For me, it might be worth it just to not have to dig through many layers of this_cpu macro definitions from time to time... :-/
Thanx, Paul
Quoted 4 lines
For now, I'm reverting this in my local tree.
-- Steve
Steven Rostedt
maintainerQuoted 7 lines · On Mon, 11 Jul 2022 17:21:28 -0700 "Paul E. McKenney" <[email protected]>…
On x86, both
srcu_read_lock()andsrcu_read_unlock()should be OK from NMI context, give or take their use of lockdep. Which is why we havesrcu_read_lock_notrace()andsrcu_read_unlock_notrace(), which do not use lockdep. Which__DO_TRACE()does in fact invoke. Ah, but you have
this: "WARN_ON_ONCE(rcuidle && in_nmi())".Because all the world is not an x86.
But since NMIs are architecture specific, we could change that to:
WARN_ON_ONCE(!srcu_nmi_safe && rcuidle && in_nmi());
and add a srcu_nmi_safe constant or macro that is 1 on architectures that srcu is safe in NMI and 0 otherwise.
Or do we care if a tracepoint happens in those architectures where it is not safe. We could then just do:
if (!srcu_nmi_safe && rcuidle && in_nmi()) return;
and just skip tracepoints that are marked rcu_idle and happen within NMI.
-- Steve
Quoted 18 lines · Steven Rostedt
Quoted 7 lines · On Mon, 11 Jul 2022 17:21:28 -0700 "Paul E. McKenney" <[email protected]>…
On x86, both
srcu_read_lock()andsrcu_read_unlock()should be OK from NMI context, give or take their use of lockdep. Which is why we havesrcu_read_lock_notrace()andsrcu_read_unlock_notrace(), which do not use lockdep. Which__DO_TRACE()does in fact invoke. Ah, but you have
this: "WARN_ON_ONCE(rcuidle && in_nmi())".Because all the world is not an x86.
But since NMIs are architecture specific, we could change that to:
WARN_ON_ONCE(!srcu_nmi_safe && rcuidle && in_nmi());and add a
srcu_nmi_safeconstant or macro that is 1 on architectures that srcu is safe in NMI and 0 otherwise.
Quoted 7 lines
Or do we care if a tracepoint happens in those architectures where it is not safe. We could then just do:
if (!srcu_nmi_safe && rcuidle && in_nmi()) return;and just skip tracepoints that are marked
rcu_idleand happen within NMI.
More generally, this is this_cpu_nmi_safe rather than just SRCU. Or could be, depending on what the architecture guys would like to guarantee. SRCU is just passing through the this_cpu*() NMI-safety property.
And in addition to in_nmi(), there is the HAVE_NMI Kconfig option:
$ git grep -w HAVE_NMI arch/Kconfig:config HAVE_NMI arch/Kconfig: depends on HAVE_NMI arch/arm/Kconfig: select HAVE_NMI arch/arm64/Kconfig: select HAVE_NMI arch/loongarch/Kconfig: select HAVE_NMI arch/mips/Kconfig: select HAVE_NMI arch/powerpc/Kconfig: select HAVE_NMI if PERF_EVENTS || (PPC64 && PPC_BOOK3S) arch/s390/Kconfig: select HAVE_NMI arch/sh/Kconfig: select HAVE_NMI arch/sparc/Kconfig: select HAVE_NMI arch/x86/Kconfig: select HAVE_NMI
So if I understand correctly, arm, loongarch, mips, powerpc, sh, and sparc are the ones that have NMIs but don't have NMI-safe this_cpu*() operations. These are the ones that would need !srcu_nmi_safe.
Or, longer term, I could make HAVE_NMI && !srcu_safe_nmi cause SRCU to use explicit atomics for srcu_read_lock_trace() and srcu_read_unlock_trace(). I am assuming that any NMI handler executing srcu_read_lock_trace() also executes the matching srcu_read_unlock_trace(). (Silly me, I know!) This assumption means that srcu_read_lock() and srcu_read_unlock() can continue with their non-atomic this_cpu_inc() ways.
