From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 14209 invoked by alias); 17 Dec 2012 16:02:30 -0000 Received: (qmail 13873 invoked by uid 22791); 17 Dec 2012 16:02:20 -0000 X-SWARE-Spam-Status: No, hits=-7.8 required=5.0 tests=AWL,BAYES_00,KHOP_RCVD_UNTRUST,KHOP_THREADED,RCVD_IN_DNSWL_HI,RCVD_IN_HOSTKARMA_W,TW_CP,TW_XZ,T_RP_MATCHES_RCVD X-Spam-Check-By: sourceware.org Received: from mga03.intel.com (HELO mga03.intel.com) (143.182.124.21) by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Mon, 17 Dec 2012 16:02:00 +0000 Received: from azsmga002.ch.intel.com ([10.2.17.35]) by azsmga101.ch.intel.com with ESMTP; 17 Dec 2012 08:01:59 -0800 X-ExtLoop1: 1 Received: from swsutil001.isw.intel.com ([10.237.237.11]) by AZSMGA002.ch.intel.com with ESMTP; 17 Dec 2012 08:01:56 -0800 Received: from ulslx001.iul.intel.com (ulslx001.iul.intel.com [172.28.207.63]) by swsutil001.isw.intel.com (8.13.6/8.13.6/MailSET/Hub) with ESMTP id qBHG1tNU019675; Mon, 17 Dec 2012 16:01:55 GMT Received: from ulslx001.iul.intel.com (localhost [127.0.0.1]) by ulslx001.iul.intel.com with ESMTP id qBHG1sJX026496; Mon, 17 Dec 2012 17:01:54 +0100 Received: (from mmetzger@localhost) by ulslx001.iul.intel.com with id qBHG1s6K026492; Mon, 17 Dec 2012 17:01:54 +0100 From: markus.t.metzger@intel.com To: jan.kratochvil@redhat.com, palves@redhat.com, tromey@redhat.com, kettenis@gnu.org Cc: gdb-patches@sourceware.org, markus.t.metzger@gmail.com, Markus Metzger Subject: [patch v6 03/12] linux, btrace: perf_event based branch tracing Date: Mon, 17 Dec 2012 16:03:00 -0000 Message-Id: <1355760101-26237-4-git-send-email-markus.t.metzger@intel.com> In-Reply-To: <1355760101-26237-1-git-send-email-markus.t.metzger@intel.com> References: <1355760101-26237-1-git-send-email-markus.t.metzger@intel.com> X-IsSubscribed: yes Mailing-List: contact gdb-patches-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: gdb-patches-owner@sourceware.org X-SW-Source: 2012-12/txt/msg00570.txt.bz2 From: Markus Metzger Implement branch tracing on Linux based on perf_event such taht it can be shared between gdb and gdbserver. The actual btrace target ops will be implemented on top. 2012-12-17 Markus Metzger * common/linux_btrace.h: New file. * common/linux_btrace.c: New file. * Makefile.in (SFILES): Add btrace.c. (HFILES_NO_SRCDIR): Add common/linux-btrace.h. (COMMON_OBS): Add btrace.o. (linux-btrace.o): New rule. gdbserver/ * Makefile.in (SFILES): Add $(srcdir)/common/linux-btrace.c. (linux_btrace_h): New variable. (linux-btrace.o): New rule. --- gdb/Makefile.in | 6 +- gdb/common/linux-btrace.c | 425 +++++++++++++++++++++++++++++++++++++++++++++ gdb/common/linux-btrace.h | 76 ++++++++ gdb/gdbserver/Makefile.in | 7 +- 4 files changed, 512 insertions(+), 2 deletions(-) create mode 100644 gdb/common/linux-btrace.c create mode 100644 gdb/common/linux-btrace.h diff --git a/gdb/Makefile.in b/gdb/Makefile.in index d6e1dbb..4ee7067 100644 --- a/gdb/Makefile.in +++ b/gdb/Makefile.in @@ -830,7 +830,7 @@ gnulib/import/stddef.in.h gnulib/import/inttypes.in.h inline-frame.h skip.h \ common/common-utils.h common/xml-utils.h common/buffer.h common/ptid.h \ common/format.h common/host-defs.h utils.h common/queue.h \ common/linux-osdata.h gdb-dlfcn.h auto-load.h probe.h stap-probe.h \ -gdb_bfd.h sparc-ravenscar-thread.h ppc-ravenscar-thread.h +gdb_bfd.h sparc-ravenscar-thread.h ppc-ravenscar-thread.h common/linux-btrace.h # Header files that already have srcdir in them, or