CoreSight

coresight@lists.linaro.org

4 participants
2387 discussions

[PATCH 0/9] perf tools: Cleanup for CoreSight trace decoding code

by Mathieu Poirier

This set serves no other purpose than doing preparatory work before support for CPU-wide trace scenarios is introduced. It cleans up a lot of code and reorganise things so that what is currently being used for per-thread decoding can be used as well for CPU-wide sessions. It is based on Leo's latest submission to the CS mailing list. I am sending this here to circumvent the constraints imposed by the merge window and will go public next week when the first release candidate for the 4.18 cycle has been minted. Leo and Rob, you guys are now familiar with the perf tools code and as such would appreciate your thougths on this. The code has been uploaded here[1]. Regards, Mathieu [1]. https://git.linaro.org/people/mathieu.poirier/coresight.git perf-opencsd-master-cpu-wide-support Mathieu Poirier (9): perf tools: Remove unused structure field "state" perf tools: Introducing cs_etm_decoder__init_dparams() perf tools: Introducing function cs_etm__init_trace_params() perf tools: Cleaning up function cs_etm__alloc_queue() perf tools: Rethink kernel address initialisation perf tools: Make cs_etm__run_decoder() queue independent perf tools: Modularize main decoder function perf tools: Modularize main packet processing loop perf tools: Modularize auxtrace_buffer fetch function tools/perf/util/cs-etm-decoder/cs-etm-decoder.h | 13 +- tools/perf/util/cs-etm.c | 339 ++++++++++++++---------- 2 files changed, 210 insertions(+), 142 deletions(-) -- 2.7.4

7 years

OpenCSD ITM support

by Athalye, Ashwini

The OpenCSD README says that ITM software trace will be added in the future. Two questions regarding this: 1. When is the support expected to be available? 2. Will the support include ITM software message decode and DWT hardware event decode as well or just the former? Thanks, Ashwini