But a quick fix that stopped the bleeding and allowed printk() to progress would be useful in the short term, regardless of whether or not in the longer term it makes sense to make srcu_read_lock_trace() and srcu_read_unlock_trace() NMI-safe.
Thoughts?
Thanx, Paul
Quoted 57 lines · Paul E. McKenney
Quoted 18 lines · Steven Rostedt
Quoted 7 lines · On Mon, 11 Jul 2022 17:21:28 -0700 "Paul E. McKenney" <[email protected]>…
On x86, both
srcu_read_lock()andsrcu_read_unlock()should be OK from NMI context, give or take their use of lockdep. Which is why we havesrcu_read_lock_notrace()andsrcu_read_unlock_notrace(), which do not use lockdep. Which__DO_TRACE()does in fact invoke. Ah, but you have
this: "WARN_ON_ONCE(rcuidle && in_nmi())".Because all the world is not an x86.
But since NMIs are architecture specific, we could change that to:
WARN_ON_ONCE(!srcu_nmi_safe && rcuidle && in_nmi());and add a
srcu_nmi_safeconstant or macro that is 1 on architectures that srcu is safe in NMI and 0 otherwise.Quoted 7 lines
Or do we care if a tracepoint happens in those architectures where it is not safe. We could then just do:
if (!srcu_nmi_safe && rcuidle && in_nmi()) return;and just skip tracepoints that are marked
rcu_idleand happen within NMI.More generally, this is
this_cpu_nmi_saferather than just SRCU. Or could be, depending on what the architecture guys would like to guarantee. SRCU is just passing through the this_cpu*() NMI-safety property.And in addition to
in_nmi(), there is theHAVE_NMIKconfigoption:$ git grep -w HAVE_NMI arch/Kconfig:config HAVE_NMI arch/Kconfig: depends on HAVE_NMI arch/arm/Kconfig: select HAVE_NMI arch/arm64/Kconfig: select HAVE_NMI arch/loongarch/Kconfig: select HAVE_NMI arch/mips/Kconfig: select HAVE_NMI arch/powerpc/Kconfig: select HAVE_NMI if PERF_EVENTS || (PPC64 && PPC_BOOK3S) arch/s390/Kconfig: select HAVE_NMI arch/sh/Kconfig: select HAVE_NMI arch/sparc/Kconfig: select HAVE_NMI arch/x86/Kconfig: select HAVE_NMISo if I understand correctly, arm, loongarch, mips, powerpc, sh, and sparc are the ones that have NMIs but don't have NMI-safe this_cpu*() operations. These are the ones that would need !srcu_nmi_safe.
Or, longer term, I could make
HAVE_NMI&& !srcu_safe_nmi cause SRCU to use explicit atomics forsrcu_read_lock_trace()andsrcu_read_unlock_trace(). I am assuming that any NMI handler executingsrcu_read_lock_trace()also executes the matchingsrcu_read_unlock_trace(). (Silly me, I know!) This assumption means thatsrcu_read_lock()andsrcu_read_unlock()can continue with their non-atomicthis_cpu_inc()ways.
Quoted 4 lines
But a quick fix that stopped the bleeding and allowed
printk()to progress would be useful in the short term, regardless of whether or not in the longer term it makes sense to makesrcu_read_lock_trace()andsrcu_read_unlock_trace()NMI-safe.
Except that doesn't rcuidle && in_nmi() imply a misplaced trace event?
Isn't it still the case that you are not supposed to have trace events in NMI handlers before RCU is watching or after it is no longer watching, just as for entry/exit code in general? Once in the body of the handler, rcuidle should be false and all should be well.
Or am I missing something here?
Thanx, Paul
Steven Rostedt
maintainerQuoted 5 lines · On Tue, 12 Jul 2022 04:49:54 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 4 lines
But a quick fix that stopped the bleeding and allowed
printk()to progress would be useful in the short term, regardless of whether or not in the longer term it makes sense to makesrcu_read_lock_trace()andsrcu_read_unlock_trace()NMI-safe.
