exit: move and extend sched_process_exit() tracepoint
Andrii Nakryiko
maintainerPATCH v2 1/1 exit: move and extend sched_process_exit() tracepoint
It is useful to be able to access current->mm at task exit to, say, record a bunch of VMA information right before the task exits (e.g., for stack symbolization reasons when dealing with short-lived processes that exit in the middle of profiling session). Currently, trace_sched_process_exit() is triggered after exit_mm() which resets current->mm to NULL making this tracepoint unsuitable for inspecting and recording task's mm_struct-related data when tracing process lifetimes.
There is a particularly suitable place, though, right after taskstats_exit() is called, but before we do exit_mm() and other exit_*() resource teardowns. taskstats performs a similar kind of accounting that some applications do with BPF, and so co-locating them seems like a good fit. So that's where trace_sched_process_exit() is moved with this patch.
Also, existing trace_sched_process_exit() tracepoint is notoriously missing group_dead flag that is certainly useful in practice and some of our production applications have to work around this. So plumb group_dead through while at it, to have a richer and more complete tracepoint.
Note that we can't use sched_process_template anymore, and so we use TRACE_EVENT()-based tracepoint definition. But all the field names and order, as well as assign and output logic remain intact. We just add one extra field at the end in backwards-compatible way.
--- include/trace/events/sched.h | 28 +++++++++++++++++++++++++--- kernel/exit.c | 2 +- 2 files changed, 26 insertions(+), 4 deletions(-)
include/trace/events/sched.h +25 -3
@@ -328,9 +328,31 @@ DEFINE_EVENT(sched_process_template, sched_process_free,/* * Tracepoint for a task exiting: */ -DEFINE_EVENT (sched_process_template, sched_process_exit, - TP_PROTO(struct task_struct *p), -TP_ARGS (p)); +TRACE_EVENT (sched_process_exit, + + TP_PROTO(struct task_struct *p, bool group_dead), + +TP_ARGS (p, group_dead), + +TP_STRUCT__entry ( + __array( char, comm, TASK_COMM_LEN ) + __field( pid_t, pid ) + __field( int, prio ) +__field ( bool, group_dead ) + ), + +TP_fast_assign ( + memcpy(__entry->comm, p->comm, TASK_COMM_LEN); + __entry->pid = p->pid; + __entry->prio = p->prio; /* XXX SCHED_DEADLINE */ + __entry->group_dead = group_dead; + ), + +TP_printk ("comm=%s pid=%d prio=%d group_dead=%s" , + __entry->comm, __entry->pid, __entry->prio, + __entry->group_dead ?"true" :"false" + ) +);/* * Tracepoint for waiting on task to unschedule:
kernel/exit.c +1 -1
@@ -937,12 +937,12 @@ void __noreturn do_exit(long code) tsk->exit_code = code;taskstats_exit (tsk, group_dead); +trace_sched_process_exit (tsk, group_dead);exit_mm ();if (group_dead)acct_process (); -trace_sched_process_exit (tsk);exit_sem (tsk);exit_shm (tsk);
Signature · 1 line
2.47.1
Quoted 27 lines · On Wed, 2 Apr 2025 11:09:25 -0700 Andrii Nakryiko <[email protected]> wrote:
It is useful to be able to access
current->mmat task exit to, say, record a bunch of VMA information right before the task exits (e.g., for stack symbolization reasons when dealing with short-lived processes that exit in the middle of profiling session). Currently,trace_sched_process_exit()is triggered afterexit_mm()which resetscurrent->mmto NULL making this tracepoint unsuitable for inspecting and recording task's mm_struct-related data when tracing process lifetimes.There is a particularly suitable place, though, right after
taskstats_exit()is called, but before we doexit_mm()and other exit_*() resource teardowns. taskstats performs a similar kind of accounting that some applications do with BPF, and so co-locating them seems like a good fit. So that's wheretrace_sched_process_exit()is moved with this patch.Also, existing
trace_sched_process_exit()tracepoint is notoriously missinggroup_deadflag that is certainly useful in practice and some of our production applications have to work around this. So plumbgroup_deadthrough while at it, to have a richer and more complete tracepoint.Note that we can't use
sched_process_templateanymore, and so we use TRACE_EVENT()-based tracepoint definition. But all the field names and order, as well as assign and output logic remain intact. We just add one extra field at the end in backwards-compatible way.
