From bc44fb5f7d3e764ed7698c835a1a0f35aba2eb3d Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 13 Mar 2009 10:42:18 +0100 Subject: x86, bts: detect size of DS fields Impact: more robust DS feature enumeration Detect the size of the pointer-type fields in the DS area configuration via the DTES64 features rather than based on the cpuid. Rename a variable to denote that size to reflect that it only covers the pointer-type fields. Add more boot-time diagnostics giving the detected size and the sizes of BTS and PEBS records. Use the size of the BTS/PEBS record to indicate that the respective feature is not available (if the record size is zero). Signed-off-by: Markus Metzger LKML-Reference: <20090313104218.A30096@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 84 +++++++++++++++++++++++++++------------------------- 1 file changed, 44 insertions(+), 40 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 87b67e3a765a..6e5ec679a0cd 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -39,7 +39,7 @@ struct ds_configuration { /* the size of one pointer-typed field in the DS structure and in the BTS and PEBS buffers in bytes; this covers the first 8 DS fields related to buffer management. */ - unsigned char sizeof_field; + unsigned char sizeof_ptr_field; /* the size of a BTS/PEBS record in bytes */ unsigned char sizeof_rec[2]; /* a series of bit-masks to control various features indexed @@ -142,14 +142,14 @@ enum ds_qualifier { static inline unsigned long ds_get(const unsigned char *base, enum ds_qualifier qual, enum ds_field field) { - base += (ds_cfg.sizeof_field * (field + (4 * qual))); + base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual))); return *(unsigned long *)base; } static inline void ds_set(unsigned char *base, enum ds_qualifier qual, enum ds_field field, unsigned long value) { - base += (ds_cfg.sizeof_field * (field + (4 * qual))); + base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual))); (*(unsigned long *)base) = value; } @@ -410,7 +410,7 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual, * Later architectures use 64bit pointers throughout, whereas earlier * architectures use 32bit pointers in 32bit mode. * - * We compute the base address for the first 8 fields based on: + * We compute the base address for the fields based on: * - the field size stored in the DS configuration * - the relative field position * @@ -441,13 +441,13 @@ enum bts_field { static inline unsigned long bts_get(const char *base, enum bts_field field) { - base += (ds_cfg.sizeof_field * field); + base += (ds_cfg.sizeof_ptr_field * field); return *(unsigned long *)base; } static inline void bts_set(char *base, enum bts_field field, unsigned long val) { - base += (ds_cfg.sizeof_field * field);; + base += (ds_cfg.sizeof_ptr_field * field);; (*(unsigned long *)base) = val; } @@ -593,6 +593,10 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, struct ds_context *context; int error; + error = -EOPNOTSUPP; + if (!ds_cfg.sizeof_rec[qual]) + goto out; + error = -EINVAL; if (!base) goto out; @@ -635,10 +639,6 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, unsigned long irq; int error; - error = -EOPNOTSUPP; - if (!ds_cfg.ctl[dsf_bts]) - goto out; - /* buffer overflow notification is not yet implemented */ error = -EOPNOTSUPP; if (ovfl) @@ -848,7 +848,8 @@ const struct pebs_trace *ds_read_pebs(struct pebs_tracer *tracer) ds_read_config(tracer->ds.context, &tracer->trace.ds, ds_pebs); tracer->trace.reset_value = - *(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8)); + *(u64 *)(tracer->ds.context->ds + + (ds_cfg.sizeof_ptr_field * 8)); return &tracer->trace; } @@ -884,7 +885,8 @@ int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value) if (!tracer) return -EINVAL; - *(u64 *)(tracer->ds.context->ds + (ds_cfg.sizeof_field * 8)) = value; + *(u64 *)(tracer->ds.context->ds + + (ds_cfg.sizeof_ptr_field * 8)) = value; return 0; } @@ -894,52 +896,54 @@ static const struct ds_configuration ds_cfg_netburst = { .ctl[dsf_bts] = (1 << 2) | (1 << 3), .ctl[dsf_bts_kernel] = (1 << 5), .ctl[dsf_bts_user] = (1 << 6), - - .sizeof_field = sizeof(long), - .sizeof_rec[ds_bts] = sizeof(long) * 3, -#ifdef __i386__ - .sizeof_rec[ds_pebs] = sizeof(long) * 10, -#else - .sizeof_rec[ds_pebs] = sizeof(long) * 18, -#endif }; static const struct ds_configuration ds_cfg_pentium_m = { .name = "Pentium M", .ctl[dsf_bts] = (1 << 6) | (1 << 7), - - .sizeof_field = sizeof(long), - .sizeof_rec[ds_bts] = sizeof(long) * 3, -#ifdef __i386__ - .sizeof_rec[ds_pebs] = sizeof(long) * 10, -#else - .sizeof_rec[ds_pebs] = sizeof(long) * 18, -#endif }; static const struct ds_configuration ds_cfg_core2_atom = { .name = "Core 2/Atom", .ctl[dsf_bts] = (1 << 6) | (1 << 7), .ctl[dsf_bts_kernel] = (1 << 9), .ctl[dsf_bts_user] = (1 << 10), - - .sizeof_field = 8, - .sizeof_rec[ds_bts] = 8 * 3, - .sizeof_rec[ds_pebs] = 8 * 18, }; static void -ds_configure(const struct ds_configuration *cfg) +ds_configure(const struct ds_configuration *cfg, + struct cpuinfo_x86 *cpu) { + unsigned long nr_pebs_fields = 0; + + printk(KERN_INFO "[ds] using %s configuration\n", cfg->name); + +#ifdef __i386__ + nr_pebs_fields = 10; +#else + nr_pebs_fields = 18; +#endif + memset(&ds_cfg, 0, sizeof(ds_cfg)); ds_cfg = *cfg; - printk(KERN_INFO "[ds] using %s configuration\n", ds_cfg.name); + ds_cfg.sizeof_ptr_field = + (cpu_has(cpu, X86_FEATURE_DTES64) ? 8 : 4); - if (!cpu_has_bts) { - ds_cfg.ctl[dsf_bts] = 0; + ds_cfg.sizeof_rec[ds_bts] = ds_cfg.sizeof_ptr_field * 3; + ds_cfg.sizeof_rec[ds_pebs] = ds_cfg.sizeof_ptr_field * nr_pebs_fields; + + if (!cpu_has(cpu, X86_FEATURE_BTS)) { + ds_cfg.sizeof_rec[ds_bts] = 0; printk(KERN_INFO "[ds] bts not available\n"); } - if (!cpu_has_pebs) + if (!cpu_has(cpu, X86_FEATURE_PEBS)) { + ds_cfg.sizeof_rec[ds_pebs] = 0; printk(KERN_INFO "[ds] pebs not available\n"); + } + + printk(KERN_INFO "[ds] sizes: address: %u bit, ", + 8 * ds_cfg.sizeof_ptr_field); + printk("bts/pebs record: %u/%u bytes\n", + ds_cfg.sizeof_rec[ds_bts], ds_cfg.sizeof_rec[ds_pebs]); WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_field)); } @@ -951,12 +955,12 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) switch (c->x86_model) { case 0x9: case 0xd: /* Pentium M */ - ds_configure(&ds_cfg_pentium_m); + ds_configure(&ds_cfg_pentium_m, c); break; case 0xf: case 0x17: /* Core2 */ case 0x1c: /* Atom */ - ds_configure(&ds_cfg_core2_atom); + ds_configure(&ds_cfg_core2_atom, c); break; case 0x1a: /* i7 */ default: @@ -969,7 +973,7 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) case 0x0: case 0x1: case 0x2: /* Netburst */ - ds_configure(&ds_cfg_netburst); + ds_configure(&ds_cfg_netburst, c); break; default: /* sorry, don't know about them */ -- cgit v1.2.3 From 8a327f6d1b05f5ce16572b4413a5df1d0e872283 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 13 Mar 2009 10:45:07 +0100 Subject: x86, bts: add selftest for BTS Perform a selftest of branch trace store when a cpu is initialized. WARN and disable branch trace store support if the selftest fails. Signed-off-by: Markus Metzger LKML-Reference: <20090313104507.A30125@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/Kconfig.debug | 9 ++ arch/x86/kernel/Makefile | 1 + arch/x86/kernel/ds.c | 21 ++++ arch/x86/kernel/ds_selftest.c | 241 ++++++++++++++++++++++++++++++++++++++++++ arch/x86/kernel/ds_selftest.h | 15 +++ 5 files changed, 287 insertions(+) create mode 100644 arch/x86/kernel/ds_selftest.c create mode 100644 arch/x86/kernel/ds_selftest.h (limited to 'arch') diff --git a/arch/x86/Kconfig.debug b/arch/x86/Kconfig.debug index fdb45df608b6..dfd74abc03f8 100644 --- a/arch/x86/Kconfig.debug +++ b/arch/x86/Kconfig.debug @@ -175,6 +175,15 @@ config IOMMU_LEAK Add a simple leak tracer to the IOMMU code. This is useful when you are debugging a buggy device driver that leaks IOMMU mappings. +config X86_DS_SELFTEST + bool "DS selftest" + default y + depends on DEBUG_KERNEL + depends on X86_DS + ---help--- + Perform Debug Store selftests at boot time. + If in doubt, say "N". + config HAVE_MMIOTRACE_SUPPORT def_bool y diff --git a/arch/x86/kernel/Makefile b/arch/x86/kernel/Makefile index 339ce35648e6..a0c9e138b008 100644 --- a/arch/x86/kernel/Makefile +++ b/arch/x86/kernel/Makefile @@ -44,6 +44,7 @@ obj-y += process.o obj-y += i387.o xsave.o obj-y += ptrace.o obj-$(CONFIG_X86_DS) += ds.o +obj-$(CONFIG_X86_DS_SELFTEST) += ds_selftest.o obj-$(CONFIG_X86_32) += tls.o obj-$(CONFIG_IA32_EMULATION) += tls.o obj-y += step.o diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 6e5ec679a0cd..51c936c1a390 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -29,6 +29,7 @@ #include #include +#include "ds_selftest.h" /* * The configuration for a particular DS hardware implementation. @@ -940,6 +941,26 @@ ds_configure(const struct ds_configuration *cfg, printk(KERN_INFO "[ds] pebs not available\n"); } + if (ds_cfg.sizeof_rec[ds_bts]) { + int error; + + error = ds_selftest_bts(); + if (error) { + WARN(1, "[ds] selftest failed. disabling bts.\n"); + ds_cfg.sizeof_rec[ds_bts] = 0; + } + } + + if (ds_cfg.sizeof_rec[ds_pebs]) { + int error; + + error = ds_selftest_pebs(); + if (error) { + WARN(1, "[ds] selftest failed. disabling pebs.\n"); + ds_cfg.sizeof_rec[ds_pebs] = 0; + } + } + printk(KERN_INFO "[ds] sizes: address: %u bit, ", 8 * ds_cfg.sizeof_ptr_field); printk("bts/pebs record: %u/%u bytes\n", diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c new file mode 100644 index 000000000000..8c46fbf38c46 --- /dev/null +++ b/arch/x86/kernel/ds_selftest.c @@ -0,0 +1,241 @@ +/* + * Debug Store support - selftest + * + * + * Copyright (C) 2009 Intel Corporation. + * Markus Metzger , 2009 + */ + +#include "ds_selftest.h" + +#include +#include + +#include + + +#define DS_SELFTEST_BUFFER_SIZE 1021 /* Intentionally chose an odd size. */ + + +static int ds_selftest_bts_consistency(const struct bts_trace *trace) +{ + int error = 0; + + if (!trace) { + printk(KERN_CONT "failed to access trace..."); + /* Bail out. Other tests are pointless. */ + return -1; + } + + if (!trace->read) { + printk(KERN_CONT "bts read not available..."); + error = -1; + } + + /* Do some sanity checks on the trace configuration. */ + if (!trace->ds.n) { + printk(KERN_CONT "empty bts buffer..."); + error = -1; + } + if (!trace->ds.size) { + printk(KERN_CONT "bad bts trace setup..."); + error = -1; + } + if (trace->ds.end != + (char *)trace->ds.begin + (trace->ds.n * trace->ds.size)) { + printk(KERN_CONT "bad bts buffer setup..."); + error = -1; + } + if ((trace->ds.top < trace->ds.begin) || + (trace->ds.end <= trace->ds.top)) { + printk(KERN_CONT "bts top out of bounds..."); + error = -1; + } + + return error; +} + +static int ds_selftest_bts_read(struct bts_tracer *tracer, + const struct bts_trace *trace, + const void *from, const void *to) +{ + const unsigned char *at; + + /* + * Check a few things which do not belong to this test. + * They should be covered by other tests. + */ + if (!trace) + return -1; + + if (!trace->read) + return -1; + + if (to < from) + return -1; + + if (from < trace->ds.begin) + return -1; + + if (trace->ds.end < to) + return -1; + + if (!trace->ds.size) + return -1; + + /* Now to the test itself. */ + for (at = from; (void *)at < to; at += trace->ds.size) { + struct bts_struct bts; + size_t index; + int error; + + if (((void *)at - trace->ds.begin) % trace->ds.size) { + printk(KERN_CONT + "read from non-integer index..."); + return -1; + } + index = ((void *)at - trace->ds.begin) / trace->ds.size; + + memset(&bts, 0, sizeof(bts)); + error = trace->read(tracer, at, &bts); + if (error < 0) { + printk(KERN_CONT + "error reading bts trace at [%lu] (0x%p)...", + index, at); + return error; + } + + switch (bts.qualifier) { + case BTS_BRANCH: + break; + default: + printk(KERN_CONT + "unexpected bts entry %llu at [%lu] (0x%p)...", + bts.qualifier, index, at); + return -1; + } + } + + return 0; +} + +int ds_selftest_bts(void) +{ + const struct bts_trace *trace; + struct bts_tracer *tracer; + int error = 0; + void *top; + unsigned char buffer[DS_SELFTEST_BUFFER_SIZE]; + + printk(KERN_INFO "[ds] bts selftest..."); + + tracer = ds_request_bts(NULL, buffer, DS_SELFTEST_BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); + if (IS_ERR(tracer)) { + error = PTR_ERR(tracer); + tracer = NULL; + + printk(KERN_CONT + "initialization failed (err: %d)...", error); + goto out; + } + + /* The return should already give us enough trace. */ + ds_suspend_bts(tracer); + + /* Let's see if we can access the trace. */ + trace = ds_read_bts(tracer); + + error = ds_selftest_bts_consistency(trace); + if (error < 0) + goto out; + + /* If everything went well, we should have a few trace entries. */ + if (trace->ds.top == trace->ds.begin) { + /* + * It is possible but highly unlikely that we got a + * buffer overflow and end up at exactly the same + * position we started from. + * Let's issue a warning, but continue. + */ + printk(KERN_CONT "no trace/overflow..."); + } + + /* Let's try to read the trace we collected. */ + error = ds_selftest_bts_read(tracer, trace, + trace->ds.begin, trace->ds.top); + if (error < 0) + goto out; + + /* + * Let's read the trace again. + * Since we suspended tracing, we should get the same result. + */ + top = trace->ds.top; + + trace = ds_read_bts(tracer); + error = ds_selftest_bts_consistency(trace); + if (error < 0) + goto out; + + if (top != trace->ds.top) { + printk(KERN_CONT "suspend not working..."); + error = -1; + goto out; + } + + /* Let's collect some more trace - see if resume is working. */ + ds_resume_bts(tracer); + ds_suspend_bts(tracer); + + trace = ds_read_bts(tracer); + + error = ds_selftest_bts_consistency(trace); + if (error < 0) + goto out; + + if (trace->ds.top == top) { + /* + * It is possible but highly unlikely that we got a + * buffer overflow and end up at exactly the same + * position we started from. + * Let's issue a warning and check the full trace. + */ + printk(KERN_CONT + "no resume progress/overflow..."); + + error = ds_selftest_bts_read(tracer, trace, + trace->ds.begin, trace->ds.end); + } else if (trace->ds.top < top) { + /* + * We had a buffer overflow - the entire buffer should + * contain trace records. + */ + error = ds_selftest_bts_read(tracer, trace, + trace->ds.begin, trace->ds.end); + } else { + /* + * It is quite likely that the buffer did not overflow. + * Let's just check the delta trace. + */ + error = ds_selftest_bts_read(tracer, trace, + top, trace->ds.top); + } + if (error < 0) + goto out; + + error = 0; + + /* The final test: release the tracer while tracing is suspended. */ + out: + ds_release_bts(tracer); + + printk(KERN_CONT "%s.\n", (error ? "failed" : "passed")); + + return error; +} + +int ds_selftest_pebs(void) +{ + return 0; +} diff --git a/arch/x86/kernel/ds_selftest.h b/arch/x86/kernel/ds_selftest.h new file mode 100644 index 000000000000..0e6e19d4c7d2 --- /dev/null +++ b/arch/x86/kernel/ds_selftest.h @@ -0,0 +1,15 @@ +/* + * Debug Store support - selftest + * + * + * Copyright (C) 2009 Intel Corporation. + * Markus Metzger , 2009 + */ + +#ifdef CONFIG_X86_DS_SELFTEST +extern int ds_selftest_bts(void); +extern int ds_selftest_pebs(void); +#else +static inline int ds_selftest_bts(void) { return 0; } +static inline int ds_selftest_pebs(void) { return 0; } +#endif /* CONFIG_X86_DS_SELFTEST */ -- cgit v1.2.3 From b8e47195451c5d3f62620b2b1b5928669afd56eb Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 13 Mar 2009 10:46:42 +0100 Subject: x86, bts: correct comment style in ds.c Correct the comment style in ds.c. Signed-off-by: Markus Metzger LKML-Reference: <20090313104642.A30149@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 79 ++++++++++++++++++++++++++-------------------------- 1 file changed, 39 insertions(+), 40 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 51c936c1a390..d9cab7168058 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -35,25 +35,22 @@ * The configuration for a particular DS hardware implementation. */ struct ds_configuration { - /* the name of the configuration */ + /* The name of the configuration. */ const char *name; - /* the size of one pointer-typed field in the DS structure and - in the BTS and PEBS buffers in bytes; - this covers the first 8 DS fields related to buffer management. */ + /* The size of pointer-typed fields in DS, BTS, and PEBS. */ unsigned char sizeof_ptr_field; - /* the size of a BTS/PEBS record in bytes */ + /* The size of a BTS/PEBS record in bytes. */ unsigned char sizeof_rec[2]; - /* a series of bit-masks to control various features indexed - * by enum ds_feature */ + /* Control bit-masks indexed by enum ds_feature. */ unsigned long ctl[dsf_ctl_max]; }; static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array); #define ds_cfg per_cpu(ds_cfg_array, smp_processor_id()) -#define MAX_SIZEOF_DS (12 * 8) /* maximal size of a DS configuration */ -#define MAX_SIZEOF_BTS (3 * 8) /* maximal size of a BTS record */ -#define DS_ALIGNMENT (1 << 3) /* BTS and PEBS buffer alignment */ +#define MAX_SIZEOF_DS (12 * 8) /* Maximal size of a DS configuration. */ +#define MAX_SIZEOF_BTS (3 * 8) /* Maximal size of a BTS record. */ +#define DS_ALIGNMENT (1 << 3) /* BTS and PEBS buffer alignment. */ #define BTS_CONTROL \ (ds_cfg.ctl[dsf_bts] | ds_cfg.ctl[dsf_bts_kernel] | ds_cfg.ctl[dsf_bts_user] |\ @@ -67,28 +64,28 @@ static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array); * to identify tracers. */ struct ds_tracer { - /* the DS context (partially) owned by this tracer */ + /* The DS context (partially) owned by this tracer. */ struct ds_context *context; - /* the buffer provided on ds_request() and its size in bytes */ + /* The buffer provided on ds_request() and its size in bytes. */ void *buffer; size_t size; }; struct bts_tracer { - /* the common DS part */ + /* The common DS part. */ struct ds_tracer ds; - /* the trace including the DS configuration */ + /* The trace including the DS configuration. */ struct bts_trace trace; - /* buffer overflow notification function */ + /* Buffer overflow notification function. */ bts_ovfl_callback_t ovfl; }; struct pebs_tracer { - /* the common DS part */ + /* The common DS part. */ struct ds_tracer ds; - /* the trace including the DS configuration */ + /* The trace including the DS configuration. */ struct pebs_trace trace; - /* buffer overflow notification function */ + /* Buffer overflow notification function. */ pebs_ovfl_callback_t ovfl; }; @@ -214,18 +211,16 @@ static inline int check_tracer(struct task_struct *task) * deallocated when the last user puts the context. */ struct ds_context { - /* pointer to the DS configuration; goes into MSR_IA32_DS_AREA */ + /* The DS configuration; goes into MSR_IA32_DS_AREA. */ unsigned char ds[MAX_SIZEOF_DS]; - /* the