Quoted 8 lines
Except that doesn't rcuidle &&
in_nmi()imply a misplaced trace event?Isn't it still the case that you are not supposed to have trace events in NMI handlers before RCU is watching or after it is no longer watching, just as for entry/exit code in general? Once in the body of the handler, rcuidle should be false and all should be well.
Or am I missing something here?
I guess the question is, can we have printk() in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.
-- Steve
Quoted 15 lines · Steven Rostedt
Quoted 12 lines · On Tue, 12 Jul 2022 04:49:54 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 4 lines
But a quick fix that stopped the bleeding and allowed
printk()to progress would be useful in the short term, regardless of whether or not in the longer term it makes sense to makesrcu_read_lock_trace()andsrcu_read_unlock_trace()NMI-safe.Except that doesn't rcuidle &&
in_nmi()imply a misplaced trace event?Isn't it still the case that you are not supposed to have trace events in NMI handlers before RCU is watching or after it is no longer watching, just as for entry/exit code in general? Once in the body of the handler, rcuidle should be false and all should be well.
Quoted 5 lines
Quoted 1 line
Or am I missing something here?
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.
As I understand it, code in such a place should be labeled noinstr. Then the call to printk() would be complained about as an illegal noinstr-to-non-noinstr call.
But where exactly is that printk()?
Thanx, Paul
Steven Rostedt
maintainerQuoted 9 lines · On Tue, 12 Jul 2022 06:49:16 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 3 lines
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.As I understand it, code in such a place should be labeled noinstr. Then the call to
printk()would be complained about as an illegal noinstr-to-non-noinstr call.But where exactly is that
printk()?
Perhaps the fix is to remove the _rcuidle() from trace_console_rcuidle(). If printk() can never be called from noinstr (aka RCU not watching).
-- Steve
Quoted 7 lines · Steven Rostedt
Quoted 4 lines · On Tue, 12 Jul 2022 06:49:16 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 3 lines
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.
Quoted 8 lines
Quoted 5 lines
As I understand it, code in such a place should be labeled noinstr. Then the call to
printk()would be complained about as an illegal noinstr-to-non-noinstr call.But where exactly is that
printk()?Perhaps the fix is to remove the
_rcuidle()fromtrace_console_rcuidle(). Ifprintk()can never be called from noinstr (aka RCU not watching).
Maybe printk() is supposed to be invoked from noinstr. It might be a special case in the tooling. I have no idea. ;-)
However, the current SRCU read-side algorithm will tolerate being invoked from noinstr as long as it is not also an NMI handler. Much though debugging tools might (or might not) complain.
Don't get me wrong, I can make SRCU tolerate being called while RCU is not watching. It is not even all that complicated. The cost is that architectures that have NMIs but do not have NMI-safe this_cpu*() operations have an SRCU reader switch from explicit smp_mb() and interrupt disabling to a cmpxchg() loop relying on the implicit barriers in cmpxchg().
For arm64, this was reportedly a win.
If it turns out that we need SRCU readers to be invoked from NMI handlers in locations where RCU is not watching, are there people who would be willing to play with a modified SRCU on the systems in question?
-ENOHARDWARE at this end.
Thanx, Paul
Petr Mladek
maintainerQuoted 12 lines · Paul E. McKenney
Quoted 11 lines · Steven Rostedt
Quoted 8 lines · On Tue, 12 Jul 2022 06:49:16 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 3 lines
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.As I understand it, code in such a place should be labeled noinstr. Then the call to
printk()would be complained about as an illegal noinstr-to-non-noinstr call.
Quoted 7 lines
Quoted 4 lines
Quoted 1 line
But where exactly is that
printk()?Perhaps the fix is to remove the
_rcuidle()fromtrace_console_rcuidle(). Ifprintk()can never be called from noinstr (aka RCU not watching).Maybe
printk()is supposed to be invoked from noinstr. It might be a special case in the tooling. I have no idea. ;-)
I think that it is ok to do _not_ support printk() in noinstr parts.
