- 24 Nov, 2009 3 commits
-
-
Steven Rostedt authored
Converting some of the scheduler trace events to use the TRACE_EVENT_TEMPLATE, DEFINE_EVENT and DEFINE_EVENT_PRINT helped to save some space: $ size kernel/sched.o-* text data bss dec hex filename 79299 6776 2520 88595 15a13 kernel/sched.o-notrace 101941 11896 2584 116421 1c6c5 kernel/sched.o-templ 104779 11896 2584 119259 1d1db kernel/sched.o-trace sched.o-notrace is without any tracepoints compiled sched.o-templ is with this patch sched.o-trace is the tracepoints before this patch The trace events converted to DEFINE_EVENT: sched_wakeup, sched_wakeup_new, sched_process_free, sched_process_exit, and sched_stat_wait. The trace events converted to DEFINE_EVENT_PRINT: sched_stat_sleep and sched_stat_iowait. Note, since the TRACE_EVENT_TEMPLATE always uses a print, the sched_stat_wait print format is defined in the template and this template is used by sched_stat_sleep and sched_stat_iowait. But the later two override the print format. Signed-off-by:Steven Rostedt <rostedt@goodmis.org>
-
Steven Rostedt authored
After creating the TRACE_EVENT_TEMPLATE I started to look at other trace points to see what duplication was made. I noticed that there are several trace points where they are almost identical except for the name and the output format. Since TRACE_EVENT_TEMPLATE was successful in bringing down the size of trace events, I added a DEFINE_EVENT_PRINT. DEFINE_EVENT_PRINT is used just like DEFINE_EVENT is. That is, the DEFINE_EVENT_PRINT also uses a TRACE_EVENT_TEMPLATE, but it allows the developer to overwrite the print format. If there are two or more TRACE_EVENTS that are identical except for the name and print, then they can be converted to use a TRACE_EVENT_TEMPLATE. Since the TRACE_EVENT_TEMPLATE already does the print output, the first trace event would have its print format held in the TRACE_EVENT_TEMPLATE and be defined with a DEFINE_EVENT. The rest will use the DEFINE_EVENT_PRINT and override the print format. Converting the sched trace points to both DEFINE_EVENT and DEFINE_EVENT_PRINT. Five were converted to DEFINE_EVENT and two were converted to DEFINE_EVENT_PRINT. I was able to get the following: $ size kernel/sched.o-* text data bss dec hex filename 79299 6776 2520 88595 15a13 kernel/sched.o-notrace 101941 11896 2584 116421 1c6c5 kernel/sched.o-templ 104779 11896 2584 119259 1d1db kernel/sched.o-trace sched.o-notrace is the scheduler compiled with no trace points. sched.o-templ is with the use of DEFINE_EVENT and DEFINE_EVENT_PRINT sched.o-trace is the current trace events. Signed-off-by: -
Steven Rostedt authored
There are some places in the kernel that define several tracepoints and they are all identical besides the name. The code to enable, disable and record is created for every trace point even if most of the code is identical. This patch adds TRACE_EVENT_TEMPLATE that lets the developer create a template TRACE_EVENT and create trace points with DEFINE_EVENT, which is based off of a given template. Each trace point used by this will share most of the code, and bring down the size of the kernel when there are several duplicate events. Usage is: TRACE_EVENT_TEMPLATE(name, proto, args, tstruct, assign, print); Which would be the same as defining a normal TRACE_EVENT. To create the trace events that the trace points will use: DEFINE_EVENT(template, name, proto, args) is done. The template is the name of the TRACE_EVENT_TEMPLATE to use. The name is the name of the trace point. The parameters proto and args must be the same as the proto and args of the template. If they are not the same, then a compile error will result. I tried hard removing this duplication but the C preprocessor is not powerful enough (or my CPP magic experience points is not at a high enough level) to not need them. A lot of trace events are coming in with new XFS development. Most of the trace points are identical except for the name. The following shows the advantage of having TRACE_EVENT_TEMPLATE: $ size fs/xfs/xfs.o.* text data bss dec hex filename 452114 2788 3520 458422 6feb6 fs/xfs/xfs.o.old 638482 38116 3744 680342 a6196 fs/xfs/xfs.o.template 996954 38116 4480 1039550 fdcbe fs/xfs/xfs.o.trace xfs.o.old is without any tracepoints. xfs.o.template uses the new TRACE_EVENT_TEMPLATE. xfs.o.trace uses the current TRACE_EVENT macros. Requested-by:Christoph Hellwig <hch@lst.de> Signed-off-by:
-
- 23 Nov, 2009 20 commits
-
-
Frederic Weisbecker authored
Fix a misplaced ifdef. We need the perf event headers also in off-case to avoid the following build error: include/linux/hw_breakpoint.h:94: error: expected declaration specifiers or '...' before 'perf_callback_t' include/linux/hw_breakpoint.h:102: error: expected declaration specifiers or '...' before 'perf_callback_t' include/linux/hw_breakpoint.h:109: error: expected declaration specifiers or '...' before 'perf_callback_t' include/linux/hw_breakpoint.h:116: error: expected declaration specifiers or '...' before 'perf_callback_t' Reported-by:
Kisskb-bot by Michael Ellerman <michael@ellerman.id.au> Signed-off-by:
Frederic Weisbecker <fweisbec@gmail.com> Cc: Prasad <prasad@linux.vnet.ibm.com> LKML-Reference: <1259011812-8093-1-git-send-email-fweisbec@gmail.com> Signed-off-by:
Ingo Molnar <mingo@elte.hu>
-
Arnaldo Carvalho de Melo authored
E.g.: [root@doppio linux-2.6-tip]# perf kmem record sleep 3s [ perf record: Woken up 2 times to write data ] [ perf record: Captured and wrote 0.804 MB perf.data (~35105 samples) ] [root@doppio linux-2.6-tip]# perf kmem --stat caller | head -10 ------------------------------------------------------------------------------ Callsite |Total_alloc/Per | Total_req/Per | Hit | Frag ------------------------------------------------------------------------------ getname/40 | 1519616/4096 | 1519616/4096 | 371| 0.000% seq_read/a2 | 987136/4096 | 987136/4096 | 241| 0.000% __netdev_alloc_skb/43 | 260368/1049 | 259968/1048 | 248| 0.154% __alloc_skb/5a | 77312/256 | 77312/256 | 302| 0.000% proc_alloc_inode/33 | 76480/632 | 76472/632 | 121| 0.010% get_empty_filp/8d | 70272/192 | 70272/192 | 366| 0.000% split_vma/8e | 42064/176 | 42064/176 | 239| 0.000% [root@doppio linux-2.6-tip]# Signed-off-by:
Arnaldo Carvalho de Melo <acme@redhat.com> Acked-by:
Pekka Enberg <penberg@cs.helsinki.fi> Cc: Eduard - Gabriel Munteanu <eduard.munteanu@linux360.ro> Cc: Frédéric Weisbecker <fweisbec@gmail.com> Cc: linux-mm@kvack.org <linux-mm@kvack.org> Cc: Li Zefan <lizf@cn.fujitsu.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Steven Rostedt <rostedt@goodmis.org> LKML-Reference: <1259005869-13487-2-git-send-email-acme@infradead.org> Signed-off-by:
-
Arnaldo Carvalho de Melo authored
So that they can be used in other tools. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
Now that we can check the buildid to see if it really matches, this can be done safely: vmlinux /boot/vmlinux /boot/vmlinux-<uts.release> /lib/modules/<uts.release>/build/vmlinux /usr/lib/debug/lib/modules/%s/vmlinux More can be added - if you know about distros that put the vmlinux somewhere else please let us know. Signed-off-by:
-
Frederic Weisbecker authored
Add the breakpoint events support with this new sysnopsis: mem:addr[:access] Where addr is a raw addr value in the kernel and access can be either [r][w][x] Example to profile tasklist_lock: $ grep tasklist_lock /proc/kallsyms ffffffff8189c000 D tasklist_lock $ perf record -e mem:0xffffffff8189c000:rw -a -f -c 1 $ perf report # Samples: 62 # # Overhead Command Shared Object Symbol # ........ ............... ............. ...... # 29.03% swapper [kernel] [k] _raw_read_trylock 29.03% swapper [kernel] [k] _raw_read_unlock 19.35% init [kernel] [k] _raw_read_trylock 19.35% init [kernel] [k] _raw_read_unlock 1.61% events/0 [kernel] [k] _raw_read_trylock 1.61% events/0 [kernel] [k] _raw_read_unlock Coming soon: - Support for symbols in the event definition. - Default period to 1 for breakpoint events because these are not high frequency events. The same thing is needed for trace events. Signed-off-by:
-
Frederic Weisbecker authored
Add the remaining necessary bits to support breakpoints created through perf syscall. We don't use the software counter interface as: - We don't need to check against recursion, this is already done in hardware breakpoints arch level. - We already know the perf event we are dealing with when the event is to be committed. Signed-off-by:
-
Frederic Weisbecker authored
Perf tools create perf events as disabled in the beginning. Breakpoints are then considered like ptrace temporary breakpoints, only meant to reserve a breakpoint slot until we get all the necessary informations from the user. In this case, we don't check the address that is breakpointed as it is NULL in the ptrace case. But perf tools don't have the same purpose, events are created disabled to wait for all events to be created before enabling all of them. We want to check the breakpoint parameters in this case. Signed-off-by:
-
Frederic Weisbecker authored
As userspace only needs the breakpoints enum types from the breakpoints headers. Signed-off-by:
-
K.Prasad authored
Attribute authorship to developers of hw-breakpoint related files. Signed-off-by:
K.Prasad <prasad@linux.vnet.ibm.com> Cc: Alan Stern <stern@rowland.harvard.edu> Cc: Frederic Weisbecker <fweisbec@gmail.com> LKML-Reference: <20091123154713.GA5593@in.ibm.com> [ v2: moved it to latest -tip ] Signed-off-by:
-
Peter Zijlstra authored
It is quite possible to call update_event_times() on a context that isn't actually running and thereby confuse the thing. perf stat was reporting !100% scale values for software counters (2e2af50b perf_events: Disable events when we detach them, solved the worst of that, but there was still some left). The thing that happens is that because we are not self-reaping (we have a caring parent) there is a time between the last schedule (out) and having do_exit() called which will detach the events. This period would be accounted as enabled,!running because the event->state==INACTIVE, even though !event->ctx->is_active. Similar issues could have been observed by calling read() on a event while the attached task was not scheduled in. Solve this by teaching update_event_times() about ctx->is_active. Signed-off-by:
Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> LKML-Reference: <1258984836.4531.480.camel@laptop> Signed-off-by:
-
Peter Zijlstra authored
Make perf_swevent_get_recursion_context return a context number and disable preemption. This could be used to remove the IRQ disable from the trace bit and index the per-cpu buffer with. Signed-off-by:
-
Peter Zijlstra authored
Move the update_event_times() call in __perf_event_exit_task() into list_del_event() because that holds the proper lock (ctx->lock) and seems a more natural place to do the last time update. Signed-off-by:
-
Peter Zijlstra authored
It appeared we did call update_event_times() on exit, but we failed to update the context time, which renders the former moot. Locking is a bit iffy, we call update_event_times under ctx->mutex instead of ctx->lock - the next patch fixes this. Signed-off-by:
-
Peter Zijlstra authored
If we leave the event in STATE_INACTIVE, any read of the event after the detach will increase the running count but not the enabled count and cause funny scaling artefacts. Signed-off-by:
-
Peter Zijlstra authored
Signed-off-by:
-
Peter Zijlstra authored
We had two almost identical functions, avoid the duplication. Signed-off-by:
-
Ingo Molnar authored