7 years

Re: Regression of 'perf record' on Hikey620

by Mathieu Poirier

On Tue, 19 Jun 2018 at 07:37, Leo Yan <leo.yan(a)linaro.org> wrote: > > Hi Mathieu, > > On Mon, Jun 18, 2018 at 07:26:20PM -0600, Mathieu Poirier wrote: > > On Mon, 18 Jun 2018 at 18:59, Leo Yan <leo.yan(a)linaro.org> wrote: > > > > > > Hi all, > > > > > > Just in case you have the same issue, when I use acme's branch perf/core > > > with latest commit e238cf2e3d2e ("perf intel-pt: Fix packet decoding of > > > CYC packets"), I can easily reproduce below failure with 'perf record' > > > command. > > > > > > If you have fixing for this, please let me know. Otherwise I will dig a > > > bit for this issue (probably relying on 'git bisect'). > > > > Haven't see this, though my tree is based on mainline 4.18-rc1. Let > > me know what bisect gives you. > > Finally I found this issue is caused by I removed 'nohlt' in kernel > command line, so CPU Idle states are enabled. > > After CPU Idle states are enabled, we cannot access the register > /sys/bus/coresight/devices/f659c000.etm/mgmt/trcauthstatus when the > CPU stays in the low power states. Interesting. > > So following this issue, I have two questions (sorry if you guys have > discussed these questions before): > > - The first one question is should we support Runtime PM (Or GenPD) in > Coresight? So when we access some component (e.g. the register > trcauthstatus) we need to ensure the corresponding power domain is > powered on properly before we access it. We rely heavily on runtime PM everywhere [1] in the code to make sure devices are accessed only when they're powered up. TRCAUTHSTATUS is a "management" register [2] and as such should not need to have the core powered up [3], which is why it is accessed with macro coresight_etm4x_reg() rather than coresight_etm4x_cross_read(). That being said implementation can differ from the guidelines. First thing to do is to try going from coresight_etm4x_reg() to coresight_etm4x_cross_read() here [4] and see if there is any changes. If that doesn't work you need to ask HiSilicon what power domain the register is in. From there get back to me and we'll talk things over. [1]. https://elixir.bootlin.com/linux/latest/source/arch/arm64/boot/dts/arm/juno… [2]. See "ARM Embedded Trace Macrocell Architecture Specifiation" (ID032614), table 7-1 on page 7-311 for details. [3]. Same document as above, table 7-4 on page 7-314 [4]. https://elixir.bootlin.com/linux/latest/source/drivers/hwtracing/coresight/… > > - The second one question is another side topic: if we support CPU > wide tracing, should we support suspend/resume flow for Coresight > ETM? So we can save and restore context for CPU respectively. When using Perf suspend/resume is already supported - tracers won't be kept on whey there is no more processes executing on a CPU (just before going to idle). Tracers will also be switched back on once the CPU has resumed, just before processes start being executed on it. Configuration of the trace unit is kept in memory. In sysFS mode I expected users to disable CPUidle, but that can be change if you have a use case. Thanks, Mathieu > > Thanks, > Leo Yan > > > > ---8<--- > > > > > > # ./perf record -e cs_etm/(a)f6402000.etf/ --per-thread uname > > > [ 240.623458] Internal error: synchronous external abort: 96000210 [#10] PREEMPT SMP > > > [ 240.631055] Modules linked in: > > > [ 240.634117] CPU: 1 PID: 2793 Comm: perf Tainted: G D 4.17.0-08502-ge238cf2 #26 > > > [ 240.642648] Hardware name: HiKey Development Board (DT) > > > [ 240.647876] pstate: 40000005 (nZcv daif -PAN -UAO) > > > [ 240.652677] pc : trcauthstatus_show+0x3c/0x78 > > > [ 240.657035] lr : trcauthstatus_show+0x34/0x78 > > > [ 240.661391] sp : ffff00000dcfbbc0 > > > [ 240.664703] x29: ffff00000dcfbbc0 x28: 0000000000000001 > > > [ 240.670019] x27: 00000000007000c0 x26: ffff800036fa4048 > > > [ 240.675335] x25: 0000000000001000 x24: ffff000008d95330 > > > [ 240.680651] x23: ffff8000372ad880 x22: ffff80003b346cb8 > > > [ 240.685966] x21: ffff800036fa8080 x20: ffff80003b346ca8 > > > [ 240.691281] x19: ffff0000099d5fb8 x18: 0000000000000000 > > > [ 240.696597] x17: 0000ffffbb44bb68 x16: ffff0000082a3908 > > > [ 240.701911] x15: 0000000000000000 x14: 0000000000000016 > > > [ 240.707227] x13: 7375746174736874 x12: 75616372742f746d > > > [ 240.712542] x11: 0000000000000020 x10: 0000000080070007 > > > [ 240.717857] x9 : 0000000000000000 x8 : ffff800036fa9080 > > > [ 240.723173] x7 : 0000000000000000 x6 : 000000000000003f > > > [ 240.728488] x5 : 0000000000000040 x4 : 0000000000000000 > > > [ 240.733803] x3 : 0000000000000000 x2 : 4d223a68697bbd00 > > > [ 240.739118] x1 : ffff80001fc30f80 x0 : 0000000000000001 > > > [ 240.744435] Process perf (pid: 2793, stack limit = 0x0000000051ca53e7) > > > [ 240.750965] Call trace: > > > [ 240.753411] trcauthstatus_show+0x3c/0x78 > > > [ 240.757423] dev_attr_show+0x3c/0x80 > > > [ 240.761002] sysfs_kf_seq_show+0xc0/0x158 > > > [ 240.765012] kernfs_seq_show+0x44/0x50 > > > [ 240.768764] seq_read+0x1cc/0x4b8 > > > [ 240.772078] kernfs_fop_read+0x13c/0x1e0 > > > [ 240.776002] __vfs_read+0x60/0x170 > > > [ 240.779403] vfs_read+0x94/0x150 > > > [ 240.782630] ksys_read+0x6c/0xd8 > > > [ 240.785856] sys_read+0x34/0x48 > > > [ 240.788998] el0_svc_naked+0x30/0x34 > > > [ 240.792575] Code: f9404c53 97f2054e f9400273 913ee273 (b9400273) > > > [ 240.798673] ---[ end trace 838ff5bf36115622 ]--- > > > > > > Message from syslogd@linaro-developer at May 23 09:50:03 ... > > > kernel:[ 240.623458] Internal error: synchronous external abort: 96000210 [#10] PREEMPT SMP > > > > > > Message from syslogd@linaro-developer at May 23 09:50:03 ... > > > kernel:[ 240.744435] Process perf (pid: 2793, stack limit = 0x0000000051ca53e7) > > > > > > Message from syslogd@linaro-developer at May 23 09:50:03 ... > > > kernel:[ 240.792575] Code: f9404c53 97f2054e f9400273 913ee273 (b9400273) > > > Segmentation fault > > > root@linaro-developer:~# > > >

