On 02/08/2023 13:25, Ruidong Tian wrote:
Hi James,
Sorry, some local patch caused inaccurate information. Please allow me to reintroduce the question:
If you use perf with 1G AUX buffer, you can get 1G perf data. Perf workload is kernel build here:
perf record -C 0 -m ,1G -e cs_etm// taskset -c 0 make
[ perf record: Captured and wrote 1025.557 MB perf.data ]
But if you use 2G AUX buffer, perf was executed unexpectedly：
perf record -C 0 -m ,2G -e cs_etm// taskset -c 0 make
[ perf record: Captured and wrote 2.615 MB perf.data ]
There are just 2.615 MB perf data rather than 2G, if you probe function "tmc_alloc_etr_buf" in
coresight_tmc module, you can find some clues:
perf probe -m coresight_tmc "tmc_alloc_etr_buf size:s64"
perf record -e probe:tmc_alloc_etr_buf -aR -- perf record -C 0 -m ,2G -e cs_etm// -o cs.data taskset -c 0 make
perf script perf 118267 [064] 4640.324670: probe:tmc_alloc_etr_buf: (ffff80007a9dce60) size_s64=-2147483648 perf 118267 [064] 4640.324681: probe:tmc_alloc_etr_buf: (ffff80007a9dce60) size_s64=1048576
It's pretty obvious what's going on here. The first call of tmc_alloc_etr_buf in alloc_etr_buf was
failed because of overflow, the second call of tmc_alloc_etr_buf just alloc 1M AUX buffer which
is default ETR buffer size rather than 2G. That is why we can just get 2.615MB ( 1M AUX data
- perf header ).
It is necessary to check the conversion from int to s64 in coresight_tmc driver. The issue[1] also
exists in coresight/perf, but it's different from this topic.
Thanks for the investigation, that makes more sense to me now. Are you able to send a v2 of the patch with an updated commit message describing these symptoms instead?
And you can also add:
Reviewed-by: James Clark james.clark@arm.com
Thanks Ruidong
On 2023/7/24 23:38, James Clark wrote:
On 14/07/2023 09:43, Ruidong Tian wrote:
Perf cs_etm session will failed when AUX buffer > 1G.
perf record -C 0 -m ,2G -e cs_etm// -- taskset -c 0 ls failed to mmap with 12 (Cannot allocate memory)
In coresight tmc driver, "nr_pages << PAGE_SHIFT" will overflow when nr_pages >= 0x80000(correspond to 1G AUX buffer). Explicit convert nr_pages to 64 bit to avoid overflow.
Hi Ruidong,
I couldn't reproduce this exact issue with the error message in the commit message. Is it not another manifestation related to this change [1]? I don't actually get any error message, but I was able to get a warning in dmesg even with [1] applied.
Does the overflow not result in a successful session but with the wrong buffer size?
I think the change makes sense, but maybe we also need a check for MAX_ORDER because I can trigger the same WARN_ON from [1]. Or maybe I'm a bit confused because of the other change and not being able to reproduce this exactly coming at the same time.
Thanks James
Signed-off-by: Ruidong Tian tianruidong@linux.alibaba.com
drivers/hwtracing/coresight/coresight-tmc-etr.c | 2 +- drivers/hwtracing/coresight/coresight-tmc.h | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/drivers/hwtracing/coresight/coresight-tmc-etr.c b/drivers/hwtracing/coresight/coresight-tmc-etr.c index 766325de0e29..1425ecd1cf78 100644 --- a/drivers/hwtracing/coresight/coresight-tmc-etr.c +++ b/drivers/hwtracing/coresight/coresight-tmc-etr.c @@ -1267,7 +1267,7 @@ alloc_etr_buf(struct tmc_drvdata *drvdata, struct perf_event *event, * than the size requested via sysfs. */ if ((nr_pages << PAGE_SHIFT) > drvdata->size) { - etr_buf = tmc_alloc_etr_buf(drvdata, (nr_pages << PAGE_SHIFT), + etr_buf = tmc_alloc_etr_buf(drvdata, ((ssize_t)nr_pages << PAGE_SHIFT), 0, node, NULL); if (!IS_ERR(etr_buf)) goto done; diff --git a/drivers/hwtracing/coresight/coresight-tmc.h b/drivers/hwtracing/coresight/coresight-tmc.h index b97da39652d2..0ee48c5ba764 100644 --- a/drivers/hwtracing/coresight/coresight-tmc.h +++ b/drivers/hwtracing/coresight/coresight-tmc.h @@ -325,7 +325,7 @@ ssize_t tmc_sg_table_get_data(struct tmc_sg_table *sg_table, static inline unsigned long tmc_sg_table_buf_size(struct tmc_sg_table *sg_table) { - return sg_table->data_pages.nr_pages << PAGE_SHIFT; + return (unsigned long)sg_table->data_pages.nr_pages << PAGE_SHIFT; } struct coresight_device *tmc_etr_get_catu_device(struct tmc_drvdata *drvdata);