From 465a454f254ee2ff7acc4aececbe31f8af046bc0 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Mon, 15 Jun 2009 12:31:37 +0200 Subject: x86, mm: Add __get_user_pages_fast() Introduce a gup_fast() variant which is usable from IRQ/NMI context. Signed-off-by: Peter Zijlstra CC: Nick Piggin Cc: Mike Galbraith Cc: Paul Mackerras Cc: Arnaldo Carvalho de Melo LKML-Reference: Signed-off-by: Ingo Molnar --- include/linux/mm.h | 6 ++++++ 1 file changed, 6 insertions(+) (limited to 'include') diff --git a/include/linux/mm.h b/include/linux/mm.h index ad613ed66ab0..b457bc047ab1 100644 --- a/include/linux/mm.h +++ b/include/linux/mm.h @@ -862,6 +862,12 @@ extern int mprotect_fixup(struct vm_area_struct *vma, int get_user_pages_fast(unsigned long start, int nr_pages, int write, struct page **pages); +/* + * doesn't attempt to fault and will return short. + */ +int __get_user_pages_fast(unsigned long start, int nr_pages, int write, + struct page **pages); + /* * A callback you can register to apply pressure to ageable caches. * -- cgit v1.2.3 From 9974458e2f9a11dbd2f4bd14fab5a79af4907b41 Mon Sep 17 00:00:00 2001 From: Paul Mackerras Date: Mon, 15 Jun 2009 21:45:16 +1000 Subject: perf_counter: Make set_perf_counter_pending() declaration common At present, every architecture that supports perf_counters has to declare set_perf_counter_pending() in its arch-specific headers. This consolidates the declarations into a single declaration in one common place, include/linux/perf_counter.h. On powerpc, we continue to provide a static inline definition of set_perf_counter_pending() in the powerpc hw_irq.h. Also, this removes from the x86 perf_counter.h the unused null definitions of {test,clear}_perf_counter_pending. Reported-by: Mike Frysinger Signed-off-by: Paul Mackerras Cc: Peter Zijlstra Cc: benh@kernel.crashing.org LKML-Reference: <18998.13388.920691.523227@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar --- arch/powerpc/include/asm/hw_irq.h | 1 - arch/powerpc/include/asm/perf_counter.h | 2 ++ arch/x86/include/asm/perf_counter.h | 5 ----- include/linux/perf_counter.h | 1 + 4 files changed, 3 insertions(+), 6 deletions(-) (limited to 'include') diff --git a/arch/powerpc/include/asm/hw_irq.h b/arch/powerpc/include/asm/hw_irq.h index 53512374e1c9..1974cf191b03 100644 --- a/arch/powerpc/include/asm/hw_irq.h +++ b/arch/powerpc/include/asm/hw_irq.h @@ -163,7 +163,6 @@ static inline unsigned long test_perf_counter_pending(void) return 0; } -static inline void set_perf_counter_pending(void) {} static inline void clear_perf_counter_pending(void) {} #endif /* CONFIG_PERF_COUNTERS */ diff --git a/arch/powerpc/include/asm/perf_counter.h b/arch/powerpc/include/asm/perf_counter.h index cc7c887705b8..b398a84edced 100644 --- a/arch/powerpc/include/asm/perf_counter.h +++ b/arch/powerpc/include/asm/perf_counter.h @@ -10,6 +10,8 @@ */ #include +#include + #define MAX_HWCOUNTERS 8 #define MAX_EVENT_ALTERNATIVES 8 #define MAX_LIMITED_HWCOUNTERS 2 diff --git a/arch/x86/include/asm/perf_counter.h b/arch/x86/include/asm/perf_counter.h index 876ed97147b3..5fb33e160ea0 100644 --- a/arch/x86/include/asm/perf_counter.h +++ b/arch/x86/include/asm/perf_counter.h @@ -84,11 +84,6 @@ union cpuid10_edx { #define MSR_ARCH_PERFMON_FIXED_CTR2 0x30b #define X86_PMC_IDX_FIXED_BUS_CYCLES (X86_PMC_IDX_FIXED + 2) -extern void set_perf_counter_pending(void); - -#define clear_perf_counter_pending() do { } while (0) -#define test_perf_counter_pending() (0) - #ifdef CONFIG_PERF_COUNTERS extern void init_hw_perf_counters(void); extern void perf_counters_lapic_init(void); diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index 