| Message ID | 20210916154635.1525-4-german.gomez@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/5] perf cs-etm: Print size using consistent format | expand |
Hi German, On Thu, Sep 16, 2021 at 04:46:34PM +0100, German Gomez wrote: > The head pointer of the AUX buffer managed by the arm_spe_pmu.c driver > is not monotonically increasing, therefore the find_snapshot callback is > needed in order to find the trace data within the AUX buffer and avoid > wasting space in the perf.data file. > > The pointer is assumed to have wrapped if the buffer contains non-zero > data at the end. If it has wrapped, the entire contents of the AUX > buffer are stored in the perf.data file. Otherwise only the data up to > the head pointer is stored. > > Reviewed-by: James Clark <james.clark@arm.com> > Signed-off-by: German Gomez <german.gomez@arm.com> > --- > tools/perf/arch/arm64/util/arm-spe.c | 145 +++++++++++++++++++++++++++ > 1 file changed, 145 insertions(+) > > diff --git a/tools/perf/arch/arm64/util/arm-spe.c b/tools/perf/arch/arm64/util/arm-spe.c > index f8b03d164b42..56785034fc84 100644 > --- a/tools/perf/arch/arm64/util/arm-spe.c > +++ b/tools/perf/arch/arm64/util/arm-spe.c > @@ -23,6 +23,7 @@ > #include "../../../util/auxtrace.h" > #include "../../../util/record.h" > #include "../../../util/arm-spe.h" > +#include <tools/libc_compat.h> // reallocarray > > #define KiB(x) ((x) * 1024) > #define MiB(x) ((x) * 1024 * 1024) > @@ -31,6 +32,8 @@ struct arm_spe_recording { > struct auxtrace_record itr; > struct perf_pmu *arm_spe_pmu; > struct evlist *evlist; > + int wrapped_cnt; > + bool *wrapped; > }; > > static void arm_spe_set_timestamp(struct auxtrace_record *itr, > @@ -299,6 +302,146 @@ static int arm_spe_snapshot_finish(struct auxtrace_record *itr) > return -EINVAL; > } > > +static int arm_spe_alloc_wrapped_array(struct arm_spe_recording *ptr, int idx) > +{ > + bool *wrapped; > + int cnt = ptr->wrapped_cnt, new_cnt, i; > + > + /* > + * No need to allocate, so return early. > + */ > + if (idx < cnt) > + return 0; > + > + /* > + * Make ptr->wrapped as big as idx. > + */ > + new_cnt = idx + 1; > + > + /* > + * Free'ed in arm_spe_recording_free(). > + */ > + wrapped = reallocarray(ptr->wrapped, new_cnt, sizeof(bool)); > + if (!wrapped) > + return -ENOMEM; > + > + /* > + * init new allocated values. > + */ > + for (i = cnt; i < new_cnt; i++) > + wrapped[i] = false; > + > + ptr->wrapped_cnt = new_cnt; > + ptr->wrapped = wrapped; > + > + return 0; > +} > + > +static bool arm_spe_buffer_has_wrapped(unsigned char *buffer, > + size_t buffer_size, u64 head) > +{ > + u64 i, watermark; > + u64 *buf = (u64 *)buffer; > + size_t buf_size = buffer_size; > + > + /* > + * Defensively handle the case where head might be continually increasing - if its value is > + * equal or greater than the size of the ring buffer, then we can safely determine it has > + * wrapped around. Otherwise, continue to detect if head might have wrapped. > + */ > + if (head >= buffer_size) > + return true; > + > + /* > + * We want to look the very last 512 byte (chosen arbitrarily) in the ring buffer. > + */ > + watermark = buf_size - 512; > + > + /* > + * The value of head is somewhere within the size of the ring buffer. This can be that there > + * hasn't been enough data to fill the ring buffer yet or the trace time was so long that > + * head has numerically wrapped around. To find we need to check if we have data at the > + * very end of the ring buffer. We can reliably do this because mmap'ed pages are zeroed > + * out and there is a fresh mapping with every new session. > + */ > + > + /* > + * head is less than 512 byte from the end of the ring buffer. > + */ > + if (head > watermark) > + watermark = head; > + > + /* > + * Speed things up by using 64 bit transactions (see "u64 *buf" above) > + */ > + watermark /= sizeof(u64); > + buf_size /= sizeof(u64); > + > + /* > + * If we find trace data at the end of the ring buffer, head has been there and has > + * numerically wrapped around at least once. > + */ > + for (i = watermark; i < buf_size; i++) > + if (buf[i]) > + return true; > + > + return false; > +} > + > +static int arm_spe_find_snapshot(struct auxtrace_record *itr, int idx, > + struct auxtrace_mmap *mm, unsigned char *data, > + u64 *head, u64 *old) > +{ > + int err; > + bool wrapped; > + struct arm_spe_recording *ptr = > + container_of(itr, struct arm_spe_recording, itr); > + > + /* > + * Allocate memory to keep track of wrapping if this is the first > + * time we deal with this *mm. > + */ > + if (idx >= ptr->wrapped_cnt) { > + err = arm_spe_alloc_wrapped_array(ptr, idx); > + if (err) > + return err; > + } > + > + /* > + * Check to see if *head has wrapped around. If it hasn't only the > + * amount of data between *head and *old is snapshot'ed to avoid > + * bloating the perf.data file with zeros. But as soon as *head has > + * wrapped around the entire size of the AUX ring buffer it taken. > + */ > + wrapped = ptr->wrapped[idx]; > + if (!wrapped && arm_spe_buffer_has_wrapped(data, mm->len, *head)) { > + wrapped = true; > + ptr->wrapped[idx] = true; > + } > + > + pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n", > + __func__, idx, (size_t)*old, (size_t)*head, mm->len); > + > + /* > + * No wrap has occurred, we can just use *head and *old. > + */ > + if (!wrapped) > + return 0; > + > + /* > + * *head has wrapped around - adjust *head and *old to pickup the > + * entire content of the AUX buffer. > + */ > + if (*head >= mm->len) { > + *old = *head - mm->len; > + } else { > + *head += mm->len; > + *old = *head - mm->len; > + } > + > + return 0; > +} > + > static u64 arm_spe_reference(struct auxtrace_record *itr __maybe_unused) > { > struct timespec ts; > @@ -313,6 +456,7 @@ static void arm_spe_recording_free(struct auxtrace_record *itr) > struct arm_spe_recording *sper = > container_of(itr, struct arm_spe_recording, itr); > > + free(sper->wrapped); > free(sper); > } > > @@ -336,6 +480,7 @@ struct auxtrace_record *arm_spe_recording_init(int *err, > sper->itr.pmu = arm_spe_pmu; > sper->itr.snapshot_start = arm_spe_snapshot_start; > sper->itr.snapshot_finish = arm_spe_snapshot_finish; > + sper->itr.find_snapshot = arm_spe_find_snapshot; If I understand correctly, this patch copies the code from cs-etm for snapshot handling. About 2 months ago, we removed the Arm cs-etm's specific snapshot callback function and directly use perf's function __auxtrace_mmap__read() to handle 'head' and 'tail' pointers. Please see the commit for details: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=2f01c200d4405c4562e45e8bb4de44a5ce37b217 Before I review more details for snapshot enabling in patches 03 and 04, could you confirm if Arm SPE can use the same way with cs-etm for snapshot handling? From my understanding, this is a better way to handle AUX buffer's 'head' and 'tail'. Thanks, Leo > sper->itr.parse_snapshot_options = arm_spe_parse_snapshot_options; > sper->itr.recording_options = arm_spe_recording_options; > sper->itr.info_priv_size = arm_spe_info_priv_size; > -- > 2.17.1 >