The structure init creates a bit memcpy, which shows up big time in perf annotate output: : ffffffff810a859d <__perf_sw_event>: 1.68 : ffffffff810a859d: 55 push %rbp 1.69 : ffffffff810a859e: 41 89 fa mov %edi,%r10d 0.01 : ffffffff810a85a1: 49 89 c9 mov %rcx,%r9 0.00 : ffffffff810a85a4: 31 c0 xor %eax,%eax 1.71 : ffffffff810a85a6: b9 16 00 00 00 mov $0x16,%ecx 0.00 : ffffffff810a85ab: 48 89 e5 mov %rsp,%rbp 0.00 : ffffffff810a85ae: 48 83 ec 60 sub $0x60,%rsp 1.52 : ffffffff810a85b2: 48 8d 7d a0 lea -0x60(%rbp),%rdi 85.20 : ffffffff810a85b6: f3 ab rep stos %eax,%es:(%rdi) None of the callees depends on the structure being pre-initialized, so only initialize ->addr. This gets rid of the memcpy overhead. Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> LKML-Reference: <new-submission> Signed-off-by: -
Ingo Molnar authored
Decreases perf overhead when function tracing is enabled, by about 50%. Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Mike Galbraith <efault@gmx.de> Cc: Paul Mackerras <paulus@samba.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> LKML-Reference: <new-submission> Signed-off-by: -
Simon Kaempflein authored
Print more accurate error message when "perf record" fails because there is no APIC support, on x86. Signed-off-by: -
Michael S. Tsirkin authored
gcc with no flags typically is a sane default for systems to use, and looking at the running kernel is probably broken for cross-builds anyway, so let's not do this. Add EXTRA_CFLAGS so that users can override default gcc mode if they want to. Signed-off-by:
Michael S. Tsirkin <mst@redhat.com> Acked-by:
Arjan van de Ven <arjan@infradead.org> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <20091122121335.GA24254@redhat.com> Signed-off-by:
-
- 22 Nov, 2009 12 commits
-
-
Arnaldo Carvalho de Melo authored
Bug introduced in 439d473b, making the initial map be used for all IPs, so that symbols outside this initial map would either be erroneously resolved or not resolve at all. Reported-by:
-
Arnaldo Carvalho de Melo authored
Suggested-by:
-
Michael S. Tsirkin authored
On error, suggest installing static libraries along with shared libraries. Signed-off-by:
-
Michael S. Tsirkin authored
Make standard error show up on console when V=2 is set. Signed-off-by:
-
Ingo Molnar authored
Fix: ERROR: "perf_swevent_put_recursion_context" [fs/ext4/ext4.ko] undefined! ERROR: "perf_swevent_get_recursion_context" [fs/ext4/ext4.ko] undefined! Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Masami Hiramatsu <mhiramat@redhat.com> Cc: Jason Baron <jbaron@redhat.com> LKML-Reference: <1258864015-10579-1-git-send-email-fweisbec@gmail.com> Signed-off-by: -
Pekka Enberg authored
The write_event() function in builtin-record.c writes out all mmap()'d DSOs including non-ELF files like GNOME resource files and such. Therefore, check for ELF_K_ELF in filename__read_build_id() before attempting to read the ELF header with gelf_getehdr(). Fixes the following error messages when running "perf kmem record": penberg@penberg-laptop:~/src/linux/tools/perf$ perf kmem record ^C[ perf record: Woken up 2 times to write data ] [ perf record: Captured and wrote 0.753 MB perf.data (~32885 samples) ] filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. filename__read_build_id: cannot get elf header. Signed-off-by:
-
Pekka Enberg authored
This patch adds support for "--sort hit" and "--sort frag" to the "perf kmem" tool. The former was already mentioned in the help text and the latter is useful for finding call-sites that exhibit worst case behavior for SLAB allocators. Signed-off-by:
-
Márton Németh authored
The buffer is first zeroed out by memset(). Then strncpy() is used to fill the content. The strncpy() function also pads the string till the end of the specified length, which is redundant. The strncpy() does not ensures that the string will be properly closed with 0. Use strlcpy() instead. The semantic match that finds this kind of pattern is as follows: (http://coccinelle.lip6.fr/) // <smpl> @@ expression buffer; expression size; expression str; @@ memset(buffer, 0, size); ... - strncpy( + strlcpy( buffer, str, sizeof(buffer) ); @@ expression buffer; expression size; expression str; @@ memset(&buffer, 0, size); ... - strncpy( + strlcpy( &buffer, str, sizeof(buffer)); @@ expression buffer; identifier field; expression size; expression str; @@ memset(buffer, 0, size); ... - strncpy( + strlcpy( buffer->field, str, sizeof(buffer->field) ); @@ expression buffer; identifier field; expression size; expression str; @@ memset(&buffer, 0, size); ... - strncpy( + strlcpy( buffer.field, str, sizeof(buffer.field)); // </smpl> On strncpy() vs strlcpy() see http://www.gratisoft.us/todd/papers/strlcpy.html . Signed-off-by:
Márton Németh <nm127@freemail.hu> Cc: Julia Lawall <julia@diku.dk> Cc: cocci@diku.dk Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Paul Mackerras <paulus@samba.org> LKML-Reference: <4B086547.5040100@freemail.hu> Signed-off-by:
-
Frederic Weisbecker authored
So that we can include this header from userspace tools, like perf tools, to get the breakpoint types and len definitions. Signed-off-by:
-
Frederic Weisbecker authored
Remove asm/processor.h and asm/debugreg.h as these headers are not used anymore in the hw-breakpoints core file. Signed-off-by:
-
Frederic Weisbecker authored
We are never in an NMI context when we commit a syscall trace to perf. So just forget about the nmi buffer there. Signed-off-by:
-
Frederic Weisbecker authored
When we commit a trace to perf, we first check if we are recursing in the same buffer so that we don't mess-up the buffer with a recursing trace. But later on, we do the same check from perf to avoid commit recursion. The recursion check is desired early before we touch the buffer but we want to do this check only once. Then export the recursion protection from perf and use it from the trace events before submitting a trace. v2: Put appropriate Reported-by tag Reported-by:
Peter Zijlstra <peterz@infradead.org> Signed-off-by:
-
- 21 Nov, 2009 5 commits
-
-
Arnaldo Carvalho de Melo authored
It better propagate errors, also if we do a simple: [root@doppio linux-2.6-tip]# perf record -R -a -f sleep 3s ; perf trace [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.182 MB perf.data (~7972 samples) ] Fatal: not an trace data file [root@doppio linux-2.6-tip]# That is what is expected, right? I.e. as we didn't specify any tracepoint event via -e, it should gracefully bail out and not SEGFAULT. Signed-off-by:
-
Arnaldo Carvalho de Melo authored
Signed-off-by:
-
Arnaldo Carvalho de Melo authored
We better call this routine after both the kernel and modules are loaded, because as it was if there weren't modules it would not be called, resulting in kernel_map->end remaining at zero, so no map would be found and consequently the kernel symtab wouldn't get loaded, i.e. no kernel symbols would be resolved. Also this fixes another case, that is when we _have_ modules, but the last map would have its ->end address not set before we loaded its symbols, which would never happen because ->end was not set. Reported-by:
-
Stephane Eranian authored
This patch fixes the default watermark value for the sampling buffer. With the existing calculation (watermark = max(PAGE_SIZE, max_size / 2)), no notification was ever received when the buffer was exactly 1 page. This was because you would never cross the threshold (there is no partial samples). In certain configuration, there was no possibilty detecting the problem because there was not enough space left to store the LOST record.In fact, there may be a more generic problem here. The kernel should ensure that there is alaways enough space to store one LOST record. This patch sets the default watermark to half the buffer size. With such limit, we are guaranteed to get a notification even with a single page buffer assuming no sample is bigger than a page. Signed-off-by:
Stephane Eranian <eranian@gmail.com> Signed-off-by:
-
Peter Zijlstra authored
We should hold event->child_mutex when iterating the inherited counters, we should hold ctx->mutex when iterating siblings. Signed-off-by:
-