which are in objdir. @@ -1952,6 +1952,10 @@ vec.o: ${srcdir}/common/vec.c $(COMPILE) $(srcdir)/common/vec.c $(POSTCOMPILE) +linux-btrace.o: ${srcdir}/common/linux-btrace.c + $(COMPILE) $(srcdir)/common/linux-btrace.c + $(POSTCOMPILE) + # # gdb/tui/ dependencies # diff --git a/gdb/common/linux-btrace.c b/gdb/common/linux-btrace.c new file mode 100644 index 0000000..1231ccc --- /dev/null +++ b/gdb/common/linux-btrace.c @@ -0,0 +1,425 @@ +/* Linux-dependent part of branch trace support for GDB, and GDBserver. + + Copyright (C) 2012 Free Software Foundation, Inc. + + Contributed by Intel Corp. + + This file is part of GDB. + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . */ + +#ifdef GDBSERVER +#include "server.h" +#else +#include "defs.h" +#endif + +#include "linux-btrace.h" +#include "common-utils.h" +#include "gdb_assert.h" + +#if HAVE_LINUX_PERF_EVENT_H + +#include +#include +#include +#include +#include +#include +#include + +#if defined(__GNUC__) +# define memory_barrier() asm volatile ("" : : : "memory") +#else +# define memory_barrier() do {} while (0) +#endif + +/* A branch trace record in perf_event. */ +struct perf_event_bts +{ + /* The linear address of the branch source. */ + uint64_t from; + + /* The linear address of the branch destination. */ + uint64_t to; +}; + +/* A perf_event branch trace sample. */ +struct perf_event_sample +{ + /* The perf_event sample header. */ + struct perf_event_header header; + + /* The perf_event branch tracing payload. */ + struct perf_event_bts bts; +}; + +/* Get the perf_event header. */ + +static inline volatile struct perf_event_mmap_page * +perf_event_header (struct btrace_target_info* tinfo) +{ + return tinfo->buffer; +} + +/* Get the size of the perf_event mmap buffer. */ + +static inline size_t +perf_event_mmap_size (const struct btrace_target_info *tinfo) +{ + /* The branch trace buffer is preceded by a configuration page. */ + return (tinfo->size + 1) * PAGE_SIZE; +} + +/* Get the size of the perf_event buffer. */ + +static inline size_t +perf_event_buffer_size (struct btrace_target_info* tinfo) +{ + return tinfo->size * PAGE_SIZE; +} + +/* Get the start address of the perf_event buffer. */ + +static inline const uint8_t * +perf_event_buffer_begin (struct btrace_target_info* tinfo) +{ + return ((const uint8_t *) tinfo->buffer) + PAGE_SIZE; +} + +/* Get the end address of the perf_event buffer. */ + +static inline const uint8_t * +perf_event_buffer_end (struct btrace_target_info* tinfo) +{ + return perf_event_buffer_begin (tinfo) + perf_event_buffer_size (tinfo); +} + +/* Check whether an address is in the kernel. */ + +static inline int +perf_event_is_kernel_addr (const struct btrace_target_info *tinfo, + uint64_t addr) +{ + uint64_t mask; + + /* If we don't know the size of a pointer, we can't check. Let's assume it's + not a kernel address in this case. */ + if (tinfo->ptr_bits == 0) + return 0; + + /* A bit mask for the most significant bit in an address. */ + mask = (uint64_t) 1 << (tinfo->ptr_bits - 1); + + /* Check whether the most significant bit in the address is set. */ + return (addr & mask) != 0; +} + +/* Check whether a perf event record should be skipped. */ + +static inline int +perf_event_skip_record (const struct btrace_target_info *tinfo, + const struct perf_event_bts *bts) +{ + /* The hardware may report branches from kernel into user space. Branches + from user into kernel space will be suppressed. We filter the