On Thu, Sep 23, 2021 at 09:50:16PM +0800, Leo Yan wrote: [...] > > @@ -336,6 +480,7 @@ struct auxtrace_record *arm_spe_recording_init(int *err, > > sper->itr.pmu = arm_spe_pmu; > > sper->itr.snapshot_start = arm_spe_snapshot_start; > > sper->itr.snapshot_finish = arm_spe_snapshot_finish; > > + sper->itr.find_snapshot = arm_spe_find_snapshot; > > If I understand correctly, this patch copies the code from cs-etm for > snapshot handling. About 2 months ago, we removed the Arm cs-etm's > specific snapshot callback function and directly use perf's function > __auxtrace_mmap__read() to handle 'head' and 'tail' pointers. Please > see the commit for details: > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=2f01c200d4405c4562e45e8bb4de44a5ce37b217 > > Before I review more details for snapshot enabling in patches 03 and > 04, could you confirm if Arm SPE can use the same way with cs-etm for > snapshot handling? From my understanding, this is a better way to > handle AUX buffer's 'head' and 'tail'. In other words, if we can only apply patch 03 and can pass the testing in patch 05, then it would be a very neat implementation. I will try to verify these patches and will get back result. Thanks, Leo
Hi Leo, On 23/09/2021 15:40, Leo Yan wrote: > On Thu, Sep 23, 2021 at 09:50:16PM +0800, Leo Yan wrote: > > [...] > >>> @@ -336,6 +480,7 @@ struct auxtrace_record *arm_spe_recording_init(int *err, >>> sper->itr.pmu = arm_spe_pmu; >>> sper->itr.snapshot_start = arm_spe_snapshot_start; >>> sper->itr.snapshot_finish = arm_spe_snapshot_finish; >>> + sper->itr.find_snapshot = arm_spe_find_snapshot; >> If I understand correctly, this patch copies the code from cs-etm for >> snapshot handling. About 2 months ago, we removed the Arm cs-etm's >> specific snapshot callback function and directly use perf's function >> __auxtrace_mmap__read() to handle 'head' and 'tail' pointers. Please >> see the commit for details: >> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=2f01c200d4405c4562e45e8bb4de44a5ce37b217 >> >> Before I review more details for snapshot enabling in patches 03 and >> 04, could you confirm if Arm SPE can use the same way with cs-etm for >> snapshot handling? From my understanding, this is a better way to >> handle AUX buffer's 'head' and 'tail'. > In other words, if we can only apply patch 03 and can pass the testing > in patch 05, then it would be a very neat implementation. > > I will try to verify these patches and will get back result. > > Thanks, > Leo The patch is indeed based on that commit. The reason behind it is that the values for *head are being wrapped in the driver side (see the macro PERF_IDX2OFF which is used at various points in /drivers/perf/arm_spe_pmu.c). If this callback is not to be added, I believe the driver needs to be updated first so that the head pointer monotonically increases like in cs-etm. Do you think this makes sense for SPE? (note that the patch will skip the wrap-around detection if this is the case, in order to handle both cases in the userspace perf tool). Thanks, German
Hi German, On Thu, Sep 30, 2021 at 01:26:15PM +0100, German Gomez wrote: [...] > The patch is indeed based on that commit. The reason behind it is that the > values for *head are being wrapped in the driver side (see the macro > PERF_IDX2OFF which is used at various points in > /drivers/perf/arm_spe_pmu.c). Yes, I noted that Arm SPE driver doesn't use monotonical increasing for AUX head. > If this callback is not to be added, I believe the driver needs to be > updated > first so that the head pointer monotonically increases like in cs-etm. Do > you think this makes sense for SPE? Please note, there have two cases should be handled for snapshot mode: - Wrap-around case, somehow function __auxtrace_mmap__read() has handled this case, see [1]; - It's possible that there have overrun case for snapshot mode, e.g. the kernel space receives multiple signals and take snapshot to save Arm SPE trace data into AUX buffer for multiple times, but the userspace tool cannot catch up to save AUX data into perf.data file. Thus the AUX head might be wrapped around for multiple times, for this case, I think monotonically increasing AUX head is the right solution to handle overrun issue. So simply say, I think the head pointer monotonically increasing is the right thing to do in Arm SPE driver. > (note that the patch will skip the wrap-around detection if this is the > case, > in order to handle both cases in the userspace perf tool). Almost agree, I read multiple times but have no idea what's the "both cases" in the last sentence. Please let me know if anything is not clear. Thanks, Leo [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/auxtrace.c#n1804
Hi Leo, Many thanks for your comments. On 04/10/2021 13:27, Leo Yan wrote: > Hi German, > > On Thu, Sep 30, 2021 at 01:26:15PM +0100, German Gomez wrote: > > [...] > > So simply say, I think the head pointer monotonically increasing is > the right thing to do in Arm SPE driver. I will talk to James about how we can proceed on this. >> (note that the patch will skip the wrap-around detection if this is the >> case, >> in order to handle both cases in the userspace perf tool). > Almost agree, I read multiple times but have no idea what's the > "both cases" in the last sentence. Apologies for the later part was not clear. What I meant to say was that in the original patch for cs-etm, it seemed to handle both cases where AUX head might be monotonically and non-monotonically increasing, so we applied the same for the Arm SPE patch. > > Please let me know if anything is not clear. > > Thanks, > Leo > > [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/auxtrace.c#n1804