7 years

Debian packaging - patches

by Wookey

I've been packaging libopenCSD for debian. That has resulted in various changes, largely to the makefiles. Most of those changes should go upstream, rather than exist only in the debian version. I'll post my issues/changes here for discussion so we can decide what is upstreamable. In the meantime anyone interested can try the debian packages here: http://wookware.org/software/repo/ either manually from: http://wookware.org/software/repo/pool/main/libo/libopencsd/ or with deb [ trusted=yes ] http://wookware.org/software/repo unstable main apt update; apt install libopencsd0 (and/or libopencsd-dev libopencsd-doc libopencsd-dbgsym) (for debian unstable, amd64 only) (package is signed, but repo isn't. Sorry) Or get the source and rebuild for a different debian-based distro/release. I've not separated all my changes into proper standalone patches yet, but some are so let's start with those. 1) The doxygen doc build doesn't find all the components because they moved. This fixes that: Index: libopencsd-0.8.0/decoder/docs/doxygen_config.dox =================================================================== --- libopencsd-0.8.0.orig/decoder/docs/doxygen_config.dox +++ libopencsd-0.8.0/decoder/docs/doxygen_config.dox @@ -765,11 +765,11 @@ WARN_LOGFILE = INPUT = ../include \ ../include/interfaces \ - ../include/etmv3 \ - ../include/etmv4 \ - ../include/ptm \ - ../include/c_api \ - ../include/stm \ + ../include/opencsd/etmv3 \ + ../include/opencsd/etmv4 \ + ../include/opencsd/ptm \ + ../include/opencsd/c_api \ + ../include/opencsd/stm \ ../include/mem_acc \ ../../README.md \ . \ 2) The makefile provides no doc-build. The patch below adds that. I didn't include an install-docs target, although if you added one with a variable to set the target dir then I'd use it. Index: libopencsd-0.8.0/decoder/build/linux/makefile =================================================================== --- libopencsd-0.8.0.orig/decoder/build/linux/makefile +++ libopencsd-0.8.0/decoder/build/linux/makefile @@ -155,6 +155,13 @@ tests: libs cd $(OCSD_ROOT)/tests/build/linux/trc_pkt_lister && make cd $(OCSD_ROOT)/tests/build/linux/c_api_pkt_print_test && make +# +# build docs +.PHONY: docs +docs: + (cd $(OCSD_ROOT/docs); doxygen doxygen_config.dox) + + ############################################################# # clean targets # @@ -176,3 +183,4 @@ clean_install: rm -f $(INSTALL_LIB_DIR)/lib$(LIB_BASE_NAME).so rm -f $(INSTALL_LIB_DIR)/lib$(LIB_CAPI_NAME).so rm -rf $(INSTALL_INCLUDE_DIR)/$(LIB_UAPI_INC_DIR) + rm -rf $(OCSD_ROOT)/docs/html 3) The static libraries are built but not installed. Now static libraries aren't much use to anyone these days, but if you build them then you might as well install them. This patch does that: Index: libopencsd-0.8.0/decoder/build/linux/makefile =================================================================== --- libopencsd-0.8.0.orig/decoder/build/linux/makefile +++ libopencsd-0.8.0/decoder/build/linux/makefile @@ -136,6 +136,8 @@ mkdir -p $(INSTALL_LIB_DIR) $(INSTALL_INCLUDE_DIR) $(INSTALL) --mode=644 $(LIB_TARGET_DIR)/lib$(LIB_BASE_NAME).so $(INSTALL_LIB_DIR)/ $(INSTALL) --mode=644 $(LIB_TARGET_DIR)/lib$(LIB_CAPI_NAME).so $(INSTALL_LIB_DIR)/ + $(INSTALL) --mode=644 $(LIB_TARGET_DIR)/lib$(LIB_BASE_NAME).a $(INSTALL_LIB_DIR)/ + $(INSTALL) --mode=644 $(LIB_TARGET_DIR)/lib$(LIB_CAPI_NAME).a $(INSTALL_LIB_DIR)/ cd $(OCSD_ROOT)/build/linux/rctdl_c_api_lib && make install_inc ################################ @@ -192,6 +194,6 @@ cd $(OCSD_ROOT)/tests/build/linux/c_api_pkt_print_test && make clean clean_install: - rm -f $(INSTALL_LIB_DIR)/lib$(LIB_BASE_NAME).so - rm -f $(INSTALL_LIB_DIR)/lib$(LIB_CAPI_NAME).so + rm -f $(INSTALL_LIB_DIR)/lib$(LIB_BASE_NAME).