From f2bf1f6f5f89d031245067512449fc889b2f4bb2 Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Wed, 6 Jun 2012 19:50:40 -0400 Subject: [PATCH 1/7] tracing: Have tracing_off() actually turn tracing off A recent update to have tracing_on/off() only affect the ftrace ring buffers instead of all ring buffers had a cut and paste error. The tracing_off() did the exact same thing as tracing_on() and would not actually turn off tracing. Unfortunately, tracing_off() is more important to be working than tracing_on() as this is a key development tool, as it lets the developer turn off tracing as soon as a problem is discovered. It is also used by panic and oops code. This bug also breaks the 'echo func:traceoff > set_ftrace_filter' Cc: # 3.4 Signed-off-by: Steven Rostedt --- kernel/trace/trace.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c index 68032c6177db..49249c28690d 100644 --- a/kernel/trace/trace.c +++ b/kernel/trace/trace.c @@ -371,7 +371,7 @@ EXPORT_SYMBOL_GPL(tracing_on); void tracing_off(void) { if (global_trace.buffer) - ring_buffer_record_on(global_trace.buffer); + ring_buffer_record_off(global_trace.buffer); /* * This flag is only looked at when buffers haven't been * allocated yet. We don't really care about the race From 80c0120a3cca30166c0ab8b24e44be67e97b79af Mon Sep 17 00:00:00 2001 From: David Ahern Date: Fri, 8 Jun 2012 11:47:51 -0300 Subject: [PATCH 2/7] perf tools: Fix endianity swapping for adds_features bitmask Based on Jiri's latest attempt: https://lkml.org/lkml/2012/5/16/61 Basically, adds_features should be byte swapped assuming unsigned longs are either 8-bytes (u64) or 4-bytes (u32). Fixes 32-bit ppc dumping 64-bit x86 feature data: ======== captured on: Sun May 20 19:23:23 2012 hostname : nxos-vdc-dev3 os release : 3.4.0-rc7+ perf version : 3.4.rc4.137.g978da3 arch : x86_64 nrcpus online : 16 nrcpus avail : 16 cpudesc : Intel(R) Xeon(R) CPU E5540 @ 2.53GHz cpuid : GenuineIntel,6,26,5 total memory : 24680324 kB ... Verified 64-bit x86 can still dump feature data for 32-bit ppc. Signed-off-by: David Ahern Reviewed-by: Jiri Olsa Cc: Corey Ashford Cc: Frederic Weisbecker Cc: Paul Mackerras Cc: Peter Zijlstra Link: http://lkml.kernel.org/r/4FBBB539.5010805@gmail.com Signed-off-by: Arnaldo Carvalho de Melo --- tools/perf/util/header.c | 16 +++++++++------- tools/perf/util/include/linux/bitops.h | 2 ++ tools/perf/util/session.c | 10 ++++++++++ tools/perf/util/session.h | 1 + 4 files changed, 22 insertions(+), 7 deletions(-) diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c index 2dd5edf161b7..4f9b247fb312 100644 --- a/tools/perf/util/header.c +++ b/tools/perf/util/header.c @@ -1942,7 +1942,6 @@ int perf_file_header__read(struct perf_file_header *header, else return -1; } else if (ph->needs_swap) { - unsigned int i; /* * feature bitmap is declared as an array of unsigned longs -- * not good since its size can differ between the host that @@ -1958,14 +1957,17 @@ int perf_file_header__read(struct perf_file_header *header, * file), punt and fallback to the original behavior -- * clearing all feature bits and setting buildid. */ - for (i = 0; i < BITS_TO_LONGS(HEADER_FEAT_BITS); ++i) - header->adds_features[i] = bswap_64(header->adds_features[i]); + mem_bswap_64(&header->adds_features, + BITS_TO_U64(HEADER_FEAT_BITS)); if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { - for (i = 0; i < BITS_TO_LONGS(HEADER_FEAT_BITS); ++i) { - header->adds_features[i] = bswap_64(header->adds_features[i]); - header->adds_features[i] = bswap_32(header->adds_features[i]); - } + /* unswap as u64 */ + mem_bswap_64(&header->adds_features, + BITS_TO_U64(HEADER_FEAT_BITS)); + + /* unswap as u32 */ + mem_bswap_32(&header->adds_features, + BITS_TO_U32(HEADER_FEAT_BITS)); } if (!test_bit(HEADER_HOSTNAME, header->adds_features)) { diff --git a/tools/perf/util/include/linux/bitops.h b/tools/perf/util/include/linux/bitops.h index f1584833bd22..587a230d2075 100644 --- a/tools/perf/util/include/linux/bitops.h +++ b/tools/perf/util/include/linux/bitops.h @@ -8,6 +8,8 @@ #define BITS_PER_LONG __WORDSIZE #define BITS_PER_BYTE 8 #define BITS_TO_LONGS(nr) DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(long)) +#define BITS_TO_U64(nr) DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(u64)) +#define BITS_TO_U32(nr) DIV_ROUND_UP(nr, BITS_PER_BYTE * sizeof(u32)) #define for_each_set_bit(bit, addr, size) \ for ((bit) = find_first_bit((addr), (size)); \ diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c index 2600916efa83..c3e399bcf18d 100644 --- a/tools/perf/util/session.c +++ b/tools/perf/util/session.c @@ -442,6 +442,16 @@ static void perf_tool__fill_defaults(struct perf_tool *tool) tool->finished_round = process_finished_round_stub; } } + +void mem_bswap_32(void *src, int byte_size) +{ + u32 *m = src; + while (byte_size > 0) { + *m = bswap_32(*m); + byte_size -= sizeof(u32); + ++m; + } +} void mem_bswap_64(void *src, int byte_size) { diff --git a/tools/perf/util/session.h b/tools/perf/util/session.h index 7a5434c00565..0c702e3f0a36 100644 --- a/tools/perf/util/session.h +++ b/tools/perf/util/session.h @@ -80,6 +80,7 @@ struct branch_info *machine__resolve_bstack(struct machine *self, bool perf_session__has_traces(struct perf_session *self, const char *msg); void mem_bswap_64(void *src, int byte_size); +void mem_bswap_32(void *src, int byte_size); void perf_event__attr_swap(struct perf_event_attr *attr); int perf_session__create_kernel_maps(struct perf_session *self); From fc3e4d077d5c7a7bc1ad5bc143895b4e070e5a8b Mon Sep 17 00:00:00 2001 From: Stephane Eranian Date: Tue, 15 May 2012 13:11:11 +0200 Subject: [PATCH 3/7] perf stat: Fix default output file The following commit: commit 56f3bae70638b33477a6015fd362ccfe354fd3ee Author: Jim Cromie Date: Wed Sep 7 17:14:00 2011 -0600 perf stat: Add --log-fd option to redirect stderr elsewhere introduced a bug in the way perf stat outputs the results by default, i.e., without the --log-fd or --output option. It would default to writing to file descriptor 0, i.e., stdin. Writing to stdin is allowed and is equivalent to writing to stdout. However, there is a major difference for any script that was already capturing the output of perf stat via redirection: perf stat >/tmp/log .... or perf stat 2>/tmp/log .... They would not capture anything anymore. They would have to do: perf stat 0>/tmp/log ... This breaks compatibility with existing scripts and does not look very natural. This patch fixes the problem by looking at output_fd only when it was modified by user (> 0). It also checks that the value if positive. Passing --log-fd 0 is ignored. I would also argue that defaulting to stderr for the results is not the right thing to do, though this patch does not address this specific issue. Signed-off-by: Stephane Eranian Cc: David Ahern Cc: Ingo Molnar Cc: Peter Zijlstra Cc: Jim Cromie Link: http://lkml.kernel.org/r/20120515111111.GA9870@quad Signed-off-by: Arnaldo Carvalho de Melo --- tools/perf/builtin-stat.