On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: [...] > > So simply say, I think the head pointer monotonically increasing is > > the right thing to do in Arm SPE driver. > > I will talk to James about how we can proceed on this. Thanks! > >> (note that the patch will skip the wrap-around detection if this is the > >> case, > >> in order to handle both cases in the userspace perf tool). > > Almost agree, I read multiple times but have no idea what's the > > "both cases" in the last sentence. > > Apologies for the later part was not clear. What I meant to say was that > in the original patch for cs-etm, it seemed to handle both cases where > AUX head might be monotonically and non-monotonically increasing, so we > applied the same for the Arm SPE patch. No worries, thanks for explanation. Leo
Hi Leo, On 06/10/2021 10:51, Leo Yan wrote: > On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: > > [...] > >>> So simply say, I think the head pointer monotonically increasing is >>> the right thing to do in Arm SPE driver. >> I will talk to James about how we can proceed on this. > Thanks! I took this offline with James and, though it looks possible to patch the SPE driver to have a monotonically increasing head pointer in order to simplify the handling in the perf tool, it could be a breaking change for users of the perf_event_open syscall that currently rely on the way it works now. An alternative way we considered to simplify the patch is to change the logic inside the find_snapshot callback so that it records the entire contents of the aux buffer every time. What do you think? Many thanks, German
On Mon, Oct 11, 2021 at 04:55:37PM +0100, German Gomez wrote: > Hi Leo, > > On 06/10/2021 10:51, Leo Yan wrote: > > On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: > > > > [...] > > > >>> So simply say, I think the head pointer monotonically increasing is > >>> the right thing to do in Arm SPE driver. > >> I will talk to James about how we can proceed on this. > > Thanks! > > I took this offline with James and, though it looks possible to patch > the SPE driver to have a monotonically increasing head pointer in order > to simplify the handling in the perf tool, it could be a breaking change > for users of the perf_event_open syscall that currently rely on the way > it works now. > > An alternative way we considered to simplify the patch is to change the > logic inside the find_snapshot callback so that it records the entire > contents of the aux buffer every time. > > What do you think? What does intel-pt do? Will
On 12/10/2021 09:19, Will Deacon wrote: > On Mon, Oct 11, 2021 at 04:55:37PM +0100, German Gomez wrote: >> Hi Leo, >> >> On 06/10/2021 10:51, Leo Yan wrote: >>> On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: >>> >>> [...] >>> >>>>> So simply say, I think the head pointer monotonically increasing is >>>>> the right thing to do in Arm SPE driver. >>>> I will talk to James about how we can proceed on this. >>> Thanks! >> >> I took this offline with James and, though it looks possible to patch >> the SPE driver to have a monotonically increasing head pointer in order >> to simplify the handling in the perf tool, it could be a breaking change >> for users of the perf_event_open syscall that currently rely on the way >> it works now. >> >> An alternative way we considered to simplify the patch is to change the >> logic inside the find_snapshot callback so that it records the entire >> contents of the aux buffer every time. >> >> What do you think? > > What does intel-pt do? Intel-pt has a wrapped head, which is why it has the intel_pt_find_snapshot() function in perf to try to not save any zeros from the buffer that haven't been written yet. (With a wrapped head pointer it's impossible to tell). Coresight has a monotonically increasing head pointer so it is possible to tell. Recently, Leo removed the Coresight version of find_snapshot() for this reason. It would be nice to do the same for SPE because that function has a heuristic and is also slow, but I imagine that not returning wrapped head pointers could break anything that expects them. James > > Will >
Hi German, On Mon, Oct 11, 2021 at 04:55:37PM +0100, German Gomez wrote: > Hi Leo, > > On 06/10/2021 10:51, Leo Yan wrote: > > On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: > > > > [...] > > > >>> So simply say, I think the head pointer monotonically increasing is > >>> the right thing to do in Arm SPE driver. > >> I will talk to James about how we can proceed on this. > > Thanks! > > I took this offline with James and, though it looks possible to patch > the SPE driver to have a monotonically increasing head pointer in order > to simplify the handling in the perf tool, it could be a breaking change > for users of the perf_event_open syscall that currently rely on the way > it works now. Here I cannot create the connection between AUX head pointer and the breakage of calling perf_event_open(). Could you elaborate what's the reason the monotonical increasing head pointer will lead to the breakage for perf_event_open()? > An alternative way we considered to simplify the patch is to change the > logic inside the find_snapshot callback so that it records the entire > contents of the aux buffer every time. > > What do you think? We cannot do this way. If we send USR2 signal with very small interval, then it's possible the hardware trace data cannot fill the full of AUX buffer. You could use below commands for the testing and should can observe it produces small chunk trace data: perf record -e arm_spe_0// -S -a -- dd if=/dev/zero of=/dev/null & PERFPID=$! sleep 1 kill -USR2 $PERFPID sleep .1 kill -USR2 $PERFPID Thanks, Leo
On Wed, Oct 13, 2021 at 08:39:16AM +0800, Leo Yan wrote: > On Mon, Oct 11, 2021 at 04:55:37PM +0100, German Gomez wrote: > > On 06/10/2021 10:51, Leo Yan wrote: > > > On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: > > > > > > [...] > > > > > >>> So simply say, I think the head pointer monotonically increasing is > > >>> the right thing to do in Arm SPE driver. > > >> I will talk to James about how we can proceed on this. > > > Thanks! > > > > I took this offline with James and, though it looks possible to patch > > the SPE driver to have a monotonically increasing head pointer in order > > to simplify the handling in the perf tool, it could be a breaking change > > for users of the perf_event_open syscall that currently rely on the way > > it works now. > > Here I cannot create the connection between AUX head pointer and the > breakage of calling perf_event_open(). > > Could you elaborate what's the reason the monotonical increasing head > pointer will lead to the breakage for perf_event_open()? It's a user-visible change in behaviour, isn't it? Therefore we risk breaking applications that rely on the current behaviour if we change it unconditionally. Given that the driver has always worked like this and it doesn't sound like it's the end of the world to deal with it in userspace (after all, it's aligned with intel-pt), then I don't think we should change it. Will