owner of the BTS and PEBS configuration, respectively */ + /* The owner of the BTS and PEBS configuration, respectively. */ struct bts_tracer *bts_master; struct pebs_tracer *pebs_master; - /* use count */ + /* Use count. */ unsigned long count; - /* a pointer to the context location inside the thread_struct - * or the per_cpu context array */ + /* Pointer to the context pointer field. */ struct ds_context **this; - /* a pointer to the task owning this context, or NULL, if the - * context is owned by a cpu */ + /* The traced task; NULL for current cpu. */ struct task_struct *task; }; @@ -350,14 +345,14 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual, unsigned long write_size, adj_write_size; /* - * write as much as possible without producing an + * Write as much as possible without producing an * overflow interrupt. * - * interrupt_threshold must either be + * Interrupt_threshold must either be * - bigger than absolute_maximum or * - point to a record between buffer_base and absolute_maximum * - * index points to a valid record. + * Index points to a valid record. */ base = ds_get(context->ds, qual, ds_buffer_base); index = ds_get(context->ds, qual, ds_index); @@ -366,8 +361,10 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual, write_end = min(end, int_th); - /* if we are already beyond the interrupt threshold, - * we fill the entire buffer */ + /* + * If we are already beyond the interrupt threshold, + * we fill the entire buffer. + */ if (write_end <= index) write_end = end; @@ -384,7 +381,7 @@ static int ds_write(struct ds_context *context, enum ds_qualifier qual, adj_write_size = write_size / ds_cfg.sizeof_rec[qual]; adj_write_size *= ds_cfg.sizeof_rec[qual]; - /* zero out trailing bytes */ + /* Zero out trailing bytes. */ memset((char *)index + write_size, 0, adj_write_size - write_size); index += adj_write_size; @@ -556,7 +553,8 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual, unsigned int flags) { unsigned long buffer, adj; - /* adjust the buffer address and size to meet alignment + /* + * Adjust the buffer address and size to meet alignment * constraints: * - buffer is double-word aligned * - size is multiple of record size @@ -578,7 +576,8 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual, trace->begin = (void *)buffer; trace->top = trace->begin; trace->end = (void *)(buffer + size); - /* The value for 'no threshold' is -1, which will set the + /* + * The value for 'no threshold' is -1, which will set the * threshold outside of the buffer, just like we want it. */ trace->ith = (void *)(buffer + size - ith); @@ -602,7 +601,7 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, if (!base) goto out; - /* we require some space to do alignment adjustments below */ + /* We require some space to do alignment adjustments below. */ error = -EINVAL; if (size < (DS_ALIGNMENT + ds_cfg.sizeof_rec[qual])) goto out; @@ -640,7 +639,7 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, unsigned long irq; int error; - /* buffer overflow notification is not yet implemented */ + /* Buffer overflow notification is not yet implemented. */ error = -EOPNOTSUPP; if (ovfl) goto out; @@ -700,7 +699,7 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, unsigned long irq; int error; - /* buffer overflow notification is not yet implemented */ + /* Buffer overflow notification is not yet implemented. */ error = -EOPNOTSUPP; if (ovfl) goto out; @@ -983,9 +982,9 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) case 0x1c: /* Atom */ ds_configure(&ds_cfg_core2_atom, c); break; - case 0x1a: /* i7 */ + case 0x1a: /* Core i7 */ default: - /* sorry, don't know about them */ + /* Sorry, don't know about them. */ break; } break; @@ -997,12 +996,12 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) ds_configure(&ds_cfg_netburst, c); break; default: - /* sorry, don't know about them */ + /* Sorry, don't know about them. */ break; } break; default: - /* sorry, don't know about them */ + /* Sorry, don't know about them. */ break; } } -- cgit v1.2.3 From e9a22d1fb94050b7d600019c32e6b672d539054b Mon Sep 17 00:00:00 2001 From: Ingo Molnar Date: Fri, 13 Mar 2009 11:54:40 +0100 Subject: x86, bts: cleanups Impact: cleanup, no code changed Cc: Markus Metzger LKML-Reference: <20090313104218.A30096@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 142 ++++++++++++++++++++++----------------- arch/x86/kernel/ds_selftest.h | 2 +- kernel/trace/trace_hw_branches.c | 6 +- kernel/trace/trace_selftest.c | 5 +- 4 files changed, 87 insertions(+), 68 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index d9cab7168058..7363e01ba082 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -19,43 +19,52 @@ * Markus Metzger , 2007-2009 */ - -#include - -#include +#include #include -#include +#include #include +#include #include -#include + +#include #include "ds_selftest.h" /* - * The configuration for a particular DS hardware implementation. + * The configuration for a particular DS hardware implementation: */ struct ds_configuration { - /* The name of the configuration. */ - const char *name; - /* The size of pointer-typed fields in DS, BTS, and PEBS. */ - unsigned char sizeof_ptr_field; - /* The size of a BTS/PEBS record in bytes. */ - unsigned char sizeof_rec[2]; - /* Control bit-masks indexed by enum ds_feature. */ - unsigned long ctl[dsf_ctl_max]; + /* The name of the configuration: */ + const char *name; + + /* The size of pointer-typed fields in DS, BTS, and PEBS: */ + unsigned char sizeof_ptr_field; + + /* The size of a BTS/PEBS record in bytes: */ + unsigned char sizeof_rec[2]; + + /* Control bit-masks indexed by enum ds_feature: */ + unsigned long ctl[dsf_ctl_max]; }; static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array); #define ds_cfg per_cpu(ds_cfg_array, smp_processor_id()) -#define MAX_SIZEOF_DS (12 * 8) /* Maximal size of a DS configuration. */ -#define MAX_SIZEOF_BTS (3 * 8) /* Maximal size of a BTS record. */ -#define DS_ALIGNMENT (1 << 3) /* BTS and PEBS buffer alignment. */ +/* Maximal size of a DS configuration: */ +#define MAX_SIZEOF_DS (12 * 8) -#define BTS_CONTROL \ - (ds_cfg.ctl[dsf_bts] | ds_cfg.ctl[dsf_bts_kernel] | ds_cfg.ctl[dsf_bts_user] |\ - ds_cfg.ctl[dsf_bts_overflow]) +/* Maximal size of a BTS record: */ +#define MAX_SIZEOF_BTS (3 * 8) +/* BTS and PEBS buffer alignment: */ +#define DS_ALIGNMENT (1 << 3) + +/* Mask of control bits in the DS MSR register: */ +#define BTS_CONTROL \ + ( ds_cfg.ctl[dsf_bts] | \ + ds_cfg.ctl[dsf_bts_kernel] | \ + ds_cfg.ctl[dsf_bts_user] | \ + ds_cfg.ctl[dsf_bts_overflow] ) /* * A BTS or PEBS tracer. @@ -65,28 +74,32 @@ static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array); */ struct ds_tracer { /* The DS context (partially) owned by this tracer. */ - struct ds_context *context; + struct ds_context *context; /* The buffer provided on ds_request() and its size in bytes. */ - void *buffer; - size_t size; + void *buffer; + size_t size; }; struct bts_tracer { - /* The common DS part. */ - struct ds_tracer ds; - /* The trace including the DS configuration. */ - struct bts_trace trace; - /* Buffer overflow notification function. */ - bts_ovfl_callback_t ovfl; + /* The common DS part: */ + struct ds_tracer ds; + + /* The trace including the DS configuration: */ + struct bts_trace trace; + + /* Buffer overflow notification function: */ + bts_ovfl_callback_t ovfl; }; struct pebs_tracer { - /* The common DS part. */ - struct ds_tracer ds; - /* The trace including the DS configuration. */ - struct pebs_trace trace; - /* Buffer overflow notification function. */ - pebs_ovfl_callback_t ovfl; + /* The common DS part: */ + struct ds_tracer ds; + + /* The trace including the DS configuration: */ + struct pebs_trace trace; + + /* Buffer overflow notification function: */ + pebs_ovfl_callback_t ovfl; }; /* @@ -95,6 +108,7 @@ struct pebs_tracer { * * The DS configuration consists of the following fields; different * architetures vary in the size of those fields. + * * - double-word aligned base linear address of the BTS buffer * - write pointer into the BTS buffer * - end linear address of the BTS buffer (one byte beyond the end of @@ -133,19 +147,20 @@ enum ds_field { }; enum ds_qualifier { - ds_bts = 0, + ds_bts = 0, ds_pebs }; -static inline unsigned long ds_get(const unsigned char *base, - enum ds_qualifier qual, enum ds_field field) +static inline unsigned long +ds_get(const unsigned char *base, enum ds_qualifier qual, enum ds_field field) { base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual))); return *(unsigned long *)base; } -static inline void ds_set(unsigned char *base, enum ds_qualifier qual, - enum ds_field field, unsigned long value) +static inline void +ds_set(unsigned char *base, enum ds_qualifier qual, enum ds_field field, + unsigned long value) { base += (ds_cfg.sizeof_ptr_field * (field + (4 * qual))); (*(unsigned long *)base) = value; @@ -157,7 +172,6 @@ static inline void ds_set(unsigned char *base, enum ds_qualifier qual, */ static DEFINE_SPINLOCK(ds_lock); - /* * We either support (system-wide) per-cpu or per-thread allocation. * We distinguish the two based on the task_struct pointer, where a @@ -211,17 +225,21 @@ static inline int check_tracer(struct task_struct *task) * deallocated when the last user puts the context. */ struct ds_context { - /* The DS configuration; goes into MSR_IA32_DS_AREA. */ - unsigned char ds[MAX_SIZEOF_DS]; - /* The owner of the BTS and PEBS configuration, respectively. */ - struct bts_tracer *bts_master; - struct pebs_tracer *pebs_master; - /* Use count. */ + /* The DS configuration; goes into MSR_IA32_DS_AREA: */ + unsigned char ds[MAX_SIZEOF_DS]; + + /* The owner of the BTS and PEBS configuration, respectively: */ + struct bts_tracer *bts_master; + struct pebs_tracer *pebs_master; + + /* Use count: */ unsigned long count; - /* Pointer to the context pointer field. */ - struct ds_context **this; - /* The traced task; NULL for current cpu. */ - struct task_struct *task; + + /* Pointer to the context pointer field: */ + struct ds_context **this; + + /* The traced task; NULL for current cpu: */ + struct task_struct *task; }; static DEFINE_PER_CPU(struct ds_context *, system_context_array); @@ -328,9 +346,9 @@ static void ds_overflow(struct ds_context *context, enum ds_qualifier qual) * The remainder of any partially written record is zeroed out. * * context: the DS context - * qual: the buffer type - * record: the data to write - * size: the size of the data + * qual: the buffer type + * record: the data to write + * size: the size of the data */ static int ds_write(struct ds_context *context, enum ds_qualifier qual, const void *record, size_t size) @@ -429,12 +447,12 @@ enum bts_field { bts_to, bts_flags, - bts_qual = bts_from, - bts_jiffies = bts_to, - bts_pid = bts_flags, + bts_qual = bts_from, + bts_jiffies = bts_to, + bts_pid = bts_flags, - bts_qual_mask = (bts_qual_max - 1), - bts_escape = ((unsigned long)-1 & ~bts_qual_mask) + bts_qual_mask = (bts_qual_max - 1), + bts_escape = ((unsigned long)-1 & ~bts_qual_mask) }; static inline unsigned long bts_get(const char *base, enum bts_field field) @@ -461,8 +479,8 @@ static inline void bts_set(char *base, enum bts_field field, unsigned long val) * * return: bytes read/written on success; -Eerrno, otherwise */ -static int bts_read(struct bts_tracer *tracer, const void *at, - struct bts_struct *out) +static int +bts_read(struct bts_tracer *tracer, const void *at, struct bts_struct *out) { if (!tracer) return -EINVAL; diff --git a/arch/x86/kernel/ds_selftest.h b/arch/x86/kernel/ds_selftest.h index 0e6e19d4c7d2..2ba8745c6663 100644 --- a/arch/x86/kernel/ds_selftest.h +++ b/arch/x86/kernel/ds_selftest.h @@ -12,4 +12,4 @@ extern int ds_selftest_pebs(void); #else static inline int ds_selftest_bts(void) { return 0; } static inline int ds_selftest_pebs(void) { return 0; } -#endif /* CONFIG_X86_DS_SELFTEST */ +#endif diff --git a/kernel/trace/trace_hw_branches.c b/kernel/trace/trace_hw_branches.c index 4ca82700c04e..8b2109a6c61c 100644 --- a/kernel/trace/trace_hw_branches.c +++ b/kernel/trace/trace_hw_branches.c @@ -1,5 +1,5 @@ /* - * h/w branch tracer for x86 based on bts + * h/w branch tracer for x86 based on BTS * * Copyright (C) 2008-2009 Intel Corporation. * Markus Metzger , 2008-2009 @@ -15,8 +15,8 @@ #include -#include "trace.h" #include "trace_output.h" +#include "trace.h" #define BTS_BUFFER_SIZE (1 << 13) @@ -197,10 +197,10 @@ static void bts_trace_print_header(struct seq_file *m) static enum print_line_t bts_trace_print_line(struct trace_iterator *iter) { + unsigned long symflags = TRACE_ITER_SYM_OFFSET; struct trace_entry *entry = iter->ent; struct trace_seq *seq = &iter->seq; struct hw_branch_entry *it; - unsigned long symflags = TRACE_ITER_SYM_OFFSET; trace_assign_type(it, entry); diff --git a/kernel/trace/trace_selftest.c b/kernel/trace/trace_selftest.c index 3c7b797d0d28..b91091267067 100644 --- a/kernel/trace/trace_selftest.c +++ b/kernel/trace/trace_selftest.c @@ -189,6 +189,7 @@ int trace_selftest_startup_dynamic_tracing(struct tracer *trace, #else # define trace_selftest_startup_dynamic_tracing(trace, tr, func) ({ 0; }) #endif /* CONFIG_DYNAMIC_FTRACE */ + /* * Simple verification test of ftrace function tracer. * Enable ftrace, sleep 1/10 second, and then read the trace @@ -698,10 +699,10 @@ int trace_selftest_startup_hw_branches(struct tracer *trace, struct trace_array *tr) { - unsigned long count; - int ret; struct trace_iterator iter; struct tracer tracer; + unsigned long count; + int ret; if (!trace->open) { printk(KERN_CONT "missing open function..."); -- cgit v1.2.3 From 79258a354e0c69be94ae2871809a195bf4a647b1 Mon Sep 17 00:00:00 2001 From: Ingo Molnar Date: Fri, 13 Mar 2009 12:02:08 +0100 Subject: x86, bts: detect size of DS fields, fix Impact: build fix One usage site was missed in the sizeof_field -> sizeof_ptr_field rename. Cc: Markus Metzger LKML-Reference: <20090313104218.A30096@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 7363e01ba082..5fd53333c1df 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -983,7 +983,7 @@ ds_configure(const struct ds_configuration *cfg, printk("bts/pebs record: %u/%u bytes\n", ds_cfg.sizeof_rec[ds_bts], ds_cfg.sizeof_rec[ds_pebs]); - WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_field)); + WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_ptr_field)); } void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) -- cgit v1.2.3 From c78a3956b982418186e40978a51636a2b43221bc Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Wed, 18 Mar 2009 19:27:00 +0100 Subject: x86, bts: use atomic memory allocation Ds_request_bts() needs to allocate memory. It uses GFP_KERNEL. Hw-branch-tracer calls ds_request_bts() within on_each_cpu(). Use atomic memory allocation to allow it to be used in that context. Signed-off-by: Markus Metzger LKML-Reference: <20090318192700.A6038@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 5fd53333c1df..b1d6e1f502fa 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -255,8 +255,13 @@ static inline struct ds_context *ds_get_context(struct task_struct *task) struct ds_context *new_context = NULL; unsigned long irq; - /* Chances are small that we already have a context. */ - new_context = kzalloc(sizeof(*new_context), GFP_KERNEL); + /* + * Chances are small that we already have a context. + * + * Contexts for per-cpu tracing are allocated using + * smp_call_function(). We must not sleep. + */ + new_context = kzalloc(sizeof(*new_context), GFP_ATOMIC); if (!new_context) return NULL; @@ -662,8 +667,12 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, if (ovfl) goto out; + /* + * Per-cpu tracing is typically requested using smp_call_function(). + * We must not sleep. + */ error = -ENOMEM; - tracer = kzalloc(sizeof(*tracer), GFP_KERNEL); + tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); if (!tracer) goto out; tracer->ovfl = ovfl; @@ -722,8 +731,12 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, if (ovfl) goto out; + /* + * Per-cpu tracing is typically requested using smp_call_function(). + * We must not sleep. + */ error = -ENOMEM; - tracer = kzalloc(sizeof(*tracer), GFP_KERNEL); + tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); if (!tracer) goto out; tracer->ovfl = ovfl; -- cgit v1.2.3 From cac94f979326212831c0ea44ed9ea1622b4f4e93 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:33 +0200 Subject: x86, bts: fix race when bts tracer is removed When the bts tracer is removed while the traced task is running, the write to clear the bts tracer pointer races with context switch code. Read the tracer once during a context switch. When a new tracer is installed, the bts tracer is set in the ds context before the tracer is initialized in order to claim the context for that tracer. This may result in write accesses using an uninitialized trace configuration when scheduling timestamps have been requested. Store active tracing flags separately and only set active flags after the tracing configuration has been initialized. Reviewed-by: Oleg Nesterov Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144548.881338000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 58 +++++++++++++++++++++++++++++++++------------------- 1 file changed, 37 insertions(+), 21 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index b1d6e1f502fa..c730155bf54d 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -89,6 +89,9 @@ struct bts_tracer { /* Buffer overflow notification function: */ bts_ovfl_callback_t ovfl; + + /* Active flags affecting trace collection. */ + unsigned int flags; }; struct pebs_tracer { @@ -799,6 +802,8 @@ void ds_suspend_bts(struct bts_tracer *tracer) if (!tracer) return; + tracer->flags = 0; + task = tracer->ds.context->task; if (!task || (task == current)) @@ -820,6 +825,8 @@ void ds_resume_bts(struct bts_tracer *tracer) if (!tracer) return; + tracer->flags = tracer->trace.ds.flags; + task = tracer->ds.context->task; control = ds_cfg.ctl[dsf_bts]; @@ -1037,43 +1044,52 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) } } +static inline void ds_take_timestamp(struct ds_context *context, + enum bts_qualifier qualifier, + struct task_struct *task) +{ + struct bts_tracer *tracer = context->bts_master; + struct bts_struct ts; + + /* Prevent compilers from reading the tracer pointer twice. */ + barrier(); + + if (!tracer || !