On 02/12/2024 11:19, James Clark wrote:
On 02/12/2024 9:46 am, Suzuki K Poulose wrote:
Hi Yicong
On 02/12/2024 09:24, Yicong Yang wrote:
From: Yicong Yang yangyicong@hisilicon.com
When trying to run perf and sysfs mode simultaneously, the WARN_ON() in tmc_etr_enable_hw() is triggered sometimes:
WARNING: CPU: 42 PID: 3911571 at drivers/hwtracing/coresight/ coresight-tmc-etr.c:1060 tmc_etr_enable_hw+0xc0/0xd8 [coresight_tmc] [..snip..] Call trace: tmc_etr_enable_hw+0xc0/0xd8 [coresight_tmc] (P) tmc_enable_etr_sink+0x11c/0x250 [coresight_tmc] (L) tmc_enable_etr_sink+0x11c/0x250 [coresight_tmc] coresight_enable_path+0x1c8/0x218 [coresight] coresight_enable_sysfs+0xa4/0x228 [coresight] enable_source_store+0x58/0xa8 [coresight] dev_attr_store+0x20/0x40 sysfs_kf_write+0x4c/0x68 kernfs_fop_write_iter+0x120/0x1b8 vfs_write+0x2c8/0x388 ksys_write+0x74/0x108 __arm64_sys_write+0x24/0x38 el0_svc_common.constprop.0+0x64/0x148 do_el0_svc+0x24/0x38 el0_svc+0x3c/0x130 el0t_64_sync_handler+0xc8/0xd0 el0t_64_sync+0x1ac/0x1b0 ---[ end trace 0000000000000000 ]---
Since the enablement of sysfs mode is separeted into two critical regions, one for sysfs buffer allocation and another for hardware enablement, it's possible to race with the perf mode. Fix this by double check whether the perf mode's been used before enabling the hardware in sysfs mode.
Thanks for the fix. Some minor comments below.
It needs a Fixes tag.
Signed-off-by: Yicong Yang yangyicong@hisilicon.com
.../hwtracing/coresight/coresight-tmc-etr.c | 30 +++++++++++++++++++ 1 file changed, 30 insertions(+)
diff --git a/drivers/hwtracing/coresight/coresight-tmc-etr.c b/ drivers/hwtracing/coresight/coresight-tmc-etr.c index ad83714ca4dc..d382d95da5ff 100644 --- a/drivers/hwtracing/coresight/coresight-tmc-etr.c +++ b/drivers/hwtracing/coresight/coresight-tmc-etr.c @@ -1230,6 +1230,36 @@ static int tmc_enable_etr_sink_sysfs(struct coresight_device *csdev) spin_lock_irqsave(&drvdata->spinlock, flags); + /* + * Since the sysfs buffer allocation and the hardware enablement is not + * in the same critical region, it's possible to race with the perf + * mode: + * [sysfs mode] [perf mode] + * tmc_etr_get_sysfs_buffer() + * spin_lock(&drvdata->spinlock) + * [sysfs buffer allocation] + * spin_unlock(&drvdata->spinlock) + * spin_lock(&drvdata->spinlock) + * tmc_etr_enable_hw() + * drvdata->etr_buf = etr_perf->etr_buf + * spin_unlock(&drvdata->spinlock) + * spin_lock(&drvdata->spinlock) + * tmc_etr_enable_hw() + * WARN_ON(drvdata->etr_buf) // WARN sicne etr_buf initialized at + * the perf side + * spin_unlock(&drvdata->spinlock) + * + * So check here before continue. + */ + if (coresight_get_mode(csdev) == CS_MODE_PERF) { + drvdata->sysfs_buf = NULL; + spin_unlock_irqrestore(&drvdata->spinlock, flags);
+ /* Free allocated memory out side of the spinlock */ + tmc_etr_free_sysfs_buf(sysfs_buf); + return -EBUSY; + }
With this in place, I think we should remove the check for CS_MODE_PERF in get_etr_sysfs_buf() to avoid confusion (which I believe opened up this race)
Suzuki
Is it not simpler to set the mode to SYSFS before allocating the buffer in the first place? Then we don't need to free if it races and can't get into the intermediate state where it's a half initialized sysfs mode. The lock doesn't need to be held the whole time, just when setting the mode.
Or maybe to make it more consistent with etm4_enable() use coresight_take_mode() outside of the lock.
And then also clean up the perf mode check in get_etr_sysfs_buf().
diff --git a/drivers/hwtracing/coresight/coresight-tmc-etr.c b/drivers/ hwtracing/coresight/coresight-tmc-etr.c index a48bb85d0e7f..29c07832127b 100644 --- a/drivers/hwtracing/coresight/coresight-tmc-etr.c +++ b/drivers/hwtracing/coresight/coresight-tmc-etr.c @@ -1219,13 +1219,17 @@ static int tmc_enable_etr_sink_sysfs(struct coresight_device *csdev) int ret = 0; unsigned long flags; struct tmc_drvdata *drvdata = dev_get_drvdata(csdev->dev.parent); - struct etr_buf *sysfs_buf = tmc_etr_get_sysfs_buffer(csdev); + struct etr_buf *sysfs_buf;
+ spin_lock_irqsave(&drvdata->spinlock, flags);
+ if (coresight_get_mode(csdev) == CS_MODE_PERF) + return -EBUSY;
+ sysfs_buf = tmc_etr_get_sysfs_buffer(csdev);
^^ This would try to loack the spinlock again. Or we should explicitly assert that the spinlock is held in the tmc_etr_get_sysfs_buffer() and release it and then allocate the buffer. Which again opens up a race for a PERF session to take it over ?
Suzuki
if (IS_ERR(sysfs_buf)) return PTR_ERR(sysfs_buf);
- spin_lock_irqsave(&drvdata->spinlock, flags);
/* * In sysFS mode we can have multiple writers per sink. Since this * sink is already enabled no memory is needed and the HW need not be