Hi Leo, Would you be ok with the current patch the way it is? In case it's of any help, I'm sharing the testing steps that James and I went through when testing this internally, if you want to add to it - Test that only a portion of the buffer is saved until there is a wraparound $ ./perf record -vvv -e arm_spe/period=148576/u -S -- taskset --cpu-list 0 stress --cpu 1 & while true; do sleep 0.2; killall -s USR2 perf; done - Test snapshot mode in CPU mode $ sudo ./perf record -vvv -C 0 -e arm_spe/period=148576/u -S -- taskset --cpu-list 0 stress --cpu 1 & - Test that auxtrace buffers correspond to an aux record - Test snapshot default sizes in sudo and user modes - Test small snapshot size $ ./perf record -vvv -e arm_spe/period=148576/u -S1000 -m16,16 -- taskset --cpu-list 0 stress --cpu 1 & If there are any concerns with the patches, please let me know and I will try to address them. Thanks, German On 13/10/2021 08:51, Will Deacon wrote: > On Wed, Oct 13, 2021 at 08:39:16AM +0800, Leo Yan wrote: >> On Mon, Oct 11, 2021 at 04:55:37PM +0100, German Gomez wrote: >>> On 06/10/2021 10:51, Leo Yan wrote: >>>> On Wed, Oct 06, 2021 at 10:35:20AM +0100, German Gomez wrote: >>>> >>>> [...] >>>> >>>>>> So simply say, I think the head pointer monotonically increasing is >>>>>> the right thing to do in Arm SPE driver. >>>>> I will talk to James about how we can proceed on this. >>>> Thanks! >>> I took this offline with James and, though it looks possible to patch >>> the SPE driver to have a monotonically increasing head pointer in order >>> to simplify the handling in the perf tool, it could be a breaking change >>> for users of the perf_event_open syscall that currently rely on the way >>> it works now. >> Here I cannot create the connection between AUX head pointer and the >> breakage of calling perf_event_open(). >> >> Could you elaborate what's the reason the monotonical increasing head >> pointer will lead to the breakage for perf_event_open()? > It's a user-visible change in behaviour, isn't it? Therefore we risk > breaking applications that rely on the current behaviour if we change it > unconditionally. > > Given that the driver has always worked like this and it doesn't sound like > it's the end of the world to deal with it in userspace (after all, it's > aligned with intel-pt), then I don't think we should change it. > > Will
Hi German, On Fri, Oct 15, 2021 at 01:33:39PM +0100, German Gomez wrote: > Hi Leo, > > Would you be ok with the current patch the way it is? Sorry for my failure to catch up the discussion. As you and Will have mentioned in other emails that it will lead to breakage if we change to monotonical increasing head, I read the code and realized the difficulty to use monotonical increasing head in Arm SPE driver. So let's use the way as this patch set is. > In case it's of > any help, I'm sharing the testing steps that James and I went through > when testing this internally, if you want to add to it > > - Test that only a portion of the buffer is saved until there is a wraparound > > $ ./perf record -vvv -e arm_spe/period=148576/u -S -- taskset --cpu-list 0 stress --cpu 1 & while true; do sleep 0.2; killall -s USR2 perf; done > > - Test snapshot mode in CPU mode > > $ sudo ./perf record -vvv -C 0 -e arm_spe/period=148576/u -S -- taskset --cpu-list 0 stress --cpu 1 & > > - Test that auxtrace buffers correspond to an aux record > - Test snapshot default sizes in sudo and user modes > - Test small snapshot size > > $ ./perf record -vvv -e arm_spe/period=148576/u -S1000 -m16,16 -- taskset --cpu-list 0 stress --cpu 1 & > > If there are any concerns with the patches, please let me know and I > will try to address them. Thanks for sharing the testing cases. Could give me a bit more time for the test at my side? And please expect I might give some comments if I think it's necessary. Thanks, Leo
On 15/10/2021 15:16, Leo Yan wrote: > Hi German, > > On Fri, Oct 15, 2021 at 01:33:39PM +0100, German Gomez wrote: > > [...] > > Thanks for sharing the testing cases. Could give me a bit more time for > the test at my side? And please expect I might give some comments if > I think it's necessary. > > Thanks, > Leo Absolutely. Please take the time you need. Many thanks, German
Hi German, Will,
On Fri, Oct 15, 2021 at 01:33:39PM +0100, German Gomez wrote:
[...]
> $ ./perf record -vvv -e arm_spe/period=148576/u -S1000 -m16,16 -- taskset --cpu-list 0 stress --cpu 1 &
When testing Arm SPE snapshot mode with the command (it's quite
similiar with up command but not exactly same):
# ./perf --debug verbose=3 record -e arm_spe/period=148576/u -C 0 -S1000 -m16,16 \
-- taskset --cpu-list 0 stress --cpu 1 &
# kill -USR2 [pid_num]
... then I wait for long time and didn't stop the perf program, then
I observed the output file contains many redundant events
PERF_RECORD_AUX. E.g. in the shared perf data file [1], you could use
below commands to see tons of the events PERF_RECORD_AUX which I only
send only one USR2 signal for taking snapshot:
# perf report -D -i perf.data --stdio | grep -E 'RECORD_AUX' | wc -l
2245787
# perf report -D -i perf.data --stdio | grep -E 'SPE'
. ... ARM SPE data: size 0x3e8 bytes
Binary file (standard input) matches
I looked into the Arm SPE driver and found it doesn't really support
free run mode for AUX ring buffer when the driver runs in snapshot
mode, the pair functions perf_aux_output_end() and
perf_aux_output_begin() are invoked when every time handle the
interrupt. The detailed flow is:
arm_spe_pmu_irq_handler()
`> arm_spe_pmu_buf_get_fault_act()
`> arm_spe_perf_aux_output_end()
`> set SPE registers
`> perf_aux_output_end()
`> arm_spe_perf_aux_output_begin()
`> perf_aux_output_begin()
`> set SPE registers
Seems to me, a possible solution is to add an extra parameter 'int
in_interrupt' for functions arm_spe_perf_aux_output_end() and
arm_spe_perf_aux_output_begin(), if this parameter is passed as 1 in
the interrupt handling, these two functions should skip invoking
perf_aux_output_end() and perf_aux_output_begin() so can avoid the
redundant perf event PERF_RECORD_AUX.
arm_spe_pmu_irq_handler()
`> arm_spe_pmu_buf_get_fault_act()
`> arm_spe_perf_aux_output_end(..., in_interrupt=1)
`> set SPE registers
`> arm_spe_perf_aux_output_begin(..., in_interrupt=1)
`> set SPE registers
P.s. I think Intel-PT has supported free run mode for snapshot mode,
so it should not generate interrupt in this mode. Thus Intel-PT can
avoid this issue, please see the code [2].