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c index 262589991ea4..07b5c7703dd1 100644 --- a/tools/perf/builtin-stat.c +++ b/tools/perf/builtin-stat.c @@ -1179,6 +1179,12 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used) fprintf(stderr, "cannot use both --output and --log-fd\n"); usage_with_options(stat_usage, options); } + + if (output_fd < 0) { + fprintf(stderr, "argument to --log-fd must be a > 0\n"); + usage_with_options(stat_usage, options); + } + if (!output) { struct timespec tm; mode = append_file ? "a" : "w"; @@ -1190,7 +1196,7 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used) } clock_gettime(CLOCK_REALTIME, &tm); fprintf(output, "# started on %s\n", ctime(&tm.tv_sec)); - } else if (output_fd != 2) { + } else if (output_fd > 0) { mode = append_file ? "a" : "w"; output = fdopen(output_fd, mode); if (!output) { From cb9dd49e11f83d548c822d7022ac180b0518b25c Mon Sep 17 00:00:00 2001 From: Arnaldo Carvalho de Melo Date: Mon, 11 Jun 2012 19:03:32 -0300 Subject: [PATCH 4/7] perf tools: Fix synthesizing tracepoint names from the perf.data headers We need to use the per event info snapshoted at record time to synthesize the events name, so do it just after reading the perf.data headers, when we already processed the /sys events data, otherwise we'll end up using the local /sys that only by sheer luck will have the same tracepoint ID -> real event association. Example: # uname -a Linux felicio.ghostprotocols.net 3.4.0-rc5+ #1 SMP Sat May 19 15:27:11 BRT 2012 x86_64 x86_64 x86_64 GNU/Linux # perf record -e sched:sched_switch usleep 1 [ perf record: Woken up 1 times to write data ] [ perf record: Captured and wrote 0.015 MB perf.data (~648 samples) ] # cat /t/events/sched/sched_switch/id 279 # perf evlist -v sched:sched_switch: sample_freq=1, type: 2, config: 279, size: 80, sample_type: 1159, read_format: 7, disabled: 1, inherit: 1, mmap: 1, comm: 1, enable_on_exec: 1, sample_id_all: 1, exclude_guest: 1 # So on the above machine the sched:sched_switch has tracepoint id 279, but on the machine were we'll analyse it it has a different id: $ cat /t/events/sched/sched_switch/id 56 $ perf evlist -i /tmp/perf.data kmem:mm_balancedirty_writeout $ cat /t/events/kmem/mm_balancedirty_writeout/id 279 With this fix: $ perf evlist -i /tmp/perf.data sched:sched_switch Reported-by: Dmitry Antipov Cc: David Ahern Cc: Frederic Weisbecker Cc: Jiri Olsa Cc: Mike Galbraith Cc: Namhyung Kim Cc: Paul Mackerras Cc: Peter Zijlstra Cc: Stephane Eranian Link: http://lkml.kernel.org/n/tip-auwks8fpuhmrdpiefs55o5oz@git.kernel.org Signed-off-by: Arnaldo Carvalho de Melo --- tools/perf/util/header.c | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+) diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c index 4f9b247fb312..e909d43cf542 100644 --- a/tools/perf/util/header.c +++ b/tools/perf/util/header.c @@ -2093,6 +2093,35 @@ static int read_attr(int fd, struct perf_header *ph, return ret <= 0 ? -1 : 0; } +static int perf_evsel__set_tracepoint_name(struct perf_evsel *evsel) +{ + struct event_format *event = trace_find_event(evsel->attr.config); + char bf[128]; + + if (event == NULL) + return -1; + + snprintf(bf, sizeof(bf), "%s:%s", event->system, event->name); + evsel->name = strdup(bf); + if (event->name == NULL) + return -1; + + return 0; +} + +static int perf_evlist__set_tracepoint_names(struct perf_evlist *evlist) +{ + struct perf_evsel *pos; + + list_for_each_entry(pos, &evlist->entries, node) { + if (pos->attr.type == PERF_TYPE_TRACEPOINT && + perf_evsel__set_tracepoint_name(pos)) + return -1; + } + + return 0; +} + int perf_session__read_header(struct perf_session *session, int fd) { struct perf_header *header = &session->header; @@ -2174,6 +2203,9 @@ int perf_session__read_header(struct perf_session *session, int fd) lseek(fd, header->data_offset, SEEK_SET); + if (perf_evlist__set_tracepoint_names(session->evlist)) + goto out_delete_evlist; + header->frozen = 1; return 0; out_errno: From 25f42985825dd93f0593efe454e54c2aa13f7830 Mon Sep 17 00:00:00 2001 From: Stephane Eranian Date: Mon, 11 Jun 2012 15:44:26 +0200 Subject: [PATCH 5/7] perf/x86: Fix broken LBR fixup code I noticed that the LBR fixups were not working