(tracer->flags & BTS_TIMESTAMPS)) + return; + + memset(&ts, 0, sizeof(ts)); + ts.qualifier = qualifier; + ts.variant.timestamp.jiffies = jiffies_64; + ts.variant.timestamp.pid = task->pid; + + bts_write(tracer, &ts); +} + /* * Change the DS configuration from tracing prev to tracing next. */ void ds_switch_to(struct task_struct *prev, struct task_struct *next) { - struct ds_context *prev_ctx = prev->thread.ds_ctx; - struct ds_context *next_ctx = next->thread.ds_ctx; + struct ds_context *prev_ctx = prev->thread.ds_ctx; + struct ds_context *next_ctx = next->thread.ds_ctx; + unsigned long debugctlmsr = next->thread.debugctlmsr; + + /* Make sure all data is read before we start. */ + barrier(); if (prev_ctx) { update_debugctlmsr(0); - if (prev_ctx->bts_master && - (prev_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) { - struct bts_struct ts = { - .qualifier = bts_task_departs, - .variant.timestamp.jiffies = jiffies_64, - .variant.timestamp.pid = prev->pid - }; - bts_write(prev_ctx->bts_master, &ts); - } + ds_take_timestamp(prev_ctx, bts_task_departs, prev); } if (next_ctx) { - if (next_ctx->bts_master && - (next_ctx->bts_master->trace.ds.flags & BTS_TIMESTAMPS)) { - struct bts_struct ts = { - .qualifier = bts_task_arrives, - .variant.timestamp.jiffies = jiffies_64, - .variant.timestamp.pid = next->pid - }; - bts_write(next_ctx->bts_master, &ts); - } + ds_take_timestamp(next_ctx, bts_task_arrives, next); wrmsrl(MSR_IA32_DS_AREA, (unsigned long)next_ctx->ds); } - update_debugctlmsr(next->thread.debugctlmsr); + update_debugctlmsr(debugctlmsr); } void ds_copy_thread(struct task_struct *tsk, struct task_struct *father) -- cgit v1.2.3 From e2b371f00a6f529f6362654239bdec8dcd510760 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:35 +0200 Subject: mm, x86, ptrace, bts: defer branch trace stopping When a ptraced task is unlinked, we need to stop branch tracing for that task. Since the unlink is called with interrupts disabled, and we need interrupts enabled to stop branch tracing, we defer the work. Collect all branch tracing related stuff in a branch tracing context. Reviewed-by: Oleg Nesterov Signed-off-by: Markus Metzger Cc: Andrew Morton Cc: Peter Zijlstra Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144550.712401000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/processor.h | 4 - arch/x86/kernel/ptrace.c | 254 ++++++++++++++++++++++++++------------- include/linux/mm.h | 3 +- include/linux/sched.h | 9 +- mm/mlock.c | 13 +- 5 files changed, 179 insertions(+), 104 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h index 34c52370f2fe..2483807e06e7 100644 --- a/arch/x86/include/asm/processor.h +++ b/arch/x86/include/asm/processor.h @@ -458,10 +458,6 @@ struct thread_struct { /* Debug Store context; see include/asm-x86/ds.h; goes into MSR_IA32_DS_AREA */ struct ds_context *ds_ctx; #endif /* CONFIG_X86_DS */ -#ifdef CONFIG_X86_PTRACE_BTS -/* the signal to send on a bts buffer overflow */ - unsigned int bts_ovfl_signal; -#endif /* CONFIG_X86_PTRACE_BTS */ }; static inline unsigned long native_get_debugreg(int regno) diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c index fe9345c967de..7c21d1e8cae7 100644 --- a/arch/x86/kernel/ptrace.c +++ b/arch/x86/kernel/ptrace.c @@ -22,6 +22,7 @@ #include #include #include +#include #include #include @@ -577,17 +578,119 @@ static int ioperm_get(struct task_struct *target, } #ifdef CONFIG_X86_PTRACE_BTS +/* + * A branch trace store context. + * + * Contexts may only be installed by ptrace_bts_config() and only for + * ptraced tasks. + * + * Contexts are destroyed when the tracee is detached from the tracer. + * The actual destruction work requires interrupts enabled, so the + * work is deferred and will be scheduled during __ptrace_unlink(). + * + * Contexts hold an additional task_struct reference on the traced + * task, as well as a reference on the tracer's mm. + * + * Ptrace already holds a task_struct for the duration of ptrace operations, + * but since destruction is deferred, it may be executed after both + * tracer and tracee exited. + */ +struct bts_context { + /* The branch trace handle. */ + struct bts_tracer *tracer; + + /* The buffer used to store the branch trace and its size. */ + void *buffer; + unsigned int size; + + /* The mm that paid for the above buffer. */ + struct mm_struct *mm; + + /* The task this context belongs to. */ + struct task_struct *task; + + /* The signal to send on a bts buffer overflow. */ + unsigned int bts_ovfl_signal; + + /* The work struct to destroy a context. */ + struct work_struct work; +}; + +static inline void alloc_bts_buffer(struct bts_context *context, + unsigned int size) +{ + void *buffer; + + buffer = alloc_locked_buffer(size); + if (buffer) { + context->buffer = buffer; + context->size = size; + context->mm = get_task_mm(current); + } +} + +static inline void free_bts_buffer(struct bts_context *context) +{ + if (!context->buffer) + return; + + kfree(context->buffer); + context->buffer = NULL; + + refund_locked_buffer_memory(context->mm, context->size); + context->size = 0; + + mmput(context->mm); + context->mm = NULL; +} + +static void free_bts_context_work(struct work_struct *w) +{ + struct bts_context *context; + + context = container_of(w, struct bts_context, work); + + ds_release_bts(context->tracer); + put_task_struct(context->task); + free_bts_buffer(context); + kfree(context); +} + +static inline void free_bts_context(struct bts_context *context) +{ + INIT_WORK(&context->work, free_bts_context_work); + schedule_work(&context->work); +} + +static inline struct bts_context *alloc_bts_context(struct task_struct *task) +{ + struct bts_context *context = kzalloc(sizeof(*context), GFP_KERNEL); + if (context) { + context->task = task; + task->bts = context; + + get_task_struct(task); + } + + return context; +} + static int ptrace_bts_read_record(struct task_struct *child, size_t index, struct bts_struct __user *out) { + struct bts_context *context; const struct bts_trace *trace; struct bts_struct bts; const unsigned char *at; int error; - trace = ds_read_bts(child->bts); + context = child->bts; + if (!context) + return -ESRCH; + + trace = ds_read_bts(context->tracer); if (!trace) - return -EPERM; + return -ESRCH; at = trace->ds.top - ((index + 1) * trace->ds.size); if ((void *)at < trace->ds.begin) @@ -596,7 +699,7 @@ static int ptrace_bts_read_record(struct task_struct *child, size_t index, if (!trace->read) return -EOPNOTSUPP; - error = trace->read(child->bts, at, &bts); + error = trace->read(context->tracer, at, &bts); if (error < 0) return error; @@ -610,13 +713,18 @@ static int ptrace_bts_drain(struct task_struct *child, long size, struct bts_struct __user *out) { + struct bts_context *context; const struct bts_trace *trace; const unsigned char *at; int error, drained = 0; - trace = ds_read_bts(child->bts); + context = child->bts; + if (!context) + return -ESRCH; + + trace = ds_read_bts(context->tracer); if (!trace) - return -EPERM; + return -ESRCH; if (!trace->read) return -EOPNOTSUPP; @@ -627,9 +735,8 @@ static int ptrace_bts_drain(struct task_struct *child, for (at = trace->ds.begin; (void *)at < trace->ds.top; out++, drained++, at += trace->ds.size) { struct bts_struct bts; - int error; - error = trace->read(child->bts, at, &bts); + error = trace->read(context->tracer, at, &bts); if (error < 0) return error; @@ -639,35 +746,18 @@ static int ptrace_bts_drain(struct task_struct *child, memset(trace->ds.begin, 0, trace->ds.n * trace->ds.size); - error = ds_reset_bts(child->bts); + error = ds_reset_bts(context->tracer); if (error < 0) return error; return drained; } -static int ptrace_bts_allocate_buffer(struct task_struct *child, size_t size) -{ - child->bts_buffer = alloc_locked_buffer(size); - if (!child->bts_buffer) - return -ENOMEM; - - child->bts_size = size; - - return 0; -} - -static void ptrace_bts_free_buffer(struct task_struct *child) -{ - free_locked_buffer(child->bts_buffer, child->bts_size); - child->bts_buffer = NULL; - child->bts_size = 0; -} - static int ptrace_bts_config(struct task_struct *child, long cfg_size, const struct ptrace_bts_config __user *ucfg) { + struct bts_context *context; struct ptrace_bts_config cfg; unsigned int flags = 0; @@ -677,28 +767,31 @@ static int ptrace_bts_config(struct task_struct *child, if (copy_from_user(&cfg, ucfg, sizeof(cfg))) return -EFAULT; - if (child->bts) { - ds_release_bts(child->bts); - child->bts = NULL; - } + context = child->bts; + if (!context) + context = alloc_bts_context(child); + if (!context) + return -ENOMEM; if (cfg.flags & PTRACE_BTS_O_SIGNAL) { if (!cfg.signal) return -EINVAL; - child->thread.bts_ovfl_signal = cfg.signal; return -EOPNOTSUPP; + context->bts_ovfl_signal = cfg.signal; } - if ((cfg.flags & PTRACE_BTS_O_ALLOC) && - (cfg.size != child->bts_size)) { - int error; + ds_release_bts(context->tracer); + context->tracer = NULL; - ptrace_bts_free_buffer(child); + if ((cfg.flags & PTRACE_BTS_O_ALLOC) && (cfg.size != context->size)) { + free_bts_buffer(context); + if (!cfg.size) + return 0; - error = ptrace_bts_allocate_buffer(child, cfg.size); - if (error < 0) - return error; + alloc_bts_buffer(context, cfg.size); + if (!context->buffer) + return -ENOMEM; } if (cfg.flags & PTRACE_BTS_O_TRACE) @@ -707,15 +800,13 @@ static int ptrace_bts_config(struct task_struct *child, if (cfg.flags & PTRACE_BTS_O_SCHED) flags |= BTS_TIMESTAMPS; - child->bts = ds_request_bts(child, child->bts_buffer, child->bts_size, - /* ovfl = */ NULL, /* th = */ (size_t)-1, - flags); - if (IS_ERR(child->bts)) { - int error = PTR_ERR(child->bts); - - ptrace_bts_free_buffer(child); - child->bts = NULL; + context->tracer = ds_request_bts(child, context->buffer, context->size, + NULL, (size_t)-1, flags); + if (unlikely(IS_ERR(context->tracer))) { + int error = PTR_ERR(context->tracer); + free_bts_buffer(context); + context->tracer = NULL; return error; } @@ -726,20 +817,25 @@ static int ptrace_bts_status(struct task_struct *child, long cfg_size, struct ptrace_bts_config __user *ucfg) { + struct bts_context *context; const struct bts_trace *trace; struct ptrace_bts_config cfg; + context = child->bts; + if (!context) + return -ESRCH; + if (cfg_size < sizeof(cfg)) return -EIO; - trace = ds_read_bts(child->bts); + trace = ds_read_bts(context->tracer); if (!trace) - return -EPERM; + return -ESRCH; memset(&cfg, 0, sizeof(cfg)); - cfg.size = trace->ds.end - trace->ds.begin; - cfg.signal = child->thread.bts_ovfl_signal; - cfg.bts_size = sizeof(struct bts_struct); + cfg.size = trace->ds.end - trace->ds.begin; + cfg.signal = context->bts_ovfl_signal; + cfg.bts_size = sizeof(struct bts_struct); if (cfg.signal) cfg.flags |= PTRACE_BTS_O_SIGNAL; @@ -758,67 +854,56 @@ static int ptrace_bts_status(struct task_struct *child, static int ptrace_bts_clear(struct task_struct *child) { + struct bts_context *context; const struct bts_trace *trace; - trace = ds_read_bts(child->bts); + context = child->bts; + if (!context) + return -ESRCH; + + trace = ds_read_bts(context->tracer); if (!trace) - return -EPERM; + return -ESRCH; memset(trace->ds.begin, 0, trace->ds.n * trace->ds.size); - return ds_reset_bts(child->bts); + return ds_reset_bts(context->tracer); } static int ptrace_bts_size(struct task_struct *child) { + struct bts_context *context; const struct bts_trace *trace; - trace = ds_read_bts(child->bts); + context = child->bts; + if (!context) + return -ESRCH; + + trace = ds_read_bts(context->tracer); if (!trace) - return -EPERM; + return -ESRCH; return (trace->ds.top - trace->ds.begin) / trace->ds.size; } -static void ptrace_bts_fork(struct task_struct *tsk) +static inline void ptrace_bts_fork(struct task_struct *tsk) { tsk->bts = NULL; - tsk->bts_buffer = NULL; - tsk->bts_size = 0; - tsk->thread.bts_ovfl_signal = 0; } -static void ptrace_bts_untrace(struct task_struct *child) +/* + * Called from __ptrace_unlink() after the child has been moved back + * to its original parent. + */ +static inline void ptrace_bts_untrace(struct task_struct *child) { if (unlikely(child->bts)) { - ds_release_bts(child->bts); + free_bts_context(child->bts); child->bts = NULL; - - /* We cannot update total_vm and locked_vm since - child's mm is already gone. But we can reclaim the - memory. */ - kfree(child->bts_buffer); - child->bts_buffer = NULL; - child->bts_size = 0; } } - -static void ptrace_bts_detach(struct task_struct *child) -{ - /* - * Ptrace_detach() races with ptrace_untrace() in case - * the child dies and is reaped by another thread. - * - * We only do the memory accounting at this point and - * leave the buffer deallocation and the bts tracer - * release to ptrace_bts_untrace() which will be called - * later on with tasklist_lock held. - */ - release_locked_buffer(child->bts_buffer, child->bts_size); -} #else static inline void ptrace_bts_fork(struct task_struct *tsk) {} -static inline void ptrace_bts_detach(struct task_struct *child) {} static inline void ptrace_bts_untrace(struct task_struct *child) {} #endif /* CONFIG_X86_PTRACE_BTS */ @@ -843,7 +928,6 @@ void ptrace_disable(struct task_struct *child) #ifdef TIF_SYSCALL_EMU clear_tsk_thread_flag(child, TIF_SYSCALL_EMU); #endif - ptrace_bts_detach(child); } #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION diff --git a/include/linux/mm.h b/include/linux/mm.h index bff1f0d475c7..64d8ed2538ae 100644 --- a/include/linux/mm.h +++ b/include/linux/mm.h @@ -13,6 +13,7 @@ #include #include #include +#include struct mempolicy; struct anon_vma; @@ -1321,6 +1322,6 @@ void vmemmap_populate_print_last(void); extern void *alloc_locked_buffer(size_t size); extern void free_locked_buffer(void *buffer, size_t size); -extern void release_locked_buffer(void *buffer, size_t size); +extern void refund_locked_buffer_memory(struct mm_struct *mm, size_t size); #endif /* __KERNEL__ */ #endif /* _LINUX_MM_H */ diff --git a/include/linux/sched.h b/include/linux/sched.h index a5b9a83065fa..52b8cd049c2e 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -96,8 +96,8 @@ struct exec_domain; struct futex_pi_state; struct robust_list_head; struct bio; -struct bts_tracer; struct fs_struct; +struct bts_context; /* * List of flags we want to share for kernel threads, @@ -1210,12 +1210,7 @@ struct task_struct { * This is the tracer handle for the ptrace BTS extension. * This field actually belongs to the ptracer task. */ - struct bts_tracer *bts; - /* - * The buffer to hold the BTS data. - */ - void *bts_buffer; - size_t bts_size; + struct bts_context *bts; #endif /* CONFIG_X86_PTRACE_BTS */ /* PID/PID hash table linkage. */ diff --git a/mm/mlock.c b/mm/mlock.c index cbe9e0581b75..749383b442c7 100644 --- a/mm/mlock.c +++ b/mm/mlock.c @@ -660,21 +660,20 @@ void *alloc_locked_buffer(size_t size) return buffer; } -void release_locked_buffer(void *buffer, size_t size) +void refund_locked_buffer_memory(struct mm_struct *mm, size_t size) { unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT; - down_write(¤t->mm->mmap_sem); + down_write(&mm->mmap_sem); - current->mm->total_vm -= pgsz; - current->mm->locked_vm -= pgsz; + mm->total_vm -= pgsz; + mm->locked_vm -= pgsz; - up_write(¤t->mm->mmap_sem); + up_write(&mm->mmap_sem); } void free_locked_buffer(void *buffer, size_t size) { - release_locked_buffer(buffer, size); - + refund_locked_buffer_memory(current->mm, size); kfree(buffer); } -- cgit v1.2.3 From 8d99b3ac2726e5edd97ad147fa5c1f2acb63a745 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:36 +0200 Subject: x86, bts: wait until traced task has been scheduled out In order to stop branch tracing for a running task, we need to first clear the branch tracing control bits before we may free the tracing buffer. If the traced task is running, the cpu might still trace that task after the branch trace control bits have cleared. Wait until the traced task has been scheduled out before proceeding. A similar problem affects the task debug store context. We first remove the context, then we need to wait until the task has been scheduled out before we can free the context memory. Reviewed-by: Oleg Nesterov Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144551.919636000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 23 +++++++++++++++++++---- 1 file changed, 19 insertions(+), 4 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index c730155bf54d..5cd137ab2672 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -299,6 +299,7 @@ static inline struct ds_context *ds_get_context(struct task_struct *task) static inline void ds_put_context(struct ds_context *context) { + struct task_struct *task; unsigned long irq; if (!context) @@ -313,14 +314,20 @@ static inline void ds_put_context(struct ds_context *context) *(context->this) = NULL; - if (context->task) - clear_tsk_thread_flag(context->task, TIF_DS_AREA_MSR); + task = context->task; + + if (task) + clear_tsk_thread_flag(task, TIF_DS_AREA_MSR); - if (!context->task || (context->task == current)) + if (!task || (task == current)) wrmsrl(MSR_IA32_DS_AREA, 0); spin_unlock_irqrestore(&ds_lock, irq); + /* The context might still be in use for context switching. */ + if (task && (task != current)) + wait_task_context_switch(task); + kfree(context); } @@ -781,15 +788,23 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, void ds_release_bts(struct bts_tracer *tracer) { + struct task_struct *task; + if (!tracer) return; + task = tracer->ds.context->task; + ds_suspend_bts(tracer); WARN_ON_ONCE(tracer->ds.context->bts_master != tracer); tracer->ds.context->bts_master = NULL; - put_tracer(tracer->ds.context->task); + /* Make sure tracing stopped and the tracer is not in use. */ + if (task && (task != current)) + wait_task_context_switch(task); + + put_tracer(task); ds_put_context(tracer->ds.context); kfree(tracer); -- cgit v1.2.3 From 38f801129ad07b9afa7f9bd3779f61b805416d8c Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:37 +0200 Subject: x86, bts: fix race between per-task and per-cpu branch tracing Per-task branch tracing installs a debug store context with the traced task. This immediately results in the branch trace control bits to be cleared for the next context switch of that task, if not set before. Either per-cpu or per-task tracing are allowed at the same time. An active per-cpu tracing would be disabled even if the per-task tracing request is rejected and the task debug store context removed. Check the tracing type (per-cpu or per-task) before installing a task debug store context. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144552.856000000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 72 ++++++++++++++++++++++++++++++---------------------- 1 file changed, 41 insertions(+), 31 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 5cd137ab2672..f03f117eff8c 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -193,12 +193,28 @@ static DEFINE_SPINLOCK(ds_lock); */ static atomic_t tracers = ATOMIC_INIT(0); -static inline void get_tracer(struct task_struct *task) +static inline int get_tracer(struct task_struct *task) { - if (task) + int error; + + spin_lock_irq(&ds_lock); + + if (task) { + error = -EPERM; + if (atomic_read(&tracers) < 0) + goto out; atomic_inc(&tracers); - else + } else { + error = -EPERM; + if (atomic_read(&tracers) > 0) + goto out; atomic_dec(&tracers); + } + + error = 0; +out: + spin_unlock_irq(&ds_lock); + return error; } static inline void put_tracer(struct task_struct *task) @@ -209,14 +225,6 @@ static inline void put_tracer(struct task_struct *task) atomic_inc(&tracers); } -static inline int check_tracer(struct task_struct *task) -{ - return task ? - (atomic_read(&tracers) >= 0) : - (atomic_read(&tracers) <= 0); -} - - /* * The DS context is either attached to a thread or to a cpu: * - in the former case, the thread_struct contains a pointer to the @@ -677,6 +685,10 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, if (ovfl) goto out; + error = get_tracer(task); + if (error < 0) + goto out; + /* * Per-cpu tracing is typically requested using smp_call_function(). * We must not sleep. @@ -684,7 +696,7 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, error = -ENOMEM; tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); if (!tracer) - goto out; + goto out_put_tracer; tracer->ovfl = ovfl; error = ds_request(&tracer->ds, &tracer->trace.ds, @@ -695,14 +707,9 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, spin_lock_irqsave(&ds_lock, irq); - error = -EPERM; - if (!check_tracer(task)) - goto out_unlock; - get_tracer(task); - error = -EPERM; if (tracer->ds.context->bts_master) - goto out_put_tracer; + goto out_unlock; tracer->ds.context->bts_master = tracer; spin_unlock_irqrestore(&ds_lock, irq); @@ -716,13 +723,13 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, return tracer; - out_put_tracer: - put_tracer(task); out_unlock: spin_unlock_irqrestore(&ds_lock, irq); ds_put_context(tracer->ds.context); out_tracer: kfree(tracer); + out_put_tracer: + put_tracer(task); out: return ERR_PTR(error); } @@ -741,6 +748,10 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, if (ovfl) goto out; + error = get_tracer(task); + if (error < 0) + goto out; + /* * Per-cpu tracing is typically requested using smp_call_function(). * We must not sleep. @@ -748,7 +759,7 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, error = -ENOMEM; tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); if (!tracer) - goto out; + goto out_put_tracer; tracer->ovfl = ovfl; error = ds_request(&tracer->ds, &tracer->trace.ds, @@ -758,14 +769,9 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, spin_lock_irqsave(&ds_lock, irq); - error = -EPERM; - if (!check_tracer(task)) - goto out_unlock; - get_tracer(task); - error = -EPERM; if (tracer->ds.context->pebs_master) - goto out_put_tracer; + goto out_unlock; tracer->ds.context->pebs_master = tracer; spin_unlock_irqrestore(&ds_lock, irq); @@ -775,13 +781,13 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, return tracer; - out_put_tracer: - put_tracer(task); out_unlock: spin_unlock_irqrestore(&ds_lock, irq); ds_put_context(tracer->ds.context); out_tracer: kfree(tracer); + out_put_tracer: + put_tracer(task); out: return ERR_PTR(error); } @@ -804,8 +810,8 @@ void ds_release_bts(struct bts_tracer *tracer) if (task && (task != current)) wait_task_context_switch(task); - put_tracer(task); ds_put_context(tracer->ds.context); + put_tracer(task); kfree(tracer); } @@ -861,16 +867,20 @@ void ds_resume_bts(struct bts_tracer *tracer) void ds_release_pebs(struct pebs_tracer *tracer) { + struct task_struct *task; + if (!tracer) return; + task = tracer->ds.context->task; + ds_suspend_pebs(tracer); WARN_ON_ONCE(tracer->ds.context->pebs_master != tracer); tracer->ds.context->pebs_master = NULL; - put_tracer(tracer->ds.context->task); ds_put_context(tracer->ds.context); + put_tracer(task); kfree(tracer); } -- cgit v1.2.3 From 15879d042164650b93d83281ad5f87ad323bfbfe Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:38 +0200 Subject: x86, bts: use trace_clock_global() for timestamps Rename the bts_struct timestamp field to event. Use trace_clock_global() for time measurement. Reported-by: Ingo Molnar Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144553.773216000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/ds.h | 4 ++-- arch/x86/kernel/ds.c | 17 +++++++++-------- 2 files changed, 11 insertions(+), 10 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/ds.h b/arch/x86/include/asm/ds.h index a8f672ba100c..772f141afb9a 100644 --- a/arch/x86/include/asm/ds.h +++ b/arch/x86/include/asm/ds.h @@ -170,9 +170,9 @@ struct bts_struct { } lbr; /* BTS_TASK_ARRIVES or BTS_TASK_DEPARTS */ struct { - __u64 jiffies; + __u64 clock; pid_t pid; - } timestamp; + } event; } variant; }; diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index f03f117eff8c..2071b992c35c 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -25,6 +25,7 @@ #include #include #include +#include #include @@ -471,7 +472,7 @@ enum bts_field { bts_flags, bts_qual = bts_from, - bts_jiffies = bts_to, + bts_clock = bts_to, bts_pid = bts_flags, bts_qual_mask = (bts_qual_max - 1), @@ -517,8 +518,8 @@ bts_read(struct bts_tracer *tracer, const void *at, struct bts_struct *out) memset(out, 0, sizeof(*out)); if ((bts_get(at, bts_qual) & ~bts_qual_mask) == bts_escape) { out->qualifier = (bts_get(at, bts_qual) & bts_qual_mask); - out->variant.timestamp.jiffies = bts_get(at, bts_jiffies); - out->variant.timestamp.pid = bts_get(at, bts_pid); + out->variant.event.clock = bts_get(at, bts_clock); + out->variant.event.pid = bts_get(at, bts_pid); } else { out->qualifier = bts_branch; out->variant.lbr.from = bts_get(at, bts_from); @@ -555,8 +556,8 @@ static int bts_write(struct bts_tracer *tracer, const struct bts_struct *in) case bts_task_arrives: case bts_task_departs: bts_set(raw, bts_qual, (bts_escape | in->qualifier)); - bts_set(raw, bts_jiffies, in->variant.timestamp.jiffies); - bts_set(raw, bts_pid, in->variant.timestamp.pid); + bts_set(raw, bts_clock, in->variant.event.clock); + bts_set(raw, bts_pid, in->variant.event.pid); break; default: return -EINVAL; @@ -1083,9 +1084,9 @@ static inline void ds_take_timestamp(struct ds_context *context, return; memset(&ts, 0, sizeof(ts)); - ts.qualifier = qualifier; - ts.variant.timestamp.jiffies = jiffies_64; - ts.variant.timestamp.pid = task->pid; + ts.qualifier = qualifier; + ts.variant.event.clock = trace_clock_global(); + ts.variant.event.pid = task->pid; bts_write(tracer, &ts); } -- cgit v1.2.3 From 35bb7600c17762bb129588c1877d2717fe325289 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:39 +0200 Subject: x86, debugctlmsr: add _on_cpu variants to debugctlmsr functions Add functions to get and set the debugctlmsr on different cpus. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144554.738772000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/processor.h | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+) (limited to 'arch') diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h index 2483807e06e7..1efeb497f1f9 100644 --- a/arch/x86/include/asm/processor.h +++ b/arch/x86/include/asm/processor.h @@ -785,6 +785,21 @@ static inline unsigned long get_debugctlmsr(void) return debugctlmsr; } +static inline unsigned long get_debugctlmsr_on_cpu(int cpu) +{ + u64 debugctlmsr = 0; + u32 val1, val2; + +#ifndef CONFIG_X86_DEBUGCTLMSR + if (boot_cpu_data.x86 < 6) + return 0; +#endif + rdmsr_on_cpu(cpu, MSR_IA32_DEBUGCTLMSR, &val1, &val2); + debugctlmsr = val1 | ((u64)val2 << 32); + + return debugctlmsr; +} + static inline void update_debugctlmsr(unsigned long debugctlmsr) { #ifndef CONFIG_X86_DEBUGCTLMSR @@ -794,6 +809,18 @@ static inline void update_debugctlmsr(unsigned long debugctlmsr) wrmsrl(MSR_IA32_DEBUGCTLMSR, debugctlmsr); } +static inline void update_debugctlmsr_on_cpu(int cpu, + unsigned long debugctlmsr) +{ +#ifndef CONFIG_X86_DEBUGCTLMSR + if (boot_cpu_data.x86 < 6) + return; +#endif + wrmsr_on_cpu(cpu, MSR_IA32_DEBUGCTLMSR, + (u32)((u64)debugctlmsr), + (u32)((u64)debugctlmsr >> 32)); +} + /* * from system description table in BIOS. Mostly for MCA use, but * others may find it useful: -- cgit v1.2.3 From de79f54f5347ad7ec6ff55ccbb6d4ab2a21f6a93 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:40 +0200 Subject: x86, bts, hw-branch-tracer: add _noirq variants to the debug store interface The hw-branch-tracer uses debug store functions from an on_each_cpu() context, which is simply wrong since the functions may sleep. Add _noirq variants for most functions, which may be called with interrupts disabled. Separate per-cpu and per-task tracing and allow per-cpu tracing to be controlled from any cpu. Make the hw-branch-tracer use the new debug store interface, synchronize with hotplug cpu event using get/put_online_cpus(), and remove the unnecessary spinlock. Make the ptrace bts and the ds selftest code use the new interface. Defer the ds selftest. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144555.658136000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/ds.h | 57 ++++- arch/x86/kernel/ds.c | 474 ++++++++++++++++++++++++++++++--------- arch/x86/kernel/ds_selftest.c | 9 +- arch/x86/kernel/ptrace.c | 5 +- kernel/trace/trace_hw_branches.c | 193 ++++++---------- 5 files changed, 492 insertions(+), 246 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/ds.h b/arch/x86/include/asm/ds.h index 772f141afb9a..413e127e567d 100644 --- a/arch/x86/include/asm/ds.h +++ b/arch/x86/include/asm/ds.h @@ -15,8 +15,8 @@ * - buffer allocation (memory accounting) * * - * Copyright (C) 2007-2008 Intel Corporation. - * Markus Metzger , 2007-2008 + * Copyright (C) 2007-2009 Intel Corporation. + * Markus Metzger , 2007-2009 */ #ifndef _ASM_X86_DS_H @@ -83,8 +83,10 @@ enum ds_feature { * The interrupt threshold is independent from the overflow callback * to allow users to use their own overflow interrupt handling mechanism. * - * task: the task to request recording for; - * NULL for per-cpu recording on the current cpu + * The function might sleep. + * + * task: the task to request recording for + * cpu: the cpu to request recording for * base: the base pointer for the (non-pageable) buffer; * size: the size of the provided buffer in bytes * ovfl: pointer to a function to be called on buffer overflow; @@ -93,19 +95,28 @@ enum ds_feature { * -1 if no interrupt threshold is requested. * flags: a bit-mask of the above flags */ -extern struct bts_tracer *ds_request_bts(struct task_struct *task, - void *base, size_t size, - bts_ovfl_callback_t ovfl, - size_t th, unsigned int flags); -extern struct pebs_tracer *ds_request_pebs(struct task_struct *task, - void *base, size_t size, - pebs_ovfl_callback_t ovfl, - size_t th, unsigned int flags); +extern struct bts_tracer *ds_request_bts_task(struct task_struct *task, + void *base, size_t size, + bts_ovfl_callback_t ovfl, + size_t th, unsigned int flags); +extern struct bts_tracer *ds_request_bts_cpu(int cpu, void *base, size_t size, + bts_ovfl_callback_t ovfl, + size_t th, unsigned int flags); +extern struct pebs_tracer *ds_request_pebs_task(struct task_struct *task, + void *base, size_t size, + pebs_ovfl_callback_t ovfl, + size_t th, unsigned int flags); +extern struct pebs_tracer *ds_request_pebs_cpu(int cpu, + void *base, size_t size, + pebs_ovfl_callback_t ovfl, + size_t th, unsigned int flags); /* * Release BTS or PEBS resources * Suspend and resume BTS or PEBS tracing * + * Must be called with irq's enabled. + * * tracer: the tracer handle returned from ds_request_~() */ extern void ds_release_bts(struct bts_tracer *tracer); @@ -115,6 +126,28 @@ extern void ds_release_pebs(struct pebs_tracer *tracer); extern void ds_suspend_pebs(struct pebs_tracer *tracer); extern void ds_resume_pebs(struct pebs_tracer *tracer); +/* + * Release BTS or PEBS resources + * Suspend and resume BTS or PEBS tracing + * + * Cpu tracers must call this on the traced cpu. + * Task tracers must call ds_release_~_noirq() for themselves. + * + * May be called with irq's disabled. + * + * Returns 0 if successful; + * -EPERM if the cpu tracer does not trace the current cpu. + * -EPERM if the task tracer does not trace itself. + * + * tracer: the tracer handle returned from ds_request_~() + */ +extern int ds_release_bts_noirq(struct bts_tracer *tracer); +extern int ds_suspend_bts_noirq(struct bts_tracer *tracer); +extern int ds_resume_bts_noirq(struct bts_tracer *tracer); +extern int ds_release_pebs_noirq(struct pebs_tracer *tracer); +extern int ds_suspend_pebs_noirq(struct pebs_tracer *tracer); +extern int ds_resume_pebs_noirq(struct pebs_tracer *tracer); + /* * The raw DS buffer state as it is used for BTS and PEBS recording. diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 2071b992c35c..21a3852abf68 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -245,60 +245,50 @@ struct ds_context { struct pebs_tracer *pebs_master; /* Use count: */ - unsigned long count; + unsigned long count; /* Pointer to the context pointer field: */ struct ds_context **this; - /* The traced task; NULL for current cpu: */ + /* The traced task; NULL for cpu tracing: */ struct task_struct *task; -}; -static DEFINE_PER_CPU(struct ds_context *, system_context_array); + /* The traced cpu; only valid if task is NULL: */ + int cpu; +}; -#define system_context per_cpu(system_context_array, smp_processor_id()) +static DEFINE_PER_CPU(struct ds_context *, cpu_context); -static inline struct ds_context *ds_get_context(struct task_struct *task) +static struct ds_context *ds_get_context(struct task_struct *task, int cpu) { struct ds_context **p_context = - (task ? &task->thread.ds_ctx : &system_context); + (task ? &task->thread.ds_ctx : &per_cpu(cpu_context, cpu)); struct ds_context *context = NULL; struct ds_context *new_context = NULL; - unsigned long irq; - /* - * Chances are small that we already have a context. - * - * Contexts for per-cpu tracing are allocated using - * smp_call_function(). We must not sleep. - */ - new_context = kzalloc(sizeof(*new_context), GFP_ATOMIC); + /* Chances are small that we already have a context. */ + new_context = kzalloc(sizeof(*new_context), GFP_KERNEL); if (!new_context) return NULL; - spin_lock_irqsave(&ds_lock, irq); + spin_lock_irq(&ds_lock); context = *p_context; - if (!context) { + if (likely(!context)) { context = new_context; context->this = p_context; context->task = task; + context->cpu = cpu; context->count = 0; - if (task) - set_tsk_thread_flag(task, TIF_DS_AREA_MSR); - - if (!task || (task == current)) - wrmsrl(MSR_IA32_DS_AREA, (unsigned long)context->ds); - *p_context = context; } context->count++; - spin_unlock_irqrestore(&ds_lock, irq); + spin_unlock_irq(&ds_lock); if (context != new_context) kfree(new_context); @@ -306,7 +296,7 @@ static inline struct ds_context *ds_get_context(struct task_struct *task) return context; } -static inline void ds_put_context(struct ds_context *context) +static void ds_put_context(struct ds_context *context) { struct task_struct *task; unsigned long irq; @@ -328,8 +318,15 @@ static inline void ds_put_context(struct ds_context *context) if (task) clear_tsk_thread_flag(task, TIF_DS_AREA_MSR); - if (!task || (task == current)) - wrmsrl(MSR_IA32_DS_AREA, 0); + /* + * We leave the (now dangling) pointer to the DS configuration in + * the DS_AREA msr. This is as good or as bad as replacing it with + * NULL - the hardware would crash if we enabled tracing. + * + * This saves us some problems with having to write an msr on a + * different cpu while preventing others from doing the same for the + * next context for that same cpu. + */ spin_unlock_irqrestore(&ds_lock, irq); @@ -340,6 +337,31 @@ static inline void ds_put_context(struct ds_context *context) kfree(context); } +static void ds_install_ds_area(struct ds_context *context) +{ + unsigned long ds; + + ds = (unsigned long)context->ds; + + /* + * There is a race between the bts master and the pebs master. + * + * The thread/cpu access is synchronized via get/put_cpu() for + * task tracing and via wrmsr_on_cpu for cpu tracing. + * + * If bts and pebs are collected for the same task or same cpu, + * the same confiuration is written twice. + */ + if (context->task) { + get_cpu(); + if (context->task == current) + wrmsrl(MSR_IA32_DS_AREA, ds); + set_tsk_thread_flag(context->task, TIF_DS_AREA_MSR); + put_cpu(); + } else + wrmsr_on_cpu(context->cpu, MSR_IA32_DS_AREA, + (u32)((u64)ds), (u32)((u64)ds >> 32)); +} /* * Call the tracer's callback on a buffer overflow. @@ -622,6 +644,7 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual, * The value for 'no threshold' is -1, which will set the * threshold outside of the buffer, just like we want it. */ + ith *= ds_cfg.sizeof_rec[qual]; trace->ith = (void *)(buffer + size - ith); trace->flags = flags; @@ -630,7 +653,7 @@ static void ds_init_ds_trace(struct ds_trace *trace, enum ds_qualifier qual, static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, enum ds_qualifier qual, struct task_struct *task, - void *base, size_t size, size_t th, unsigned int flags) + int cpu, void *base, size_t size, size_t th) { struct ds_context *context; int error; @@ -643,7 +666,7 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, if (!base) goto out; - /* We require some space to do alignment adjustments below. */ + /* We need space for alignment adjustments in ds_init_ds_trace(). */ error = -EINVAL; if (size < (DS_ALIGNMENT + ds_cfg.sizeof_rec[qual])) goto out; @@ -660,25 +683,27 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, tracer->size = size; error = -ENOMEM; - context = ds_get_context(task); + context = ds_get_context(task, cpu); if (!context) goto out; tracer->context = context; - ds_init_ds_trace(trace, qual, base, size, th, flags); + /* + * Defer any tracer-specific initialization work for the context until + * context ownership has been clarified. + */ error = 0; out: return error; } -struct bts_tracer *ds_request_bts(struct task_struct *task, - void *base, size_t size, - bts_ovfl_callback_t ovfl, size_t th, - unsigned int flags) +static struct bts_tracer *ds_request_bts(struct task_struct *task, int cpu, + void *base, size_t size, + bts_ovfl_callback_t ovfl, size_t th, + unsigned int flags) { struct bts_tracer *tracer; - unsigned long irq; int error; /* Buffer overflow notification is not yet implemented. */ @@ -690,42 +715,46 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, if (error < 0) goto out; - /* - * Per-cpu tracing is typically requested using smp_call_function(). - * We must not sleep. - */ error = -ENOMEM; - tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); + tracer = kzalloc(sizeof(*tracer), GFP_KERNEL); if (!tracer) goto out_put_tracer; tracer->ovfl = ovfl; + /* Do some more error checking and acquire a tracing context. */ error = ds_request(&tracer->ds, &tracer->trace.ds, - ds_bts, task, base, size, th, flags); + ds_bts, task, cpu, base, size, th); if (error < 0) goto out_tracer; - - spin_lock_irqsave(&ds_lock, irq); + /* Claim the bts