Thanks,
Leo
[1] https://people.linaro.org/~leo.yan/spe/snapshot_test/perf.data
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/events/intel/pt.c#n753
On Thu, Sep 16, 2021 at 04:46:34PM +0100, German Gomez wrote: [...] > +static int arm_spe_find_snapshot(struct auxtrace_record *itr, int idx, > + struct auxtrace_mmap *mm, unsigned char *data, > + u64 *head, u64 *old) > +{ > + int err; > + bool wrapped; > + struct arm_spe_recording *ptr = > + container_of(itr, struct arm_spe_recording, itr); > + > + /* > + * Allocate memory to keep track of wrapping if this is the first > + * time we deal with this *mm. > + */ > + if (idx >= ptr->wrapped_cnt) { > + err = arm_spe_alloc_wrapped_array(ptr, idx); > + if (err) > + return err; > + } > + > + /* > + * Check to see if *head has wrapped around. If it hasn't only the > + * amount of data between *head and *old is snapshot'ed to avoid > + * bloating the perf.data file with zeros. But as soon as *head has > + * wrapped around the entire size of the AUX ring buffer it taken. > + */ > + wrapped = ptr->wrapped[idx]; > + if (!wrapped && arm_spe_buffer_has_wrapped(data, mm->len, *head)) { > + wrapped = true; > + ptr->wrapped[idx] = true; > + } > + > + pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n", > + __func__, idx, (size_t)*old, (size_t)*head, mm->len); > + > + /* > + * No wrap has occurred, we can just use *head and *old. > + */ > + if (!wrapped) > + return 0; > + > + /* > + * *head has wrapped around - adjust *head and *old to pickup the > + * entire content of the AUX buffer. > + */ > + if (*head >= mm->len) { > + *old = *head - mm->len; > + } else { > + *head += mm->len; > + *old = *head - mm->len; > + } > + > + return 0; > +} If run a test case (the test is pasted at the end of the reply), I can get quite different AUX trace data with passing different wait period before sending the first USR2 signal. # sh test_arm_spe_snapshot.sh 2 Couldn't synthesize bpf events. stress: info: [5768] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd [ perf record: Woken up 3 times to write data ] [ perf record: Captured and wrote 2.833 MB perf.data ] # sh test_arm_spe_snapshot.sh 10 Couldn't synthesize bpf events. stress: info: [5776] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd [ perf record: Woken up 3 times to write data ] [ perf record: Captured and wrote 24.356 MB perf.data ] The first command passes argument '2' so the test will wait for 2 seconds before send USR2 signal for snapshot, and the perf data file is 2.833 MB (so this means the Arm SPE trace data is about 2MB) for three snapshots. In the second command, the argument '10' means it will wait for 10 seconds before sending the USR2 signals, and every time it records the trace data from the full AUX buffer (8MB), at the end it gets 24MB AUX trace data. The issue happens in the second command, waiting for 10 seconds leads to the *full* AUX ring buffer is filled by Arm SPE, so the function arm_spe_buffer_has_wrapped() always return back true for this case. Afterwards, arm_spe_find_snapshot() doesn't respect the passed old header (from '*old') and assumes the trace data size is 'mm->len'. To allow arm_spe_buffer_has_wrapped() to work properly, I think we need to clean up the top 8 bytes of the AUX buffer in Arm SPE driver when start the PMU event (please note, this change has an assumption that is meantioned in another email that suggests to remove redundant PERF_RECORD_AUX events so the function arm_spe_perf_aux_output_begin() is invoked only once when start PMU event, so we can use the top 8 bytes in AUX buffer to indicate trace is wrap around or not). diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c index d44bcc29d99c..eb35f85d0efb 100644 --- a/drivers/perf/arm_spe_pmu.c +++ b/drivers/perf/arm_spe_pmu.c @@ -493,6 +493,16 @@ static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle, if (limit) limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT); + /* + * Cleanup the top 8 bytes for snapshot mode; these 8 bytes are + * used to indicate if trace data is wrap around if they are not + * zero. + */ + if (buf->snapshot) { + void *tail = buf->base + (buf->nr_pages << PAGE_SHIFT) - 8; + memset(tail, 0x0, 8); + } + limit += (u64)buf->base; base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf); write_sysreg_s(base, SYS_PMBPTR_EL1); Thanks, Leo ---8<--- #!/bin/sh ./perf record -e arm_spe/period=148576/u -C 0 -S -m8M,8M -- taskset --cpu-list 0 stress --cpu 1 & PERFPID=$! echo "sleep $1 seconds" > /sys/kernel/debug/tracing/trace_marker # Wait for perf program sleep $1 # Send signal to snapshot trace data kill -USR2 $PERFPID sleep .03 kill -USR2 $PERFPID sleep .03 kill -USR2 $PERFPID echo "Stop snapshot" > /sys/kernel/debug/tracing/trace_marker kill $PERFPID wait $PERFPID
On Sun, Oct 17, 2021 at 08:05:46PM +0800, Leo Yan wrote: [...] > To allow arm_spe_buffer_has_wrapped() to work properly, I think we > need to clean up the top 8 bytes of the AUX buffer in Arm SPE driver > when start the PMU event (please note, this change has an assumption > that is meantioned in another email that suggests to remove redundant > PERF_RECORD_AUX events so the function arm_spe_perf_aux_output_begin() > is invoked only once when start PMU event, so we can use the top 8 > bytes in AUX buffer to indicate trace is wrap around or not). > > > diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c > index d44bcc29d99c..eb35f85d0efb 100644 > --- a/drivers/perf/arm_spe_pmu.c > +++ b/drivers/perf/arm_spe_pmu.c > @@ -493,6 +493,16 @@ static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle, > if (limit) > limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT); > > + /* > + * Cleanup the top 8 bytes for snapshot mode; these 8 bytes are > + * used to indicate if trace data is wrap around if they are not > + * zero. > + */ > + if (buf->snapshot) { > + void *tail = buf->base + (buf->nr_pages << PAGE_SHIFT) - 8; > + memset(tail, 0x0, 8); Here need to add below code for flushing data cache: flush_dcache_range((unsigned long)tail, (unsigned long)tail+8); Sorry for spamming. Leo > + } > + > limit += (u64)buf->base; > base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf); > write_sysreg_s(base, SYS_PMBPTR_EL1);