Quoted 5 lines
Signed-off-by Andrii Nakryiko <[email protected]>--- include/trace/events/sched.h | 28 +++++++++++++++++++++++++--- kernel/exit.c | 2 +- 2 files changed, 26 insertions(+), 4 deletions(-)
-- Steve
Quoted 4 lines · Andrii Nakryiko
kernel/exit.c void __noreturn do_exit(long code) · line 945
Quoted 10 lines
kernel/exit.c +1
-1@@ -937,12 +937,12 @@ void __noreturn do_exit(long code) tsk->exit_code = code;taskstats_exit (tsk, group_dead); +trace_sched_process_exit (tsk, group_dead);exit_mm ();if (group_dead)acct_process (); -trace_sched_process_exit (tsk);
I see nothing wrong in this change.
(and I think that the current placement of trace_sched_process_exit() was
more or less "random").
Quoted 16 lines · Andrii Nakryiko
It is useful to be able to access
current->mmat task exit to, say, record a bunch of VMA information right before the task exits (e.g., for stack symbolization reasons when dealing with short-lived processes that exit in the middle of profiling session). Currently,trace_sched_process_exit()is triggered afterexit_mm()which resetscurrent->mmto NULL making this tracepoint unsuitable for inspecting and recording task's mm_struct-related data when tracing process lifetimes.There is a particularly suitable place, though, right after
taskstats_exit()is called, but before we doexit_mm()and other exit_*() resource teardowns. taskstats performs a similar kind of accounting that some applications do with BPF, and so co-locating them seems like a good fit. So that's wheretrace_sched_process_exit()is moved with this patch.
Quoted 8 lines
Also, existing
trace_sched_process_exit()tracepoint is notoriously missinggroup_deadflag that is certainly useful in practice and some of our production applications have to work around this. So plumbgroup_deadthrough while at it, to have a richer and more complete tracepoint.Note that we can't use
sched_process_templateanymore, and so we use TRACE_EVENT()-based tracepoint definition.
But all the field names and
Quoted 34 lines
order, as well as assign and output logic remain intact. We just add one extra field at the end in backwards-compatible way.
Signed-off-by Andrii Nakryiko <[email protected]>--- include/trace/events/sched.h | 28 +++++++++++++++++++++++++--- kernel/exit.c | 2 +- 2 files changed, 26 insertions(+), 4 deletions(-)include/trace/events/sched.h +14
-3@@ -328,9 +328,31 @@ DEFINE_EVENT(sched_process_template, sched_process_free,/* * Tracepoint for a task exiting: */ -DEFINE_EVENT (sched_process_template, sched_process_exit, - TP_PROTO(struct task_struct *p), -TP_ARGS (p)); +TRACE_EVENT(sched_process_exit, + + TP_PROTO(struct task_struct *p, bool group_dead), + + TP_ARGS(p, group_dead), + + TP_STRUCT__entry( + __array( char, comm, TASK_COMM_LEN ) + __field( pid_t, pid ) + __field( int, prio ) + __field( bool, group_dead ) + ), + + TP_fast_assign(
include/trace/events/sched.h DEFINE_EVENT(sched_process_template, sched_process_free, · line 354
Quoted 10 lines
include/trace/events/sched.h +10
-0+memcpy (__entry->comm, p->comm,TASK_COMM_LEN ); + __entry->pid = p->pid; + __entry->prio = p->prio;/* XXX SCHED_DEADLINE */ + __entry->group_dead = group_dead; + ), + +TP_printk ("comm=%s pid=%d prio=%d group_dead=%s" , + __entry->comm, __entry->pid, __entry->prio, + __entry->group_dead ?"true" :"false" + )
This feels really fragile, could you please at least add a comment that points out that this is basically an extension of sched_process_template, and that it should remain a subset of it, or something to that end?