{so,a} + rm -f $(INSTALL_LIB_DIR)/lib$(LIB_CAPI_NAME).{so,a} rm -rf $(INSTALL_INCLUDE_DIR)/$(LIB_UAPI_INC_DIR) 4) The makefile arbitrarily restricts the build to arm and x86 architectures. The only reason for this is in order to build into a particular named directory, and to set -m32/-m64 flags for x86 which should be defaulted correctly on any sensible toolchain or cross-toolchain anyway. This code can build, and be used, on any arch and the makefile should allow that, so this should be fixed. I don't see the need to build into a host-arch named PLAT_DIR directory. Is there one? It only does one build at a time, whether crossing or not, so I see no reason not to use a fixed name for this dir, such as 'builddir'. This has no effect on the ability to cross or not. If you really want to change the name of PLAT_DIR then find out the HOST triplet and use that. On debian this is dpkg-architecture -q DEB_HOST_GNU_TYPE, (which is also the same prefix as would be used to specify the toolchain prefix for crossing). But like I say, I think a fixed builddir is actually all you need unless I'm missing something. So this patch lets the build work on any arch: Index: libopencsd-0.8.0/decoder/build/linux/makefile =================================================================== --- libopencsd-0.8.0.orig/decoder/build/linux/makefile +++ libopencsd-0.8.0/decoder/build/linux/makefile @@ -92,27 +92,6 @@ BUILD_VARIANT=rel endif -# platform bit size variant -ifeq ($(ARCH),x86) - MFLAG:="-m32" - BIT_VARIANT=32 -else ifeq ($(ARCH),x86_64) - MFLAG:="-m64" - BIT_VARIANT=64 -else ifeq ($(ARCH),arm) - BIT_VARIANT=-arm -else ifeq ($(ARCH),arm64) - BIT_VARIANT=-arm64 -else ifeq ($(ARCH),aarch64) - BIT_VARIANT=-arm64 -else ifeq ($(ARCH),aarch32) - BIT_VARIANT=-arm -endif - -MASTER_CC_FLAGS += $(MFLAG) -MASTER_CPP_FLAGS += $(MFLAG) -MASTER_LINKER_FLAGS += $(MFLAG) - # export build flags export MASTER_CC_FLAGS export MASTER_CPP_FLAGS @@ -120,7 +99,7 @@ export MASTER_LINKER_FLAGS export MASTER_LIB_FLAGS # target directories -export PLAT_DIR=linux$(BIT_VARIANT)/$(BUILD_VARIANT) +export PLAT_DIR=builddir export LIB_TARGET_DIR=$(OCSD_LIB_ROOT)/$(PLAT_DIR) export LIB_TEST_TARGET_DIR=$(OCSD_TESTS)/lib/$(PLAT_DIR) export BIN_TEST_TARGET_DIR=$(OCSD_TESTS)/bin/$(PLAT_DIR) All these patches included as files too. Wookey -- Principal hats: Linaro, Debian, Wookware, ARM http://wookware.org/

7 years

OpenCSD library v0.9.0 released.

by Mike Leach

Latest version of OpenCSD library v0.9.0 is now released in the github repository. https://github.com/Linaro/OpenCSD This contains: - Updates for improved client performance - including perf. - Additional documentation consisting of Programmers Guide, and Generic Trace Output Packet reference. Regards Mike -- Mike Leach Principal Engineer, ARM Ltd. Manchester Design Centre. UK

