Message ID | 20230413161725.195417-1-alexghiti@rivosinc.com (mailing list archive) |
---|---|
Headers | show |
Series | riscv: Allow userspace to directly access perf counters | expand |
On Thu, Apr 13, 2023 at 9:17 AM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > riscv used to allow direct access to cycle/time/instret counters, > bypassing the perf framework, this patchset intends to allow the user to > mmap any counter when accessed through perf. But we can't break the > existing behaviour so we introduce a sysctl perf_user_access like arm64 > does, which defaults to the legacy mode described above. > > The core of this patchset lies in patch 4, the first 3 patches are > simple fixes. > > base-commit-tag: v6.3-rc1 > > Alexandre Ghiti (4): > perf: Fix wrong comment about default event_idx > include: riscv: Fix wrong include guard in riscv_pmu.h > riscv: Make legacy counter enum match the HW numbering > riscv: Enable perf counters user access only through perf Presumably the test also needs patching: https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/tests/mmap-basic.c?h=perf-tools-next#n287 Thanks, Ian > Documentation/admin-guide/sysctl/kernel.rst | 23 +++- > arch/riscv/include/asm/perf_event.h | 3 + > arch/riscv/kernel/Makefile | 2 +- > arch/riscv/kernel/perf_event.c | 65 +++++++++++ > drivers/perf/riscv_pmu.c | 42 ++++++++ > drivers/perf/riscv_pmu_legacy.c | 24 ++++- > drivers/perf/riscv_pmu_sbi.c | 113 ++++++++++++++++++-- > include/linux/perf/riscv_pmu.h | 9 +- > include/linux/perf_event.h | 3 +- > tools/lib/perf/mmap.c | 65 +++++++++++ > 10 files changed, 332 insertions(+), 17 deletions(-) > create mode 100644 arch/riscv/kernel/perf_event.c > > -- > 2.37.2 >
On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > riscv used to allow direct access to cycle/time/instret counters, > bypassing the perf framework, this patchset intends to allow the user to > mmap any counter when accessed through perf. But we can't break the > existing behaviour so we introduce a sysctl perf_user_access like arm64 > does, which defaults to the legacy mode described above. > It would be good provide additional direction for user space packages: The legacy behavior is supported for now in order to avoid breaking existing software. However, reading counters directly without perf interaction may provide incorrect values which the userspace software must avoid. We are hoping that the user space packages which read the cycle/instret directly, will move to the proper interface eventually if they actually need it. Most of the users are supposed to read "time" instead of "cycle" if they intend to read timestamps. The legacy sysctl option will be removed in the future. The plan is that the distros will set the default option to SYSCTL_USER_ACCESS which enables user counters only through perf sooner (as soon as they make sure the packages built for that distro don't read cycle/instret) directly. > The core of this patchset lies in patch 4, the first 3 patches are > simple fixes. > > base-commit-tag: v6.3-rc1 > > Alexandre Ghiti (4): > perf: Fix wrong comment about default event_idx > include: riscv: Fix wrong include guard in riscv_pmu.h > riscv: Make legacy counter enum match the HW numbering > riscv: Enable perf counters user access only through perf > > Documentation/admin-guide/sysctl/kernel.rst | 23 +++- > arch/riscv/include/asm/perf_event.h | 3 + > arch/riscv/kernel/Makefile | 2 +- > arch/riscv/kernel/perf_event.c | 65 +++++++++++ > drivers/perf/riscv_pmu.c | 42 ++++++++ > drivers/perf/riscv_pmu_legacy.c | 24 ++++- > drivers/perf/riscv_pmu_sbi.c | 113 ++++++++++++++++++-- > include/linux/perf/riscv_pmu.h | 9 +- > include/linux/perf_event.h | 3 +- > tools/lib/perf/mmap.c | 65 +++++++++++ > 10 files changed, 332 insertions(+), 17 deletions(-) > create mode 100644 arch/riscv/kernel/perf_event.c > > -- > 2.37.2 >
From: Atish Patra > Sent: 13 April 2023 20:18 > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > riscv used to allow direct access to cycle/time/instret counters, > > bypassing the perf framework, this patchset intends to allow the user to > > mmap any counter when accessed through perf. But we can't break the > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > does, which defaults to the legacy mode described above. > > > > It would be good provide additional direction for user space packages: > > The legacy behavior is supported for now in order to avoid breaking > existing software. > However, reading counters directly without perf interaction may > provide incorrect values which > the userspace software must avoid. We are hoping that the user space > packages which > read the cycle/instret directly, will move to the proper interface > eventually if they actually need it. > Most of the users are supposed to read "time" instead of "cycle" if > they intend to read timestamps. If you are trying to measure the performance of short code fragments then you need pretty much raw access directly to the cycle/clock count register. I've done this on x86 to compare the actual cycle times of different implementations of the IP checksum loop (and compare them to the theoretical limit). The perf framework just added far too much latency, only directly reading the cpu registers gave anything like reliable (and consistent) answers. Clearly process switches (especially cpu migrations) cause problems, but they are obviously invalid values and can be ignored. So while a lot of uses may be 'happy' with the values the perf framework gives, sometimes you do need to directly read the relevant registers. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > From: Atish Patra > > Sent: 13 April 2023 20:18 > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > bypassing the perf framework, this patchset intends to allow the user to > > > mmap any counter when accessed through perf. But we can't break the > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > does, which defaults to the legacy mode described above. > > > > > > > It would be good provide additional direction for user space packages: > > > > The legacy behavior is supported for now in order to avoid breaking > > existing software. > > However, reading counters directly without perf interaction may > > provide incorrect values which > > the userspace software must avoid. We are hoping that the user space > > packages which > > read the cycle/instret directly, will move to the proper interface > > eventually if they actually need it. > > Most of the users are supposed to read "time" instead of "cycle" if > > they intend to read timestamps. > > If you are trying to measure the performance of short code > fragments then you need pretty much raw access directly to > the cycle/clock count register. > > I've done this on x86 to compare the actual cycle times > of different implementations of the IP checksum loop > (and compare them to the theoretical limit). > The perf framework just added far too much latency, > only directly reading the cpu registers gave anything > like reliable (and consistent) answers. > This series allows direct access to the counters once configured through the perf. Earlier the cycle/instret counters are directly exposed to the userspace without kernel/perf frameworking knowing when/which user space application is reading it. That has security implications. With this series applied, the user space application just needs to configure the event (cycle/instret) through perf syscall. Once configured, the userspace application can find out the counter information from the mmap & directly read the counter. There is no latency while reading the counters. This mechanism allows stop/clear the counters when the requesting task is not running. It also takes care of context switching which may result in invalid values as you mentioned below. This is nothing new and all other arch (x86, ARM64) allow user space counter read through the same mechanism. Here is the relevant upstream discussion: https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ ARM64: https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access example usage in x86: https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > Clearly process switches (especially cpu migrations) cause > problems, but they are obviously invalid values and can > be ignored. > > So while a lot of uses may be 'happy' with the values the > perf framework gives, sometimes you do need to directly > read the relevant registers. > > David > > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales)
On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > From: Atish Patra > > > Sent: 13 April 2023 20:18 > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > mmap any counter when accessed through perf. But we can't break the > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > existing software. > > > However, reading counters directly without perf interaction may > > > provide incorrect values which > > > the userspace software must avoid. We are hoping that the user space > > > packages which > > > read the cycle/instret directly, will move to the proper interface > > > eventually if they actually need it. > > > Most of the users are supposed to read "time" instead of "cycle" if > > > they intend to read timestamps. > > > > If you are trying to measure the performance of short code > > fragments then you need pretty much raw access directly to > > the cycle/clock count register. > > > > I've done this on x86 to compare the actual cycle times > > of different implementations of the IP checksum loop > > (and compare them to the theoretical limit). > > The perf framework just added far too much latency, > > only directly reading the cpu registers gave anything > > like reliable (and consistent) answers. > > > > This series allows direct access to the counters once configured > through the perf. > Earlier the cycle/instret counters are directly exposed to the > userspace without kernel/perf frameworking knowing > when/which user space application is reading it. That has security implications. > > With this series applied, the user space application just needs to > configure the event (cycle/instret) through perf syscall. > Once configured, the userspace application can find out the counter > information from the mmap & directly > read the counter. There is no latency while reading the counters. > > This mechanism allows stop/clear the counters when the requesting task > is not running. It also takes care of context switching > which may result in invalid values as you mentioned below. This is > nothing new and all other arch (x86, ARM64) allow user space > counter read through the same mechanism. > > Here is the relevant upstream discussion: > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > ARM64: > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > example usage in x86: > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c The canonical implementation of this should be: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 which is updated in these patches but the tests are not: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 Which appears to be an oversight. The tests display some differences between x86 and aarch64 that have assumed userspace hardware counter access, and everything else that it is assumed don't. Thanks, Ian > > Clearly process switches (especially cpu migrations) cause > > problems, but they are obviously invalid values and can > > be ignored. > > > > So while a lot of uses may be 'happy' with the values the > > perf framework gives, sometimes you do need to directly > > read the relevant registers. > > > > David > > > > - > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > Registration No: 1397386 (Wales) > > > > -- > Regards, > Atish
On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@google.com> wrote: > > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > > > From: Atish Patra > > > > Sent: 13 April 2023 20:18 > > > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > > mmap any counter when accessed through perf. But we can't break the > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > > existing software. > > > > However, reading counters directly without perf interaction may > > > > provide incorrect values which > > > > the userspace software must avoid. We are hoping that the user space > > > > packages which > > > > read the cycle/instret directly, will move to the proper interface > > > > eventually if they actually need it. > > > > Most of the users are supposed to read "time" instead of "cycle" if > > > > they intend to read timestamps. > > > > > > If you are trying to measure the performance of short code > > > fragments then you need pretty much raw access directly to > > > the cycle/clock count register. > > > > > > I've done this on x86 to compare the actual cycle times > > > of different implementations of the IP checksum loop > > > (and compare them to the theoretical limit). > > > The perf framework just added far too much latency, > > > only directly reading the cpu registers gave anything > > > like reliable (and consistent) answers. > > > > > > > This series allows direct access to the counters once configured > > through the perf. > > Earlier the cycle/instret counters are directly exposed to the > > userspace without kernel/perf frameworking knowing > > when/which user space application is reading it. That has security implications. > > > > With this series applied, the user space application just needs to > > configure the event (cycle/instret) through perf syscall. > > Once configured, the userspace application can find out the counter > > information from the mmap & directly > > read the counter. There is no latency while reading the counters. > > > > This mechanism allows stop/clear the counters when the requesting task > > is not running. It also takes care of context switching > > which may result in invalid values as you mentioned below. This is > > nothing new and all other arch (x86, ARM64) allow user space > > counter read through the same mechanism. > > > > Here is the relevant upstream discussion: > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > > > ARM64: > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > > > example usage in x86: > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > > The canonical implementation of this should be: > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 Thanks for sharing the libperf implementation. > which is updated in these patches but the tests are not: > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 > Which appears to be an oversight. The tests display some differences Yes. It's an oversight. We should make sure that perf mmap tests pass for RISC-V as well. > between x86 and aarch64 that have assumed userspace hardware counter > access, and everything else that it is assumed don't. > > Thanks, > Ian > > > > Clearly process switches (especially cpu migrations) cause > > > problems, but they are obviously invalid values and can > > > be ignored. > > > > > > So while a lot of uses may be 'happy' with the values the > > > perf framework gives, sometimes you do need to directly > > > read the relevant registers. > > > > > > David > > > > > > - > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > > Registration No: 1397386 (Wales) > > > > > > > > -- > > Regards, > > Atish -- Regards, Atish