1b3118a1023a..eccae437fe37 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -604,6 +604,7 @@ extern void perf_counter_task_tick(struct task_struct *task, int cpu); extern int perf_counter_init_task(struct task_struct *child); extern void perf_counter_exit_task(struct task_struct *child); extern void perf_counter_free_task(struct task_struct *task); +extern void set_perf_counter_pending(void); extern void perf_counter_do_pending(void); extern void perf_counter_print_debug(void); extern void __perf_disable(void); -- cgit v1.2.3 From 7522060c95395f479ee4a6af3bbf9e097e92e48f Mon Sep 17 00:00:00 2001 From: Ingo Molnar Date: Thu, 18 Jun 2009 08:00:17 +0200 Subject: perf report: Add validation of call-chain entries Add boundary checks for call-chain events. In case of corrupted entries we could crash otherwise. Cc: Peter Zijlstra Cc: Mike Galbraith Cc: Paul Mackerras Cc: Arnaldo Carvalho de Melo LKML-Reference: Signed-off-by: Ingo Molnar --- include/linux/perf_counter.h | 20 ++++++------ tools/perf/builtin-report.c | 74 +++++++++++++++++++++++++++----------------- 2 files changed, 56 insertions(+), 38 deletions(-) (limited to 'include') diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index eccae437fe37..a7d3a61a59b7 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -337,6 +337,16 @@ enum perf_event_type { */ }; +#define MAX_STACK_DEPTH 255 + +struct perf_callchain_entry { + __u16 nr; + __u16 hv; + __u16 kernel; + __u16 user; + __u64 ip[MAX_STACK_DEPTH]; +}; + #ifdef __KERNEL__ /* * Kernel-internal data types and definitions: @@ -652,16 +662,6 @@ extern void perf_counter_fork(struct task_struct *tsk); extern void perf_counter_task_migration(struct task_struct *task, int cpu); -#define MAX_STACK_DEPTH 255 - -struct perf_callchain_entry { - u16 nr; - u16 hv; - u16 kernel; - u16 user; - u64 ip[MAX_STACK_DEPTH]; -}; - extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs); extern int sysctl_perf_counter_paranoid; diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 986834623b43..e14e98676171 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -39,6 +39,8 @@ static int dump_trace = 0; #define cdprintf(x...) do { if (dump_trace) color_fprintf(stdout, color, x); } while (0) static int verbose; +#define eprintf(x...) do { if (verbose) fprintf(stderr, x); } while (0) + static int full_paths; static unsigned long page_size; @@ -47,14 +49,6 @@ static unsigned long mmap_window = 32; static char *parent_pattern = "^sys_|^do_page_fault"; static regex_t parent_regex; -struct ip_chain_event { - __u16 nr; - __u16 hv; - __u16 kernel; - __u16 user; - __u64 ips[]; -}; - struct ip_event { struct perf_event_header header; __u64 ip; @@ -131,15 +125,11 @@ static struct dso *dsos__findnew(const char *name) nr = dso__load(dso, NULL, verbose); if (nr < 0) { - if (verbose) - fprintf(stderr, "Failed to open: %s\n", name); + eprintf("Failed to open: %s\n", name); goto out_delete_dso; } - if (!nr && verbose) { - fprintf(stderr, - "No symbols found in: %s, maybe install a debug package?\n", - name); - } + if (!nr) + eprintf("No symbols found in: %s, maybe install a debug package?