Details

Reviewers

kib
markj

Summary

This adds two new counters modes:

PMC_MODE_ST -- system-wide tracing
PMC_MODE_TT -- thread virtual tracing

Four new PMC ioctl calls:

LOG_KERNEL_MAP

Used to request kernel mappings to be send over pmc log. This is required in order to configure address ranges filter prior to starting system-wide PMC trace.

THREAD_UNSUSPEND

Used to release the thread. As user thread can call mmap() any time, we put thread on suspend, then we ship mmap result over PMC log, so user can get idea about mapping location (for libc etc) and reconfigure PMC address ranges prior to thread execution.

TRACE_READ

Used to get current fill of the trace buffers.

TRACE_CONFIG

Used to reconfigure tracing address ranges after PMC allocation and prior starting PMC.

We now also create /dev/pmcX per each cpu core, so user can access trace buffers.

Example usage:

pmctrace -u -e branches uname
pmctrace -u -e branches -i libc.so.7 -f memset sleep 1
sudo pmctrace -s -e branches -i kernel -f cpu_switch

Other parts of patch:
part 2
part 3
part 4
part 5
part 6
part 7
part 8
part 9

Test Plan

Tested on ThinkPad X1 Carbon (Kabylake)
Tested on DragonBoard 410c (ARM64)

Diff Detail

Lint

Lint Skipped

Unit

Tests Skipped

Event Timeline

There are a very large number of changes, so older changes are hidden. Show Older Changes

kib added inline comments.Mar 2 2018, 11:26 PM

sys/dev/hwpmc/hwpmc_cs.c
159	Don''t you need to zero page if PG_ZERO is not set in flags ?

Add assert to ensure buffer deallocation happens in the same thread as allocation
Zero the pages if required. vm_page_alloc_contig does not guarantee that page will be zeroed with VM_ALLOC_ZERO flag set
Iterate over pages properly using vm_page_next

br marked an inline comment as done.Mar 5 2018, 6:47 PM

br added inline comments.

sys/dev/hwpmc/hwpmc_cs.c
149	Wanted to ensure that object will be destroyed after we stop tracing operation.
155	Thanks I moved taking a reference few lines later. Not sure however yet if we need to take additional reference. It works fine with and without taking additional reference.
158	Thanks I switched to vm_page_next() for iteration.
159	Yes, thanks.
191	Added.

kib added inline comments.Mar 5 2018, 7:22 PM

sys/dev/hwpmc/hwpmc_cs.c
155	You might just leak the object with the additional reference.

Don't take additional reference for VM object.
Suppress ETF buffer fill readings. ETF is a small FIFO hardware buffer we are not accessing in software

br added inline comments.Mar 6 2018, 2:43 PM

sys/dev/hwpmc/hwpmc_cs.c
155	I removed taking additional reference. In works fine. Thanks!

andrew added inline comments.Mar 6 2018, 4:24 PM

sys/arm64/coresight/coresight-dynamic-replicator.c
100–103	Why are you using function pointers and not a new device interface?
sys/arm64/coresight/coresight-etm4x.c
81	Why do we need the `isb` here & below?
140	Is this comment > 80 characters? It's difficult to tell in the phabricator interface.
180	What are these magic values?
182	This should be at the top of the function
184	Should this be under bootverbose?
207	What's `1 << 0`?

Add coresight device interface. Use it instead of function pointers
Wrap long lines
Unmagic values
Print debugging information if bootverbose set
Add ETM_DEBUG macro

Include coresight_if.m to patch

br added inline comments.Mar 7 2018, 3:14 PM

sys/arm64/coresight/coresight-dynamic-replicator.c
100–103	I switched to new device interface. Thanks
sys/arm64/coresight/coresight-etm4x.c
81	I think we don't need this. I took it from linux during early debugging. They use it to ensure everyone sees the write or to ensure order
140	Fixed
180	Fixed
182	Thanks. Fixed
184	I added ETM_DEBUG macro for this. Thanks
207	Fixed. Thanks

Support for Coresight devices splitted-out to a separate review https://reviews.freebsd.org/D14618

Regenerate patch

Initialize ev to NULL before processing command line. This fixes operation using Intel PT

Regenerate patch

Regenerate patch (opencsd committed to HEAD)

Regenerate patch

split out changes

add context

br added a child revision: D15088: HWPMC tracing support (3) -- libpmc.Apr 16 2018, 1:11 PM

br retitled this revision from HWPMC tracing support to HWPMC tracing support (1) -- main.Apr 16 2018, 1:28 PM

br added child revisions: D15089: HWPMC tracing support (2) -- VM, D15090: HWPMC tracing support (4) -- ARM Coresight, D15091: HWPMC tracing support (5) -- Intel PT, D15092: HWPMC tracing support (6) -- pmctrace main, D15094: HWPMC tracing support (7) -- pmctrace Intel PT, D15095: HWPMC tracing support (8) -- pmctrace ARM Coresight.Apr 16 2018, 1:36 PM

check return value for pmc_vm_initialize(md);