former to + provide a consistent branch trace excluding kernel. */ + return perf_event_is_kernel_addr (tinfo, bts->from); +} + +/* Perform a few consistency checks on a perf event sample record. This is + meant to catch cases when we get out of sync with the perf event stream. */ + +static inline int +perf_event_sample_ok (const struct perf_event_sample *sample) +{ + if (sample->header.type != PERF_RECORD_SAMPLE) + return 0; + + if (sample->header.size != sizeof (*sample)) + return 0; + + return 1; +} + +/* Branch trace is collected in a circular buffer [begin; end) as pairs of from + and to addresses (plus a header). + + Start points into that buffer at the next sample position. + We read the collected samples backwards from start. + + While reading the samples, we convert the information into a list of blocks. + For two adjacent samples s1 and s2, we form a block b such that b.begin = + s1.to and b.end = s2.from. + + In case the buffer overflows during sampling, one sample may have its lower + part at the end and its upper part at the beginning of the buffer. */ + +static VEC (btrace_block_s) * +perf_event_read_bts (struct btrace_target_info* tinfo, const uint8_t *begin, + const uint8_t *end, const uint8_t *start) +{ + VEC (btrace_block_s) *btrace = NULL; + struct perf_event_sample sample; + size_t read = 0, size = (end - begin); + struct btrace_block block = { 0, 0 }; + + gdb_assert (begin <= start); + gdb_assert (start <= end); + + /* The buffer may contain a partial record as its last entry (i.e. when the + buffer size is not a multiple of the sample size). */ + read = sizeof (sample) - 1; + + for (; read < size; read += sizeof (sample)) + { + const struct perf_event_sample *psample; + + /* Find the next perf_event sample in a backwards traversal. */ + start -= sizeof (sample); + + /* If we're still inside the buffer, we're done. */ + if (begin <= start) + psample = (const struct perf_event_sample *) start; + else + { + int missing; + + /* We're to the left of the ring buffer, we will wrap around and + reappear at the very right of the ring buffer. */ + + missing = (begin - start); + start = (end - missing); + + /* If the entire sample is missing, we're done. */ + if (missing == sizeof (sample)) + psample = (const struct perf_event_sample *) start; + else + { + uint8_t *stack; + + /* The sample wrapped around. The lower part is at the end and + the upper part is at the beginning of the buffer. */ + stack = (uint8_t *) &sample; + + /* Copy the two parts so we have a contiguous sample. */ + memcpy (stack, start, missing); + memcpy (stack + missing, begin, sizeof (sample) - missing); + + psample = &sample; + } + } + + if (perf_event_sample_ok (psample) == 0) + { + warning (_("Branch trace may be incomplete.")); + break; + } + + if (perf_event_skip_record (tinfo, &psample->bts) != 0) + continue; + + /* We found a valid sample, so we can complete the current block. */ + block.begin = psample->bts.to; + + VEC_safe_push (btrace_block_s, btrace, &block); + + /* Start the next block. */ + block.end = psample->bts.from; + } + + return btrace; +} + +/* See linux-btrace.h. */ + +int +linux_supports_btrace (void) +{ + return 1; +} + +/* See linux-btrace.h. */ + +int +linux_btrace_has_changed (struct btrace_target_info *tinfo) +{ + volatile struct perf_event_mmap_page *header = perf_event_header (tinfo); + + return header->data_head != tinfo->data_head; +} + +/* See linux-btrace.h. */ + +struct btrace_target_info * +linux_enable_btrace (ptid_t ptid) +{ + struct btrace_target_info *tinfo; + int pid; + + tinfo = xzalloc (sizeof (*tinfo)); + tinfo->attr.size = sizeof (tinfo->attr); + + tinfo->attr.type = PERF_TYPE_HARDWARE; + tinfo->attr.config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS; + tinfo->attr.sample_period = 1; + + /* We sample from and to address. */ + tinfo->attr.sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_ADDR; + + tinfo->attr.exclude_kernel = 1; + tinfo->attr.exclude_hv = 1; + tinfo->attr.exclude_idle = 1; + + tinfo->ptr_bits = 0; + + pid = ptid_get_lwp (ptid); + if (pid == 0) + pid = ptid_get_pid (ptid); + + errno = 0; + tinfo->file = syscall (SYS_perf_event_open, &tinfo->attr, pid, -1, -1, 0); + if (tinfo->file < 0) + goto err; + + /* We hard-code the trace buffer size. + At some later time, we should make this configurable. */ + tinfo->size = 1; + tinfo->buffer = mmap (NULL, perf_event_mmap_size (tinfo), + PROT_READ, MAP_SHARED, tinfo->file, 0); + if (tinfo->buffer == MAP_FAILED) + goto err_file; + + return tinfo; + +err_file: + close (tinfo->file); + +err: + xfree (tinfo); + return NULL; +} + +/* See linux-btrace.h. */ + +int +linux_disable_btrace (struct btrace_target_info *tinfo) +{ + int errcode; + + errno = 0; + errcode = munmap (tinfo->buffer, perf_event_mmap_size (tinfo)); + if (errcode != 0) + return errno; + + close (tinfo->file); + xfree (tinfo); + + return 0; +} + +/* See linux-btrace.h. */ + +VEC (btrace_block_s) * +linux_read_btrace (struct btrace_target_info *tinfo) +{ + VEC (btrace_block_s) *btrace = NULL; + volatile struct perf_event_mmap_page *header; + const uint8_t *begin, *end, *start; + unsigned long data_head, retries = 5; + size_t buffer_size; + + header = perf_event_header (tinfo); + buffer_size = perf_event_buffer_size (tinfo); + + /* We may need to retry reading the trace. See below. */ + while (retries--) + { + data_head = header->data_head; + + /* Make sure the trace data is up-to-date. */ + memory_barrier (); + + /* If there's new trace, let's read it. */ + if (data_head != tinfo->data_head) + { + /* Data_head keeps growing; the buffer itself is circular. */ + begin = perf_event_buffer_begin (tinfo); + start = begin + data_head % buffer_size; + + if (data_head <= buffer_size) + end = start; + else + end = perf_event_buffer_end (tinfo); + + btrace = perf_event_read_bts (tinfo, begin, end, start); + } + + /* The stopping thread notifies its ptracer before it is scheduled out. + On multi-core systems, the debugger might therefore run while the + kernel might be writing the last branch trace records. + + Let's check whether the data head moved while we read the trace. */ + if (data_head == header->data_head) + break; + } + + tinfo->data_head = data_head; + + return btrace; +} + +#else /* !HAVE_LINUX_PERF_EVENT_H */ + +/* See linux-btrace.h. */ + +int +linux_supports_btrace (void) +{ + return 0; +} + +/* See linux-btrace.h. */ + +int +linux_btrace_has_changed (struct btrace_target_info *tinfo) +{ + return 0; +} + +/* See linux-btrace.h. */ + +struct btrace_target_info * +linux_enable_btrace (ptid_t ptid) +{ + return NULL; +} + +/* See linux-btrace.h. */ + +int +linux_disable_btrace (struct btrace_target_info *tinfo) +{ + return ENOSYS; +} + +/* See linux-btrace.h. */ + +VEC (btrace_block_s) * +linux_read_btrace (struct btrace_target_info *tinfo) +{ + return NULL; +} + +#endif /* !HAVE_LINUX_PERF_EVENT_H */ diff --git a/gdb/common/linux-btrace.h