\n", name); dsos__add(dso); @@ -844,7 +834,7 @@ static struct symbol *call__match(struct symbol *sym) static int hist_entry__add(struct thread *thread, struct map *map, struct dso *dso, - struct symbol *sym, __u64 ip, struct ip_chain_event *chain, + struct symbol *sym, __u64 ip, struct perf_callchain_entry *chain, char level, __u64 count) { struct rb_node **p = &hist.rb_node; @@ -868,7 +858,7 @@ hist_entry__add(struct thread *thread, struct map *map, struct dso *dso, __u64 ip; for (i = 0; i < chain->kernel; i++) { - ip = chain->ips[nr + i]; + ip = chain->ip[nr + i]; dso = kernel_dso; sym = resolve_symbol(thread, NULL, &dso, &ip); entry.parent = call__match(sym); @@ -878,7 +868,7 @@ hist_entry__add(struct thread *thread, struct map *map, struct dso *dso, nr += i; for (i = 0; i < chain->user; i++) { - ip = chain->ips[nr + i]; + ip = chain->ip[nr + i]; sym = resolve_symbol(thread, NULL, NULL, &ip); entry.parent = call__match(sym); if (entry.parent) @@ -1080,6 +1070,30 @@ static unsigned long total = 0, total_fork = 0, total_unknown = 0; +static int validate_chain(struct perf_callchain_entry *chain, event_t *event) +{ + unsigned int chain_size; + + if (chain->nr > MAX_STACK_DEPTH) + return -1; + if (chain->hv > MAX_STACK_DEPTH) + return -1; + if (chain->kernel > MAX_STACK_DEPTH) + return -1; + if (chain->user > MAX_STACK_DEPTH) + return -1; + if (chain->hv + chain->kernel + chain->user != chain->nr) + return -1; + + chain_size = event->header.size; + chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event; + + if (chain->nr*sizeof(__u64) > chain_size) + return -1; + + return 0; +} + static int process_overflow_event(event_t *event, unsigned long offset, unsigned long head) { @@ -1091,7 +1105,7 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head) __u64 period = 1; struct map *map = NULL; void *more_data = event->ip.__more_data; - struct ip_chain_event *chain = NULL; + struct perf_callchain_entry *chain = NULL; if (event->header.type & PERF_SAMPLE_PERIOD) { period = *(__u64 *)more_data; @@ -1111,21 +1125,26 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head) chain = (void *)more_data; - if (dump_trace) { - dprintf("... chain: u:%d, k:%d, nr:%d\n", - chain->user, - chain->kernel, - chain->nr); + dprintf("... chain: u:%d, k:%d, nr:%d\n", + chain->user, + chain->kernel, + chain->nr); + if (validate_chain(chain, event) < 0) { + eprintf("call-chain problem with event, skipping it.\n"); + return 0; + } + + if (dump_trace) { for (i = 0; i < chain->nr; i++) - dprintf("..... %2d: %016Lx\n", i, chain->ips[i]); + dprintf("..... %2d: %016Lx\n", i, chain->ip[i]); } } dprintf(" ... thread: %s:%d\n", thread->comm, thread->pid); if (thread == NULL) { - fprintf(stderr, "problem processing %d event, skipping it.\n", + eprintf("problem processing %d event, skipping it.\n", event->header.type); return -1; } @@ -1153,8 +1172,7 @@ process_overflow_event(event_t *event, unsigned long offset, unsigned long head) struct symbol *sym = resolve_symbol(thread, &map, &dso, &ip); if (hist_entry__add(thread, map, dso, sym, ip, chain, level, period)) { - fprintf(stderr, - "problem incrementing symbol count, skipping event\n"); + eprintf("problem incrementing symbol count, skipping event\n"); return -1; } } -- cgit v1.2.3 From 43a21ea81a2400992561146327c4785ce7f7be38 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Wed, 25 Mar 2009 19:39:37 +0100 Subject: perf_counter: Add event overlow handling Alternative method of mmap() data output handling that provides better overflow management and a more reliable data stream. Unlike the previous method, that didn't have any user->kernel feedback and relied on userspace keeping up, this method relies on userspace writing its last read position into the control page. It will ensure new output doesn't overwrite not-yet read events, new events for which there is no space left are lost and the overflow counter is incremented, providing exact event loss numbers. Signed-off-by: Peter Zijlstra Cc: Mike Galbraith Cc: Paul Mackerras Cc: Arnaldo Carvalho de Melo LKML-Reference: Signed-off-by: Ingo Molnar --- include/linux/perf_counter.h | 40 +++++++--- kernel/perf_counter.c | 185 ++++++++++++++++++++++++++++++------------- 2 files changed, 158 insertions(+), 67 deletions(-) (limited to 'include') diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index a7d3a61a59b7..0765e8e69843 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -236,10 +236,16 @@ struct perf_counter_mmap_page { /* * Control data for the mmap() data buffer. * - * User-space reading this value should issue an rmb(), on SMP capable - * platforms, after reading this value -- see perf_counter_wakeup(). + * User-space reading the @data_head value should issue an rmb(), on + * SMP capable platforms, after reading this value -- see + * perf_counter_wakeup(). + * + * When the mapping is PROT_WRITE the @data_tail value should be + * written by userspace to reflect the last read data. In this case + * the kernel will not over-write unread data. */ __u64 data_head; /* head in the data section */ + __u64 data_tail; /* user-space written tail */ }; #define PERF_EVENT_MISC_CPUMODE_MASK (3 << 0) @@ -273,6 +279,15 @@ enum perf_event_type { */ PERF_EVENT_MMAP = 1, + /* + * struct { + * struct perf_event_header header; + * u64 id; + * u64 lost; + * }; + */ + PERF_EVENT_LOST = 2, + /* * struct { * struct perf_event_header header; @@ -313,26 +328,26 @@ enum perf_event_type { /* * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field - * will be PERF_RECORD_* + * will be PERF_SAMPLE_* * * struct { * struct perf_event_header header; * - * { u64 ip; } && PERF_RECORD_IP - * { u32 pid, tid; } && PERF_RECORD_TID - * { u64 time; } && PERF_RECORD_TIME - * { u64 addr; } && PERF_RECORD_ADDR - * { u64 config; } && PERF_RECORD_CONFIG - * { u32 cpu, res; } && PERF_RECORD_CPU + * { u64 ip; } && PERF_SAMPLE_IP + * { u32 pid, tid; } && PERF_SAMPLE_TID + * { u64 time; } && PERF_SAMPLE_TIME + * { u64 addr; } && PERF_SAMPLE_ADDR + * { u64 config; } && PERF_SAMPLE_CONFIG + * { u32 cpu, res; } && PERF_SAMPLE_CPU * * { u64 nr; - * { u64 id, val; } cnt[nr]; } && PERF_RECORD_GROUP + * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP * * { u16 nr, * hv, * kernel, * user; - * u64 ips[nr]; } && PERF_RECORD_CALLCHAIN + * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN * }; */ }; @@ -424,6 +439,7 @@ struct file; struct perf_mmap_data { struct rcu_head rcu_head; int nr_pages; /* nr of data pages */ + int writable; /* are we writable */ int nr_locked; /* nr pages mlocked */ atomic_t poll; /* POLL_ for wakeups */ @@ -433,8 +449,8 @@ struct perf_mmap_data { atomic_long_t done_head; /* completed head */ atomic_t lock; /* concurrent writes */ - atomic_t wakeup; /* needs a wakeup */ + atomic_t lost; /* nr records lost */ struct perf_counter_mmap_page *user_page; void *data_pages[0]; diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index 109a95723859..7e9108efd305 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -1794,6 +1794,12 @@ static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) struct perf_mmap_data *data; int ret = VM_FAULT_SIGBUS; + if (vmf->flags & FAULT_FLAG_MKWRITE) { + if (vmf->pgoff == 0) + ret = 0; + return ret; + } + rcu_read_lock(); data = rcu_dereference(counter->data); if (!data) @@ -1807,9 +1813,16 @@ static int perf_mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf) if ((unsigned)nr > data->nr_pages) goto unlock; + if (vmf->flags & FAULT_FLAG_WRITE) + goto unlock; + vmf->page = virt_to_page(data->data_pages[nr]); } + get_page(vmf->page); + vmf->page->mapping = vma->vm_file->f_mapping; + vmf->page->index = vmf->pgoff; + ret = 0; unlock: rcu_read_unlock(); @@ -1862,6 +1875,14 @@ fail: return -ENOMEM; } +static void perf_mmap_free_page(unsigned long addr) +{ + struct page *page = virt_to_page(addr); + + page->mapping = NULL; + __free_page(page); +} + static void __perf_mmap_data_free(struct rcu_head *rcu_head) { struct perf_mmap_data *data; @@ -1869,9 +1890,10 @@ static void __perf_mmap_data_free(struct rcu_head *rcu_head) data = container_of(rcu_head, struct perf_mmap_data, rcu_head); - free_page((unsigned long)data->user_page); + perf_mmap_free_page((unsigned long)data->user_page); for (i = 0; i < data->nr_pages; i++) - free_page((unsigned long)data->data_pages[i]); + perf_mmap_free_page((unsigned long)data->data_pages[i]); + kfree(data); } @@ -1908,9 +1930,10 @@ static void perf_mmap_close(struct vm_area_struct *vma) } static struct vm_operations_struct perf_mmap_vmops = { - .open = perf_mmap_open, - .close = perf_mmap_close, - .fault = perf_mmap_fault, + .open = perf_mmap_open, + .close = perf_mmap_close, + .fault = perf_mmap_fault, + .page_mkwrite = perf_mmap_fault, }; static int perf_mmap(struct file *file, struct vm_area_struct *vma) @@ -1924,7 +1947,7 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma) long user_extra, extra; int ret = 0; - if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE)) + if (!(vma->vm_flags & VM_SHARED)) return -EINVAL; vma_size = vma->vm_end - vma->vm_start; @@ -1983,10 +2006,12 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma) atomic_long_add(user_extra, &user->locked_vm); vma->vm_mm->locked_vm += extra; counter->data->nr_locked = extra; + if (vma->vm_flags & VM_WRITE) + counter->data->writable = 1; + unlock: mutex_unlock(&counter->mmap_mutex); - vma->vm_flags &= ~VM_MAYWRITE; vma->vm_flags |= VM_RESERVED; vma->vm_ops = &perf_mmap_vmops; @@ -2163,11 +2188,38 @@ struct perf_output_handle { unsigned long head; unsigned long offset; int nmi; - int overflow; + int sample; int locked; unsigned long flags; }; +static bool perf_output_space(struct perf_mmap_data *data, + unsigned int offset, unsigned int head) +{ + unsigned long tail; + unsigned long mask; + + if (!data->writable) + return true; + + mask = (data->nr_pages << PAGE_SHIFT) - 1; + /* + * Userspace could choose to issue a mb() before updating the tail + * pointer. So that all reads will be completed before the write is + * issued. + */ + tail = ACCESS_ONCE(data->user_page->data_tail); + smp_rmb(); + + offset = (offset - tail) & mask; + head = (head - tail) & mask; + + if ((int)(head - offset) < 0) + return false; + + return true; +} + static void perf_output_wakeup(struct perf_output_handle *handle) { atomic_set(&handle->data->poll, POLL_IN); @@ -2258,12 +2310,57 @@ out: local_irq_restore(handle->flags); } +static void perf_output_copy(struct perf_output_handle *handle, + const void *buf, unsigned int len) +{ + unsigned int pages_mask; + unsigned int offset; + unsigned int size; + void **pages; + + offset = handle->offset; + pages_mask = handle->data->nr_pages - 1; + pages = handle->data->data_pages; + + do { + unsigned int page_offset; + int nr; + + nr = (offset >> PAGE_SHIFT) & pages_mask; + page_offset = offset & (PAGE_SIZE - 1); + size = min_t(unsigned int, PAGE_SIZE - page_offset, len); + + memcpy(pages[nr] + page_offset, buf, size); + + len -= size; + buf += size; + offset += size; + } while (len); + + handle->offset = offset; + + /* + * Check we didn't copy past our reservation window, taking the + * possible unsigned int wrap into account. + */ + WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0); +} + +#define perf_output_put(handle, x) \ + perf_output_copy((handle), &(x), sizeof(x)) + static