Regenerate patch

br added a reviewer: mmacy.May 29 2018, 10:46 AM

br edited the summary of this revision. (Show Details)May 29 2018, 10:49 AM

merge changes from upstream

regenerate

br edited the summary of this revision. (Show Details)Jun 1 2018, 5:05 PM

merge changes from upstream

merge from upstream

br edited the summary of this revision. (Show Details)Aug 2 2018, 4:38 PM

regenerate

fix style

Add a comment on why we pause the thread

regenerate patch

markj added inline comments.Dec 27 2018, 5:12 PM

sys/dev/hwpmc/hwpmc_mod.c
1568	Use bool literals, "true"/"false".
1588	Isn't this somewhat fragile? If pmctrace crashes before unsuspending the thread, won't the thread stay stuck? ptrace(), for instance, will send SIGKILL to the target process if the tracing process dies.
5243	This should come earlier. This case is only meant to be executed if initialization completes successfully, but the pmc_vm_initialize() call introduces a new error case.

Use bool literals
Move pmc_vm_initialize() just after pmc_md_initialize()

br added inline comments.Jan 14 2019, 11:00 AM

sys/dev/hwpmc/hwpmc_mod.c
1588	Yes it is fragile. Any crash of pmctrace leaves the tracing pid on a pause. I am not sure how to avoid this. At least a warning to user must be added

markj added inline comments.Jan 16 2019, 7:22 PM

sys/dev/hwpmc/hwpmc_mod.c
1588	Why is it necessary to suspend all threads in the process? At this point we have not yet returned from mmap(), so other application threads should not be accessing the newly mapped region.

Use thread_suspend_one() instead of thread_suspend_switch()

br marked 2 inline comments as done.Jan 23 2019, 2:04 PM

br added inline comments.

sys/dev/hwpmc/hwpmc_mod.c
1588	It is not necessary, so I changed to thread_suspend_one()

use sleep(9) instead of thread_suspend_one

Regenerate patch.

br edited the summary of this revision. (Show Details)Oct 7 2019, 4:22 PM

br edited the test plan for this revision. (Show Details)

br edited the summary of this revision. (Show Details)Oct 8 2019, 3:45 PM

HWPMC tracing support (1) -- mainNeeds ReviewPublicActions

Details

Diff Detail

Event Timeline

Revision ContentsChangeset List

Diff 39967

lib/Makefile

lib/libipt/Makefile

lib/libopencsd/Makefile

lib/libpmc/libpmc.c

lib/libpmc/pmc.h

lib/libpmcstat/libpmcstat_image.c

sys/amd64/include/pmc_mdep.h

sys/arm64/coresight/coresight-cmd.c

sys/arm64/coresight/coresight-cpu-debug.c

sys/arm64/coresight/coresight-dynamic-replicator.c

sys/arm64/coresight/coresight-etm4x.h

sys/arm64/coresight/coresight-etm4x.c

sys/arm64/coresight/coresight-funnel.h

sys/arm64/coresight/coresight-funnel.c

sys/arm64/coresight/coresight-tmc.h

sys/arm64/coresight/coresight-tmc.c

sys/arm64/coresight/coresight.h

sys/arm64/coresight/coresight.c

sys/conf/files

sys/dev/hwpmc/hwpmc_arm64.c

sys/dev/hwpmc/hwpmc_core.c

sys/dev/hwpmc/hwpmc_cs.h

sys/dev/hwpmc/hwpmc_cs.c

sys/dev/hwpmc/hwpmc_intel.c

sys/dev/hwpmc/hwpmc_mod.c

sys/dev/hwpmc/hwpmc_pt.h

sys/dev/hwpmc/hwpmc_pt.c

sys/dev/hwpmc/hwpmc_vm.h

sys/dev/hwpmc/hwpmc_vm.c

sys/dev/hwpmc/pmc_events.h

sys/kern/vfs_vnops.c

sys/modules/hwpmc/Makefile

sys/sys/pmc.h

sys/x86/include/specialreg.h

usr.sbin/Makefile

usr.sbin/pmctrace/Makefile

usr.sbin/pmctrace/pmctrace.h

usr.sbin/pmctrace/pmctrace.c

usr.sbin/pmctrace/pmctrace_cs.h

usr.sbin/pmctrace/pmctrace_cs.c

usr.sbin/pmctrace/pmctrace_pt.h

usr.sbin/pmctrace/pmctrace_pt.c

HWPMC tracing support (1) -- main
Needs ReviewPublic
Actions

Revision Contents
Changeset List