Hi Leo, Yeah I agree the redundant AUX events are adding unnecessary bloat to the perf.data file... We actually cam across this when doing one of the test cases. Sorry for not reporting it! Could we patch the driver in a separate patch set? Or do you think this is critical for the purposes of this one? Thanks, German On 17/10/2021 07:13, Leo Yan wrote: > Hi German, Will, > > On Fri, Oct 15, 2021 at 01:33:39PM +0100, German Gomez wrote: > > [...] > >> $ ./perf record -vvv -e arm_spe/period=148576/u -S1000 -m16,16 -- taskset --cpu-list 0 stress --cpu 1 & > When testing Arm SPE snapshot mode with the command (it's quite > similiar with up command but not exactly same): > > # ./perf --debug verbose=3 record -e arm_spe/period=148576/u -C 0 -S1000 -m16,16 \ > -- taskset --cpu-list 0 stress --cpu 1 & > # kill -USR2 [pid_num] > > ... then I wait for long time and didn't stop the perf program, then > I observed the output file contains many redundant events > PERF_RECORD_AUX. E.g. in the shared perf data file [1], you could use > below commands to see tons of the events PERF_RECORD_AUX which I only > send only one USR2 signal for taking snapshot: > > # perf report -D -i perf.data --stdio | grep -E 'RECORD_AUX' | wc -l > 2245787 > > # perf report -D -i perf.data --stdio | grep -E 'SPE' > . ... ARM SPE data: size 0x3e8 bytes > Binary file (standard input) matches > > I looked into the Arm SPE driver and found it doesn't really support > free run mode for AUX ring buffer when the driver runs in snapshot > mode, the pair functions perf_aux_output_end() and > perf_aux_output_begin() are invoked when every time handle the > interrupt. The detailed flow is: > > arm_spe_pmu_irq_handler() > `> arm_spe_pmu_buf_get_fault_act() > `> arm_spe_perf_aux_output_end() > `> set SPE registers > `> perf_aux_output_end() > `> arm_spe_perf_aux_output_begin() > `> perf_aux_output_begin() > `> set SPE registers > > Seems to me, a possible solution is to add an extra parameter 'int > in_interrupt' for functions arm_spe_perf_aux_output_end() and > arm_spe_perf_aux_output_begin(), if this parameter is passed as 1 in > the interrupt handling, these two functions should skip invoking > perf_aux_output_end() and perf_aux_output_begin() so can avoid the > redundant perf event PERF_RECORD_AUX. > > arm_spe_pmu_irq_handler() > `> arm_spe_pmu_buf_get_fault_act() > `> arm_spe_perf_aux_output_end(..., in_interrupt=1) > `> set SPE registers > `> arm_spe_perf_aux_output_begin(..., in_interrupt=1) > `> set SPE registers > > P.s. I think Intel-PT has supported free run mode for snapshot mode, > so it should not generate interrupt in this mode. Thus Intel-PT can > avoid this issue, please see the code [2]. > > Thanks, > Leo > > [1] https://people.linaro.org/~leo.yan/spe/snapshot_test/perf.data > [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/events/intel/pt.c#n753
Hi German, On Tue, Oct 19, 2021 at 10:23:01AM +0100, German Gomez wrote: > Hi Leo, > > Yeah I agree the redundant AUX events are adding unnecessary bloat to > the perf.data file... We actually cam across this when doing one of the > test cases. Sorry for not reporting it! No worries. > Could we patch the driver in a separate patch set? Or do you think this > is critical for the purposes of this one? Yeah, we can take low priority for the redundant AUX events issue. Please take a look for the issue mentioned in another email for recording trace data with wrong size. I think the issue for wrong snapshot trace size should have a fixing in Arm SPE driver, and the fixing need to be verified with the perf patches. After that I am fine for merging the perf patches (and you could upstream kernel driver patches separately). How about you think? Thanks, Leo
Hi Leo, On 17/10/2021 13:05, Leo Yan wrote: > On Thu, Sep 16, 2021 at 04:46:34PM +0100, German Gomez wrote: > > [...] > > If run a test case (the test is pasted at the end of the reply), I > can get quite different AUX trace data with passing different wait > period before sending the first USR2 signal. > > # sh test_arm_spe_snapshot.sh 2 > Couldn't synthesize bpf events. > stress: info: [5768] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd > [ perf record: Woken up 3 times to write data ] > [ perf record: Captured and wrote 2.833 MB perf.data ] > > # sh test_arm_spe_snapshot.sh 10 > Couldn't synthesize bpf events. > stress: info: [5776] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd > [ perf record: Woken up 3 times to write data ] > [ perf record: Captured and wrote 24.356 MB perf.data ] > > The first command passes argument '2' so the test will wait for 2 > seconds before send USR2 signal for snapshot, and the perf data file is > 2.833 MB (so this means the Arm SPE trace data is about 2MB) for three > snapshots. In the second command, the