7 years

New ETR driver on GitHub

by Mathieu Poirier

Good afternoon to all, For the 4.18 cycle I pushed Suzuki's new ETR driver implementation to GitHub's perf-opencsd repository [1]. It includes support for SG and contiguous mode, all handled automatically without the need for user intervention. I gave it a couple of spin and things work as advertised. So it's up there now - please try it out in your setup while I review the patches on the public mailing list. Note that this has no correlation with the support for CPU-wide scenarios patchset I sent out last week. Regards, Mathieu [1]. https://github.com/Linaro/perf-opencsd/commits/master

7 years

[PATCH 00/12] perf tools: Add support for CoreSight CPU-wide session

by Mathieu Poirier

This set adds support for CoreSight CPU-wide trace sessions. It borrows most of its code from the per-thread implementation with exception that range packets are processed and synthesised according to the time the trace they contain has been executed. This is done using the timestamp and contextID feature available on ETM4x tracers (ETM3x/PTM aren't addressed yet). Decoding between processors is done in chronological order using a min heap. Of special interest is the way timestamp packets are used to account for temporal execution of traced instructions. Since a timestamp typically happen after range packets have been recorded, the timestamp from the previous range is used as the start time of the current range. When a timestamp for the previous range doesn't exist (i.e start of trace or discontinuity) the start time is estimated. Open question: -------------- At this time the implementation supports tracing a single CPU since the only HW we have exhibit an N:1 source/sink topology. The HW itself does support collecting traces from more than one source but using the feature in this way could be very confusing and mislead users. For example the following: # perf record -e cs_etm/20070000.etr/ -C 2,3 application1 would end up tracing everyting that is happening on CPU 2 and 3 for as long as appliation1 is executing. Because the HW doesn't give us an interrupt when buffers are full, traces from one CPU could easily clobber traces from the other, giving the impression that nothing was executed on the latter. So this would work: # perf record -e cs_etm/20070000.etr/ -C 3 application1 I am open to discussion on the topic should someone think of something. As with the cleanup set this code has been uploaded here [1]. Thanks, Mathieu [1].https://git.linaro.org/people/mathieu.poirier/coresight.git perf-opencsd-master-cpu-wide-support Mathieu Poirier (12): perf tools: Add defines for CONTEXTID configuration perf tools: Configure contextID tracing in CPU-wide mode perf tools: Configure timestsamp generation in CPU-wide mode perf tools: Configure SWITCH_EVENTS in CPU-wide mode perf tools: Add handling of itrace start events perf tools: Add handling of switch-CPU-wide events perf tools: Linking PE contextID with perf thread mechanic perf tools: Allocate decoder tree as needed perf tools: Make cs_etm__dump_event() work with CPU-wide scenarios perf tools: Add notion of time to the decoding code perf tools: Make function cs_etm_decoder__clear_buffer() public perf tools: Add support for CPU-wide trace scenarios include/linux/coresight-pmu.h | 2 + tools/include/linux/coresight-pmu.h | 2 + tools/perf/arch/arm/util/cs-etm.c | 174 ++++++++++-- tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 140 +++++++++- tools/perf/util/cs-etm-decoder/cs-etm-decoder.h | 4 +- tools/perf/util/cs-etm.c | 334 ++++++++++++++++++++++-- tools/perf/util/cs-etm.h | 17 ++ 7 files changed, 623 insertions(+), 50 deletions(-) -- 2.7.4