Hi Ian, On Tue, Apr 18, 2023 at 10:30 PM Atish Patra <atishp@atishpatra.org> wrote: > > On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@google.com> wrote: > > > > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > > > > > From: Atish Patra > > > > > Sent: 13 April 2023 20:18 > > > > > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > > > mmap any counter when accessed through perf. But we can't break the > > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > > > existing software. > > > > > However, reading counters directly without perf interaction may > > > > > provide incorrect values which > > > > > the userspace software must avoid. We are hoping that the user space > > > > > packages which > > > > > read the cycle/instret directly, will move to the proper interface > > > > > eventually if they actually need it. > > > > > Most of the users are supposed to read "time" instead of "cycle" if > > > > > they intend to read timestamps. > > > > > > > > If you are trying to measure the performance of short code > > > > fragments then you need pretty much raw access directly to > > > > the cycle/clock count register. > > > > > > > > I've done this on x86 to compare the actual cycle times > > > > of different implementations of the IP checksum loop > > > > (and compare them to the theoretical limit). > > > > The perf framework just added far too much latency, > > > > only directly reading the cpu registers gave anything > > > > like reliable (and consistent) answers. > > > > > > > > > > This series allows direct access to the counters once configured > > > through the perf. > > > Earlier the cycle/instret counters are directly exposed to the > > > userspace without kernel/perf frameworking knowing > > > when/which user space application is reading it. That has security implications. > > > > > > With this series applied, the user space application just needs to > > > configure the event (cycle/instret) through perf syscall. > > > Once configured, the userspace application can find out the counter > > > information from the mmap & directly > > > read the counter. There is no latency while reading the counters. > > > > > > This mechanism allows stop/clear the counters when the requesting task > > > is not running. It also takes care of context switching > > > which may result in invalid values as you mentioned below. This is > > > nothing new and all other arch (x86, ARM64) allow user space > > > counter read through the same mechanism. > > > > > > Here is the relevant upstream discussion: > > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > > > > > ARM64: > > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > > > > > example usage in x86: > > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > > > > The canonical implementation of this should be: > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 > > Thanks for sharing the libperf implementation. > > > which is updated in these patches but the tests are not: > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 > > Which appears to be an oversight. The tests display some differences > > Yes. It's an oversight. We should make sure that perf mmap tests pass > for RISC-V as well. Yes, that's an oversight, I had a local test adapted from this one but forgot to update it afterwards, I'll do that in the next version. Thanks for your quick feedbacks and sorry for being late, Alex > > > > between x86 and aarch64 that have assumed userspace hardware counter > > access, and everything else that it is assumed don't. > > > > Thanks, > > Ian > > > > > > Clearly process switches (especially cpu migrations) cause > > > > problems, but they are obviously invalid values and can > > > > be ignored. > > > > > > > > So while a lot of uses may be 'happy' with the values the > > > > perf framework gives, sometimes you do need to directly > > > > read the relevant registers. > > > > > > > > David > > > > > > > > - > > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > > > Registration No: 1397386 (Wales) > > > > > > > > > > > > -- > > > Regards, > > > Atish > > > > -- > Regards, > Atish
On Wed, Apr 19, 2023 at 2:21 AM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > Hi Ian, > > On Tue, Apr 18, 2023 at 10:30 PM Atish Patra <atishp@atishpatra.org> wrote: > > > > On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@google.com> wrote: > > > > > > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > > > > > > > From: Atish Patra > > > > > > Sent: 13 April 2023 20:18 > > > > > > > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > > > > mmap any counter when accessed through perf. But we can't break the > > > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > > > > existing software. > > > > > > However, reading counters directly without perf interaction may > > > > > > provide incorrect values which > > > > > > the userspace software must avoid. We are hoping that the user space > > > > > > packages which > > > > > > read the cycle/instret directly, will move to the proper interface > > > > > > eventually if they actually need it. > > > > > > Most of the users are supposed to read "time" instead of "cycle" if > > > > > > they intend to read timestamps. > > > > > > > > > > If you are trying to measure the performance of short code > > > > > fragments then you need pretty much raw access directly to > > > > > the cycle/clock count register. > > > > > > > > > > I've done this on x86 to compare the actual cycle times > > > > > of different implementations of the IP checksum loop > > > > > (and compare them to the theoretical limit). > > > > > The perf framework just added far too much latency, > > > > > only directly reading the cpu registers gave anything > > > > > like reliable (and consistent) answers. > > > > > > > > > > > > > This series allows direct access to the counters once configured > > > > through the perf. > > > > Earlier the cycle/instret counters are directly exposed to the > > > > userspace without kernel/perf frameworking knowing > > > > when/which user space application is reading it. That has security implications. > > > > > > > > With this series applied, the user