Quoted 4 lines
However, the current SRCU read-side algorithm will tolerate being invoked from noinstr as long as it is not also an NMI handler. Much though debugging tools might (or might not) complain.
Quoted 8 lines
Don't get me wrong, I can make SRCU tolerate being called while RCU is not watching. It is not even all that complicated. The cost is that architectures that have NMIs but do not have NMI-safe this_cpu*() operations have an SRCU reader switch from explicit
smp_mb()and interrupt disabling to acmpxchg()loop relying on the implicit barriers incmpxchg().For arm64, this was reportedly a win.
IMHO, the tracepoint in printk() is not worth slowing down other important fast paths.
The tracepoint was moved into vprintk_store() in 5.19-rc1. It used to be in console_unlock() before. The previous location was not reliable by definition. Old messages might be overridden by new ones before they reach console. Also messages in NMI context used to be stored in per-CPU buffers. There was even bigger risk that they would not reach the console.
Best Regards,
Petr
Quoted 39 lines · Petr Mladek
Quoted 12 lines · Paul E. McKenney
Quoted 11 lines · Steven Rostedt
Quoted 8 lines · On Tue, 12 Jul 2022 06:49:16 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 3 lines
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.As I understand it, code in such a place should be labeled noinstr. Then the call to
printk()would be complained about as an illegal noinstr-to-non-noinstr call.Quoted 7 lines
Quoted 4 lines
Quoted 1 line
But where exactly is that
printk()?Perhaps the fix is to remove the
_rcuidle()fromtrace_console_rcuidle(). Ifprintk()can never be called from noinstr (aka RCU not watching).Maybe
printk()is supposed to be invoked from noinstr. It might be a special case in the tooling. I have no idea. ;-)I think that it is ok to do _not_ support
printk()in noinstr parts.Quoted 4 lines
However, the current SRCU read-side algorithm will tolerate being invoked from noinstr as long as it is not also an NMI handler. Much though debugging tools might (or might not) complain.
Quoted 8 lines
Don't get me wrong, I can make SRCU tolerate being called while RCU is not watching. It is not even all that complicated. The cost is that architectures that have NMIs but do not have NMI-safe this_cpu*() operations have an SRCU reader switch from explicit
smp_mb()and interrupt disabling to acmpxchg()loop relying on the implicit barriers incmpxchg().For arm64, this was reportedly a win.
IMHO, the tracepoint in
printk()is not worth slowing down other important fast paths.
Quoted 6 lines
The tracepoint was moved into
vprintk_store()in 5.19-rc1. It used to be inconsole_unlock()before. The previous location was not reliable by definition. Old messages might be overridden by new ones before they reach console. Also messages in NMI context used to be stored in per-CPU buffers. There was even bigger risk that they would not reach the console.
Fair enough, works for me!
Thanx, Paul
Petr Mladek
maintainerQuoted 23 lines · Paul E. McKenney
Quoted 22 lines · Petr Mladek
Quoted 2 lines · Paul E. McKenney
Maybe
printk()is supposed to be invoked from noinstr. It might be a special case in the tooling. I have no idea. ;-)I think that it is ok to do _not_ support
printk()in noinstr parts.Quoted 4 lines
However, the current SRCU read-side algorithm will tolerate being invoked from noinstr as long as it is not also an NMI handler. Much though debugging tools might (or might not) complain.
Quoted 8 lines
Don't get me wrong, I can make SRCU tolerate being called while RCU is not watching. It is not even all that complicated. The cost is that architectures that have NMIs but do not have NMI-safe this_cpu*() operations have an SRCU reader switch from explicit
smp_mb()and interrupt disabling to acmpxchg()loop relying on the implicit barriers incmpxchg().For arm64, this was reportedly a win.
IMHO, the tracepoint in
printk()is not worth slowing down other important fast paths.
Quoted 8 lines
Quoted 6 lines
The tracepoint was moved into
vprintk_store()in 5.19-rc1. It used to be inconsole_unlock()before. The previous location was not reliable by definition. Old messages might be overridden by new ones before they reach console. Also messages in NMI context used to be stored in per-CPU buffers. There was even bigger risk that they would not reach the console.Fair enough, works for me!