7 years

Regression of 'perf record' on Hikey620

by Leo Yan

Hi all, Just in case you have the same issue, when I use acme's branch perf/core with latest commit e238cf2e3d2e ("perf intel-pt: Fix packet decoding of CYC packets"), I can easily reproduce below failure with 'perf record' command. If you have fixing for this, please let me know. Otherwise I will dig a bit for this issue (probably relying on 'git bisect'). Thanks, Leo Yan ---8<--- # ./perf record -e cs_etm/(a)f6402000.etf/ --per-thread uname [ 240.623458] Internal error: synchronous external abort: 96000210 [#10] PREEMPT SMP [ 240.631055] Modules linked in: [ 240.634117] CPU: 1 PID: 2793 Comm: perf Tainted: G D 4.17.0-08502-ge238cf2 #26 [ 240.642648] Hardware name: HiKey Development Board (DT) [ 240.647876] pstate: 40000005 (nZcv daif -PAN -UAO) [ 240.652677] pc : trcauthstatus_show+0x3c/0x78 [ 240.657035] lr : trcauthstatus_show+0x34/0x78 [ 240.661391] sp : ffff00000dcfbbc0 [ 240.664703] x29: ffff00000dcfbbc0 x28: 0000000000000001 [ 240.670019] x27: 00000000007000c0 x26: ffff800036fa4048 [ 240.675335] x25: 0000000000001000 x24: ffff000008d95330 [ 240.680651] x23: ffff8000372ad880 x22: ffff80003b346cb8 [ 240.685966] x21: ffff800036fa8080 x20: ffff80003b346ca8 [ 240.691281] x19: ffff0000099d5fb8 x18: 0000000000000000 [ 240.696597] x17: 0000ffffbb44bb68 x16: ffff0000082a3908 [ 240.701911] x15: 0000000000000000 x14: 0000000000000016 [ 240.707227] x13: 7375746174736874 x12: 75616372742f746d [ 240.712542] x11: 0000000000000020 x10: 0000000080070007 [ 240.717857] x9 : 0000000000000000 x8 : ffff800036fa9080 [ 240.723173] x7 : 0000000000000000 x6 : 000000000000003f [ 240.728488] x5 : 0000000000000040 x4 : 0000000000000000 [ 240.733803] x3 : 0000000000000000 x2 : 4d223a68697bbd00 [ 240.739118] x1 : ffff80001fc30f80 x0 : 0000000000000001 [ 240.744435] Process perf (pid: 2793, stack limit = 0x0000000051ca53e7) [ 240.750965] Call trace: [ 240.753411] trcauthstatus_show+0x3c/0x78 [ 240.757423] dev_attr_show+0x3c/0x80 [ 240.761002] sysfs_kf_seq_show+0xc0/0x158 [ 240.765012] kernfs_seq_show+0x44/0x50 [ 240.768764] seq_read+0x1cc/0x4b8 [ 240.772078] kernfs_fop_read+0x13c/0x1e0 [ 240.776002] __vfs_read+0x60/0x170 [ 240.779403] vfs_read+0x94/0x150 [ 240.782630] ksys_read+0x6c/0xd8 [ 240.785856] sys_read+0x34/0x48 [ 240.788998] el0_svc_naked+0x30/0x34 [ 240.792575] Code: f9404c53 97f2054e f9400273 913ee273 (b9400273) [ 240.798673] ---[ end trace 838ff5bf36115622 ]--- Message from syslogd@linaro-developer at May 23 09:50:03 ... kernel:[ 240.623458] Internal error: synchronous external abort: 96000210 [#10] PREEMPT SMP Message from syslogd@linaro-developer at May 23 09:50:03 ... kernel:[ 240.744435] Process perf (pid: 2793, stack limit = 0x0000000051ca53e7) Message from syslogd@linaro-developer at May 23 09:50:03 ... kernel:[ 240.792575] Code: f9404c53 97f2054e f9400273 913ee273 (b9400273) Segmentation fault root@linaro-developer:~#