space application just needs to > > > > configure the event (cycle/instret) through perf syscall. > > > > Once configured, the userspace application can find out the counter > > > > information from the mmap & directly > > > > read the counter. There is no latency while reading the counters. > > > > > > > > This mechanism allows stop/clear the counters when the requesting task > > > > is not running. It also takes care of context switching > > > > which may result in invalid values as you mentioned below. This is > > > > nothing new and all other arch (x86, ARM64) allow user space > > > > counter read through the same mechanism. > > > > > > > > Here is the relevant upstream discussion: > > > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > > > > > > > ARM64: > > > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > > > > > > > example usage in x86: > > > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > > > > > > The canonical implementation of this should be: > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 > > > > Thanks for sharing the libperf implementation. > > > > > which is updated in these patches but the tests are not: > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 > > > Which appears to be an oversight. The tests display some differences > > > > Yes. It's an oversight. We should make sure that perf mmap tests pass > > for RISC-V as well. > > Yes, that's an oversight, I had a local test adapted from this one but > forgot to update it afterwards, I'll do that in the next version. > > Thanks for your quick feedbacks and sorry for being late, > > Alex Thanks Alex, there was an equally likely chance that I wasn't understanding things :-) Is there any information on RISC-V PMU testing? I know ParanLee is interested. It'd be awesome to have something say on: https://perf.wiki.kernel.org/index.php/Main_Page on how to run tests, perhaps on QEMU or known to work boards. We can also just drop a link on there if there is information. We can also add the RISC-V PMU information to the links here: https://perf.wiki.kernel.org/index.php/Useful_Links Thanks, Ian > > > > > > > > between x86 and aarch64 that have assumed userspace hardware counter > > > access, and everything else that it is assumed don't. > > > > > > Thanks, > > > Ian > > > > > > > > Clearly process switches (especially cpu migrations) cause > > > > > problems, but they are obviously invalid values and can > > > > > be ignored. > > > > > > > > > > So while a lot of uses may be 'happy' with the values the > > > > > perf framework gives, sometimes you do need to directly > > > > > read the relevant registers. > > > > > > > > > > David > > > > > > > > > > - > > > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > > > > Registration No: 1397386 (Wales) > > > > > > > > > > > > > > > > -- > > > > Regards, > > > > Atish > > > > > > > > -- > > Regards, > > Atish
On Wed, Apr 19, 2023 at 11:13 PM Ian Rogers <irogers@google.com> wrote: > > On Wed, Apr 19, 2023 at 2:21 AM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > Hi Ian, > > > > On Tue, Apr 18, 2023 at 10:30 PM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@google.com> wrote: > > > > > > > > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > > > > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > > > > > > > > > From: Atish Patra > > > > > > > Sent: 13 April 2023 20:18 > > > > > > > > > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > > > > > mmap any counter when accessed through perf. But we can't break the > > > > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > > > > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > > > > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > > > > > existing software. > > > > > > > However, reading counters directly without perf interaction may > > > > > > > provide incorrect values which > > > > > > > the userspace software must avoid. We are hoping that the user space > > > > > > > packages which > > > > > > > read the cycle/instret directly, will move to the proper interface > > > > > > > eventually if they actually need it. > > > > > > > Most of the users are supposed to read "time" instead of "cycle" if > > > > > > > they intend to read timestamps. > > > > > > > > > > > > If you are trying to measure the performance of short code > > > > > > fragments then you need pretty much raw access directly to > > > > > > the cycle/clock count register. > > > > > > > > > > > > I've done this on x86 to compare the actual cycle times > > > > > > of different implementations of the IP checksum loop > > > > > > (and compare them to the theoretical limit). > > > > > > The perf framework just added far too much latency, > > > > > > only directly reading the cpu registers gave anything > > > > > > like reliable (and consistent) answers. > > > > > > > > > > > > > > > > This series allows direct access to the counters once configured > > > > > through the perf. > > > > > Earlier the cycle/instret counters are directly exposed to the > > > > > userspace without kernel/perf frameworking knowing > > > > > when/which user space application is reading it. That has security implications. > > > > > > > > > > With this series applied, the user space application just needs to > > > > > configure the event (cycle/instret) through perf syscall. > > > > > Once configured, the userspace application can find out the counter > > > > > information from the mmap & directly > > > > > read the counter. There is no latency while reading the counters. > > > > > > > > > > This mechanism allows stop/clear the counters when the requesting task > > > > > is not running. It also takes care of context switching > > > > > which may result in invalid values as you mentioned below. This is > > > > > nothing new and all other arch (x86, ARM64) allow user space > > > > > counter read through the same mechanism. > > > > > > > > > > Here is the relevant upstream discussion: > > > > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > > > > > > > > > ARM64: > > > > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > > > > > > > > > example usage in x86: > > > > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > > > > > > > > The canonical implementation of this should be: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 > > > > > > Thanks for sharing the libperf implementation. > > > > > > > which is updated in these patches but the tests are not: > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 > > > > Which appears to be an oversight. The tests display some differences > > > > > > Yes. It's an oversight. We should make sure that perf mmap tests pass > > > for RISC-V as well. > > > > Yes, that's an oversight, I had a local test adapted from this one but > > forgot to update it afterwards, I'll do that in the next version. > > > > Thanks for your quick feedbacks and sorry for being late, > > > > Alex > > Thanks Alex, there was an equally likely chance that I wasn't > understanding things :-) Is there any information on RISC-V PMU > testing? I know ParanLee is interested. It'd be awesome to have Are you looking for something specific to RISC-V general or perf on RISC-V? All the RISC-V PMU patches have been upstream for a while (both in the Qemu & Linux kernel). Perf should work out-of-the box when you boot the latest kernel in the latest version of the Qemu. Initial KVM[1] patches support got merged during the last merge window. It doesn't support event sampling yet. We are working on that. [1] https://lore.kernel.org/lkml/20230207095529.1787260-1-atishp@rivosinc.com/ > something say on: > https://perf.wiki.kernel.org/index.php/Main_Page > on how to run tests, perhaps on QEMU or known to work boards. We can > also just drop a link on there if there is information. We can also > add the RISC-V PMU information to the links here: > https://perf.wiki.kernel.org/index.php/Useful_Links > I did not see any arch specific information there. Let us know what would be good to add there and we would be happy to add. > Thanks, > Ian > > > > > > > > > > > > > between x86 and aarch64 that have assumed userspace hardware counter > > > > access, and everything else that it is assumed don't. > > > > > > > > Thanks, > > > > Ian > > > > > > > > > > Clearly process switches (especially cpu migrations) cause > > > > > > problems, but they are obviously invalid values and can > > > > > > be ignored. > > > > > > > > > > > > So while a lot of uses may be 'happy' with the values the > > > > > > perf framework gives, sometimes you do need to directly > > > > > > read the relevant registers. > > > > > > > > > > > > David > > > > > > > > > > > > - > > > > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > > > > > Registration No: 1397386 (Wales) > > > > > > > > > > > > > > > > > > > > -- > > > > > Regards, > > > > > Atish > > > > > > > > > > > > -- > > > Regards, > > > Atish
On Wed, Apr 19, 2023 at 4:22 PM Atish Patra <atishp@atishpatra.org> wrote: > > On Wed, Apr 19, 2023 at 11:13 PM Ian Rogers <irogers@google.com> wrote: > > > > On Wed, Apr 19, 2023 at 2:21 AM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > Hi Ian, > > > > > > On Tue, Apr 18, 2023 at 10:30 PM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > > > On Tue, Apr 18, 2023 at 11:46 PM Ian Rogers <irogers@google.com> wrote: > > > > > > > > > > On Tue, Apr 18, 2023 at 9:43 AM Atish Patra <atishp@atishpatra.org> wrote: > > > > > > > > > > > > On Fri, Apr 14, 2023 at 2:40 AM David Laight <David.Laight@aculab.com> wrote: > > > > > > > > > > > > > > From: Atish Patra > > > > > > > > Sent: 13 April 2023 20:18 > > > > > > > > > > > > > > > > On Thu, Apr 13, 2023 at 9:47 PM Alexandre Ghiti <alexghiti@rivosinc.com> wrote: > > > > > > > > > > > > > > > > > > riscv used to allow direct access to cycle/time/instret counters, > > > > > > > > > bypassing the perf framework, this patchset intends to allow the user to > > > > > > > > > mmap any counter when accessed through perf. But we can't break the > > > > > > > > > existing behaviour so we introduce a sysctl perf_user_access like arm64 > > > > > > > > > does, which defaults to the legacy mode described above. > > > > > > > > > > > > > > > > > > > > > > > > > It would be good provide additional direction for user space packages: > > > > > > > > > > > > > > > > The legacy behavior is supported for now in order to avoid breaking > > > > > > > > existing software. > > > > > > > > However, reading counters directly without perf interaction may > > > > > > > > provide incorrect values which > > > > > > > > the userspace software must avoid. We are hoping that the user space > > > > > > > > packages which > > > > > > > > read the cycle/instret directly, will move to the proper interface > > > > > > > > eventually if they actually need it. > > > > > > > > Most