Message ID | 98bfca4cbaa24462994bcb533d365414@hyperstone.com |
---|---|
State | New |
Headers | show |
Series | sd: sd_open: prevent device removal during sd_open | expand |
On Thu, Sep 02, 2021 at 01:57:13PM +0000, Christian L?hle wrote: > sd and parent devices must not be removed as sd_open checks for events > > sd_need_revalidate and sd_revalidate_disk traverse the device path > to check for event changes. If during this, e.g. the scsi host is being > removed and its resources freed, this traversal crashes. > Locking with scan_mutex for just a scsi disk open may seem blunt, but there > does not seem to be a more granular option. Also opening /dev/sdX directly > happens rarely enough that this shouldn't cause any issues. Can you please root cause how the device could not be valid, as that should not happen? > > The issue occurred on an older kernel with the following trace: > stack segment: 0000 [#1] PREEMPT SMP PTI > CPU: 1 PID: 121457 Comm: python3 Not tainted 4.14.238hyLinux #1 .. preferably with a current mainline kernel as things changed a lot in this area.
>On Thu, Sep 02, 2021 at 01:57:13PM +0000, Christian L?hle wrote: >> sd and parent devices must not be removed as sd_open checks for events >> >> sd_need_revalidate and sd_revalidate_disk traverse the device path >> to check for event changes. If during this, e.g. the scsi host is being >> removed and its resources freed, this traversal crashes. >> Locking with scan_mutex for just a scsi disk open may seem blunt, but there >> does not seem to be a more granular option. Also opening /dev/sdX directly >> happens rarely enough that this shouldn't cause any issues. > > Can you please root cause how the device could not be valid, as that > should not happen? The device is being removed physically, the driver e.g. uas calls scsi_remove_host while user-space opens the /dev/sdX. >> >> The issue occurred on an older kernel with the following trace: >> stack segment: 0000 [#1] PREEMPT SMP PTI >> CPU: 1 PID: 121457 Comm: python3 Not tainted 4.14.238hyLinux #1 > > .. preferably with a current mainline kernel as things changed a lot > in this area. I will, but might take a while to get a setup to reproduce the issue. (The setup where this occurred actually powerfails the usb device repreatedly.)= Hyperstone GmbH | Line-Eid-Strasse 3 | 78467 Konstanz Managing Directors: Dr. Jan Peter Berns. Commercial register of local courts: Freiburg HRB381782
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c index 610ebba0d66e..ad4da985a473 100644 --- a/drivers/scsi/sd.c +++ b/drivers/scsi/sd.c @@ -1436,6 +1436,16 @@ static int sd_open(struct block_device *bdev, fmode_t mode) if (!scsi_block_when_processing_errors(sdev)) goto error_out; + /* + * Checking for changes to the device must not race with the device + * or its parent host being removed, so lock until sd_open returns. + */ + mutex_lock(&sdev->host->scan_mutex); + if (sdev->sdev_state != SDEV_RUNNING) { + retval = -ERESTARTSYS; + goto unlock_scan_error_out; + } + if (sd_need_revalidate(bdev, sdkp)) sd_revalidate_disk(bdev->bd_disk); @@ -1444,7 +1454,7 @@ static int sd_open(struct block_device *bdev, fmode_t mode) */ retval = -ENOMEDIUM; if (sdev->removable && !sdkp->media_present && !(mode & FMODE_NDELAY)) - goto error_out; + goto unlock_scan_error_out; /* * If the device has the write protect tab set, have the open fail @@ -1452,7 +1462,7 @@ static int sd_open(struct block_device *bdev, fmode_t mode) */ retval = -EROFS; if (sdkp->write_prot && (mode & FMODE_WRITE)) - goto error_out; + goto unlock_scan_error_out; /* * It is possible that the disk changing stuff resulted in @@ -1462,15 +1472,19 @@ static int sd_open(struct block_device *bdev, fmode_t mode) */ retval = -ENXIO; if (!scsi_device_online(sdev)) - goto error_out; + goto unlock_scan_error_out; if ((atomic_inc_return(&sdkp->openers) == 1) && sdev->removable) { if (scsi_block_when_processing_errors(sdev)) scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT); } + mutex_unlock(&sdev->host->scan_mutex); return 0; +unlock_scan_error_out: + mutex_unlock(&sdev->host->scan_mutex); + error_out: scsi_disk_put(sdkp); return retval;
sd and parent devices must not be removed as sd_open checks for events sd_need_revalidate and sd_revalidate_disk traverse the device path to check for event changes. If during this, e.g. the scsi host is being removed and its resources freed, this traversal crashes. Locking with scan_mutex for just a scsi disk open may seem blunt, but there does not seem to be a more granular option. Also opening /dev/sdX directly happens rarely enough that this shouldn't cause any issues. The issue occurred on an older kernel with the following trace: stack segment: 0000 [#1] PREEMPT SMP PTI CPU: 1 PID: 121457 Comm: python3 Not tainted 4.14.238hyLinux #1 Hardware name: ASUS All Series/H81M-D, BIOS 0601 02/20/2014 task: ffff888213dbb700 task.stack: ffffc90008c14000 RIP: 0010:kobject_get_path+0x2a/0xe0 ... Call Trace: kobject_uevent_env+0xe6/0x550 disk_check_events+0x101/0x160 disk_clear_events+0x75/0x100 check_disk_change+0x22/0x60 sd_open+0x70/0x170 [sd_mod] __blkdev_get+0x3fd/0x4b0 ? get_empty_filp+0x57/0x1b0 blkdev_get+0x11b/0x330 ? bd_acquire+0xc0/0xc0 do_dentry_open+0x1ef/0x320 ? __inode_permission+0x85/0xc0 path_openat+0x5cb/0x1500 ? terminate_walk+0xeb/0x100 do_filp_open+0x9b/0x110 ? __check_object_size+0xb4/0x190 ? do_sys_open+0x1bd/0x250 do_sys_open+0x1bd/0x250 do_syscall_64+0x67/0x120 entry_SYSCALL_64_after_hwframe+0x41/0xa6 and this commit fixed that issue, as there has been no other such synchronization in place since then, the issue should still be present in recent kernels. Signed-off-by: Christian Loehle <cloehle@hyperstone.com> --- drivers/scsi/sd.c | 20 +++++++++++++++++--- 1 file changed, 17 insertions(+), 3 deletions(-)