[tip:perf/core] perf: Avoid horrible stack usage
Commit-ID: 86038c5ea81b519a8a1fcfcd5e4599aab0cdd119 Gitweb: http://git.kernel.org/tip/86038c5ea81b519a8a1fcfcd5e4599aab0cdd119 Author: Peter Zijlstra (Intel) AuthorDate: Tue, 16 Dec 2014 12:47:34 +0100 Committer: Ingo Molnar CommitDate: Wed, 14 Jan 2015 15:11:45 +0100 perf: Avoid horrible stack usage Both Linus (most recent) and Steve (a while ago) reported that perf related callbacks have massive stack bloat. The problem is that software events need a pt_regs in order to properly report the event location and unwind stack. And because we could not assume one was present we allocated one on stack and filled it with minimal bits required for operation. Now, pt_regs is quite large, so this is undesirable. Furthermore it turns out that most sites actually have a pt_regs pointer available, making this even more onerous, as the stack space is pointless waste. This patch addresses the problem by observing that software events have well defined nesting semantics, therefore we can use static per-cpu storage instead of on-stack. Linus made the further observation that all but the scheduler callers of perf_sw_event() have a pt_regs available, so we change the regular perf_sw_event() to require a valid pt_regs (where it used to be optional) and add perf_sw_event_sched() for the scheduler. We have a scheduler specific call instead of a more generic _noregs() like construct because we can assume non-recursion from the scheduler and thereby simplify the code further (_noregs would have to put the recursion context call inline in order to assertain which __perf_regs element to use). One last note on the implementation of perf_trace_buf_prepare(); we allow .regs = NULL for those cases where we already have a pt_regs pointer available and do not need another. Reported-by: Linus Torvalds Reported-by: Steven Rostedt Signed-off-by: Peter Zijlstra (Intel) Cc: Arnaldo Carvalho de Melo Cc: Javi Merino Cc: Linus Torvalds Cc: Mathieu Desnoyers Cc: Oleg Nesterov Cc: Paul Mackerras Cc: Petr Mladek Cc: Steven Rostedt Cc: Tom Zanussi Cc: Vaibhav Nagarnaik Link: http://lkml.kernel.org/r/20141216115041.gw3...@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- include/linux/ftrace_event.h| 2 +- include/linux/perf_event.h | 28 +--- include/trace/ftrace.h | 7 --- kernel/events/core.c| 23 +-- kernel/sched/core.c | 2 +- kernel/trace/trace_event_perf.c | 4 +++- kernel/trace/trace_kprobe.c | 4 ++-- kernel/trace/trace_syscalls.c | 4 ++-- kernel/trace/trace_uprobe.c | 2 +- 9 files changed, 52 insertions(+), 24 deletions(-) diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h index 0bebb5c..d36f68b 100644 --- a/include/linux/ftrace_event.h +++ b/include/linux/ftrace_event.h @@ -595,7 +595,7 @@ extern int ftrace_profile_set_filter(struct perf_event *event, int event_id, char *filter_str); extern void ftrace_profile_free_filter(struct perf_event *event); extern void *perf_trace_buf_prepare(int size, unsigned short type, - struct pt_regs *regs, int *rctxp); + struct pt_regs **regs, int *rctxp); static inline void perf_trace_buf_submit(void *raw_data, int size, int rctx, u64 addr, diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index 4f7a61c..3a7bd80 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -665,6 +665,7 @@ static inline int is_software_event(struct perf_event *event) extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; +extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64); extern void __perf_sw_event(u32, u64, struct pt_regs *, u64); #ifndef perf_arch_fetch_caller_regs @@ -689,14 +690,25 @@ static inline void perf_fetch_caller_regs(struct pt_regs *regs) static __always_inline void perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { - struct pt_regs hot_regs; + if (static_key_false(_swevent_enabled[event_id])) + __perf_sw_event(event_id, nr, regs, addr); +} + +DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]); +/* + * 'Special' version for the scheduler, it hard assumes no recursion, + * which is guaranteed by us not actually scheduling inside other swevents + * because those disable preemption. + */ +static __always_inline void +perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) +{ if (static_key_false(_swevent_enabled[event_id])) { - if (!regs) { - perf_fetch_caller_regs(_regs); - regs = _regs; - } - __perf_sw_event(event_id, nr, regs, addr); + struct pt_regs *regs = this_cpu_ptr(&__perf_regs[0]); + + perf_fetch_caller_regs(regs); + ___perf_sw_event(event_id, nr, regs, addr); }
[tip:perf/core] perf: Avoid horrible stack usage