Thanks,
Ingo
Steven Rostedt
maintainerQuoted 4 lines · On Thu, 3 Apr 2025 15:54:22 +0200 Ingo Molnar <[email protected]> wrote:
This feels really fragile, could you please at least add a comment that points out that this is basically an extension of
sched_process_template, and that it should remain a subset of it, or something to that end?
Is there any dependency on this?
The reason to use the templates is because it saves memory. Each TRACE_EVENT() can add ~5k (which a TRACE_EVENT() is really just a DECLARE_EVENT_CLASS() + DEFINE_EVENT() for a single event).
Each DEFINE_EVENT() just adds around 250 bytes. Hence, if you have multiple events that share the same fields and output, it's much more memory efficient to use the CLASS and EVENT logic then making each their own TRACE_EVENT().
I don't know of any other dependency to why this was a template other than to save memory.
-- Steve
Quoted 8 lines · Steven Rostedt
Quoted 4 lines · On Thu, 3 Apr 2025 15:54:22 +0200 Ingo Molnar <[email protected]> wrote:
This feels really fragile, could you please at least add a comment that points out that this is basically an extension of
sched_process_template, and that it should remain a subset of it, or something to that end?
Quoted 4 lines
Is there any dependency on this?
I don't know of any other dependency to why this was a template other than to save memory.
Uhm, to state the obvious: to not replicate the same definitions over and over again three times times, for 3 scheduler tracepoints that share the record format?
Removing just a single sched_process_template use bloats the source and adds in potential fragility:
2 files changed, 26 insertions(+), 4 deletions(-)
So my request is to please at least add a comment that points the reader to the shared record format between sched_process_exit and the other two tracepoints.
Thanks,
Ingo
Quoted 21 lines · Ingo Molnar
Quoted 8 lines · Steven Rostedt
Quoted 4 lines · On Thu, 3 Apr 2025 15:54:22 +0200 Ingo Molnar <[email protected]> wrote:
This feels really fragile, could you please at least add a comment that points out that this is basically an extension of
sched_process_template, and that it should remain a subset of it, or something to that end?Quoted 4 lines
Is there any dependency on this?
I don't know of any other dependency to why this was a template other than to save memory.
Uhm, to state the obvious: to not replicate the same definitions over and over again three times times, for 3 scheduler tracepoints that share the record format?
Quoted 8 lines
Removing just a single
sched_process_templateuse bloats the source and adds in potential fragility:2 files changed, 26 insertions(+), 4 deletions(-)So my request is to please at least add a comment that points the reader to the shared record format between
sched_process_exitand the other two tracepoints.
Sounds good, no problem. I'll send a follow up patch which Andrew can fold, if he prefers.
Quoted 4 lines
Thanks,
Ingo
[tip: sched/core] sched/tracepoints: Move and extend the sched_process_exit() tracepoint
In reply to Andrii NakryikoThe following commit has been merged into the sched/core branch of tip:
The following commit has been merged into the sched/core branch of tip:
Commit-ID: 3e816361e94a0e79b1aabf44abec552e9698b196 Gitweb: https://git.kernel.org/tip/3e816361e94a0e79b1aabf44abec552e9698b196 Author: Andrii Nakryiko <[email protected]> AuthorDate: Wed, 02 Apr 2025 11:09:25 -07:00 Committer: Ingo Molnar <[email protected]> CommitterDate: Fri, 04 Apr 2025 10:30:19 +02:00
sched/tracepoints: Move and extend the sched_process_exit() tracepoint
It is useful to be able to access current->mm at task exit to, say, record a bunch of VMA information right before the task exits (e.g., for stack symbolization reasons when dealing with short-lived processes that exit in the middle of profiling session). Currently, trace_sched_process_exit() is triggered after exit_mm() which resets current->mm to NULL making this tracepoint unsuitable for inspecting and recording task's mm_struct-related data when tracing process lifetimes.