The remaining question is how to make the code safe and calm down the warning.
My understanding is that Peter Zijlstra wants to reduce the scope of the rcuidle code even more in the future. So, we could do something like:
From 24c3517dedf2a30efabe72871c188fbfffffd397 Mon Sep 17 00:00:00 2001
From: Petr Mladek <[email protected]>
Date: Thu, 14 Jul 2022 14:54:12 +0200
Subject: [PATCH] printk: Make console tracepoint safe in NMI() context
The commit 701850dc0c31bfadf75a0 ("printk, tracing: fix console tracepoint") moved the tracepoint from console_unlock() to vprintk_store(). As a result, it might be called in any context and triggered the following warning:
WARNING: CPU: 1 PID: 16462 at include/trace/events/printk.h:10 printk_sprint+0x81/0xda Modules linked in: ppdev parport_pc parport CPU: 1 PID: 16462 Comm: event_benchmark Not tainted 5.19.0-rc5-test+ #5 Hardware name: MSI MS-7823/CSM-H87M-G43 (MS-7823), BIOS V1.6 02/22/2014 EIP: printk_sprint+0x81/0xda Code: 89 d8 e8 88 fc 33 00 e9 02 00 00 00 eb 6b 64 a1 a4 b8 91 c1 e8 fd d6 ff ff 84 c0 74 5c 64 a1 14 08 92 c1 a9 00 00 f0 00 74 02 <0f> 0b 64 ff 05 14 08 92 c1 b8 e0 c4 6b c1 e8 a5 dc 00 00 89 c7 e8 EAX: 80110001 EBX: c20a52f8 ECX: 0000000c EDX: 6d203036 ESI: 3df6004c EDI: 00000000 EBP: c61fbd7c ESP: c61fbd70 DS: 007b ES: 007b FS: 00d8 GS: 0000 SS: 0068 EFLAGS: 00010006 CR0: 80050033 CR2: b7efc000 CR3: 05b80000 CR4: 001506f0 Call Trace: vprintk_store+0x24b/0x2ff vprintk+0x37/0x4d _printk+0x14/0x16 nmi_handle+0x1ef/0x24e ? find_next_bit.part.0+0x13/0x13 ? find_next_bit.part.0+0x13/0x13 ? function_trace_call+0xd8/0xd9 default_do_nmi+0x57/0x1af ? trace_hardirqs_off_finish+0x2a/0xd9 ? to_kthread+0xf/0xf exc_nmi+0x9b/0xf4 asm_exc_nmi+0xae/0x29c
It comes from:
#define __DO_TRACE(name, args, cond, rcuidle) \
[...]
/* srcu can't be used from NMI */ \ WARN_ON_ONCE(rcuidle && in_nmi()); \
It might be possible to make srcu working in NMI. But it would be slower on some architectures. It is not worth doing it just because of this tracepoint.
It would be possible to disable this tracepoint in NMI or in rcuidle context. Where the rcuidle context looks more rare and thus more acceptable to be ignored.
Alternative solution would be to move the tracepoint back to console code. But the location is less reliable by definition. Also the synchronization against other tracing messages is much worse.
Let's ignore the tracepoint in rcuidle context as the least evil solution.
There seems to be three possibilities.
