Re: [PATCH] scsi/sg: don't grab scsi host module reference

[Yu Kuai <yukuai1@xxxxxxxxxxxxxxx>]

Hi,

在 2023/07/05 10:16, Yu Kuai 写道:
> Hi,
>
> 在 2023/07/05 2:51, Benjamin Block 写道:
> > On Tue, Jul 04, 2023 at 07:04:00PM +0200, Marc Hartmayer wrote:
> > > On Thu, Jun 22, 2023 at 12:01 AM +0800, Yu Kuai 
> > > <yukuai1@xxxxxxxxxxxxxxx> wrote:
> > > > From: Yu Kuai <yukuai3@xxxxxxxxxx>
> > > > diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
> > > > index 2433eeef042a..dcb73787c29d 100644
> > > > --- a/drivers/scsi/sg.c
> > > > +++ b/drivers/scsi/sg.c
> > > > @@ -1497,7 +1497,7 @@ sg_add_device(struct device *cl_dev)
> > > >       int error;
> > > >       unsigned long iflags;
> > > > -    error = scsi_device_get(scsidp);
> > > > +    error = blk_get_queue(scsidp->request_queue);
> > > >       if (error)
> > > >           return error;
> >
> > Might be interesting as well. Marc showed me a `dmesg` snipped earlier
> > from when the bind fails:
> >
> >    [   15.441817] scsi host2: scsi_eh_2: sleeping
> >    [   15.441899] scsi_debug:sdebug_driver_probe: scsi_debug: trim 
> > poll_queues to 0. poll_q/nr_hw = (0/1)
> >    [   15.441907] scsi host2: scsi_debug: version 0191 [20210520]
> >                     dev_size_mb=8, opts=0x0, submit_queues=1, 
> > statistics=0
> >    [   15.442078] scsi host2: scsi_scan_host_selected: 
> > <4294967295:4294967295:18446744073709551615>
> >    [   15.442267] scsi 2:0:0:0: scsi scan: INQUIRY pass 1 length 36
> >    [   15.442286] scsi 2:0:0:0: scsi scan: INQUIRY successful with 
> > code 0x0
> >    [   15.442296] scsi 2:0:0:0: scsi scan: INQUIRY pass 2 length 96
> >    [   15.442308] scsi 2:0:0:0: scsi scan: INQUIRY successful with 
> > code 0x0
> >    [   15.442317] scsi 2:0:0:0: Direct-Access     Linux    
> > scsi_debug       0191 PQ: 0 ANSI: 7
> >    [   15.442554] scsi 2:0:0:0: Power-on or device reset occurred
> >    [   15.442560] scsi 2:0:0:0: tag#50 Done: SUCCESS Result: 
> > hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
> >    [   15.442565] scsi 2:0:0:0: tag#50 CDB: Report supported operation 
> > codes a3 0c 01 88 00 00 00 00 00 14 00 00
> >    [   15.442569] scsi 2:0:0:0: tag#50 Sense Key : Unit Attention 
> > [current]
> >    [   15.442573] scsi 2:0:0:0: tag#50 Add. Sense: Power on occurred
> >
> > The bind should happend around here somewhere I think.
> >
> >    [   15.472680] sd 2:0:0:0: scsi scan: Sending REPORT LUNS to (try 0)
> >    [   15.472703] sd 2:0:0:0: scsi scan: REPORT LUNS successful (try 
> > 0) result 0x0
> >    [   15.472706] sd 2:0:0:0: scsi scan: REPORT LUN scan
> >    [   15.472709] sd 2:0:0:0: scsi scan: device exists on 2:0:0:0
> >    [   15.492874] sd 2:0:0:0: [sdi] 16384 512-byte logical blocks: 
> > (8.39 MB/8.00 MiB)
> >    [   15.502853] sd 2:0:0:0: [sdi] Write Protect is off
> >    [   15.502856] sd 2:0:0:0: [sdi] Mode Sense: 73 00 10 08
> >    [   15.522819] sd 2:0:0:0: [sdi] Write cache: enabled, read cache: 
> > enabled, supports DPO and FUA
> >    [   15.552773] sd 2:0:0:0: [sdi] Preferred minimum I/O size 512 bytes
> >    [   15.552776] sd 2:0:0:0: [sdi] Optimal transfer size 524288 bytes
> >    [   15.575373] sd 2:0:0:0: [sdi] tag#62 Done: SUCCESS Result: 
> > hostbyte=DID_OK driverbyte=DRIVER_OK cmd_age=0s
> >    [   15.575377] sd 2:0:0:0: [sdi] tag#62 CDB: Inquiry 12 01 b9 00 04 00
> >    [   15.575380] sd 2:0:0:0: [sdi] tag#62 Sense Key : Illegal Request 
> > [current]
> >    [   15.575383] sd 2:0:0:0: [sdi] tag#62 Add. Sense: Invalid field 
> > in cdb
> >    [   15.645749] sd 2:0:0:0: [sdi] Attached SCSI disk
> >
> > But we don't even see the `sg_alloc: dev=...` message that is logged in
> > `sg_alloc()`. And between the change above and the call to `sg_alloc()`,
> > there is only the character device allocation; and if that failed, it
> > would print an error. So either the bind is never even tried, or the new
> > `blk_get_queue()` fails to get a reference.
> >      Which is odd, since the only way that would happen is, if the queue
> > was marked dying; but we see that the stack is using it for LUN probing
> > in `sd`.
>
> Yes, if scsi_device_get() works fine, but blk_get_queue() has problems,
> it seems to me that sg_add_device() can be called with scsi_device queue
> mark dying? This is odd, but I'm not sure if it's the case.

Sorry that I totally messed how blk_get_queue() is called, it returns
true on success and false on failure, not 0 on success and errno on
failure.

Sorry for all the troble.

Kuai
> Thanks,
> Kuai
> > > This change (bisected) triggers a regression in our KVM on s390x CI. The
> > > symptom is that a “scsi_debug device” does not bind to the scsi_generic
> > > driver. On s390x you can reproduce the problem as follows (I have not
> > > tested on x86):
> > >
> > > With this patch applied:
> > >
> > > $ sudo modprobe scsi_debug
> >
> > One more thing maybe worth mentioning: in the kernel configuration we
> > use in the CI we have `sg` built-in. I guess most have it built as
> > module.
> >
> > > $ # Get the 'scsi_host,channel,target_number,LUN' tuple for the 
> > > scsi_debug device
> > > $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> > > [0:0:0:0]
> > > $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
> > > stat: cannot statx '/sys/bus/scsi/devices/0:0:0:0/scsi_generic': No 
> > > such file or directory
> > >
> > >
> > > Patch reverted:
> > >
> > > $ sudo modprobe scsi_debug
> > > $ lsscsi |grep scsi_debug |awk '{ print $1 }'
> > > [0:0:0:0]
> > > $ sudo stat /sys/bus/scsi/devices/0:0:0:0/scsi_generic
> > >    File: /sys/bus/scsi/devices/0:0:0:0/scsi_generic
> > >    Size: 0             Blocks: 0          IO Block: 4096   directory
> > > Device: 0,20    Inode: 12155       Links: 3
> > > …
> >
> > That's all I got from looking at it earlier, so far.
>
> .