7 years

[PATCH v1 1/6] perf cs-etm: Introduce invalid address macro

by Leo Yan

This patch introduces invalid address macro and uses it to replace dummy value '0xdeadbeefdeadbeefUL'. Signed-off-by: Leo Yan <leo.yan(a)linaro.org> --- tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 8 ++++---- tools/perf/util/cs-etm-decoder/cs-etm-decoder.h | 2 ++ 2 files changed, 6 insertions(+), 4 deletions(-) diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c index 4d5fc37..24aabf0 100644 --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c @@ -261,8 +261,8 @@ static void cs_etm_decoder__clear_buffer(struct cs_etm_decoder *decoder) decoder->tail = 0; decoder->packet_count = 0; for (i = 0; i < MAX_BUFFER; i++) { - decoder->packet_buffer[i].start_addr = 0xdeadbeefdeadbeefUL; - decoder->packet_buffer[i].end_addr = 0xdeadbeefdeadbeefUL; + decoder->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR; + decoder->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR; decoder->packet_buffer[i].last_instr_taken_branch = false; decoder->packet_buffer[i].exc = false; decoder->packet_buffer[i].exc_ret = false; @@ -295,8 +295,8 @@ cs_etm_decoder__buffer_packet(struct cs_etm_decoder *decoder, decoder->packet_buffer[et].exc = false; decoder->packet_buffer[et].exc_ret = false; decoder->packet_buffer[et].cpu = *((int *)inode->priv); - decoder->packet_buffer[et].start_addr = 0xdeadbeefdeadbeefUL; - decoder->packet_buffer[et].end_addr = 0xdeadbeefdeadbeefUL; + decoder->packet_buffer[et].start_addr = CS_ETM_INVAL_ADDR; + decoder->packet_buffer[et].end_addr = CS_ETM_INVAL_ADDR; if (decoder->packet_count == MAX_BUFFER - 1) return OCSD_RESP_WAIT; diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h index 743f5f4..cb31b19 100644 --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h @@ -13,6 +13,8 @@ #include <linux/types.h> #include <stdio.h> +#define CS_ETM_INVAL_ADDR 0xdeadbeefdeadbeefUL + struct cs_etm_decoder; struct cs_etm_buffer { -- 2.7.4