--- kernel/printk/printk.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-)
kernel/printk/printk.c +9 -1
@@ -2108,7 +2108,15 @@ static u16 printk_sprint(char *text, u16 size, int facility, } } -trace_console_rcuidle (text, text_len); +/* +* trace_console_idle() is not working in NMI. printk() +* is used more often in NMI than in rcuidle context. +* Choose the less evil solution here. +* +* smp_processor_id() is reliable in rcuidle context. +*/ +if (!rcu_is_idle_cpu (smp_processor_id ())) +trace_console (text, text_len);return text_len; }
Signature · 1 line
2.35.3
Steven Rostedt
maintainerQuoted 10 lines · On Thu, 14 Jul 2022 16:53:24 +0200 Petr Mladek <[email protected]> wrote:
kernel/printk/printk.c +2
-1@@ -2108,7 +2108,15 @@ static u16 printk_sprint(char *text, u16 size, int facility, } } -trace_console_rcuidle (text, text_len); +/* +* trace_console_idle() is not working in NMI. printk()
kernel/printk/printk.c static u16 printk_sprint(char *text, u16 size, int facility, · line 2122
Quoted 10 lines
kernel/printk/printk.c +7
-0+* is used more often in NMI than in rcuidle context. +* Choose the less evil solution here. +* +* smp_processor_id() is reliable in rcuidle context. +*/ +if (!rcu_is_idle_cpu (smp_processor_id ())) +trace_console (text, text_len);return text_len; }
Although printk is not really a fast path, you could do this and avoid the check when the trace event is not active:
(Not even compiled tested)
Tweaked the comment, and used raw_smp_processor_id() as I'm not sure we are in a preempt disabled context, and we don't care if we are not.
-- Steve
include/trace/events/printk.h +10 -1
@@ -7,11 +7,20 @@#include <linux/tracepoint.h> -TRACE_EVENT(console, +TRACE_EVENT_CONDITION(console, TP_PROTO(const char *text, size_t len), TP_ARGS(text, len), + /* + * trace_console_rcuidle() is not working in NMI. printk() + * is used more often in NMI than in rcuidle context. + * Choose the less evil solution here. + * + * raw_smp_processor_id() is reliable in rcuidle context. + */ + TP_CONDITION(!rcu_is_idle_cpu(raw_smp_processor_id())), + TP_STRUCT__entry( __dynamic_array(char, msg, len + 1) ),
Petr Mladek
maintainerQuoted 7 lines · Steven Rostedt
Although printk is not really a fast path, you could do this and avoid the check when the trace event is not active:
(Not even compiled tested)
Tweaked the comment, and used
raw_smp_processor_id()as I'm not sure we are in a preempt disabled context, and we don't care if we are not.
Makes sense.
Quoted 16 lines
include/trace/events/printk.h +3
-1@@ -7,11 +7,20 @@#include <linux/tracepoint.h> -TRACE_EVENT(console, +TRACE_EVENT_CONDITION(console, TP_PROTO(const char *text, size_t len), TP_ARGS(text, len), +/* +* trace_console_rcuidle() is not working in NMI. printk()
include/trace/events/printk.h line 26
Quoted 10 lines
include/trace/events/printk.h +7
-0+* is used more often in NMI than in rcuidle context. +* Choose the less evil solution here. +* +* raw_smp_processor_id() is reliable in rcuidle context. +*/ +TP_CONDITION (!rcu_is_idle_cpu (raw_smp_processor_id ())), +TP_STRUCT__entry (__dynamic_array (char, msg, len +1 ) ),
It looks better. I am going to test it and send as a proper patch for review.
Thanks for help.
Best Regards,
Petr
Petr Mladek
maintainerQuoted 7 lines · Steven Rostedt
Quoted 4 lines · On Tue, 12 Jul 2022 06:49:16 -0700 "Paul E. McKenney" <[email protected]>…
Quoted 3 lines
I guess the question is, can we have
printk()in such a place? Because this tracepoint is attached to printk and where ever printk is done so is this tracepoint.
Quoted 8 lines
Quoted 5 lines
As I understand it, code in such a place should be labeled noinstr. Then the call to
printk()would be complained about as an illegal noinstr-to-non-noinstr call.But where exactly is that
printk()?Perhaps the fix is to remove the
_rcuidle()fromtrace_console_rcuidle(). Ifprintk()can never be called from noinstr (aka RCU not watching).
Good point!
My understanding is that printk() should not get called when rcu is not watching. But it might need to reduce the scope of the code when rcu is not watching.
PeterZ actually removed _rcuidle() in this trace a patchset, see https://lore.kernel.org/r/[email protected].
Adding Peter into Cc.
Best Regards,
Petr