int perf_output_begin(struct perf_output_handle *handle, struct perf_counter *counter, unsigned int size, - int nmi, int overflow) + int nmi, int sample) { struct perf_mmap_data *data; unsigned int offset, head; + int have_lost; + struct { + struct perf_event_header header; + u64 id; + u64 lost; + } lost_event; /* * For inherited counters we send all the output towards the parent. @@ -2276,19 +2373,25 @@ static int perf_output_begin(struct perf_output_handle *handle, if (!data) goto out; - handle->data = data; - handle->counter = counter; - handle->nmi = nmi; - handle->overflow = overflow; + handle->data = data; + handle->counter = counter; + handle->nmi = nmi; + handle->sample = sample; if (!data->nr_pages) goto fail; + have_lost = atomic_read(&data->lost); + if (have_lost) + size += sizeof(lost_event); + perf_output_lock(handle); do { offset = head = atomic_long_read(&data->head); head += size; + if (unlikely(!perf_output_space(data, offset, head))) + goto fail; } while (atomic_long_cmpxchg(&data->head, offset, head) != offset); handle->offset = offset; @@ -2297,55 +2400,27 @@ static int perf_output_begin(struct perf_output_handle *handle, if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT)) atomic_set(&data->wakeup, 1); + if (have_lost) { + lost_event.header.type = PERF_EVENT_LOST; + lost_event.header.misc = 0; + lost_event.header.size = sizeof(lost_event); + lost_event.id = counter->id; + lost_event.lost = atomic_xchg(&data->lost, 0); + + perf_output_put(handle, lost_event); + } + return 0; fail: - perf_output_wakeup(handle); + atomic_inc(&data->lost); + perf_output_unlock(handle); out: rcu_read_unlock(); return -ENOSPC; } -static void perf_output_copy(struct perf_output_handle *handle, - const void *buf, unsigned int len) -{ - unsigned int pages_mask; - unsigned int offset; - unsigned int size; - void **pages; - - offset = handle->offset; - pages_mask = handle->data->nr_pages - 1; - pages = handle->data->data_pages; - - do { - unsigned int page_offset; - int nr; - - nr = (offset >> PAGE_SHIFT) & pages_mask; - page_offset = offset & (PAGE_SIZE - 1); - size = min_t(unsigned int, PAGE_SIZE - page_offset, len); - - memcpy(pages[nr] + page_offset, buf, size); - - len -= size; - buf += size; - offset += size; - } while (len); - - handle->offset = offset; - - /* - * Check we didn't copy past our reservation window, taking the - * possible unsigned int wrap into account. - */ - WARN_ON_ONCE(((long)(handle->head - handle->offset)) < 0); -} - -#define perf_output_put(handle, x) \ - perf_output_copy((handle), &(x), sizeof(x)) - static void perf_output_end(struct perf_output_handle *handle) { struct perf_counter *counter = handle->counter; @@ -2353,7 +2428,7 @@ static void perf_output_end(struct perf_output_handle *handle) int wakeup_events = counter->attr.wakeup_events; - if (handle->overflow && wakeup_events) { + if (handle->sample && wakeup_events) { int events = atomic_inc_return(&data->events); if (events >= wakeup_events) { atomic_sub(wakeup_events, &data->events); @@ -2958,7 +3033,7 @@ static void perf_log_throttle(struct perf_counter *counter, int enable) } /* - * Generic counter overflow handling. + * Generic counter overflow handling, sampling. */ int perf_counter_overflow(struct perf_counter *counter, int nmi, -- cgit v1.2.3 From f9188e023c248d73f5b4a589b480e065c1864068 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 18 Jun 2009 22:20:52 +0200 Subject: perf_counter: Make callchain samples extensible Before exposing upstream tools to a callchain-samples ABI, tidy it up to make it more extensible in the future: Use markers in the IP chain to denote context, use (u64)-1..