7 years

[RFT v4 1/3] perf cs-etm: Generate branch sample for exception packet

by Leo Yan

The exception packet appears as one element with 'elem_type' == OCSD_GEN_TRC_ELEM_EXCEPTION or OCSD_GEN_TRC_ELEM_EXCEPTION_RET, which present for exception entry and exit respectively. The decoder set packet fields 'packet->exc' and 'packet->exc_ret' to indicate the exception packets; but exception packets don't have dedicated sample type and shares the same sample type CS_ETM_RANGE with normal instruction packets. As result, the exception packets are taken as normal instruction packets and this introduces confusion to mix different packet types. Furthermore, these instruction range packets will be processed for branch sample only when 'packet->last_instr_taken_branch' is true, otherwise they will be omitted, this can introduce mess for exception and exception returning due we don't have complete address range info for context switching. To process exception packets properly, this patch introduce two new sample type: CS_ETM_EXCEPTION and CS_ETM_EXCEPTION_RET; for these two kind packets, they will be handled by cs_etm__exception(). The func cs_etm__exception() forces to set previous CS_ETM_RANGE packet flag 'prev_packet->last_instr_taken_branch' to true, this matches well with the program flow when the exception is trapped from user space to kernel space, no matter if the most recent flow has branch taken or not; this is also safe for returning to user space after exception handling. After exception packets have their own sample type, the packet fields 'packet->exc' and 'packet->exc_ret' aren't needed anymore, so remove them. Signed-off-by: Leo Yan <leo.yan(a)linaro.org> --- tools/perf/util/cs-etm-decoder/cs-etm-decoder.c | 26 +++++++++++++++++------ tools/perf/util/cs-etm-decoder/cs-etm-decoder.h | 10 ++++----- tools/perf/util/cs-etm.c | 28 +++++++++++++++++++++++++ 3 files changed, 53 insertions(+), 11 deletions(-) diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c index 24aabf0..c1715ff 100644 --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.c @@ -264,8 +264,6 @@ static void cs_etm_decoder__clear_buffer(struct cs_etm_decoder *decoder) decoder->packet_buffer[i].start_addr = CS_ETM_INVAL_ADDR; decoder->packet_buffer[i].end_addr = CS_ETM_INVAL_ADDR; decoder->packet_buffer[i].last_instr_taken_branch = false; - decoder->packet_buffer[i].exc = false; - decoder->packet_buffer[i].exc_ret = false; decoder->packet_buffer[i].cpu = INT_MIN; } } @@ -292,8 +290,6 @@ cs_etm_decoder__buffer_packet(struct cs_etm_decoder *decoder, decoder->packet_count++; decoder->packet_buffer[et].sample_type = sample_type; - decoder->packet_buffer[et].exc = false; - decoder->packet_buffer[et].exc_ret = false; decoder->packet_buffer[et].cpu = *((int *)inode->priv); decoder->packet_buffer[et].start_addr = CS_ETM_INVAL_ADDR; decoder->packet_buffer[et].end_addr = CS_ETM_INVAL_ADDR; @@ -345,6 +341,22 @@ cs_etm_decoder__buffer_trace_on(struct cs_etm_decoder *decoder, CS_ETM_TRACE_ON); } +static ocsd_datapath_resp_t +cs_etm_decoder__buffer_exception(struct cs_etm_decoder *decoder, + const uint8_t trace_chan_id) +{ + return cs_etm_decoder__buffer_packet(decoder, trace_chan_id, + CS_ETM_EXCEPTION); +} + +static ocsd_datapath_resp_t +cs_etm_decoder__buffer_exception_ret(struct cs_etm_decoder *decoder, + const uint8_t trace_chan_id) +{ + return cs_etm_decoder__buffer_packet(decoder, trace_chan_id, + CS_ETM_EXCEPTION_RET); +} + static ocsd_datapath_resp_t cs_etm_decoder__gen_trace_elem_printer( const void *context, const ocsd_trc_index_t indx __maybe_unused, @@ -370,10 +382,12 @@ static ocsd_datapath_resp_t cs_etm_decoder__gen_trace_elem_printer( trace_chan_id); break; case OCSD_GEN_TRC_ELEM_EXCEPTION: - decoder->packet_buffer[decoder->tail].exc = true; + resp = cs_etm_decoder__buffer_exception(decoder, + trace_chan_id); break; case OCSD_GEN_TRC_ELEM_EXCEPTION_RET: - decoder->packet_buffer[decoder->tail].exc_ret = true; + resp = cs_etm_decoder__buffer_exception_ret(decoder, + trace_chan_id); break; case OCSD_GEN_TRC_ELEM_PE_CONTEXT: case OCSD_GEN_TRC_ELEM_EO_TRACE: diff --git a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h index 108dc9d..cb57756 100644 --- a/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h +++ b/tools/perf/util/cs-etm-decoder/cs-etm-decoder.h @@ -25,9 +25,11 @@ struct cs_etm_buffer { }; enum cs_etm_sample_type { - CS_ETM_EMPTY = 0, - CS_ETM_RANGE = 1 << 0, - CS_ETM_TRACE_ON = 1 << 1, + CS_ETM_EMPTY = 0, + CS_ETM_RANGE = 1 << 0, + CS_ETM_TRACE_ON = 1 << 1, + CS_ETM_EXCEPTION = 1 << 2, + CS_ETM_EXCEPTION_RET = 1 << 3, }; struct cs_etm_packet { @@ -35,8 +37,6 @@ struct cs_etm_packet { u64 start_addr; u64 end_addr; u8 last_instr_taken_branch; - u8 exc; - u8 exc_ret; int cpu; }; diff --git a/tools/perf/util/cs-etm.c b/tools/perf/util/cs-etm.c index 2ae6402..b85100b 100644 --- a/tools/perf/util/cs-etm.c +++ b/tools/perf/util/cs-etm.c @@ -942,6 +942,25 @@ static int cs_etm__sample(struct cs_etm_queue *etmq) return 0; } +static int cs_etm__exception(struct cs_etm_queue *etmq) +{ + /* + * When the exception packet is inserted, whether the last instruction + * in previous range packet is taken branch or not, we need to force + * to set 'prev_packet->last_instr_taken_branch' to true. This ensures + * to generate branch sample for the instruction range before the + * exception is trapped to kernel or before the exception returning. + * + * The exception packet includes the dummy address values, so don't + * swap PACKET with PREV_PACKET. This keeps PREV_PACKET to be useful + * for generating instruction and branch samples. + */ + if (etmq->prev_packet->sample_type == CS_ETM_RANGE) + etmq->prev_packet->last_instr_taken_branch = true; + + return 0; +} + static int cs_etm__flush(struct cs_etm_queue *etmq) { int err = 0; @@ -1057,6 +1076,15 @@ static int cs_etm__run_decoder(struct cs_etm_queue *etmq) */ cs_etm__sample(etmq); break; + case CS_ETM_EXCEPTION: + case CS_ETM_EXCEPTION_RET: + /* + * If the exception packet is coming, + * make sure the previous instruction + * range packet to be handled properly. + */ + cs_etm__exception(etmq); + break; case CS_ETM_TRACE_ON: /* * Discontinuity in trace, flush -- 2.7.4

7 years

Jump to page:

2025

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

CoreSight