part of the tracing context we acquired above. */ + spin_lock_irq(&ds_lock); error = -EPERM; if (tracer->ds.context->bts_master) goto out_unlock; tracer->ds.context->bts_master = tracer; - spin_unlock_irqrestore(&ds_lock, irq); + spin_unlock_irq(&ds_lock); + /* + * Now that we own the bts part of the context, let's complete the + * initialization for that part. + */ + ds_init_ds_trace(&tracer->trace.ds, ds_bts, base, size, th, flags); + ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts); + ds_install_ds_area(tracer->ds.context); tracer->trace.read = bts_read; tracer->trace.write = bts_write; - ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_bts); + /* Start tracing. */ ds_resume_bts(tracer); return tracer; out_unlock: - spin_unlock_irqrestore(&ds_lock, irq); + spin_unlock_irq(&ds_lock); ds_put_context(tracer->ds.context); out_tracer: kfree(tracer); @@ -735,13 +764,27 @@ struct bts_tracer *ds_request_bts(struct task_struct *task, return ERR_PTR(error); } -struct pebs_tracer *ds_request_pebs(struct task_struct *task, - void *base, size_t size, - pebs_ovfl_callback_t ovfl, size_t th, - unsigned int flags) +struct bts_tracer *ds_request_bts_task(struct task_struct *task, + void *base, size_t size, + bts_ovfl_callback_t ovfl, + size_t th, unsigned int flags) +{ + return ds_request_bts(task, 0, base, size, ovfl, th, flags); +} + +struct bts_tracer *ds_request_bts_cpu(int cpu, void *base, size_t size, + bts_ovfl_callback_t ovfl, + size_t th, unsigned int flags) +{ + return ds_request_bts(NULL, cpu, base, size, ovfl, th, flags); +} + +static struct pebs_tracer *ds_request_pebs(struct task_struct *task, int cpu, + void *base, size_t size, + pebs_ovfl_callback_t ovfl, size_t th, + unsigned int flags) { struct pebs_tracer *tracer; - unsigned long irq; int error; /* Buffer overflow notification is not yet implemented. */ @@ -753,37 +796,43 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, if (error < 0) goto out; - /* - * Per-cpu tracing is typically requested using smp_call_function(). - * We must not sleep. - */ error = -ENOMEM; - tracer = kzalloc(sizeof(*tracer), GFP_ATOMIC); + tracer = kzalloc(sizeof(*tracer), GFP_KERNEL); if (!tracer) goto out_put_tracer; tracer->ovfl = ovfl; + /* Do some more error checking and acquire a tracing context. */ error = ds_request(&tracer->ds, &tracer->trace.ds, - ds_pebs, task, base, size, th, flags); + ds_pebs, task, cpu, base, size, th); if (error < 0) goto out_tracer; - spin_lock_irqsave(&ds_lock, irq); + /* Claim the pebs part of the tracing context we acquired above. */ + spin_lock_irq(&ds_lock); error = -EPERM; if (tracer->ds.context->pebs_master) goto out_unlock; tracer->ds.context->pebs_master = tracer; - spin_unlock_irqrestore(&ds_lock, irq); + spin_unlock_irq(&ds_lock); + /* + * Now that we own the pebs part of the context, let's complete the + * initialization for that part. + */ + ds_init_ds_trace(&tracer->trace.ds, ds_pebs, base, size, th, flags); ds_write_config(tracer->ds.context, &tracer->trace.ds, ds_pebs); + ds_install_ds_area(tracer->ds.context); + + /* Start tracing. */ ds_resume_pebs(tracer); return tracer; out_unlock: - spin_unlock_irqrestore(&ds_lock, irq); + spin_unlock_irq(&ds_lock); ds_put_context(tracer->ds.context); out_tracer: kfree(tracer); @@ -793,16 +842,26 @@ struct pebs_tracer *ds_request_pebs(struct task_struct *task, return ERR_PTR(error); } -void ds_release_bts(struct bts_tracer *tracer) +struct pebs_tracer *ds_request_pebs_task(struct task_struct *task, + void *base, size_t size, + pebs_ovfl_callback_t ovfl, + size_t th, unsigned int flags) { - struct task_struct *task; + return ds_request_pebs(task, 0, base, size, ovfl, th, flags); +} - if (!tracer) - return; +struct pebs_tracer *ds_request_pebs_cpu(int cpu, void *base, size_t size, + pebs_ovfl_callback_t ovfl, + size_t th, unsigned int flags) +{ + return ds_request_pebs(NULL, cpu, base, size, ovfl, th, flags); +} - task = tracer->ds.context->task; +static void ds_free_bts(struct bts_tracer *tracer) +{ + struct task_struct *task; - ds_suspend_bts(tracer); + task = tracer->ds.context->task; WARN_ON_ONCE(tracer->ds.context->bts_master != tracer); tracer->ds.context->bts_master = NULL; @@ -817,9 +876,69 @@ void ds_release_bts(struct bts_tracer *tracer) kfree(tracer); } +void ds_release_bts(struct bts_tracer *tracer) +{ + might_sleep(); + + if (!tracer) + return; + + ds_suspend_bts(tracer); + ds_free_bts(tracer); +} + +int ds_release_bts_noirq(struct bts_tracer *tracer) +{ + struct task_struct *task; + unsigned long irq; + int error; + + if (!tracer) + return 0; + + task = tracer->ds.context->task; + + local_irq_save(irq); + + error = -EPERM; + if (!task && + (tracer->ds.context->cpu != smp_processor_id())) + goto out; + + error = -EPERM; + if (task && (task != current)) + goto out; + + ds_suspend_bts_noirq(tracer); + ds_free_bts(tracer); + + error = 0; + out: + local_irq_restore(irq); + return error; +} + +static void update_task_debugctlmsr(struct task_struct *task, + unsigned long debugctlmsr) +{ + task->thread.debugctlmsr = debugctlmsr; + + get_cpu(); + if (task == current) + update_debugctlmsr(debugctlmsr); + + if (task->thread.debugctlmsr) + set_tsk_thread_flag(task, TIF_DEBUGCTLMSR); + else + clear_tsk_thread_flag(task, TIF_DEBUGCTLMSR); + put_cpu(); +} + void ds_suspend_bts(struct bts_tracer *tracer) { struct task_struct *task; + unsigned long debugctlmsr; + int cpu; if (!tracer) return; @@ -827,29 +946,60 @@ void ds_suspend_bts(struct bts_tracer *tracer) tracer->flags = 0; task = tracer->ds.context->task; + cpu = tracer->ds.context->cpu; - if (!task || (task == current)) - update_debugctlmsr(get_debugctlmsr() & ~BTS_CONTROL); + WARN_ON(!task && irqs_disabled()); - if (task) { - task->thread.debugctlmsr &= ~BTS_CONTROL; + debugctlmsr = (task ? + task->thread.debugctlmsr : + get_debugctlmsr_on_cpu(cpu)); + debugctlmsr &= ~BTS_CONTROL; - if (!task->thread.debugctlmsr) - clear_tsk_thread_flag(task, TIF_DEBUGCTLMSR); - } + if (task) + update_task_debugctlmsr(task, debugctlmsr); + else + update_debugctlmsr_on_cpu(cpu, debugctlmsr); } -void ds_resume_bts(struct bts_tracer *tracer) +int ds_suspend_bts_noirq(struct bts_tracer *tracer) { struct task_struct *task; - unsigned long control; + unsigned long debugctlmsr, irq; + int cpu, error = 0; if (!tracer) - return; + return 0; - tracer->flags = tracer->trace.ds.flags; + tracer->flags = 0; task = tracer->ds.context->task; + cpu = tracer->ds.context->cpu; + + local_irq_save(irq); + + error = -EPERM; + if (!task && (cpu != smp_processor_id())) + goto out; + + debugctlmsr = (task ? + task->thread.debugctlmsr : + get_debugctlmsr()); + debugctlmsr &= ~BTS_CONTROL; + + if (task) + update_task_debugctlmsr(task, debugctlmsr); + else + update_debugctlmsr(debugctlmsr); + + error = 0; + out: + local_irq_restore(irq); + return error; +} + +static unsigned long ds_bts_control(struct bts_tracer *tracer) +{ + unsigned long control; control = ds_cfg.ctl[dsf_bts]; if (!(tracer->trace.ds.flags & BTS_KERNEL)) @@ -857,25 +1007,77 @@ void ds_resume_bts(struct bts_tracer *tracer) if (!(tracer->trace.ds.flags & BTS_USER)) control |= ds_cfg.ctl[dsf_bts_user]; - if (task) { - task->thread.debugctlmsr |= control; - set_tsk_thread_flag(task, TIF_DEBUGCTLMSR); - } - - if (!task || (task == current)) - update_debugctlmsr(get_debugctlmsr() | control); + return control; } -void ds_release_pebs(struct pebs_tracer *tracer) +void ds_resume_bts(struct bts_tracer *tracer) { struct task_struct *task; + unsigned long debugctlmsr; + int cpu; if (!tracer) return; + tracer->flags = tracer->trace.ds.flags; + task = tracer->ds.context->task; + cpu = tracer->ds.context->cpu; - ds_suspend_pebs(tracer); + WARN_ON(!task && irqs_disabled()); + + debugctlmsr = (task ? + task->thread.debugctlmsr : + get_debugctlmsr_on_cpu(cpu)); + debugctlmsr |= ds_bts_control(tracer); + + if (task) + update_task_debugctlmsr(task, debugctlmsr); + else + update_debugctlmsr_on_cpu(cpu, debugctlmsr); +} + +int ds_resume_bts_noirq(struct bts_tracer *tracer) +{ + struct task_struct *task; + unsigned long debugctlmsr, irq; + int cpu, error = 0; + + if (!tracer) + return 0; + + tracer->flags = tracer->trace.ds.flags; + + task = tracer->ds.context->task; + cpu = tracer->ds.context->cpu; + + local_irq_save(irq); + + error = -EPERM; + if (!task && (cpu != smp_processor_id())) + goto out; + + debugctlmsr = (task ? + task->thread.debugctlmsr : + get_debugctlmsr()); + debugctlmsr |= ds_bts_control(tracer); + + if (task) + update_task_debugctlmsr(task, debugctlmsr); + else + update_debugctlmsr(debugctlmsr); + + error = 0; + out: + local_irq_restore(irq); + return error; +} + +static void ds_free_pebs(struct pebs_tracer *tracer) +{ + struct task_struct *task; + + task = tracer->ds.context->task; WARN_ON_ONCE(tracer->ds.context->pebs_master != tracer); tracer->ds.context->pebs_master = NULL; @@ -886,16 +1088,68 @@ void ds_release_pebs(struct pebs_tracer *tracer) kfree(tracer); } +void ds_release_pebs(struct pebs_tracer *tracer) +{ + might_sleep(); + + if (!tracer) + return; + + ds_suspend_pebs(tracer); + ds_free_pebs(tracer); +} + +int ds_release_pebs_noirq(struct pebs_tracer *tracer) +{ + struct task_struct *task; + unsigned long irq; + int error; + + if (!tracer) + return 0; + + task = tracer->ds.context->task; + + local_irq_save(irq); + + error = -EPERM; + if (!task && + (tracer->ds.context->cpu != smp_processor_id())) + goto out; + + error = -EPERM; + if (task && (task != current)) + goto out; + + ds_suspend_pebs_noirq(tracer); + ds_free_pebs(tracer); + + error = 0; + out: + local_irq_restore(irq); + return error; +} + void ds_suspend_pebs(struct pebs_tracer *tracer) { } +int ds_suspend_pebs_noirq(struct pebs_tracer *tracer) +{ + return 0; +} + void ds_resume_pebs(struct pebs_tracer *tracer) { } +int ds_resume_pebs_noirq(struct pebs_tracer *tracer) +{ + return 0; +} + const struct bts_trace *ds_read_bts(struct bts_tracer *tracer) { if (!tracer) @@ -1004,26 +1258,6 @@ ds_configure(const struct ds_configuration *cfg, printk(KERN_INFO "[ds] pebs not available\n"); } - if (ds_cfg.sizeof_rec[ds_bts]) { - int error; - - error = ds_selftest_bts(); - if (error) { - WARN(1, "[ds] selftest failed. disabling bts.\n"); - ds_cfg.sizeof_rec[ds_bts] = 0; - } - } - - if (ds_cfg.sizeof_rec[ds_pebs]) { - int error; - - error = ds_selftest_pebs(); - if (error) { - WARN(1, "[ds] selftest failed. disabling pebs.\n"); - ds_cfg.sizeof_rec[ds_pebs] = 0; - } - } - printk(KERN_INFO "[ds] sizes: address: %u bit, ", 8 * ds_cfg.sizeof_ptr_field); printk("bts/pebs record: %u/%u bytes\n", @@ -1127,3 +1361,29 @@ void ds_copy_thread(struct task_struct *tsk, struct task_struct *father) void ds_exit_thread(struct task_struct *tsk) { } + +static __init int ds_selftest(void) +{ + if (ds_cfg.sizeof_rec[ds_bts]) { + int error; + + error = ds_selftest_bts(); + if (error) { + WARN(1, "[ds] selftest failed. disabling bts.\n"); + ds_cfg.sizeof_rec[ds_bts] = 0; + } + } + + if (ds_cfg.sizeof_rec[ds_pebs]) { + int error; + + error = ds_selftest_pebs(); + if (error) { + WARN(1, "[ds] selftest failed. disabling pebs.\n"); + ds_cfg.sizeof_rec[ds_pebs] = 0; + } + } + + return 0; +} +device_initcall(ds_selftest); diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index 8c46fbf38c46..e5a263c8a14c 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -10,11 +10,12 @@ #include #include +#include #include -#define DS_SELFTEST_BUFFER_SIZE 1021 /* Intentionally chose an odd size. */ +#define BUFFER_SIZE 1021 /* Intentionally chose an odd size. */ static int ds_selftest_bts_consistency(const struct bts_trace *trace) @@ -125,12 +126,12 @@ int ds_selftest_bts(void) struct bts_tracer *tracer; int error = 0; void *top; - unsigned char buffer[DS_SELFTEST_BUFFER_SIZE]; + unsigned char buffer[BUFFER_SIZE]; printk(KERN_INFO "[ds] bts selftest..."); - tracer = ds_request_bts(NULL, buffer, DS_SELFTEST_BUFFER_SIZE, - NULL, (size_t)-1, BTS_KERNEL); + tracer = ds_request_bts_cpu(smp_processor_id(), buffer, BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); if (IS_ERR(tracer)) { error = PTR_ERR(tracer); tracer = NULL; diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c index 7c21d1e8cae7..adbb24322d8f 100644 --- a/arch/x86/kernel/ptrace.c +++ b/arch/x86/kernel/ptrace.c @@ -800,8 +800,9 @@ static int ptrace_bts_config(struct task_struct *child, if (cfg.flags & PTRACE_BTS_O_SCHED) flags |= BTS_TIMESTAMPS; - context->tracer = ds_request_bts(child, context->buffer, context->size, - NULL, (size_t)-1, flags); + context->tracer = + ds_request_bts_task(child, context->buffer, context->size, + NULL, (size_t)-1, flags); if (unlikely(IS_ERR(context->tracer))) { int error = PTR_ERR(context->tracer); diff --git a/kernel/trace/trace_hw_branches.c b/kernel/trace/trace_hw_branches.c index 8b2109a6c61c..50565d8cd2ed 100644 --- a/kernel/trace/trace_hw_branches.c +++ b/kernel/trace/trace_hw_branches.c @@ -4,7 +4,6 @@ * Copyright (C) 2008-2009 Intel Corporation. * Markus Metzger , 2008-2009 */ -#include #include #include #include @@ -21,168 +20,113 @@ #define BTS_BUFFER_SIZE (1 << 13) -/* - * The tracer lock protects the below per-cpu tracer array. - * It needs to be held to: - * - start tracing on all cpus - * - stop tracing on all cpus - * - start tracing on a single hotplug cpu - * - stop tracing on a single hotplug cpu - * - read the trace from all cpus - * - read the trace from a single cpu - */ -static DEFINE_SPINLOCK(bts_tracer_lock); static DEFINE_PER_CPU(struct bts_tracer *, tracer); static DEFINE_PER_CPU(unsigned char[BTS_BUFFER_SIZE], buffer); #define this_tracer per_cpu(tracer, smp_processor_id()) -#define this_buffer per_cpu(buffer, smp_processor_id()) static int trace_hw_branches_enabled __read_mostly; static int trace_hw_branches_suspended __read_mostly; static struct trace_array *hw_branch_trace __read_mostly; -/* - * Initialize the tracer for the current cpu. - * The argument is ignored. - * - * pre: bts_tracer_lock must be locked. - */ -static void bts_trace_init_cpu(void *arg) +static void bts_trace_init_cpu(int cpu) { - if (this_tracer) - ds_release_bts(this_tracer); + per_cpu(tracer, cpu) = + ds_request_bts_cpu(cpu, per_cpu(buffer, cpu), BTS_BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); - this_tracer = ds_request_bts(NULL, this_buffer, BTS_BUFFER_SIZE, - NULL, (size_t)-1, BTS_KERNEL); - if (IS_ERR(this_tracer)) { - this_tracer = NULL; - return; - } + if (IS_ERR(per_cpu(tracer, cpu))) + per_cpu(tracer, cpu) = NULL; } static int bts_trace_init(struct trace_array *tr) { - int cpu, avail; - - spin_lock(&bts_tracer_lock); + int cpu; hw_branch_trace = tr; + trace_hw_branches_enabled = 0; - on_each_cpu(bts_trace_init_cpu, NULL, 1); - - /* Check on how many cpus we could enable tracing */ - avail = 0; - for_each_online_cpu(cpu) - if (per_cpu(tracer, cpu)) - avail++; + get_online_cpus(); + for_each_online_cpu(cpu) { + bts_trace_init_cpu(cpu); - trace_hw_branches_enabled = (avail ? 1 : 0); + if (likely(per_cpu(tracer, cpu))) + trace_hw_branches_enabled = 1; + } trace_hw_branches_suspended = 0; - - spin_unlock(&bts_tracer_lock); - + put_online_cpus(); /* If we could not enable tracing on a single cpu, we fail. */ - return avail ? 0 : -EOPNOTSUPP; -} - -/* - * Release the tracer for the current cpu. - * The argument is ignored. - * - * pre: bts_tracer_lock must be locked. - */ -static void bts_trace_release_cpu(void *arg) -{ - if (this_tracer) { - ds_release_bts(this_tracer); - this_tracer = NULL; - } + return trace_hw_branches_enabled ? 0 : -EOPNOTSUPP; } static void bts_trace_reset(struct trace_array *tr) { - spin_lock(&bts_tracer_lock); + int cpu; - on_each_cpu(bts_trace_release_cpu, NULL, 1); + get_online_cpus(); + for_each_online_cpu(cpu) { + if (likely(per_cpu(tracer, cpu))) { + ds_release_bts(per_cpu(tracer, cpu)); + per_cpu(tracer, cpu) = NULL; + } + } trace_hw_branches_enabled = 0; trace_hw_branches_suspended = 0; - - spin_unlock(&bts_tracer_lock); -} - -/* - * Resume tracing on the current cpu. - * The argument is ignored. - * - * pre: bts_tracer_lock must be locked. - */ -static void bts_trace_resume_cpu(void *arg) -{ - if (this_tracer) - ds_resume_bts(this_tracer); + put_online_cpus(); } static void bts_trace_start(struct trace_array *tr) { - spin_lock(&bts_tracer_lock); + int cpu; - on_each_cpu(bts_trace_resume_cpu, NULL, 1); + get_online_cpus(); + for_each_online_cpu(cpu) + if (likely(per_cpu(tracer, cpu))) + ds_resume_bts(per_cpu(tracer, cpu)); trace_hw_branches_suspended = 0; - - spin_unlock(&bts_tracer_lock); -} - -/* - * Suspend tracing on the current cpu. - * The argument is ignored. - * - * pre: bts_tracer_lock must be locked. - */ -static void bts_trace_suspend_cpu(void *arg) -{ - if (this_tracer) - ds_suspend_bts(this_tracer); + put_online_cpus(); } static void bts_trace_stop(struct trace_array *tr) { - spin_lock(&bts_tracer_lock); + int cpu; - on_each_cpu(bts_trace_suspend_cpu, NULL, 1); + get_online_cpus(); + for_each_online_cpu(cpu) + if (likely(per_cpu(tracer, cpu))) + ds_suspend_bts(per_cpu(tracer, cpu)); trace_hw_branches_suspended = 1; - - spin_unlock(&bts_tracer_lock); + put_online_cpus(); } static int __cpuinit bts_hotcpu_handler(struct notifier_block *nfb, unsigned long action, void *hcpu) { - unsigned int cpu = (unsigned long)hcpu; - - spin_lock(&bts_tracer_lock); - - if (!trace_hw_branches_enabled) - goto out; + int cpu = (long)hcpu; switch (action) { case CPU_ONLINE: case CPU_DOWN_FAILED: - smp_call_function_single(cpu, bts_trace_init_cpu, NULL, 1); - - if (trace_hw_branches_suspended) - smp_call_function_single(cpu, bts_trace_suspend_cpu, - NULL, 1); + /* The notification is sent with interrupts enabled. */ + if (trace_hw_branches_enabled) { + bts_trace_init_cpu(cpu); + + if (trace_hw_branches_suspended && + likely(per_cpu(tracer, cpu))) + ds_suspend_bts(per_cpu(tracer, cpu)); + } break; + case CPU_DOWN_PREPARE: - smp_call_function_single(cpu, bts_trace_release_cpu, NULL, 1); - break; + /* The notification is sent with interrupts enabled. */ + if (likely(per_cpu(tracer, cpu))) { + ds_release_bts(per_cpu(tracer, cpu)); + per_cpu(tracer, cpu) = NULL; + } } - out: - spin_unlock(&bts_tracer_lock); return NOTIFY_DONE; } @@ -274,7 +218,7 @@ static void trace_bts_at(const struct bts_trace *trace, void *at) /* * Collect the trace on the current cpu and write it into the ftrace buffer. * - * pre: bts_tracer_lock must be locked + * pre: tracing must be suspended on the current cpu */ static void trace_bts_cpu(void *arg) { @@ -291,10 +235,9 @@ static void trace_bts_cpu(void *arg) if (unlikely(!this_tracer)) return; - ds_suspend_bts(this_tracer); trace = ds_read_bts(this_tracer); if (!trace) - goto out; + return; for (at = trace->ds.top; (void *)at < trace->ds.end; at += trace->ds.size) @@ -303,18 +246,27 @@ static void trace_bts_cpu(void *arg) for (at = trace->ds.begin; (void *)at < trace->ds.top; at += trace->ds.size) trace_bts_at(trace, at); - -out: - ds_resume_bts(this_tracer); } static void trace_bts_prepare(struct trace_iterator *iter) { - spin_lock(&bts_tracer_lock); + int cpu; + get_online_cpus(); + for_each_online_cpu(cpu) + if (likely(per_cpu(tracer, cpu))) + ds_suspend_bts(per_cpu(tracer, cpu)); + /* + * We need to collect the trace on the respective cpu since ftrace + * implicitly adds the record for the current cpu. + * Once that is more flexible, we could collect the data from any cpu. + */ on_each_cpu(trace_bts_cpu, iter->tr, 1); - spin_unlock(&bts_tracer_lock); + for_each_online_cpu(cpu) + if (likely(per_cpu(tracer, cpu))) + ds_resume_bts(per_cpu(tracer, cpu)); + put_online_cpus(); } static void trace_bts_close(struct trace_iterator *iter) @@ -324,12 +276,11 @@ static void trace_bts_close(struct trace_iterator *iter) void trace_hw_branch_oops(void) { - spin_lock(&bts_tracer_lock); - - if (trace_hw_branches_enabled) + if (this_tracer) { + ds_suspend_bts_noirq(this_tracer); trace_bts_cpu(hw_branch_trace); - - spin_unlock(&bts_tracer_lock); + ds_resume_bts_noirq(this_tracer); + } } struct tracer bts_tracer __read_mostly = -- cgit v1.2.3 From 353afeea24cc51aafc0ff21a72ec740b6f0af50c Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:42 +0200 Subject: x86, ds: fix compiler warning Size_t is defined differently on i386 and x86_64. Change type to avoid compiler warning. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144557.523964000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds_selftest.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index e5a263c8a14c..e1ba5101b576 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -87,7 +87,7 @@ static int ds_selftest_bts_read(struct bts_tracer *tracer, /* Now to the test itself. */ for (at = from; (void *)at < to; at += trace->ds.size) { struct bts_struct bts; - size_t index; + unsigned long index; int error; if (((void *)at - trace->ds.begin) % trace->ds.size) { -- cgit v1.2.3 From 84f201139245c30777ff858e71b8d7e134b8c3ed Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:43 +0200 Subject: x86, ds: fix bounds check in ds selftest Fix a bad bounds check in the debug store selftest. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144558.450027000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds_selftest.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index e1ba5101b576..cccc19a38f6d 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -47,8 +47,13 @@ static int ds_selftest_bts_consistency(const struct bts_trace *trace) printk(KERN_CONT "bad bts buffer setup..."); error = -1; } + /* + * We allow top in [begin; end], since its not clear when the + * overflow adjustment happens: after the increment or before the + * write. + */ if ((trace->ds.top < trace->ds.begin) || - (trace->ds.end <= trace->ds.top)) { + (trace->ds.end < trace->ds.top)) { printk(KERN_CONT "bts top out of bounds..."); error = -1; } -- cgit v1.2.3 From 01f6569ece6915616f6cae1d7d8b46ab8da9c1bd Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:44 +0200 Subject: x86, ds: selftest each cpu Perform debug store selftests on each cpu. Cover both the normal and the _noirq variant of the debug store interface. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144559.394583000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds_selftest.c | 182 +++++++++++++++++++++++++++++++----------- 1 file changed, 135 insertions(+), 47 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index cccc19a38f6d..599a96300628 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -11,13 +11,21 @@ #include #include #include +#include #include -#define BUFFER_SIZE 1021 /* Intentionally chose an odd size. */ +#define BUFFER_SIZE 521 /* Intentionally chose an odd size. */ +struct ds_selftest_bts_conf { + struct bts_tracer *tracer; + int error; + int (*suspend)(struct bts_tracer *); + int (*resume)(struct bts_tracer *); +}; + static int ds_selftest_bts_consistency(const struct bts_trace *trace) { int error = 0; @@ -125,36 +133,32 @@ static int ds_selftest_bts_read(struct bts_tracer *tracer, return 0; } -int ds_selftest_bts(void) +static void ds_selftest_bts_cpu(void *arg) { + struct ds_selftest_bts_conf *conf = arg; const struct bts_trace *trace; - struct bts_tracer *tracer; - int error = 0; void *top; - unsigned char buffer[BUFFER_SIZE]; - printk(KERN_INFO "[ds] bts selftest..."); - - tracer = ds_request_bts_cpu(smp_processor_id(), buffer, BUFFER_SIZE, - NULL, (size_t)-1, BTS_KERNEL); - if (IS_ERR(tracer)) { - error = PTR_ERR(tracer); - tracer = NULL; + if (IS_ERR(conf->tracer)) { + conf->error = PTR_ERR(conf->tracer); + conf->tracer = NULL; printk(KERN_CONT - "initialization failed (err: %d)...", error); - goto out; + "initialization failed (err: %d)...", conf->error); + return; } - /* The return should already give us enough trace. */ - ds_suspend_bts(tracer); + /* We should meanwhile have enough trace. */ + conf->error = conf->suspend(conf->tracer); + if (conf->error < 0) + return; /* Let's see if we can access the trace. */ - trace = ds_read_bts(tracer); + trace = ds_read_bts(conf->tracer); - error = ds_selftest_bts_consistency(trace); - if (error < 0) - goto out; + conf->error = ds_selftest_bts_consistency(trace); + if (conf->error < 0) + return; /* If everything went well, we should have a few trace entries. */ if (trace->ds.top == trace->ds.begin) { @@ -168,10 +172,11 @@ int ds_selftest_bts(void) } /* Let's try to read the trace we collected. */ - error = ds_selftest_bts_read(tracer, trace, + conf->error = + ds_selftest_bts_read(conf->tracer, trace, trace->ds.begin, trace->ds.top); - if (error < 0) - goto out; + if (conf->error < 0) + return; /* * Let's read the trace again. @@ -179,26 +184,31 @@ int ds_selftest_bts(void) */ top = trace->ds.top; - trace = ds_read_bts(tracer); - error = ds_selftest_bts_consistency(trace); - if (error < 0) - goto out; + trace = ds_read_bts(conf->tracer); + conf->error = ds_selftest_bts_consistency(trace); + if (conf->error < 0) + return; if (top != trace->ds.top) { printk(KERN_CONT "suspend not working..."); - error = -1; - goto out; + conf->error = -1; + return; } /* Let's collect some more trace - see if resume is working. */ - ds_resume_bts(tracer); - ds_suspend_bts(tracer); + conf->error = conf->resume(conf->tracer); + if (conf->error < 0) + return; + + conf->error = conf->suspend(conf->tracer); + if (conf->error < 0) + return; - trace = ds_read_bts(tracer); + trace = ds_read_bts(conf->tracer); - error = ds_selftest_bts_consistency(trace); - if (error < 0) - goto out; + conf->error = ds_selftest_bts_consistency(trace); + if (conf->error < 0) + return; if (trace->ds.top == top) { /* @@ -210,35 +220,113 @@ int ds_selftest_bts(void) printk(KERN_CONT "no resume progress/overflow..."); - error = ds_selftest_bts_read(tracer, trace, + conf->error = + ds_selftest_bts_read(conf->tracer, trace, trace->ds.begin, trace->ds.end); } else if (trace->ds.top < top) { /* * We had a buffer overflow - the entire buffer should * contain trace records. */ - error = ds_selftest_bts_read(tracer, trace, + conf->error = + ds_selftest_bts_read(conf->tracer, trace, trace->ds.begin, trace->ds.end); } else { /* * It is quite likely that the buffer did not overflow. * Let's just check the delta trace. */ - error = ds_selftest_bts_read(tracer, trace, - top, trace->ds.top); + conf->error = + ds_selftest_bts_read(conf->tracer, trace, top, + trace->ds.top); } - if (error < 0) - goto out; + if (conf->error < 0) + return; - error = 0; + conf->error = 0; +} - /* The final test: release the tracer while tracing is suspended. */ - out: - ds_release_bts(tracer); +static int ds_suspend_bts_wrap(struct bts_tracer *tracer) +{ + ds_suspend_bts(tracer); + return 0; +} + +static int ds_resume_bts_wrap(struct bts_tracer *tracer) +{ + ds_resume_bts(tracer); + return 0; +} - printk(KERN_CONT "%s.\n", (error ? "failed" : "passed")); +static void ds_release_bts_noirq_wrap(void *tracer) +{ + (void)ds_release_bts_noirq(tracer); +} - return error; +static int ds_selftest_bts_bad_release_noirq(int cpu, + struct bts_tracer *tracer) +{ + int error = -EPERM; + + /* Try to release the tracer on the wrong cpu. */ + get_cpu(); + if (cpu != smp_processor_id()) { + error = ds_release_bts_noirq(tracer); + if (error != -EPERM) + printk(KERN_CONT "release on wrong cpu..."); + } + put_cpu(); + + return error ? 0 : -1; +} + +int ds_selftest_bts(void) +{ + struct ds_selftest_bts_conf conf; + unsigned char buffer[BUFFER_SIZE]; + int cpu; + + printk(KERN_INFO "[ds] bts selftest..."); + conf.error = 0; + + get_online_cpus(); + for_each_online_cpu(cpu) { + conf.suspend = ds_suspend_bts_wrap; + conf.resume = ds_resume_bts_wrap; + conf.tracer = + ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); + ds_selftest_bts_cpu(&conf); + ds_release_bts(conf.tracer); + if (conf.error < 0) + goto out; + + conf.suspend = ds_suspend_bts_noirq; + conf.resume = ds_resume_bts_noirq; + conf.tracer = + ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); + smp_call_function_single(cpu, ds_selftest_bts_cpu, &conf, 1); + if (conf.error >= 0) { + conf.error = + ds_selftest_bts_bad_release_noirq(cpu, + conf.tracer); + /* We must not release the tracer twice. */ + if (conf.error < 0) + conf.tracer = NULL; + } + smp_call_function_single(cpu, ds_release_bts_noirq_wrap, + conf.tracer, 1); + if (conf.error < 0) + goto out; + } + + conf.error = 0; + out: + put_online_cpus(); + printk(KERN_CONT "%s.\n", (conf.error ? "failed" : "passed")); + + return conf.error; } int ds_selftest_pebs(void) -- cgit v1.2.3 From 3a68eef945b234f286406d96dc690fe17863c203 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:45 +0200 Subject: x86, ds: add task tracing selftest Add selftests to cover per-task branch tracing. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144600.329346000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds_selftest.c | 71 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 71 insertions(+) (limited to 'arch') diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index 599a96300628..a40b2533c71e 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -280,10 +280,51 @@ static int ds_selftest_bts_bad_release_noirq(int cpu, return error ? 0 : -1; } +static int ds_selftest_bts_bad_request_cpu(int cpu, void *buffer) +{ + struct bts_tracer *tracer; + int error; + + /* Try to request cpu tracing while task tracing is active. */ + tracer = ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE, NULL, + (size_t)-1, BTS_KERNEL); + error = PTR_ERR(tracer); + if (!IS_ERR(tracer)) { + ds_release_bts(tracer); + error = 0; + } + + if (error != -EPERM) + printk(KERN_CONT "cpu/task tracing overlap..."); + + return error ? 0 : -1; +} + +static int ds_selftest_bts_bad_request_task(void *buffer) +{ + struct bts_tracer *tracer; + int error; + + /* Try to request cpu tracing while task tracing is active. */ + tracer = ds_request_bts_task(current, buffer, BUFFER_SIZE, NULL, + (size_t)-1, BTS_KERNEL); + error = PTR_ERR(tracer); + if (!IS_ERR(tracer)) { + error = 0; + ds_release_bts(tracer); + } + + if (error != -EPERM) + printk(KERN_CONT "task/cpu tracing overlap..."); + + return error ? 0 : -1; +} + int ds_selftest_bts(void) { struct ds_selftest_bts_conf conf; unsigned char buffer[BUFFER_SIZE]; + unsigned long irq; int cpu; printk(KERN_INFO "[ds] bts selftest..."); @@ -297,6 +338,8 @@ int ds_selftest_bts(void) ds_request_bts_cpu(cpu, buffer, BUFFER_SIZE, NULL, (size_t)-1, BTS_KERNEL); ds_selftest_bts_cpu(&conf); + if (conf.error >= 0) + conf.error = ds_selftest_bts_bad_request_task(buffer); ds_release_bts(conf.tracer); if (conf.error < 0) goto out; @@ -315,12 +358,40 @@ int ds_selftest_bts(void) if (conf.error < 0) conf.tracer = NULL; } + if (conf.error >= 0) + conf.error = ds_selftest_bts_bad_request_task(buffer); smp_call_function_single(cpu, ds_release_bts_noirq_wrap, conf.tracer, 1); if (conf.error < 0) goto out; } + conf.suspend = ds_suspend_bts_wrap; + conf.resume = ds_resume_bts_wrap; + conf.tracer = + ds_request_bts_task(current, buffer, BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); + ds_selftest_bts_cpu(&conf); + if (conf.error >= 0) + conf.error = ds_selftest_bts_bad_request_cpu(0, buffer); + ds_release_bts(conf.tracer); + if (conf.error < 0) + goto out; + + conf.suspend = ds_suspend_bts_noirq; + conf.resume = ds_resume_bts_noirq; + conf.tracer = + ds_request_bts_task(current, buffer, BUFFER_SIZE, + NULL, (size_t)-1, BTS_KERNEL); + local_irq_save(irq); + ds_selftest_bts_cpu(&conf); + if (conf.error >= 0) + conf.error = ds_selftest_bts_bad_request_cpu(0, buffer); + ds_release_bts_noirq(conf.tracer); + local_irq_restore(irq); + if (conf.error < 0) + goto out; + conf.error = 0; out: put_online_cpus(); -- cgit v1.2.3 From 2311f0de21c17b2a8b960677a9cccfbfa52beb35 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:46 +0200 Subject: x86, ds: add leakage warning Add a warning in case a debug store context is not removed before the task it is attached to is freed. Remove the old warning at thread exit. It is too early. Declare the debug store context field in thread_struct unconditionally. Remove ds_copy_thread() and ds_exit_thread() and do the work directly in process*.c. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144601.254472000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/ds.h | 9 --------- arch/x86/include/asm/processor.h | 4 +--- arch/x86/kernel/ds.c | 10 ---------- arch/x86/kernel/process.c | 5 +++-- arch/x86/kernel/process_32.c | 3 ++- arch/x86/kernel/process_64.c | 3 ++- 6 files changed, 8 insertions(+), 26 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/ds.h b/arch/x86/include/asm/ds.h index 413e127e567d..149e5208e967 100644 --- a/arch/x86/include/asm/ds.h +++ b/arch/x86/include/asm/ds.h @@ -285,21 +285,12 @@ extern void __cpuinit ds_init_intel(struct cpuinfo_x86 *); */ extern void ds_switch_to(struct task_struct *prev, struct task_struct *next); -/* - * Task clone/init and cleanup work - */ -extern void ds_copy_thread(struct task_struct *tsk, struct task_struct *father); -extern void ds_exit_thread(struct task_struct *tsk); - #else /* CONFIG_X86_DS */ struct cpuinfo_x86; static inline void __cpuinit ds_init_intel(struct cpuinfo_x86 *ignored) {} static inline void ds_switch_to(struct task_struct *prev, struct task_struct *next) {} -static inline void ds_copy_thread(struct task_struct *tsk, - struct task_struct *father) {} -static inline void ds_exit_thread(struct task_struct *tsk) {} #endif /* CONFIG_X86_DS */ #endif /* _ASM_X86_DS_H */ diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h index 1efeb497f1f9..7c39de7e709a 100644 --- a/arch/x86/include/asm/processor.h +++ b/arch/x86/include/asm/processor.h @@ -454,10 +454,8 @@ struct thread_struct { unsigned io_bitmap_max; /* MSR_IA32_DEBUGCTLMSR value to switch in if TIF_DEBUGCTLMSR is set. */ unsigned long debugctlmsr; -#ifdef CONFIG_X86_DS -/* Debug Store context; see include/asm-x86/ds.h; goes into MSR_IA32_DS_AREA */ + /* Debug Store context; see asm/ds.h */ struct ds_context *ds_ctx; -#endif /* CONFIG_X86_DS */ }; static inline unsigned long native_get_debugreg(int regno) diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 21a3852abf68..71cab3b62dce 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -1352,16 +1352,6 @@ void ds_switch_to(struct task_struct *prev, struct task_struct *next) update_debugctlmsr(debugctlmsr); } -void ds_copy_thread(struct task_struct *tsk, struct task_struct *father) -{ - clear_tsk_thread_flag(tsk, TIF_DS_AREA_MSR); - tsk->thread.ds_ctx = NULL; -} - -void ds_exit_thread(struct task_struct *tsk) -{ -} - static __init int ds_selftest(void) { if (ds_cfg.sizeof_rec[ds_bts]) { diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c index ca989158e847..fb5dfb891f0f 100644 --- a/arch/x86/kernel/process.c +++ b/arch/x86/kernel/process.c @@ -14,6 +14,7 @@ #include #include #include +#include unsigned long idle_halt; EXPORT_SYMBOL(idle_halt); @@ -45,6 +46,8 @@ void free_thread_xstate(struct task_struct *tsk) kmem_cache_free(task_xstate_cachep, tsk->thread.xstate); tsk->thread.xstate = NULL; } + + WARN(tsk->thread.ds_ctx, "leaking DS context\n"); } void free_thread_info(struct thread_info *ti) @@ -83,8 +86,6 @@ void exit_thread(void) put_cpu(); kfree(bp); } - - ds_exit_thread(current); } void flush_thread(void) diff --git a/arch/x86/kernel/process_32.c b/arch/x86/kernel/process_32.c index 76f8f84043a2..b5e4bfef4472 100644 --- a/arch/x86/kernel/process_32.c +++ b/arch/x86/kernel/process_32.c @@ -290,7 +290,8 @@ int copy_thread(unsigned long clone_flags, unsigned long sp, p->thread.io_bitmap_max = 0; } - ds_copy_thread(p, current); + clear_tsk_thread_flag(p, TIF_DS_AREA_MSR); + p->thread.ds_ctx = NULL; clear_tsk_thread_flag(p, TIF_DEBUGCTLMSR); p->thread.debugctlmsr = 0; diff --git a/arch/x86/kernel/process_64.c b/arch/x86/kernel/process_64.c index b751a41392b1..5a1a1de292ec 100644 --- a/arch/x86/kernel/process_64.c +++ b/arch/x86/kernel/process_64.c @@ -335,7 +335,8 @@ int copy_thread(unsigned long clone_flags, unsigned long sp, goto out; } - ds_copy_thread(p, me); + clear_tsk_thread_flag(p, TIF_DS_AREA_MSR); + p->thread.ds_ctx = NULL; clear_tsk_thread_flag(p, TIF_DEBUGCTLMSR); p->thread.debugctlmsr = 0; -- cgit v1.2.3 From ee811517a5604aa63fae803b7c044712699e1303 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:47 +0200 Subject: x86, ds: use single debug store cpu configuration Use a single configuration for all cpus. Reported-by: Ingo Molnar Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144602.191165000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 71cab3b62dce..443f415441da 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -47,9 +47,8 @@ struct ds_configuration { /* Control bit-masks indexed by enum ds_feature: */ unsigned long ctl[dsf_ctl_max]; }; -static DEFINE_PER_CPU(struct ds_configuration, ds_cfg_array); +static struct ds_configuration ds_cfg __read_mostly; -#define ds_cfg per_cpu(ds_cfg_array, smp_processor_id()) /* Maximal size of a DS configuration: */ #define MAX_SIZEOF_DS (12 * 8) @@ -1268,6 +1267,10 @@ ds_configure(const struct ds_configuration *cfg, void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) { + /* Only configure the first cpu. Others are identical. */ + if (ds_cfg.name) + return; + switch (c->x86) { case 0x6: switch (c->x86_model) { -- cgit v1.2.3 From 0f4814065ff8c24ca8bfd75c9b73502be152c287 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:48 +0200 Subject: x86, ptrace: add bts context unconditionally Add the ptrace bts context field to task_struct unconditionally. Initialize the field directly in copy_process(). Remove all the unneeded functionality used to initialize that field. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144603.292754000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/ptrace.h | 9 ++++----- arch/x86/kernel/ptrace.c | 20 +------------------- include/linux/ptrace.h | 10 ---------- include/linux/sched.h | 2 -- kernel/fork.c | 4 ++-- kernel/ptrace.c | 10 ---------- 6 files changed, 7 insertions(+), 48 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/ptrace.h b/arch/x86/include/asm/ptrace.h index e304b66abeea..5cdd19f20b5b 100644 --- a/arch/x86/include/asm/ptrace.h +++ b/arch/x86/include/asm/ptrace.h @@ -235,12 +235,11 @@ extern int do_get_thread_area(struct task_struct *p, int idx, extern int do_set_thread_area(struct task_struct *p, int idx, struct user_desc __user *info, int can_allocate); -extern void x86_ptrace_untrace(struct task_struct *); -extern void x86_ptrace_fork(struct task_struct *child, - unsigned long clone_flags); +#ifdef CONFIG_X86_PTRACE_BTS +extern void ptrace_bts_untrace(struct task_struct *tsk); -#define arch_ptrace_untrace(tsk) x86_ptrace_untrace(tsk) -#define arch_ptrace_fork(child, flags) x86_ptrace_fork(child, flags) +#define arch_ptrace_untrace(tsk) ptrace_bts_untrace(tsk) +#endif /* CONFIG_X86_PTRACE_BTS */ #endif /* __KERNEL__ */ diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c index adbb24322d8f..b32a8ee53381 100644 --- a/arch/x86/kernel/ptrace.c +++ b/arch/x86/kernel/ptrace.c @@ -887,37 +887,19 @@ static int ptrace_bts_size(struct task_struct *child) return (trace->ds.top - trace->ds.begin) / trace->ds.size; } -static inline void ptrace_bts_fork(struct task_struct *tsk) -{ - tsk->bts = NULL; -} - /* * Called from __ptrace_unlink() after the child has been moved back * to its original parent. */ -static inline void ptrace_bts_untrace(struct task_struct *child) +void ptrace_bts_untrace(struct task_struct *child) { if (unlikely(child->bts)) { free_bts_context(child->bts); child->bts = NULL; } } -#else -static inline void ptrace_bts_fork(struct task_struct *tsk) {} -static inline void ptrace_bts_untrace(struct task_struct *child) {} #endif /* CONFIG_X86_PTRACE_BTS */ -void x86_ptrace_fork(struct task_struct *child, unsigned long clone_flags) -{ - ptrace_bts_fork(child); -} - -void x86_ptrace_untrace(struct task_struct *child) -{ - ptrace_bts_untrace(child); -} - /* * Called by kernel/ptrace.c when detaching.. * diff --git a/include/linux/ptrace.h b/include/linux/ptrace.h index 67c15653fc23..59e133d39d50 100644 --- a/include/linux/ptrace.h +++ b/include/linux/ptrace.h @@ -95,7 +95,6 @@ extern void __ptrace_link(struct task_struct *child, struct task_struct *new_parent); extern void __ptrace_unlink(struct task_struct *child); extern void exit_ptrace(struct task_struct *tracer); -extern void ptrace_fork(struct task_struct *task, unsigned long clone_flags); #define PTRACE_MODE_READ 1 #define PTRACE_MODE_ATTACH 2 /* Returns 0 on success, -errno on denial. */ @@ -327,15 +326,6 @@ static inline void user_enable_block_step(struct task_struct *task) #define arch_ptrace_untrace(task) do { } while (0) #endif -#ifndef arch_ptrace_fork -/* - * Do machine-specific work to initialize a new task. - * - * This is called from copy_process(). - */ -#define arch_ptrace_fork(child, clone_flags) do { } while (0) -#endif - extern int task_current_syscall(struct task_struct *target, long *callno, unsigned long args[6], unsigned int maxargs, unsigned long *sp, unsigned long *pc); diff --git a/include/linux/sched.h b/include/linux/sched.h index 52b8cd049c2e..451186a22ef5 100644 --- a/include/linux/sched.h +++ b/include/linux/sched.h @@ -1205,13 +1205,11 @@ struct task_struct { struct list_head ptraced; struct list_head ptrace_entry; -#ifdef CONFIG_X86_PTRACE_BTS /* * This is the tracer handle for the ptrace BTS extension. * This field actually belongs to the ptracer task. */ struct bts_context *bts; -#endif /* CONFIG_X86_PTRACE_BTS */ /* PID/PID hash table linkage. */ struct pid_link pids[PIDTYPE_MAX]; diff --git a/kernel/fork.c b/kernel/fork.c index 660c2b8765bc..69bde7a22e9b 100644 --- a/kernel/fork.c +++ b/kernel/fork.c @@ -1086,8 +1086,8 @@ static struct task_struct *copy_process(unsigned long clone_flags, #ifdef CONFIG_DEBUG_MUTEXES p->blocked_on = NULL; /* not blocked yet */ #endif - if (unlikely(current->ptrace)) - ptrace_fork(p, clone_flags); + + p->bts = NULL; /* Perform scheduler related setup. Assign this task to a CPU. */ sched_fork(p, clone_flags); diff --git a/kernel/ptrace.c b/kernel/ptrace.c index aaad0ec34194..321127d965c2 100644 --- a/kernel/ptrace.c +++ b/kernel/ptrace.c @@ -26,16 +26,6 @@ #include -/* - * Initialize a new task whose father had been ptraced. - * - * Called from copy_process(). - */ -void ptrace_fork(struct task_struct *child, unsigned long clone_flags) -{ - arch_ptrace_fork(child, clone_flags); -} - /* * ptrace a task: make the debugger its new parent and * move it to the ptrace list. -- cgit v1.2.3 From 6047550d3d26fed88b18a208b31f8b90b5ef3e9b Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:49 +0200 Subject: x86, ds: dont use TIF_DEBUGCTLMSR Debug store already uses TIF_DS_AREA_MSR to trigger debug store context switch handling. No need to use TIF_DEBUGCTLMSR, as well. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144604.256645000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 5 ----- 1 file changed, 5 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 443f415441da..cab28320dac7 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -925,11 +925,6 @@ static void update_task_debugctlmsr(struct task_struct *task, get_cpu(); if (task == current) update_debugctlmsr(debugctlmsr); - - if (task->thread.debugctlmsr) - set_tsk_thread_flag(task, TIF_DEBUGCTLMSR); - else - clear_tsk_thread_flag(task, TIF_DEBUGCTLMSR); put_cpu(); } -- cgit v1.2.3 From 608780a9048efa3e85fbc4d8649b26805cc588aa Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:50 +0200 Subject: x86, ds: fix bad ds_reset_pebs() Ds_reset_pebs() passed the wrong qualifier to a shared function resulting in a reset of bts, rather than pebs. Reported-by: Stephane Eranian Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144605.206510000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index cab28320dac7..ebfb0fde8e6f 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -1186,7 +1186,7 @@ int ds_reset_pebs(struct pebs_tracer *tracer) tracer->trace.ds.top = tracer->trace.ds.begin; - ds_set(tracer->ds.context->ds, ds_bts, ds_index, + ds_set(tracer->ds.context->ds, ds_pebs, ds_index, (unsigned long)tracer->trace.ds.top); return 0; -- cgit v1.2.3 From 150f5164c1258e05b7dea16f29e592f354c48f34 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:51 +0200 Subject: x86, ds: allow small debug store buffers Check the buffer size more precisely to allow buffers for exactly one element provided the base address is already properly aligned. Add a debug store selftest. Reported-by: Stephane Eranian Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144606.139137000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds.c | 9 +++++++-- arch/x86/kernel/ds_selftest.c | 6 +++--- 2 files changed, 10 insertions(+), 5 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index ebfb0fde8e6f..4e05157506aa 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -656,6 +656,7 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, { struct ds_context *context; int error; + size_t req_size; error = -EOPNOTSUPP; if (!ds_cfg.sizeof_rec[qual]) @@ -665,9 +666,13 @@ static int ds_request(struct ds_tracer *tracer, struct ds_trace *trace, if (!base) goto out; - /* We need space for alignment adjustments in ds_init_ds_trace(). */ + req_size = ds_cfg.sizeof_rec[qual]; + /* We might need space for alignment adjustments. */ + if (!IS_ALIGNED((unsigned long)base, DS_ALIGNMENT)) + req_size += DS_ALIGNMENT; + error = -EINVAL; - if (size < (DS_ALIGNMENT + ds_cfg.sizeof_rec[qual])) + if (size < req_size) goto out; if (th != (size_t)-1) { diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index a40b2533c71e..5f104a0ace66 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -16,8 +16,8 @@ #include -#define BUFFER_SIZE 521 /* Intentionally chose an odd size. */ - +#define BUFFER_SIZE 521 /* Intentionally chose an odd size. */ +#define SMALL_BUFFER_SIZE 24 /* A single bts entry. */ struct ds_selftest_bts_conf { struct bts_tracer *tracer; @@ -381,7 +381,7 @@ int ds_selftest_bts(void) conf.suspend = ds_suspend_bts_noirq; conf.resume = ds_resume_bts_noirq; conf.tracer = - ds_request_bts_task(current, buffer, BUFFER_SIZE, + ds_request_bts_task(current, buffer, SMALL_BUFFER_SIZE, NULL, (size_t)-1, BTS_KERNEL); local_irq_save(irq); ds_selftest_bts_cpu(&conf); -- cgit v1.2.3 From 017bc617657c928cb9a0c45a7a7e9f4e66695347 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 3 Apr 2009 16:43:52 +0200 Subject: x86, ds: support Core i7 Add debug store support for Core i7. Core i7 adds a reset value for each performance counter and a new PEBS record format. Signed-off-by: Markus Metzger Cc: roland@redhat.com Cc: eranian@googlemail.com Cc: oleg@redhat.com Cc: juan.villacis@intel.com Cc: ak@linux.jf.intel.com LKML-Reference: <20090403144607.088997000@intel.com> Signed-off-by: Ingo Molnar --- arch/x86/include/asm/ds.h | 12 ++++++--- arch/x86/kernel/ds.c | 69 ++++++++++++++++++++++++++++++++++++++++++----- 2 files changed, 71 insertions(+), 10 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/ds.h b/arch/x86/include/asm/ds.h index 149e5208e967..70dac199b093 100644 --- a/arch/x86/include/asm/ds.h +++ b/arch/x86/include/asm/ds.h @@ -234,8 +234,12 @@ struct bts_trace { struct pebs_trace { struct ds_trace ds; - /* the PEBS reset value */ - unsigned long long reset_value; + /* the number of valid counters in the below array */ + unsigned int counters; + +#define MAX_PEBS_COUNTERS 4 + /* the counter reset value */ + unsigned long long counter_reset[MAX_PEBS_COUNTERS]; }; @@ -270,9 +274,11 @@ extern int ds_reset_pebs(struct pebs_tracer *tracer); * Returns 0 on success; -Eerrno on error * * tracer: the tracer handle returned from ds_request_pebs() + * counter: the index of the counter * value: the new counter reset value */ -extern int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value); +extern int ds_set_pebs_reset(struct pebs_tracer *tracer, + unsigned int counter, u64 value); /* * Initialization diff --git a/arch/x86/kernel/ds.c b/arch/x86/kernel/ds.c index 4e05157506aa..48bfe1386038 100644 --- a/arch/x86/kernel/ds.c +++ b/arch/x86/kernel/ds.c @@ -44,6 +44,9 @@ struct ds_configuration { /* The size of a BTS/PEBS record in bytes: */ unsigned char sizeof_rec[2]; + /* The number of pebs counter reset values in the DS structure. */ + unsigned char nr_counter_reset; + /* Control bit-masks indexed by enum ds_feature: */ unsigned long ctl[dsf_ctl_max]; }; @@ -51,7 +54,7 @@ static struct ds_configuration ds_cfg __read_mostly; /* Maximal size of a DS configuration: */ -#define MAX_SIZEOF_DS (12 * 8) +#define MAX_SIZEOF_DS 0x80 /* Maximal size of a BTS record: */ #define MAX_SIZEOF_BTS (3 * 8) @@ -59,6 +62,12 @@ static struct ds_configuration ds_cfg __read_mostly; /* BTS and PEBS buffer alignment: */ #define DS_ALIGNMENT (1 << 3) +/* Number of buffer pointers in DS: */ +#define NUM_DS_PTR_FIELDS 8 + +/* Size of a pebs reset value in DS: */ +#define PEBS_RESET_FIELD_SIZE 8 + /* Mask of control bits in the DS MSR register: */ #define BTS_CONTROL \ ( ds_cfg.ctl[dsf_bts] | \ @@ -1164,9 +1173,12 @@ const struct pebs_trace *ds_read_pebs(struct pebs_tracer *tracer) return NULL; ds_read_config(tracer->ds.context, &tracer->trace.ds, ds_pebs); - tracer->trace.reset_value = - *(u64 *)(tracer->ds.context->ds + - (ds_cfg.sizeof_ptr_field * 8)); + + tracer->trace.counters = ds_cfg.nr_counter_reset; + memcpy(tracer->trace.counter_reset, + tracer->ds.context->ds + + (NUM_DS_PTR_FIELDS * ds_cfg.sizeof_ptr_field), + ds_cfg.nr_counter_reset * PEBS_RESET_FIELD_SIZE); return &tracer->trace; } @@ -1197,13 +1209,18 @@ int ds_reset_pebs(struct pebs_tracer *tracer) return 0; } -int ds_set_pebs_reset(struct pebs_tracer *tracer, u64 value) +int ds_set_pebs_reset(struct pebs_tracer *tracer, + unsigned int counter, u64 value) { if (!tracer) return -EINVAL; + if (ds_cfg.nr_counter_reset < counter) + return -EINVAL; + *(u64 *)(tracer->ds.context->ds + - (ds_cfg.sizeof_ptr_field * 8)) = value; + (NUM_DS_PTR_FIELDS * ds_cfg.sizeof_ptr_field) + + (counter * PEBS_RESET_FIELD_SIZE)) = value; return 0; } @@ -1213,16 +1230,26 @@ static const struct ds_configuration ds_cfg_netburst = { .ctl[dsf_bts] = (1 << 2) | (1 << 3), .ctl[dsf_bts_kernel] = (1 << 5), .ctl[dsf_bts_user] = (1 << 6), + .nr_counter_reset = 1, }; static const struct ds_configuration ds_cfg_pentium_m = { .name = "Pentium M", .ctl[dsf_bts] = (1 << 6) | (1 << 7), + .nr_counter_reset = 1, }; static const struct ds_configuration ds_cfg_core2_atom = { .name = "Core 2/Atom", .ctl[dsf_bts] = (1 << 6) | (1 << 7), .ctl[dsf_bts_kernel] = (1 << 9), .ctl[dsf_bts_user] = (1 << 10), + .nr_counter_reset = 1, +}; +static const struct ds_configuration ds_cfg_core_i7 = { + .name = "Core i7", + .ctl[dsf_bts] = (1 << 6) | (1 << 7), + .ctl[dsf_bts_kernel] = (1 << 9), + .ctl[dsf_bts_user] = (1 << 10), + .nr_counter_reset = 4, }; static void @@ -1239,6 +1266,32 @@ ds_configure(const struct ds_configuration *cfg, nr_pebs_fields = 18; #endif + /* + * Starting with version 2, architectural performance + * monitoring supports a format specifier. + */ + if ((cpuid_eax(0xa) & 0xff) > 1) { + unsigned long perf_capabilities, format; + + rdmsrl(MSR_IA32_PERF_CAPABILITIES, perf_capabilities); + + format = (perf_capabilities >> 8) & 0xf; + + switch (format) { + case 0: + nr_pebs_fields = 18; + break; + case 1: + nr_pebs_fields = 22; + break; + default: + printk(KERN_INFO + "[ds] unknown PEBS format: %lu\n", format); + nr_pebs_fields = 0; + break; + } + } + memset(&ds_cfg, 0, sizeof(ds_cfg)); ds_cfg = *cfg; @@ -1262,7 +1315,7 @@ ds_configure(const struct ds_configuration *cfg, printk("bts/pebs record: %u/%u bytes\n", ds_cfg.sizeof_rec[ds_bts], ds_cfg.sizeof_rec[ds_pebs]); - WARN_ON_ONCE(MAX_SIZEOF_DS < (12 * ds_cfg.sizeof_ptr_field)); + WARN_ON_ONCE(MAX_PEBS_COUNTERS < ds_cfg.nr_counter_reset); } void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) @@ -1284,6 +1337,8 @@ void __cpuinit ds_init_intel(struct cpuinfo_x86 *c) ds_configure(&ds_cfg_core2_atom, c); break; case 0x1a: /* Core i7 */ + ds_configure(&ds_cfg_core_i7, c); + break; default: /* Sorry, don't know about them. */ break; -- cgit v1.2.3 From e71e99c294058a61b7a8b9bb6da2f745ac51aa4f Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Wed, 25 Mar 2009 14:30:04 -0400 Subject: x86, function-graph: only save return values on x86_64 Impact: speed up The return to handler portion of the function graph tracer should only need to save the return values. The caller already saved off the registers that the callee can modify. The returning function already saved the registers it modified. When we call our own trace function it too will save the registers that the callee must restore. There's no reason to save off anything more that the registers used to return the values. Note, I did a complete kernel build with this modification and the function graph tracer running on x86_64. Signed-off-by: Steven Rostedt Signed-off-by: Ingo Molnar --- arch/x86/kernel/entry_64.S | 19 +++---------------- 1 file changed, 3 insertions(+), 16 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/entry_64.S b/arch/x86/kernel/entry_64.S index a331ec38af9e..1ac99865591c 100644 --- a/arch/x86/kernel/entry_64.S +++ b/arch/x86/kernel/entry_64.S @@ -147,27 +147,14 @@ END(ftrace_graph_caller) GLOBAL(return_to_handler) subq $80, %rsp + /* Save the return values */ movq %rax, (%rsp) - movq %rcx, 8(%rsp) - movq %rdx, 16(%rsp) - movq %rsi, 24(%rsp) - movq %rdi, 32(%rsp) - movq %r8, 40(%rsp) - movq %r9, 48(%rsp) - movq %r10, 56(%rsp) - movq %r11, 64(%rsp) + movq %rdx, 8(%rsp) call ftrace_return_to_handler movq %rax, 72(%rsp) - movq 64(%rsp), %r11 - movq 56(%rsp), %r10 - movq 48(%rsp), %r9 - movq 40(%rsp), %r8 - movq 32(%rsp), %rdi - movq 24(%rsp), %rsi - movq 16(%rsp), %rdx - movq 8(%rsp), %rcx + movq 8(%rsp), %rdx movq (%rsp), %rax addq $72, %rsp retq -- cgit v1.2.3 From d7285c6b5c54397fdf112c2fb98ee43193173aa9 Mon Sep 17 00:00:00 2001 From: Chris Wright Date: Thu, 23 Apr 2009 10:21:38 -0700 Subject: x86: use native register access for native tlb flushing currently these are paravirtulaized, doesn't appear any callers rely on this (no pv_ops backends are using native_tlb and overriding cr3/4 access). [ Impact: fix lockdep warning with paravirt and function tracer ] Signed-off-by: Chris Wright LKML-Reference: <20090423172138.GR3036@sequoia.sous-sol.org> Signed-off-by: Steven Rostedt --- arch/x86/include/asm/tlbflush.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) (limited to 'arch') diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h index d3539f998f88..e2927c5f45b1 100644 --- a/arch/x86/include/asm/tlbflush.h +++ b/arch/x86/include/asm/tlbflush.h @@ -17,7 +17,7 @@ static inline void __native_flush_tlb(void) { - write_cr3(read_cr3()); + native_write_cr3(native_read_cr3()); } static inline void __native_flush_tlb_global(void) @@ -32,11 +32,11 @@ static inline void __native_flush_tlb_global(void) */ raw_local_irq_save(flags); - cr4 = read_cr4(); + cr4 = native_read_cr4(); /* clear PGE */ - write_cr4(cr4 & ~X86_CR4_PGE); + native_write_cr4(cr4 & ~X86_CR4_PGE); /* write old PGE again and flush TLBs */ - write_cr4(cr4); + native_write_cr4(cr4); raw_local_irq_restore(flags); } -- cgit v1.2.3 From 782cc5ae6331d63b4febaa312c9d14493aafa9b8 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 24 Apr 2009 09:43:09 +0200 Subject: x86, ds: fix buffer alignment in debug store selftest The debug store selftest code uses a stack-allocated buffer, which is not necessarily correctly aligned. For tests using a buffer to hold a single entry, the buffer that is passed to ds_request must already be suitably aligned. Pass a suitably aligned portion of the bigger buffer. [ Impact: fix hw-branch-tracer self-test failure ] Signed-off-by: Markus Metzger Cc: markus.t.metzger@gmail.com LKML-Reference: <20090424094309.A30145@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ds_selftest.