anymore on programs where they used to. I tracked this down to a recent change to copy_from_user_nmi(): db0dc75d640 ("perf/x86: Check user address explicitly in copy_from_user_nmi()") This commit added a call to __range_not_ok() to the copy_from_user_nmi() routine. The problem is that the logic of the test must be reversed. __range_not_ok() returns 0 if the range is VALID. We want to return early from copy_from_user_nmi() if the range is NOT valid. Signed-off-by: Stephane Eranian Signed-off-by: Peter Zijlstra Acked-by: Arun Sharma Link: http://lkml.kernel.org/r/20120611134426.GA7542@quad Signed-off-by: Ingo Molnar --- arch/x86/lib/usercopy.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/arch/x86/lib/usercopy.c b/arch/x86/lib/usercopy.c index 677b1ed184c9..4f74d94c8d97 100644 --- a/arch/x86/lib/usercopy.c +++ b/arch/x86/lib/usercopy.c @@ -22,7 +22,7 @@ copy_from_user_nmi(void *to, const void __user *from, unsigned long n) void *map; int ret; - if (__range_not_ok(from, n, TASK_SIZE) == 0) + if (__range_not_ok(from, n, TASK_SIZE)) return len; do { From a70270468234749741c5893ae78e5bb524771402 Mon Sep 17 00:00:00 2001 From: Don Zickus Date: Wed, 13 Jun 2012 09:35:48 -0400 Subject: [PATCH 6/7] watchdog: Quiet down the boot messages A bunch of bugzillas have complained how noisy the nmi_watchdog is during boot-up especially with its expected failure cases (like virt and bios resource contention). This is my attempt to quiet them down and keep it less confusing for the end user. What I did is print the message for cpu0 and save it for future comparisons. If future cpus have an identical message as cpu0, then don't print the redundant info. However, if a future cpu has a different message, happily print that loudly. Before the change, you would see something like: ..TIMER: vector=0x30 apic1=0 pin1=2 apic2=-1 pin2=-1 CPU0: Intel(R) Core(TM)2 Quad CPU Q9550 @ 2.83GHz stepping 0a Performance Events: PEBS fmt0+, Core2 events, Intel PMU driver. ... version: 2 ... bit width: 40 ... generic registers: 2 ... value mask: 000000ffffffffff ... max period: 000000007fffffff ... fixed-purpose events: 3 ... event mask: 0000000700000003 NMI watchdog enabled, takes one hw-pmu counter. Booting Node 0, Processors #1 NMI watchdog enabled, takes one hw-pmu counter. #2 NMI watchdog enabled, takes one hw-pmu counter. #3 Ok. NMI watchdog enabled, takes one hw-pmu counter. Brought up 4 CPUs Total of 4 processors activated (22607.24 BogoMIPS). After the change, it is simplified to: ..TIMER: vector=0x30 apic1=0 pin1=2 apic2=-1 pin2=-1 CPU0: Intel(R) Core(TM)2 Quad CPU Q9550 @ 2.83GHz stepping 0a Performance Events: PEBS fmt0+, Core2 events, Intel PMU driver. ... version: 2 ... bit width: 40 ... generic registers: 2 ... value mask: 000000ffffffffff ... max period: 000000007fffffff ... fixed-purpose events: 3 ... event mask: 0000000700000003 NMI watchdog: enabled on all CPUs, permanently consumes one hw-PMU counter. Booting Node 0, Processors #1 #2 #3 Ok. Brought up 4 CPUs V2: little changes based on Joe Perches' feedback V3: printk cleanup based on Ingo's feedback; checkpatch fix V4: keep printk as one long line V5: Ingo fix ups Reported-and-tested-by: Nathan Zimmer Signed-off-by: Don Zickus Cc: nzimmer@sgi.com Cc: joe@perches.com Link: http://lkml.kernel.org/r/1339594548-17227-1-git-send-email-dzickus@redhat.com Signed-off-by: Ingo Molnar --- kernel/watchdog.c | 19 ++++++++++++++++++- 1 file changed, 18 insertions(+), 1 deletion(-) diff --git a/kernel/watchdog.c b/kernel/watchdog.c index e5e1d85b8c7c..4b1dfba70f7c 100644 --- a/kernel/watchdog.c +++ b/kernel/watchdog.c @@ -372,6 +372,13 @@ static int watchdog(void *unused) #ifdef CONFIG_HARDLOCKUP_DETECTOR +/* + * People like the simple