b/gdb/common/linux-btrace.h new file mode 100644 index 0000000..b0f8459 --- /dev/null +++ b/gdb/common/linux-btrace.h @@ -0,0 +1,76 @@ +/* Linux-dependent part of branch trace support for GDB, and GDBserver. + + Copyright (C) 2012 Free Software Foundation, Inc. + + Contributed by Intel Corp. + + This file is part of GDB. + + This program is free software; you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation; either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . */ + +#ifndef LINUX_BTRACE_H +#define LINUX_BTRACE_H + +#include "btrace-common.h" +#include "config.h" +#include "vec.h" +#include "ptid.h" +#include +#include + +#if HAVE_LINUX_PERF_EVENT_H +# include +#endif + +/* Branch trace target information per thread. */ +struct btrace_target_info +{ +#if HAVE_LINUX_PERF_EVENT_H + /* The Linux perf_event configuration for collecting the branch trace. */ + struct perf_event_attr attr; + + /* The mmap configuration mapping the branch trace perf_event buffer. + + file .. the file descriptor + buffer .. the mmapped memory buffer + size .. the buffer's size in pages without the configuration page + data_head .. the data head from the last read */ + int file; + void *buffer; + size_t size; + unsigned long data_head; +#endif /* HAVE_LINUX_PERF_EVENT_H */ + + /* The size of a pointer in bits for this thread. + The information is used to identify kernel addresses in order to skip + records from/to kernel space. */ + int ptr_bits; +}; + +/* Check whether branch tracing is supported. */ +extern int linux_supports_btrace (void); + +/* Enable branch tracing for @ptid. */ +extern struct btrace_target_info *linux_enable_btrace (ptid_t ptid); + +/* Disable branch tracing and deallocate @tinfo. */ +extern int linux_disable_btrace (struct btrace_target_info *tinfo); + +/* Check whether there is new trace data available. */ +extern int linux_btrace_has_changed (struct btrace_target_info *); + +/* Read branch trace data. */ +extern VEC (btrace_block_s) *linux_read_btrace (struct btrace_target_info *); + +#endif /* LINUX_BTRACE_H */ diff --git a/gdb/gdbserver/Makefile.in b/gdb/gdbserver/Makefile.in index c4f1a2d..cdb6918 100644 --- a/gdb/gdbserver/Makefile.in +++ b/gdb/gdbserver/Makefile.in @@ -155,7 +155,7 @@ SFILES= $(srcdir)/gdbreplay.c $(srcdir)/inferiors.c $(srcdir)/dll.c \ $(srcdir)/common/vec.c $(srcdir)/common/gdb_vecs.c \ $(srcdir)/common/common-utils.c $(srcdir)/common/xml-utils.c \ $(srcdir)/common/linux-osdata.c $(srcdir)/common/ptid.c \ - $(srcdir)/common/buffer.c + $(srcdir)/common/buffer.c $(srcdir)/common/linux-btrace.c DEPFILES = @GDBSERVER_DEPFILES@ @@ -423,6 +423,8 @@ signals_h = $(srcdir)/../../include/gdb/signals.h $(signals_def) ptid_h = $(srcdir)/../common/ptid.h ax_h = $(srcdir)/ax.h agent_h = $(srcdir)/../common/agent.h +linux_btrace_h = $(srcdir)/../common/linux-btrace.h \ + $(srcdir)/../common/btrace-common.h $(vec_h) $(ptid_h) linux_osdata_h = $(srcdir)/../common/linux-osdata.h vec_h = $(srcdir)/../common/vec.h gdb_vecs_h = $(srcdir)/../common/gdb_vecs.h @@ -542,6 +544,9 @@ agent.o: ../common/agent.c $(COMPILE) $< $(POSTCOMPILE) +linux-btrace.o: ../common/linux-btrace.c $(linux_btrace_h) $(server_h) + $(CC) -c $(CPPFLAGS) $(INTERNAL_CFLAGS) $< + # We build vasprintf with -DHAVE_CONFIG_H because we want that unit to # include our config.h file. Otherwise, some system headers do not get # included, and the compiler emits a warning about implicitly defined -- 1.7.6.5