argument '10' means it will wait > for 10 seconds before sending the USR2 signals, and every time it records > the trace data from the full AUX buffer (8MB), at the end it gets 24MB > AUX trace data. > > The issue happens in the second command, waiting for 10 seconds leads > to the *full* AUX ring buffer is filled by Arm SPE, so the function > arm_spe_buffer_has_wrapped() always return back true for this case. > Afterwards, arm_spe_find_snapshot() doesn't respect the passed old > header (from '*old') and assumes the trace data size is 'mm->len'. Returning the entire contents of the buffer once the first wrap-around was detected was the intention of the patch, so I don't currently see it as wrong. What were the values you were expecting to see in the test? If the handling of snapshot mode by the perf tool can be improved after upstreaming the changes to the driver, we could submit a followup patch after that has been fixed. > > To allow arm_spe_buffer_has_wrapped() to work properly, I think we > need to clean up the top 8 bytes of the AUX buffer in Arm SPE driver > when start the PMU event (please note, this change has an assumption > that is meantioned in another email that suggests to remove redundant > PERF_RECORD_AUX events so the function arm_spe_perf_aux_output_begin() > is invoked only once when start PMU event, so we can use the top 8 > bytes in AUX buffer to indicate trace is wrap around or not). > > > diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c > index d44bcc29d99c..eb35f85d0efb 100644 > --- a/drivers/perf/arm_spe_pmu.c > +++ b/drivers/perf/arm_spe_pmu.c > @@ -493,6 +493,16 @@ static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle, > if (limit) > limit |= BIT(SYS_PMBLIMITR_EL1_E_SHIFT); > > + /* > + * Cleanup the top 8 bytes for snapshot mode; these 8 bytes are > + * used to indicate if trace data is wrap around if they are not > + * zero. > + */ > + if (buf->snapshot) { > + void *tail = buf->base + (buf->nr_pages << PAGE_SHIFT) - 8; > + memset(tail, 0x0, 8); > + } > + > limit += (u64)buf->base; > base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf); > write_sysreg_s(base, SYS_PMBPTR_EL1); > > Thanks, > Leo I will try these and the other driver changes and discuss them with the team internally, thanks! > > ---8<--- > > #!/bin/sh > > ./perf record -e arm_spe/period=148576/u -C 0 -S -m8M,8M -- taskset --cpu-list 0 stress --cpu 1 & > > PERFPID=$! > > echo "sleep $1 seconds" > /sys/kernel/debug/tracing/trace_marker > > # Wait for perf program > sleep $1 > > # Send signal to snapshot trace data > kill -USR2 $PERFPID > sleep .03 > kill -USR2 $PERFPID > sleep .03 > kill -USR2 $PERFPID > > echo "Stop snapshot" > /sys/kernel/debug/tracing/trace_marker > > kill $PERFPID > wait $PERFPID
On Tue, Oct 19, 2021 at 06:34:24PM +0100, German Gomez wrote: > Hi Leo, > > On 17/10/2021 13:05, Leo Yan wrote: > > On Thu, Sep 16, 2021 at 04:46:34PM +0100, German Gomez wrote: > > > > [...] > > > > If run a test case (the test is pasted at the end of the reply), I > > can get quite different AUX trace data with passing different wait > > period before sending the first USR2 signal. > > > > # sh test_arm_spe_snapshot.sh 2 > > Couldn't synthesize bpf events. > > stress: info: [5768] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd > > [ perf record: Woken up 3 times to write data ] > > [ perf record: Captured and wrote 2.833 MB perf.data ] > > > > # sh test_arm_spe_snapshot.sh 10 > > Couldn't synthesize bpf events. > > stress: info: [5776] dispatching hogs: 1 cpu, 0 io, 0 vm, 0 hdd > > [ perf record: Woken up 3 times to write data ] > > [ perf record: Captured and wrote 24.356 MB perf.data ] > > > > The first command passes argument '2' so the test will wait for 2 > > seconds before send USR2 signal for snapshot, and the perf data file is > > 2.833 MB (so this means the Arm SPE trace data is about 2MB) for three > > snapshots. In the second command, the argument '10' means it will wait > > for 10 seconds before sending the USR2 signals, and every time it records > > the trace data from the full AUX buffer (8MB), at the end it gets 24MB > > AUX trace data. > > > > The issue happens in the second command, waiting for 10 seconds leads > > to the *full* AUX ring buffer is filled by Arm SPE, so the function > > arm_spe_buffer_has_wrapped() always return back true for this case. > > Afterwards, arm_spe_find_snapshot() doesn't respect the passed old > > header (from '*old') and assumes the trace data size is 'mm->len'. > > Returning the entire contents of the buffer once the first wrap-around > was detected was the intention of the patch, so I don't currently see it > as wrong. What were the values you were expecting to see in the test? I expect the second command takes three snapshots: the first time it should record AUX trace data with full buffer size (8MB) after waiting for 10 seconds, and later two times will take small AUX trace data since the interval (0.03s) is short and Arm SPE has not filled the full AUX buffer. > If the handling of snapshot mode by the perf tool can be improved after > upstreaming the changes to the driver, we could submit a followup patch > after that has been fixed. Okay, I understand now the main concern is for kernel driver changes, this patch for perf tool is fine for me: Reviewed-by: Leo Yan <leo.yan@linaro.org> Tested-by: Leo Yan <leo.yan@linaro.org> [...] > I will try these and the other driver changes and discuss them with the > team internally, thanks! Thanks a lot!