-4095 range for these context markers because we use them for ERR_PTR(), so these addresses are unlikely to be mapped. Signed-off-by: Peter Zijlstra Cc: Mike Galbraith Cc: Paul Mackerras Cc: Arnaldo Carvalho de Melo LKML-Reference: Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_counter.c | 29 ++++++----------------------- include/linux/perf_counter.h | 28 +++++++++++++++++----------- 2 files changed, 23 insertions(+), 34 deletions(-) (limited to 'include') diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c index ce1ae3f1f86c..76dfef23f789 100644 --- a/arch/x86/kernel/cpu/perf_counter.c +++ b/arch/x86/kernel/cpu/perf_counter.c @@ -1555,9 +1555,9 @@ const struct pmu *hw_perf_counter_init(struct perf_counter *counter) */ static inline -void callchain_store(struct perf_callchain_entry *entry, unsigned long ip) +void callchain_store(struct perf_callchain_entry *entry, u64 ip) { - if (entry->nr < MAX_STACK_DEPTH) + if (entry->nr < PERF_MAX_STACK_DEPTH) entry->ip[entry->nr++] = ip; } @@ -1602,22 +1602,10 @@ static const struct stacktrace_ops backtrace_ops = { static void perf_callchain_kernel(struct pt_regs *regs, struct perf_callchain_entry *entry) { - unsigned long bp; - char *stack; - int nr = entry->nr; - + callchain_store(entry, PERF_CONTEXT_KERNEL); callchain_store(entry, regs->ip); - stack = ((char *)regs + sizeof(struct pt_regs)); -#ifdef CONFIG_FRAME_POINTER - get_bp(bp); -#else - bp = 0; -#endif - - dump_trace(NULL, regs, (void *)&stack, bp, &backtrace_ops, entry); - - entry->kernel = entry->nr - nr; + dump_trace(NULL, regs, NULL, 0, &backtrace_ops, entry); } /* @@ -1669,16 +1657,16 @@ perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry) { struct stack_frame frame; const void __user *fp; - int nr = entry->nr; if (!user_mode(regs)) regs = task_pt_regs(current); fp = (void __user *)regs->bp; + callchain_store(entry, PERF_CONTEXT_USER); callchain_store(entry, regs->ip); - while (entry->nr < MAX_STACK_DEPTH) { + while (entry->nr < PERF_MAX_STACK_DEPTH) { frame.next_frame = NULL; frame.return_address = 0; @@ -1691,8 +1679,6 @@ perf_callchain_user(struct pt_regs *regs, struct perf_callchain_entry *entry) callchain_store(entry, frame.return_address); fp = frame.next_frame; } - - entry->user = entry->nr - nr; } static void @@ -1728,9 +1714,6 @@ struct perf_callchain_entry *perf_callchain(struct pt_regs *regs) entry = &__get_cpu_var(irq_entry); entry->nr = 0; - entry->hv = 0; - entry->kernel = 0; - entry->user = 0; perf_do_callchain(regs, entry); diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index 0765e8e69843..e7e7e0242767 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -343,23 +343,22 @@ enum perf_event_type { * { u64 nr; * { u64 id, val; } cnt[nr]; } && PERF_SAMPLE_GROUP * - * { u16 nr, - * hv, - * kernel, - * user; + * { u64 nr, * u64 ips[nr]; } && PERF_SAMPLE_CALLCHAIN * }; */ }; -#define MAX_STACK_DEPTH 255 +enum perf_callchain_context { + PERF_CONTEXT_HV = (__u64)-32, + PERF_CONTEXT_KERNEL = (__u64)-128, + PERF_CONTEXT_USER = (__u64)-512, -struct perf_callchain_entry { - __u16 nr; - __u16 hv; - __u16 kernel; - __u16 user; - __u64 ip[MAX_STACK_DEPTH]; + PERF_CONTEXT_GUEST = (__u64)-2048, + PERF_CONTEXT_GUEST_KERNEL = (__u64)-2176, + PERF_CONTEXT_GUEST_USER = (__u64)-2560, + + PERF_CONTEXT_MAX = (__u64)-4095, }; #ifdef __KERNEL__ @@ -381,6 +380,13 @@ struct perf_callchain_entry { #include #include +#define PERF_MAX_STACK_DEPTH 255 + +struct perf_callchain_entry { + __u64 nr; + __u64 ip[PERF_MAX_STACK_DEPTH]; +}; + struct task_struct; /** -- cgit v1.2.3 From