There is a particularly suitable place, though, right after taskstats_exit() is called, but before we do exit_mm() and other exit_*() resource teardowns. taskstats performs a similar kind of accounting that some applications do with BPF, and so co-locating them seems like a good fit. So that's where trace_sched_process_exit() is moved with this patch.
Also, existing trace_sched_process_exit() tracepoint is notoriously missing group_dead flag that is certainly useful in practice and some of our production applications have to work around this. So plumb group_dead through while at it, to have a richer and more complete tracepoint.
Note that we can't use sched_process_template anymore, and so we use TRACE_EVENT()-based tracepoint definition. But all the field names and order, as well as assign and output logic remain intact. We just add one extra field at the end in backwards-compatible way.
Document the dependency to sched_process_template anyway.
--- include/trace/events/sched.h | 34 ++++++++++++++++++++++++++++++---- kernel/exit.c | 2 +- 2 files changed, 31 insertions(+), 5 deletions(-)
include/trace/events/sched.h +30 -4
@@ -326,11 +326,37 @@ DEFINE_EVENT(sched_process_template, sched_process_free, TP_ARGS(p));/* -* Tracepoint for a task exiting: +* Tracepoint for a task exiting. +* Note, it's a superset of sched_process_template and should be kept +* compatible as much as possible. sched_process_exits has an extra +* `group_dead` argument, so sched_process_template can't be used, +* unfortunately, just like sched_migrate_task above. */ -DEFINE_EVENT (sched_process_template, sched_process_exit, - TP_PROTO(struct task_struct *p), -TP_ARGS (p)); +TRACE_EVENT (sched_process_exit, + + TP_PROTO(struct task_struct *p, bool group_dead), + +TP_ARGS (p, group_dead), + +TP_STRUCT__entry ( + __array( char, comm, TASK_COMM_LEN ) + __field( pid_t, pid ) + __field( int, prio ) +__field ( bool, group_dead ) + ), + +TP_fast_assign ( + memcpy(__entry->comm, p->comm, TASK_COMM_LEN); + __entry->pid = p->pid; + __entry->prio = p->prio; /* XXX SCHED_DEADLINE */ + __entry->group_dead = group_dead; + ), + +TP_printk ("comm=%s pid=%d prio=%d group_dead=%s" , + __entry->comm, __entry->pid, __entry->prio, + __entry->group_dead ?"true" :"false" + ) +);/* * Tracepoint for waiting on task to unschedule:
kernel/exit.c +1 -1
@@ -936,12 +936,12 @@ void __noreturn do_exit(long code) tsk->exit_code = code;taskstats_exit (tsk, group_dead); +trace_sched_process_exit (tsk, group_dead);exit_mm ();if (group_dead)acct_process (); -trace_sched_process_exit (tsk);exit_sem (tsk);exit_shm (tsk);
Re: [tip: sched/core] sched/tracepoints: Move and extend the sched_process_exit() tracepoint
Quoted 39 lines · On Fri, Apr 4, 2025 at 1:37 AM tip-bot2 for Andrii Nakryiko…
The following commit has been merged into the sched/core branch of tip:
Commit-ID: 3e816361e94a0e79b1aabf44abec552e9698b196 Gitweb: https://git.kernel.org/tip/3e816361e94a0e79b1aabf44abec552e9698b196 Author: Andrii Nakryiko <[email protected]> AuthorDate: Wed, 02 Apr 2025 11:09:25 -07:00 Committer: Ingo Molnar <[email protected]> CommitterDate: Fri, 04 Apr 2025 10:30:19 +02:00sched/tracepoints: Move and extend the
sched_process_exit()tracepointIt is useful to be able to access
current->mmat task exit to, say, record a bunch of VMA information right before the task exits (e.g., for stack symbolization reasons when dealing with short-lived processes that exit in the middle of profiling session). Currently,trace_sched_process_exit()is triggered afterexit_mm()which resetscurrent->mmto NULL making this tracepoint unsuitable for inspecting and recording task's mm_struct-related data when tracing process lifetimes.There is a particularly suitable place, though, right after
taskstats_exit()is called, but before we doexit_mm()and other exit_*() resource teardowns. taskstats performs a similar kind of accounting that some applications do with BPF, and so co-locating them seems like a good fit. So that's wheretrace_sched_process_exit()is moved with this patch.Also, existing
trace_sched_process_exit()tracepoint is notoriously missinggroup_deadflag that is certainly useful in practice and some of our production applications have to work around this. So plumbgroup_deadthrough while at it, to have a richer and more complete tracepoint.Note that we can't use
sched_process_templateanymore, and so we use TRACE_EVENT()-based tracepoint definition. But all the field names and order, as well as assign and output logic remain intact. We just add one extra field at the end in backwards-compatible way.
