On Mon, 2017-03-13 at 14:55 -0700, James Bottomley wrote:
> On Mon, 2017-03-13 at 20:33 +0000, Bart Van Assche wrote:
> > On Mon, 2017-03-13 at 12:23 -0700, James Bottomley wrote:
> > > On Mon, 2017-03-13 at 18:49 +0000, Bart Van Assche wrote:
> > > > diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
> > > > index 7bfbcfa7af40..b3bb49d06943 100644
> > > > --- a/drivers/scsi/scsi.c
> > > > +++ b/drivers/scsi/scsi.c
> > > > @@ -602,7 +602,7 @@ EXPORT_SYMBOL(scsi_device_get);
> > > > */
> > > > void scsi_device_put(struct scsi_device *sdev)
> > > > {
> > > > - module_put(sdev->host->hostt->module);
> > > > + module_put(sdev->hostt->module);
> > > > put_device(&sdev->sdev_gendev);
> > > > }
> > > > EXPORT_SYMBOL(scsi_device_put);
> > > > diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
> > > > index 6f7128f49c30..7134487abbb1 100644
> > > > --- a/drivers/scsi/scsi_scan.c
> > > > +++ b/drivers/scsi/scsi_scan.c
> > > > @@ -227,6 +227,7 @@ static struct scsi_device
> > > > *scsi_alloc_sdev(struct
> > > > scsi_target *starget,
> > > > sdev->model = scsi_null_device_strs;
> > > > sdev->rev = scsi_null_device_strs;
> > > > sdev->host = shost;
> > > > + sdev->hostt = shost->hostt;
> > > > sdev->queue_ramp_up_period = SCSI_DEFAULT_RAMP_UP_PERIOD;
> > > > sdev->id = starget->id;
> > > > sdev->lun = lun;
> > > > diff --git a/include/scsi/scsi_device.h
> > > > b/include/scsi/scsi_device.h
> > > > index 6f22b39f1b0c..cda620ed5922 100644
> > > > --- a/include/scsi/scsi_device.h
> > > > +++ b/include/scsi/scsi_device.h
> > > > @@ -82,6 +82,7 @@ struct scsi_event {
> > > >
> > > > struct scsi_device {
> > > > struct Scsi_Host *host;
> > > > + struct scsi_host_template *hostt;
> > > > struct request_queue *request_queue;
> > > >
> > >
> > > The apparent assumption behind this patch is that sdev->host can be
> > > freed but the sdev will still exist? That shouldn't be correct:
> > > the
> > > rule for struct devices is that the child always holds the parent
> > > and
> > > the host is parented (albeit not necessarily directly) to the sdev,
> > > so
> > > it looks like something has gone wrong if the host had been freed
> > > before the sdev.
> >
> > Hello James,
> >
> > scsi_remove_host() decreases the sdev reference count but does not
> > wait until the sdev release work has finished. This is why the SCSI
> > host can already have disappeared before the last scsi_device_put()
> > call occurs.
>
> This is true, but I don't see how it can cause the host to be freed
> before the sdev. The memory for struct Scsi_Host is freed in the
> shost_gendev release routine, which should be pinned by the parent
> traversal from sdev. So it should not be possible for
> scsi_host_dev_release() to be called before
> scsi_device_dev_release_usercontext() becase the latter has the final
> put of the parent device.
Hello Hannes,
The following crash only occurs with async aborts enabled:
general protection fault: 0000 [#1] SMP
RIP: 0010:scsi_device_put+0xb/0x30
Call Trace:
scsi_disk_put+0x2d/0x40
sd_release+0x3d/0xb0
__blkdev_put+0x29e/0x360
blkdev_put+0x49/0x170
dm_put_table_device+0x58/0xc0 [dm_mod]
dm_put_device+0x70/0xc0 [dm_mod]
free_priority_group+0x92/0xc0 [dm_multipath]
free_multipath+0x70/0xc0 [dm_multipath]
multipath_dtr+0x19/0x20 [dm_multipath]
dm_table_destroy+0x67/0x120 [dm_mod]
dev_suspend+0xde/0x240 [dm_mod]
ctl_ioctl+0x1f5/0x520 [dm_mod]
dm_ctl_ioctl+0xe/0x20 [dm_mod]
do_vfs_ioctl+0x8f/0x700
SyS_ioctl+0x3c/0x70
entry_SYSCALL_64_fastpath+0x18/0xad
I hadn't seen this crash before kernel v4.11-rc1 but with kernel v4.11-rc1
and later I see it if I let the srp-test scripts run for a few minutes. The
patch I used to disable async aborts on my test setup is as follows:
diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
index f2cafae150bc..9075e126d6c8 100644
--- a/drivers/scsi/scsi_error.c
+++ b/drivers/scsi/scsi_error.c
@@ -171,6 +171,7 @@ scmd_eh_abort_handler(struct work_struct *work)
}
}
+#if 0
/**
* scsi_abort_command - schedule a command abort
* @scmd: scmd to abort.
@@ -219,6 +220,12 @@ scsi_abort_command(struct scsi_cmnd *scmd)
queue_delayed_work(shost->tmf_work_q, &scmd->abort_work, HZ / 100);
return SUCCESS;
}
+#else
+static int scsi_abort_command(struct scsi_cmnd *scmd)
+{
+ return FAILED;
+}
+#endif
/**
* scsi_eh_scmd_add - add scsi cmd to error handling.
Bart.