clean cpu node info on boot. + * Reduce the watchdog noise by only printing messages + * that are different from what cpu0 displayed. + */ +static unsigned long cpu0_err; + static int watchdog_nmi_enable(int cpu) { struct perf_event_attr *wd_attr; @@ -390,11 +397,21 @@ static int watchdog_nmi_enable(int cpu) /* Try to register using hardware perf events */ event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); + + /* save cpu0 error for future comparision */ + if (cpu == 0 && IS_ERR(event)) + cpu0_err = PTR_ERR(event); + if (!IS_ERR(event)) { - pr_info("enabled, takes one hw-pmu counter.\n"); + /* only print for cpu0 or different than cpu0 */ + if (cpu == 0 || cpu0_err) + pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); goto out_save; } + /* skip displaying the same error again */ + if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) + return PTR_ERR(event); /* vary the KERN level based on the returned errno */ if (PTR_ERR(event) == -EOPNOTSUPP) From 9c5da09d266ca9b32eb16cf940f8161d949c2fe5 Mon Sep 17 00:00:00 2001 From: Salman Qazi Date: Thu, 14 Jun 2012 15:31:09 -0700 Subject: [PATCH 7/7] perf: Use css_tryget() to avoid propping up css refcount An rmdir pushes css's ref count to zero. However, if the associated directory is open at the time, the dentry ref count is non-zero. If the fd for this directory is then passed into perf_event_open, it does a css_get(). This bounces the ref count back up from zero. This is a problem by itself. But what makes it turn into a crash is the fact that we end up doing an extra dput, since we perform a dput when css_put sees the ref count go down to zero. css_tryget() does not fall into that trap. So, we use that instead. Reproduction test-case for the bug: #include #include #include #include #include #include #include #include #include #define PERF_FLAG_PID_CGROUP (1U << 2) int perf_event_open(struct perf_event_attr *hw_event_uptr, pid_t pid, int cpu, int group_fd, unsigned long flags) { return syscall(__NR_perf_event_open,hw_event_uptr, pid, cpu, group_fd, flags); } /* * Directly poke at the perf_event bug, since it's proving hard to repro * depending on where in the kernel tree. what moved? */ int main(int argc, char **argv) { int fd; struct perf_event_attr attr; memset(&attr, 0, sizeof(attr)); attr.exclude_kernel = 1; attr.size = sizeof(attr); mkdir("/dev/cgroup/perf_event/blah", 0777); fd = open("/dev/cgroup/perf_event/blah", O_RDONLY); perror("open"); rmdir("/dev/cgroup/perf_event/blah"); sleep(2); perf_event_open(&attr, fd, 0, -1, PERF_FLAG_PID_CGROUP); perror("perf_event_open"); close(fd); return 0; } Signed-off-by: Salman Qazi Signed-off-by: Peter Zijlstra Acked-by: Tejun Heo Link: http://lkml.kernel.org/r/20120614223108.1025.2503.stgit@dungbeetle.mtv.corp.google.com Signed-off-by: Ingo Molnar --- kernel/events/core.c | 10 +++++++--- 1 file changed, 7 insertions(+), 3 deletions(-) diff --git a/kernel/events/core.c b/kernel/events/core.c index f85c0154b333..d7d71d6ec972 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -253,9 +253,9 @@ perf_cgroup_match(struct perf_event *event) return !event->cgrp || event->cgrp == cpuctx->cgrp; } -static inline void perf_get_cgroup(struct perf_event *event) +static inline bool perf_tryget_cgroup(struct perf_event *event) { - css_get(&event->cgrp->css); + return css_tryget(&event->cgrp->css); } static inline void perf_put_cgroup(struct perf_event *event) @@ -484,7 +484,11 @@ static inline int perf_cgroup_connect(int fd, struct perf_event *event, event->cgrp = cgrp; /* must be done before we fput() the file */ - perf_get_cgroup(event); + if (!perf_tryget_cgroup(event)) { + event->cgrp = NULL; + ret = -ENOENT; + goto out; + } /* * all events in a group must monitor