Quoted 4 lines
Document the dependency to
sched_process_templateanyway.Signed-off-by Andrii Nakryiko <[email protected]>Signed-off-by Ingo Molnar <[email protected]>
Adding Andrew.
Seems like my patch was applied both by Andrew ([0], [1]) and Ingo. Andew, would it be possible to drop those from your tree and keep the one in Ingo's tip/sched/core? Thanks!
[0] https://git.kernel.org/pub/scm/linux/kernel/git/akpm/25-new.git/tree/patches/exit-move-and-extend-sched_process_exit-tracepoint.patch
[1] https://git.kernel.org/pub/scm/linux/kernel/git/akpm/25-new.git/tree/patches/exit-move-and-extend-sched_process_exit-tracepoint-fix.patch
Quoted 73 lines
Acked-by Oleg Nesterov <[email protected]>Cc Peter Zijlstra <[email protected]>--- include/trace/events/sched.h | 34 ++++++++++++++++++++++++++++++---- kernel/exit.c | 2 +- 2 files changed, 31 insertions(+), 5 deletions(-)include/trace/events/sched.h +30
-4@@ -326,11 +326,37 @@ DEFINE_EVENT(sched_process_template, sched_process_free, TP_ARGS(p));/* -* Tracepoint for a task exiting: +* Tracepoint for a task exiting. +* Note, it's a superset of sched_process_template and should be kept +* compatible as much as possible. sched_process_exits has an extra +* `group_dead` argument, so sched_process_template can't be used, +* unfortunately, just like sched_migrate_task above. */ -DEFINE_EVENT (sched_process_template, sched_process_exit, - TP_PROTO(struct task_struct *p), -TP_ARGS (p)); +TRACE_EVENT (sched_process_exit, + + TP_PROTO(struct task_struct *p, bool group_dead), + +TP_ARGS (p, group_dead), + +TP_STRUCT__entry ( + __array( char, comm, TASK_COMM_LEN ) + __field( pid_t, pid ) + __field( int, prio ) +__field ( bool, group_dead ) + ), + +TP_fast_assign ( + memcpy(__entry->comm, p->comm, TASK_COMM_LEN); + __entry->pid = p->pid; + __entry->prio = p->prio; /* XXX SCHED_DEADLINE */ + __entry->group_dead = group_dead; + ), + +TP_printk ("comm=%s pid=%d prio=%d group_dead=%s" , + __entry->comm, __entry->pid, __entry->prio, + __entry->group_dead ?"true" :"false" + ) +);/* * Tracepoint for waiting on task to unschedule: kernel/exit.c +1
-1@@ -936,12 +936,12 @@ void __noreturn do_exit(long code) tsk->exit_code = code;taskstats_exit (tsk, group_dead); +trace_sched_process_exit (tsk, group_dead);exit_mm ();if (group_dead)acct_process (); -trace_sched_process_exit (tsk);exit_sem (tsk);exit_shm (tsk);