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ds_selftest.c b/arch/x86/kernel/ds_selftest.c index 5f104a0ace66..6bc7c199ab99 100644 --- a/arch/x86/kernel/ds_selftest.c +++ b/arch/x86/kernel/ds_selftest.c @@ -323,13 +323,15 @@ static int ds_selftest_bts_bad_request_task(void *buffer) int ds_selftest_bts(void) { struct ds_selftest_bts_conf conf; - unsigned char buffer[BUFFER_SIZE]; + unsigned char buffer[BUFFER_SIZE], *small_buffer; unsigned long irq; int cpu; printk(KERN_INFO "[ds] bts selftest..."); conf.error = 0; + small_buffer = (unsigned char *)ALIGN((unsigned long)buffer, 8) + 8; + get_online_cpus(); for_each_online_cpu(cpu) { conf.suspend = ds_suspend_bts_wrap; @@ -381,7 +383,7 @@ int ds_selftest_bts(void) conf.suspend = ds_suspend_bts_noirq; conf.resume = ds_resume_bts_noirq; conf.tracer = - ds_request_bts_task(current, buffer, SMALL_BUFFER_SIZE, + ds_request_bts_task(current, small_buffer, SMALL_BUFFER_SIZE, NULL, (size_t)-1, BTS_KERNEL); local_irq_save(irq); ds_selftest_bts_cpu(&conf); -- cgit v1.2.3 From 7e0bfad24d85de7cf2202a7b0ce51de11a077b21 Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 24 Apr 2009 09:44:48 +0200 Subject: x86, bts: reenable ptrace branch trace support The races found by Oleg Nesterov have been fixed. Reenable branch trace support. Signed-off-by: Markus Metzger Acked-by: Oleg Nesterov LKML-Reference: <20090424094448.A30216@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/Kconfig.cpu | 1 - 1 file changed, 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/Kconfig.cpu b/arch/x86/Kconfig.cpu index 8130334329c0..924e156a85ab 100644 --- a/arch/x86/Kconfig.cpu +++ b/arch/x86/Kconfig.cpu @@ -506,7 +506,6 @@ config X86_PTRACE_BTS bool "Branch Trace Store" default y depends on X86_DEBUGCTLMSR - depends on BROKEN ---help--- This adds a ptrace interface to the hardware's branch trace store. -- cgit v1.2.3 From 1cb81b143fa8f0e4629f10690862e2e52ca792ff Mon Sep 17 00:00:00 2001 From: Markus Metzger Date: Fri, 24 Apr 2009 09:51:43 +0200 Subject: x86, bts, mm: clean up buffer allocation The current mm interface is asymetric. One function allocates a locked buffer, another function only refunds the memory. Change this to have two functions for accounting and refunding locked memory, respectively; and do the actual buffer allocation in ptrace. [ Impact: refactor BTS buffer allocation code ] Signed-off-by: Markus Metzger Acked-by: Andrew Morton Cc: Peter Zijlstra LKML-Reference: <20090424095143.A30265@sedona.ch.intel.com> Signed-off-by: Ingo Molnar --- arch/x86/kernel/ptrace.c | 39 ++++++++++++++++++++++++++------------- include/linux/mm.h | 6 ++++-- mm/mlock.c | 36 +++++++++++++++++------------------- 3 files changed, 47 insertions(+), 34 deletions(-) (limited to 'arch') diff --git a/arch/x86/kernel/ptrace.c b/arch/x86/kernel/ptrace.c index d5252ae6c520..09ecbde91c13 100644 --- a/arch/x86/kernel/ptrace.c +++ b/arch/x86/kernel/ptrace.c @@ -617,17 +617,28 @@ struct bts_context { struct work_struct work; }; -static inline void alloc_bts_buffer(struct bts_context *context, - unsigned int size) +static int alloc_bts_buffer(struct bts_context *context, unsigned int size) { - void *buffer; + void *buffer = NULL; + int err = -ENOMEM; - buffer = alloc_locked_buffer(size); - if (buffer) { - context->buffer = buffer; - context->size = size; - context->mm = get_task_mm(current); - } + err = account_locked_memory(current->mm, current->signal->rlim, size); + if (err < 0) + return err; + + buffer = kzalloc(size, GFP_KERNEL); + if (!buffer) + goto out_refund; + + context->buffer = buffer; + context->size = size; + context->mm = get_task_mm(current); + + return 0; + + out_refund: + refund_locked_memory(current->mm, size); + return err; } static inline void free_bts_buffer(struct bts_context *context) @@ -638,7 +649,7 @@ static inline void free_bts_buffer(struct bts_context *context) kfree(context->buffer); context->buffer = NULL; - refund_locked_buffer_memory(context->mm, context->size); + refund_locked_memory(context->mm, context->size); context->size = 0; mmput(context->mm); @@ -786,13 +797,15 @@ static int ptrace_bts_config(struct task_struct *child, context->tracer = NULL; if ((cfg.flags & PTRACE_BTS_O_ALLOC) && (cfg.size != context->size)) { + int err; + free_bts_buffer(context); if (!cfg.size) return 0; - alloc_bts_buffer(context, cfg.size); - if (!context->buffer) - return -ENOMEM; + err = alloc_bts_buffer(context, cfg.size); + if (err < 0) + return err; } if (cfg.flags & PTRACE_BTS_O_TRACE) diff --git a/include/linux/mm.h b/include/linux/mm.h index a3963ba23a6d..009eabd3c21c 100644 --- a/include/linux/mm.h +++ b/include/linux/mm.h @@ -19,6 +19,7 @@ struct anon_vma; struct file_ra_state; struct user_struct; struct writeback_control; +struct rlimit; #ifndef CONFIG_DISCONTIGMEM /* Don't use mapnrs, do it properly */ extern unsigned long max_mapnr; @@ -1319,7 +1320,8 @@ int vmemmap_populate_basepages(struct page *start_page, int vmemmap_populate(struct page *start_page, unsigned long pages, int node); void vmemmap_populate_print_last(void); -extern void *alloc_locked_buffer(size_t size); -extern void refund_locked_buffer_memory(struct mm_struct *mm, size_t size); +extern int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim, + size_t size); +extern void refund_locked_memory(struct mm_struct *mm, size_t size); #endif /* __KERNEL__ */ #endif /* _LINUX_MM_H */ diff --git a/mm/mlock.c b/mm/mlock.c index 28be15ead9c1..ac130433c7d3 100644 --- a/mm/mlock.c +++ b/mm/mlock.c @@ -629,38 +629,36 @@ void user_shm_unlock(size_t size, struct user_struct *user) free_uid(user); } -void *alloc_locked_buffer(size_t size) +int account_locked_memory(struct mm_struct *mm, struct rlimit *rlim, + size_t size) { - unsigned long rlim, vm, pgsz; - void *buffer = NULL; + unsigned long lim, vm, pgsz; + int error = -ENOMEM; pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT; - down_write(¤t->mm->mmap_sem); - - rlim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT; - vm = current->mm->total_vm + pgsz; - if (rlim < vm) - goto out; + down_write(&mm->mmap_sem); - rlim = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT; - vm = current->mm->locked_vm + pgsz; - if (rlim < vm) + lim = rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT; + vm = mm->total_vm + pgsz; + if (lim < vm) goto out; - buffer = kzalloc(size, GFP_KERNEL); - if (!buffer) + lim = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT; + vm = mm->locked_vm + pgsz; + if (lim < vm) goto out; - current->mm->total_vm += pgsz; - current->mm->locked_vm += pgsz; + mm->total_vm += pgsz; + mm->locked_vm += pgsz; + error = 0; out: - up_write(¤t->mm->mmap_sem); - return buffer; + up_write(&mm->mmap_sem); + return error; } -void refund_locked_buffer_memory(struct mm_struct *mm, size_t size) +void refund_locked_memory(struct mm_struct *mm, size_t size) { unsigned long pgsz = PAGE_ALIGN(size) >> PAGE_SHIFT; -- cgit v1.2.3 From 0492e1bb8fe7d122901c9f3af75e537d4129712e Mon Sep 17 00:00:00 2001 From: Stuart Bennett Date: Tue, 28 Apr 2009 20:17:49 +0100 Subject: tracing: x86, mmiotrace: code consistency/legibility improvement kmmio_probe being *p and kmmio_fault_page being sometimes *f and sometimes *p is not helpful. [ Impact: cleanup ] Signed-off-by: Stuart Bennett Acked-by: Pekka Paalanen Cc: Steven Rostedt LKML-Reference: <1240946271-7083-3-git-send-email-stuart@freedesktop.org> Signed-off-by: Ingo Molnar --- arch/x86/mm/kmmio.c | 34 +++++++++++++++++----------------- 1 file changed, 17 insertions(+), 17 deletions(-) (limited to 'arch') diff --git a/arch/x86/mm/kmmio.c b/arch/x86/mm/kmmio.c index 4f115e00486b..869181a917d9 100644 --- a/arch/x86/mm/kmmio.c +++ b/arch/x86/mm/kmmio.c @@ -97,13 +97,13 @@ static struct kmmio_probe *get_kmmio_probe(unsigned long addr) static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page) { struct list_head *head; - struct kmmio_fault_page *p; + struct kmmio_fault_page *f; page &= PAGE_MASK; head = kmmio_page_list(page); - list_for_each_entry_rcu(p, head, list) { - if (p->page == page) - return p; + list_for_each_entry_rcu(f, head, list) { + if (f->page == page) + return f; } return NULL; } @@ -439,12 +439,12 @@ static void rcu_free_kmmio_fault_pages(struct rcu_head *head) head, struct kmmio_delayed_release, rcu); - struct kmmio_fault_page *p = dr->release_list; - while (p) { - struct kmmio_fault_page *next = p->release_next; - BUG_ON(p->count); - kfree(p); - p = next; + struct kmmio_fault_page *f = dr->release_list; + while (f) { + struct kmmio_fault_page *next = f->release_next; + BUG_ON(f->count); + kfree(f); + f = next; } kfree(dr); } @@ -453,19 +453,19 @@ static void remove_kmmio_fault_pages(struct rcu_head *head) { struct kmmio_delayed_release *dr = container_of(head, struct kmmio_delayed_release, rcu); - struct kmmio_fault_page *p = dr->release_list; + struct kmmio_fault_page *f = dr->release_list; struct kmmio_fault_page **prevp = &dr->release_list; unsigned long flags; spin_lock_irqsave(&kmmio_lock, flags); - while (p) { - if (!p->count) { - list_del_rcu(&p->list); - prevp = &p->release_next; + while (f) { + if (!f->count) { + list_del_rcu(&f->list); + prevp = &f->release_next; } else { - *prevp = p->release_next; + *prevp = f->release_next; } - p = p->release_next; + f = f->release_next; } spin_unlock_irqrestore(&kmmio_lock, flags); -- cgit v1.2.3 From 46e91d00b1165b14b484aa33800e1bba0794ae1a Mon Sep 17 00:00:00 2001 From: Stuart Bennett Date: Tue, 28 Apr 2009 20:17:50 +0100 Subject: tracing: x86, mmiotrace: refactor clearing/restore of page presence * change function names to clear_* from set_*: in reality we only clear and restore page presence, and never unconditionally set present. Using clear_*({true, false}, ...) is therefore more honest than set_*({false, true}, ...) * upgrade presence storage to pteval_t: doing user-space tracing will require saving and manipulation of the _PAGE_PROTNONE bit, in addition to the existing _PAGE_PRESENT changes, and having multiple bools stored and passed around does not seem optimal [ Impact: refactor, clean up mmiotrace code ] Signed-off-by: Stuart Bennett Acked-by: Pekka Paalanen Cc: Steven Rostedt LKML-Reference: <1240946271-7083-4-git-send-email-stuart@freedesktop.org> Signed-off-by: Ingo Molnar --- arch/x86/mm/kmmio.c | 43 ++++++++++++++++++++++--------------------- 1 file changed, 22 insertions(+), 21 deletions(-) (limited to 'arch') diff --git a/arch/x86/mm/kmmio.c b/arch/x86/mm/kmmio.c index 869181a917d9..a769d1a2d93b 100644 --- a/arch/x86/mm/kmmio.c +++ b/arch/x86/mm/kmmio.c @@ -32,7 +32,7 @@ struct kmmio_fault_page { struct list_head list; struct kmmio_fault_page *release_next; unsigned long page; /* location of the fault page */ - bool old_presence; /* page presence prior to arming */ + pteval_t old_presence; /* page presence prior to arming */ bool armed; /* @@ -108,49 +108,51 @@ static struct kmmio_fault_page *get_kmmio_fault_page(unsigned long page) return NULL; } -static void set_pmd_presence(pmd_t *pmd, bool present, bool *old) +static void clear_pmd_presence(pmd_t *pmd, bool clear, pmdval_t *old) { pmdval_t v = pmd_val(*pmd); - *old = !!(v & _PAGE_PRESENT); - v &= ~_PAGE_PRESENT; - if (present) - v |= _PAGE_PRESENT; + if (clear) { + *old = v & _PAGE_PRESENT; + v &= ~_PAGE_PRESENT; + } else /* presume this has been called with clear==true previously */ + v |= *old; set_pmd(pmd, __pmd(v)); } -static void set_pte_presence(pte_t *pte, bool present, bool *old) +static void clear_pte_presence(pte_t *pte, bool clear, pteval_t *old) { pteval_t v = pte_val(*pte); - *old = !!(v & _PAGE_PRESENT); - v &= ~_PAGE_PRESENT; - if (present) - v |= _PAGE_PRESENT; + if (clear) { + *old = v & _PAGE_PRESENT; + v &= ~_PAGE_PRESENT; + } else /* presume this has been called with clear==true previously */ + v |= *old; set_pte_atomic(pte, __pte(v)); } -static int set_page_presence(unsigned long addr, bool present, bool *old) +static int clear_page_presence(struct kmmio_fault_page *f, bool clear) { unsigned int level; - pte_t *pte = lookup_address(addr, &level); + pte_t *pte = lookup_address(f->page, &level); if (!pte) { - pr_err("kmmio: no pte for page 0x%08lx\n", addr); + pr_err("kmmio: no pte for page 0x%08lx\n", f->page); return -1; } switch (level) { case PG_LEVEL_2M: - set_pmd_presence((pmd_t *)pte, present, old); + clear_pmd_presence((pmd_t *)pte, clear, &f->old_presence); break; case PG_LEVEL_4K: - set_pte_presence(pte, present, old); + clear_pte_presence(pte, clear, &f->old_presence); break; default: pr_err("kmmio: unexpected page level 0x%x.\n", level); return -1; } - __flush_tlb_one(addr); + __flush_tlb_one(f->page); return 0; } @@ -171,9 +173,9 @@ static int arm_kmmio_fault_page(struct kmmio_fault_page *f) WARN_ONCE(f->armed, KERN_ERR "kmmio page already armed.\n"); if (f->armed) { pr_warning("kmmio double-arm: page 0x%08lx, ref %d, old %d\n", - f->page, f->count, f->old_presence); + f->page, f->count, !!f->old_presence); } - ret = set_page_presence(f->page, false, &f->old_presence); + ret = clear_page_presence(f, true); WARN_ONCE(ret < 0, KERN_ERR "kmmio arming 0x%08lx failed.\n", f->page); f->armed = true; return ret; @@ -182,8 +184,7 @@ static int arm_kmmio_fault_page(struct kmmio_fault_page *f) /** Restore the given page to saved presence state. */ static void disarm_kmmio_fault_page(struct kmmio_fault_page *f) { - bool tmp; - int ret = set_page_presence(f->page, f->old_presence, &tmp); + int ret = clear_page_presence(f, false); WARN_ONCE(ret < 0, KERN_ERR "kmmio disarming 0x%08lx failed.\n", f->page); f->armed = false; -- cgit v1.2.3 From 0f9a623dd6c9b5b4dd00c232f29525bfc7a8ecf2 Mon Sep 17 00:00:00 2001 From: Stuart Bennett Date: Tue, 28 Apr 2009 20:17:51 +0100 Subject: tracing: x86, mmiotrace: only register for die notifier when tracer active Follow up to afcfe024aebd74b0984a41af9a34e009cf5badaf in Linus' tree ("x86: mmiotrace: quieten spurious warning message") Signed-off-by: Stuart Bennett Acked-by: Pekka Paalanen Cc: Steven Rostedt LKML-Reference: <1240946271-7083-5-git-send-email-stuart@freedesktop.org> Signed-off-by: Ingo Molnar --- arch/x86/mm/kmmio.c | 27 ++++++++++++++++++++++----- arch/x86/mm/mmio-mod.c | 2 ++ include/linux/mmiotrace.h | 2 ++ 3 files changed, 26 insertions(+), 5 deletions(-) (limited to 'arch') diff --git a/arch/x86/mm/kmmio.c b/arch/x86/mm/kmmio.c index a769d1a2d93b..256ce643b0ba 100644 --- a/arch/x86/mm/kmmio.c +++ b/arch/x86/mm/kmmio.c @@ -311,7 +311,12 @@ static int post_kmmio_handler(unsigned long condition, struct pt_regs *regs) struct kmmio_context *ctx = &get_cpu_var(kmmio_ctx); if (!ctx->active) { - pr_debug("kmmio: spurious debug trap on CPU %d.\n", + /* + * debug traps without an active context are due to either + * something external causing them (f.e. using a debugger while + * mmio tracing enabled), or erroneous behaviour + */ + pr_warning("kmmio: unexpected debug trap on CPU %d.\n", smp_processor_id()); goto out; } @@ -529,8 +534,8 @@ void unregister_kmmio_probe(struct kmmio_probe *p) } EXPORT_SYMBOL(unregister_kmmio_probe); -static int kmmio_die_notifier(struct notifier_block *nb, unsigned long val, - void *args) +static int +kmmio_die_notifier(struct notifier_block *nb, unsigned long val, void *args) { struct die_args *arg = args; @@ -545,11 +550,23 @@ static struct notifier_block nb_die = { .notifier_call = kmmio_die_notifier }; -static int __init init_kmmio(void) +int kmmio_init(void) { int i; + for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) INIT_LIST_HEAD(&kmmio_page_table[i]); + return register_die_notifier(&nb_die); } -fs_initcall(init_kmmio); /* should be before device_initcall() */ + +void kmmio_cleanup(void) +{ + int i; + + unregister_die_notifier(&nb_die); + for (i = 0; i < KMMIO_PAGE_TABLE_SIZE; i++) { + WARN_ONCE(!list_empty(&kmmio_page_table[i]), + KERN_ERR "kmmio_page_table not empty at cleanup, any further tracing will leak memory.\n"); + } +} diff --git a/arch/x86/mm/mmio-mod.c b/arch/x86/mm/mmio-mod.c index c9342ed8b402..132772a8ec57 100644 --- a/arch/x86/mm/mmio-mod.c +++ b/arch/x86/mm/mmio-mod.c @@ -451,6 +451,7 @@ void enable_mmiotrace(void) if (nommiotrace) pr_info(NAME "MMIO tracing disabled.\n"); + kmmio_init(); enter_uniprocessor(); spin_lock_irq(&trace_lock); atomic_inc(&mmiotrace_enabled); @@ -473,6 +474,7 @@ void disable_mmiotrace(void) clear_trace_list(); /* guarantees: no more kmmio callbacks */ leave_uniprocessor(); + kmmio_cleanup(); pr_info(NAME "disabled.\n"); out: mutex_unlock(&mmiotrace_mutex); diff --git a/include/linux/mmiotrace.h b/include/linux/mmiotrace.h index 3d1b7bde1283..97491f78b08c 100644 --- a/include/linux/mmiotrace.h +++ b/include/linux/mmiotrace.h @@ -30,6 +30,8 @@ extern unsigned int kmmio_count; extern int register_kmmio_probe(struct kmmio_probe *p); extern void unregister_kmmio_probe(struct kmmio_probe *p); +extern int kmmio_init(void); +extern void kmmio_cleanup(void); #ifdef CONFIG_MMIOTRACE /* kmmio is active by some kmmio_probes? */ -- cgit v1.2.3 From 29a679754b1a2581ee456eada6c2de7ce95068bb Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Thu, 14 May 2009 23:19:09 -0400 Subject: x86/stacktrace: return 0 instead of -1 for stack ops If we return -1 in the ops->stack for the stacktrace saving, we end up breaking out of the loop if the stack we are tracing is in the exception stack. This causes traces like: -0 [002] 34263.745825: raise_softirq_irqoff <-__blk_complete_request -0 [002] 34263.745826: <= 0 <= 0 <= 0 <= 0 <= 0 <= 0 <= 0 By returning "0" instead, the irq stack is saved as well, and we see: -0 [003] 883.280992: raise_softirq_irqoff <-__hrtimer_star t_range_ns -0 [003] 883.280992: <= hrtimer_start_range_ns <= tick_nohz_restart_sched_tick <= cpu_idle <= start_secondary <= <= 0 <= 0 [ Impact: record stacks from interrupts ] Signed-off-by: Steven Rostedt --- arch/x86/kernel/stacktrace.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch') diff --git a/arch/x86/kernel/stacktrace.c b/arch/x86/kernel/stacktrace.c index f7bddc2e37d1..4aaf7e48394f 100644 --- a/arch/x86/kernel/stacktrace.c +++ b/arch/x86/kernel/stacktrace.c @@ -20,7 +20,7 @@ save_stack_warning_symbol(void *data, char *msg, unsigned long symbol) static int save_stack_stack(void *data, char *name) { - return -1; + return 0; } static void save_stack_address(void *data, unsigned long addr, int reliable) -- cgit v1.2.3 From 511b01bdf64ad8a38414096eab283c7784aebfc4 Mon Sep 17 00:00:00 2001 From: Ingo Molnar Date: Thu, 11 Jun 2009 00:32:00 +0200 Subject: Revert "x86, bts: reenable ptrace branch trace support" This reverts commit 7e0bfad24d85de7cf2202a7b0ce51de11a077b21. A late objection to the ABI has arrived: http://lkml.org/lkml/2009/6/10/253 Keep the ABI disabled out of caution, to not create premature user-space expectations. While the hw-branch-tracing variant uses and tests the BTS code. Cc: Peter Zijlstra Cc: Markus Metzger Cc: Oleg Nesterov Signed-off-by: Ingo Molnar --- arch/x86/Kconfig.cpu | 1 + 1 file changed, 1 insertion(+) (limited to 'arch') diff --git a/arch/x86/Kconfig.cpu b/arch/x86/Kconfig.cpu index 924e156a85ab..8130334329c0 100644 --- a/arch/x86/Kconfig.cpu +++ b/arch/x86/Kconfig.cpu @@ -506,6 +506,7 @@ config X86_PTRACE_BTS bool "Branch Trace Store" default y depends on X86_DEBUGCTLMSR + depends on BROKEN ---help--- This adds a ptrace interface to the hardware's branch trace store. -- cgit v1.2.3