of the users are supposed to read "time" instead of "cycle" if > > > > > > > > they intend to read timestamps. > > > > > > > > > > > > > > If you are trying to measure the performance of short code > > > > > > > fragments then you need pretty much raw access directly to > > > > > > > the cycle/clock count register. > > > > > > > > > > > > > > I've done this on x86 to compare the actual cycle times > > > > > > > of different implementations of the IP checksum loop > > > > > > > (and compare them to the theoretical limit). > > > > > > > The perf framework just added far too much latency, > > > > > > > only directly reading the cpu registers gave anything > > > > > > > like reliable (and consistent) answers. > > > > > > > > > > > > > > > > > > > This series allows direct access to the counters once configured > > > > > > through the perf. > > > > > > Earlier the cycle/instret counters are directly exposed to the > > > > > > userspace without kernel/perf frameworking knowing > > > > > > when/which user space application is reading it. That has security implications. > > > > > > > > > > > > With this series applied, the user space application just needs to > > > > > > configure the event (cycle/instret) through perf syscall. > > > > > > Once configured, the userspace application can find out the counter > > > > > > information from the mmap & directly > > > > > > read the counter. There is no latency while reading the counters. > > > > > > > > > > > > This mechanism allows stop/clear the counters when the requesting task > > > > > > is not running. It also takes care of context switching > > > > > > which may result in invalid values as you mentioned below. This is > > > > > > nothing new and all other arch (x86, ARM64) allow user space > > > > > > counter read through the same mechanism. > > > > > > > > > > > > Here is the relevant upstream discussion: > > > > > > https://lore.kernel.org/lkml/Y7wLa7I2hlz3rKw%2F@hirez.programming.kicks-ass.net/T/ > > > > > > > > > > > > ARM64: > > > > > > https://docs.kernel.org/arm64/perf.html?highlight=perf_user_access#perf-userspace-pmu-hardware-counter-access > > > > > > > > > > > > example usage in x86: > > > > > > https://github.com/andikleen/pmu-tools/blob/master/jevents/rdpmc.c > > > > > > > > > > The canonical implementation of this should be: > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/perf/mmap.c#n400 > > > > > > > > Thanks for sharing the libperf implementation. > > > > > > > > > which is updated in these patches but the tests are not: > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/tests/mmap-basic.c#n287 > > > > > Which appears to be an oversight. The tests display some differences > > > > > > > > Yes. It's an oversight. We should make sure that perf mmap tests pass > > > > for RISC-V as well. > > > > > > Yes, that's an oversight, I had a local test adapted from this one but > > > forgot to update it afterwards, I'll do that in the next version. > > > > > > Thanks for your quick feedbacks and sorry for being late, > > > > > > Alex > > > > Thanks Alex, there was an equally likely chance that I wasn't > > understanding things :-) Is there any information on RISC-V PMU > > testing? I know ParanLee is interested. It'd be awesome to have > > Are you looking for something specific to RISC-V general or perf on RISC-V? > All the RISC-V PMU patches have been upstream for a while (both in the > Qemu & Linux kernel). > Perf should work out-of-the box when you boot the latest kernel in the > latest version of the Qemu. > > Initial KVM[1] patches support got merged during the last merge > window. It doesn't support > event sampling yet. We are working on that. > > [1] https://lore.kernel.org/lkml/20230207095529.1787260-1-atishp@rivosinc.com/ Cool, it'd be nice to have a recipe for this from x86 Linux :-) > > something say on: > > https://perf.wiki.kernel.org/index.php/Main_Page > > on how to run tests, perhaps on QEMU or known to work boards. We can > > also just drop a link on there if there is information. We can also > > add the RISC-V PMU information to the links here: > > https://perf.wiki.kernel.org/index.php/Useful_Links > > > > I did not see any arch specific information there. Let us know what > would be good to > add there and we would be happy to add. I was specifically thinking under Manuals where the Intel, AMD and ARM manuals are, links to the RISC-V documentation could be added. Thanks, Ian > > Thanks, > > Ian > > > > > > > > > > > > > > > > > > between x86 and aarch64 that have assumed userspace hardware counter > > > > > access, and everything else that it is assumed don't. > > > > > > > > > > Thanks, > > > > > Ian > > > > > > > > > > > > Clearly process switches (especially cpu migrations) cause > > > > > > > problems, but they are obviously invalid values and can > > > > > > > be ignored. > > > > > > > > > > > > > > So while a lot of uses may be 'happy' with the values the > > > > > > > perf framework gives, sometimes you do need to directly > > > > > > > read the relevant registers. > > > > > > > > > > > > > > David > > > > > > > > > > > > > > - > > > > > > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > > > > > > Registration No: 1397386 (Wales) > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > Regards, > > > > > > Atish > > > > > > > > > > > > > > > > -- > > > > Regards, > > > > Atish > > > > -- > Regards, > Atish