Hi Leo, On 17/10/2021 07:13, Leo Yan wrote: > [...] > > I looked into the Arm SPE driver and found it doesn't really support > free run mode for AUX ring buffer when the driver runs in snapshot > mode, the pair functions perf_aux_output_end() and > perf_aux_output_begin() are invoked when every time handle the > interrupt. The detailed flow is: > > arm_spe_pmu_irq_handler() > `> arm_spe_pmu_buf_get_fault_act() > `> arm_spe_perf_aux_output_end() > `> set SPE registers > `> perf_aux_output_end() > `> arm_spe_perf_aux_output_begin() > `> perf_aux_output_begin() > `> set SPE registers > > Seems to me, a possible solution is to add an extra parameter 'int > in_interrupt' for functions arm_spe_perf_aux_output_end() and > arm_spe_perf_aux_output_begin(), if this parameter is passed as 1 in > the interrupt handling, these two functions should skip invoking > perf_aux_output_end() and perf_aux_output_begin() so can avoid the > redundant perf event PERF_RECORD_AUX. > > arm_spe_pmu_irq_handler() > `> arm_spe_pmu_buf_get_fault_act() > `> arm_spe_perf_aux_output_end(..., in_interrupt=1) > `> set SPE registers > `> arm_spe_perf_aux_output_begin(..., in_interrupt=1) > `> set SPE registers I brought the issue of the redundant AUX events to the team, and we know of at least one tool in Arm relying on these events in snapshot mode. So we think that changing this behavior of the driver might not be easy to do right now. > > P.s. I think Intel-PT has supported free run mode for snapshot mode, > so it should not generate interrupt in this mode. Thus Intel-PT can > avoid this issue, please see the code [2]. > > Thanks, > Leo > > [1] https://people.linaro.org/~leo.yan/spe/snapshot_test/perf.data > [2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/events/intel/pt.c#n753 Thanks, German
diff --git a/tools/perf/arch/arm64/util/arm-spe.c b/tools/perf/arch/arm64/util/arm-spe.c index f8b03d164b42..56785034fc84 100644 --- a/tools/perf/arch/arm64/util/arm-spe.c +++ b/tools/perf/arch/arm64/util/arm-spe.c @@ -23,6 +23,7 @@ #include "../../../util/auxtrace.h" #include "../../../util/record.h" #include "../../../util/arm-spe.h" +#include <tools/libc_compat.h> // reallocarray #define KiB(x) ((x) * 1024) #define MiB(x) ((x) * 1024 * 1024) @@ -31,6 +32,8 @@ struct arm_spe_recording { struct auxtrace_record itr; struct perf_pmu *arm_spe_pmu; struct evlist *evlist; + int wrapped_cnt; + bool *wrapped; }; static void arm_spe_set_timestamp(struct auxtrace_record *itr, @@ -299,6 +302,146 @@ static int arm_spe_snapshot_finish(struct auxtrace_record *itr) return -EINVAL; } +static int arm_spe_alloc_wrapped_array(struct arm_spe_recording *ptr, int idx) +{ + bool *wrapped; + int cnt = ptr->wrapped_cnt, new_cnt, i; + + /* + * No need to allocate, so return early. + */ + if (idx < cnt) + return 0; + + /* + * Make ptr->wrapped as big as idx. + */ + new_cnt = idx + 1; + + /* + * Free'ed in arm_spe_recording_free(). + */ + wrapped = reallocarray(ptr->wrapped, new_cnt, sizeof(bool)); + if (!wrapped) + return -ENOMEM; + + /* + * init new allocated values. + */ + for (i = cnt; i < new_cnt; i++) + wrapped[i] = false; + + ptr->wrapped_cnt = new_cnt; + ptr->wrapped = wrapped; + + return 0; +} + +static bool arm_spe_buffer_has_wrapped(unsigned char *buffer, + size_t buffer_size, u64 head) +{ + u64 i, watermark; + u64 *buf = (u64 *)buffer; + size_t buf_size = buffer_size; + + /* + * Defensively handle the case where head might be continually increasing - if its value is + * equal or greater than the size of the ring buffer, then we can safely determine it has + * wrapped around. Otherwise, continue to detect if head might have wrapped. + */ + if (head >= buffer_size) + return true; + + /* + * We want to look the very last 512 byte (chosen arbitrarily) in the ring buffer. + */ + watermark = buf_size - 512; + + /* + * The value of head is somewhere within the size of the ring buffer. This can be that there + * hasn't been enough data to fill the ring buffer yet or the trace time was so long that + * head has numerically wrapped around. To find we need to check if we have data at the + * very end of the ring buffer. We can reliably do this because mmap'ed pages are zeroed + * out and there is a fresh mapping with every new session. + */ + + /* + * head is less than 512 byte from the end of the ring buffer. + */ + if (head > watermark) + watermark = head; + + /* + * Speed things up by using 64 bit transactions (see "u64 *buf" above) + */ + watermark /= sizeof(u64); + buf_size /= sizeof(u64); + + /* + * If we find trace data at the end of the ring buffer, head has been there and has + * numerically wrapped around at least once. + */ + for (i = watermark; i < buf_size; i++) + if (buf[i]) + return true; + + return false; +} + +static int arm_spe_find_snapshot(struct auxtrace_record *itr, int idx, + struct auxtrace_mmap *mm, unsigned char *data, + u64 *head, u64 *old) +{ + int err; + bool wrapped; + struct arm_spe_recording *ptr = + container_of(itr, struct arm_spe_recording, itr); + + /* + * Allocate memory to keep track of wrapping if this is the first + * time we deal with this *mm. + */ + if (idx >= ptr->wrapped_cnt) { + err = arm_spe_alloc_wrapped_array(ptr, idx); + if (err) + return err; + } + + /* + * Check to see if *head has wrapped around. If it hasn't only the + * amount of data between *head and *old is snapshot'ed to avoid + * bloating the perf.data file with zeros. But as soon as *head has + * wrapped around the entire size of the AUX ring buffer it taken. + */ + wrapped = ptr->wrapped[idx]; + if (!wrapped && arm_spe_buffer_has_wrapped(data, mm->len, *head)) { + wrapped = true; + ptr->wrapped[idx] = true; + } + + pr_debug3("%s: mmap index %d old head %zu new head %zu size %zu\n", + __func__, idx, (size_t)*old, (size_t)*head, mm->len); + + /* + * No wrap has occurred, we can just use *head and *old. + */ + if (!wrapped) + return 0; + + /* + * *head has wrapped around - adjust *head and *old to pickup the + * entire content of the AUX buffer. + */ + if (*head >= mm->len) { + *old = *head - mm->len; + } else { + *head += mm->len; + *old = *head - mm->len; + } + + return 0; +} + static u64 arm_spe_reference(struct auxtrace_record *itr __maybe_unused) { struct timespec ts; @@ -313,6 +456,7 @@ static void arm_spe_recording_free(struct auxtrace_record *itr) struct arm_spe_recording *sper = container_of(itr, struct arm_spe_recording, itr); + free(sper->wrapped); free(sper); } @@ -336,6 +480,7 @@ struct auxtrace_record *arm_spe_recording_init(int *err, sper->itr.pmu = arm_spe_pmu; sper->itr.snapshot_start = arm_spe_snapshot_start; sper->itr.snapshot_finish = arm_spe_snapshot_finish; + sper->itr.find_snapshot = arm_spe_find_snapshot; sper->itr.parse_snapshot_options = arm_spe_parse_snapshot_options; sper->itr.recording_options = arm_spe_recording_options; sper->itr.info_priv_size = arm_spe_info_priv_size;