Commit-ID: 86038c5ea81b519a8a1fcfcd5e4599aab0cdd119 Gitweb: http://git.kernel.org/tip/86038c5ea81b519a8a1fcfcd5e4599aab0cdd119 Author: Peter Zijlstra (Intel) pet...@infradead.org AuthorDate: Tue, 16 Dec 2014 12:47:34 +0100 Committer: Ingo Molnar mi...@kernel.org CommitDate: Wed, 14 Jan 2015 15:11:45 +0100 perf: Avoid horrible stack usage Both Linus (most recent) and Steve (a while ago) reported that perf related callbacks have massive stack bloat. The problem is that software events need a pt_regs in order to properly report the event location and unwind stack. And because we could not assume one was present we allocated one on stack and filled it with minimal bits required for operation. Now, pt_regs is quite large, so this is undesirable. Furthermore it turns out that most sites actually have a pt_regs pointer available, making this even more onerous, as the stack space is pointless waste. This patch addresses the problem by observing that software events have well defined nesting semantics, therefore we can use static per-cpu storage instead of on-stack. Linus made the further observation that all but the scheduler callers of perf_sw_event() have a pt_regs available, so we change the regular perf_sw_event() to require a valid pt_regs (where it used to be optional) and add perf_sw_event_sched() for the scheduler. We have a scheduler specific call instead of a more generic _noregs() like construct because we can assume non-recursion from the scheduler and thereby simplify the code further (_noregs would have to put the recursion context call inline in order to assertain which __perf_regs element to use). One last note on the implementation of perf_trace_buf_prepare(); we allow .regs = NULL for those cases where we already have a pt_regs pointer available and do not need another. Reported-by: Linus Torvalds torva...@linux-foundation.org Reported-by: Steven Rostedt rost...@goodmis.org Signed-off-by: Peter Zijlstra (Intel) pet...@infradead.org Cc: Arnaldo Carvalho de Melo a...@kernel.org Cc: Javi Merino javi.mer...@arm.com Cc: Linus Torvalds torva...@linux-foundation.org Cc: Mathieu Desnoyers mathieu.desnoy...@efficios.com Cc: Oleg Nesterov o...@redhat.com Cc: Paul Mackerras pau...@samba.org Cc: Petr Mladek pmla...@suse.cz Cc: Steven Rostedt rost...@goodmis.org Cc: Tom Zanussi tom.zanu...@linux.intel.com Cc: Vaibhav Nagarnaik vnagarn...@google.com Link: http://lkml.kernel.org/r/20141216115041.gw3...@twins.programming.kicks-ass.net Signed-off-by: Ingo Molnar mi...@kernel.org --- include/linux/ftrace_event.h| 2 +- include/linux/perf_event.h | 28 +--- include/trace/ftrace.h | 7 --- kernel/events/core.c| 23 +-- kernel/sched/core.c | 2 +- kernel/trace/trace_event_perf.c | 4 +++- kernel/trace/trace_kprobe.c | 4 ++-- kernel/trace/trace_syscalls.c | 4 ++-- kernel/trace/trace_uprobe.c | 2 +- 9 files changed, 52 insertions(+), 24 deletions(-) diff --git a/include/linux/ftrace_event.h b/include/linux/ftrace_event.h index 0bebb5c..d36f68b 100644 --- a/include/linux/ftrace_event.h +++ b/include/linux/ftrace_event.h @@ -595,7 +595,7 @@ extern int ftrace_profile_set_filter(struct perf_event *event, int event_id, char *filter_str); extern void ftrace_profile_free_filter(struct perf_event *event); extern void *perf_trace_buf_prepare(int size, unsigned short type, - struct pt_regs *regs, int *rctxp); + struct pt_regs **regs, int *rctxp); static inline void perf_trace_buf_submit(void *raw_data, int size, int rctx, u64 addr, diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index 4f7a61c..3a7bd80 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -665,6 +665,7 @@ static inline int is_software_event(struct perf_event *event) extern struct static_key perf_swevent_enabled[PERF_COUNT_SW_MAX]; +extern void ___perf_sw_event(u32, u64, struct pt_regs *, u64); extern void __perf_sw_event(u32, u64, struct pt_regs *, u64); #ifndef perf_arch_fetch_caller_regs @@ -689,14 +690,25 @@ static inline void perf_fetch_caller_regs(struct pt_regs *regs) static __always_inline void perf_sw_event(u32 event_id, u64 nr, struct pt_regs *regs, u64 addr) { - struct pt_regs hot_regs; + if (static_key_false(perf_swevent_enabled[event_id])) + __perf_sw_event(event_id, nr, regs, addr); +} + +DECLARE_PER_CPU(struct pt_regs, __perf_regs[4]); +/* + * 'Special' version for the scheduler, it hard assumes no recursion, + * which is guaranteed by us not actually scheduling inside other swevents + * because those disable preemption. + */ +static __always_inline void +perf_sw_event_sched(u32 event_id, u64 nr, u64 addr) +{ if (static_key_false(perf_swevent_enabled[event_id])) { - if (!regs) { -