e5289d4a181fb6c0b7a7607649af2ffdc491335c Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Fri, 19 Jun 2009 13:22:51 +0200 Subject: perf_counter: Simplify and fix task migration counting The task migrations counter was causing rare and hard to decypher memory corruptions under load. After a day of debugging and bisection we found that the problem was introduced with: 3f731ca: perf_counter: Fix cpu migration counter Turning them off fixes the crashes. Incidentally, the whole perf_counter_task_migration() logic can be done simpler as well, by injecting a proper sw-counter event. This cleanup also fixed the crashes. The precise failure mode is not completely clear yet, but we are clearly not unhappy about having a fix ;-) Signed-off-by: Peter Zijlstra Cc: Mike Galbraith Cc: Paul Mackerras Cc: Corey Ashford Cc: Marcelo Tosatti Cc: Arnaldo Carvalho de Melo LKML-Reference: Signed-off-by: Ingo Molnar --- include/linux/perf_counter.h | 4 ---- kernel/perf_counter.c | 23 +---------------------- kernel/sched.c | 3 ++- 3 files changed, 3 insertions(+), 27 deletions(-) (limited to 'include') diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h index e7e7e0242767..89698d8aba5c 100644 --- a/include/linux/perf_counter.h +++ b/include/linux/perf_counter.h @@ -682,8 +682,6 @@ static inline void perf_counter_mmap(struct vm_area_struct *vma) extern void perf_counter_comm(struct task_struct *tsk); extern void perf_counter_fork(struct task_struct *tsk); -extern void perf_counter_task_migration(struct task_struct *task, int cpu); - extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs); extern int sysctl_perf_counter_paranoid; @@ -724,8 +722,6 @@ static inline void perf_counter_mmap(struct vm_area_struct *vma) { } static inline void perf_counter_comm(struct task_struct *tsk) { } static inline void perf_counter_fork(struct task_struct *tsk) { } static inline void perf_counter_init(void) { } -static inline void perf_counter_task_migration(struct task_struct *task, - int cpu) { } #endif #endif /* __KERNEL__ */ diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c index 7e9108efd305..8d4f0dd41c22 100644 --- a/kernel/perf_counter.c +++ b/kernel/perf_counter.c @@ -124,7 +124,7 @@ void perf_enable(void) static void get_ctx(struct perf_counter_context *ctx) { - atomic_inc(&ctx->refcount); + WARN_ON(!atomic_inc_not_zero(&ctx->refcount)); } static void free_ctx(struct rcu_head *head) @@ -3467,27 +3467,6 @@ static const struct pmu perf_ops_task_clock = { .read = task_clock_perf_counter_read, }; -/* - * Software counter: cpu migrations - */ -void perf_counter_task_migration(struct task_struct *task, int cpu) -{ - struct perf_cpu_context *cpuctx = &per_cpu(perf_cpu_context, cpu); - struct perf_counter_context *ctx; - - perf_swcounter_ctx_event(&cpuctx->ctx, PERF_TYPE_SOFTWARE, - PERF_COUNT_SW_CPU_MIGRATIONS, - 1, 1, NULL, 0); - - ctx = perf_pin_task_context(task); - if (ctx) { - perf_swcounter_ctx_event(ctx, PERF_TYPE_SOFTWARE, - PERF_COUNT_SW_CPU_MIGRATIONS, - 1, 1, NULL, 0); - perf_unpin_context(ctx); - } -} - #ifdef CONFIG_EVENT_PROFILE void perf_tpcounter_event(int event_id) { diff --git a/kernel/sched.c b/kernel/sched.c index 8fb88a906aaa..f46540b359c0 100644 --- a/kernel/sched.c +++ b/kernel/sched.c @@ -1978,7 +1978,8 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu) if (task_hot(p, old_rq->clock, NULL)) schedstat_inc(p, se.nr_forced2_migrations); #endif - perf_counter_task_migration(p, new_cpu); + perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS, + 1, 1, NULL, 0); } p->se.vruntime -= old_cfsrq->min_vruntime - new_cfsrq->min_vruntime; -- cgit v1.2.3