OOPS in scsi generic stuff 2.4.10-pre6
Quoted 3 lines · [email protected] <<nameless>> wrote:
The ksymoops output. Let me know if there is anything else I can offer to help? This happened when ripping a cd with cdparanoia on an IDE drive with the ide-scsi stuff.
....
Quoted 4 lines
Quoted 1 line
Quoted 1 line
EIP;
c01b7688<generic_unplug_device+8/30> <=====Trace;
c01f8fe6<sg_common_write+1d6/1f0>
Trace;c01f9bc0<sg_cmd_done_bh+0/280> #### bizarre
Trace;c01f8c16<sg_write+256/280>
generic_unplug_device() was an addition into the sg driver by Jens Axboe. Under heavy stress testing I have also received an oops from this function.
It is there because the tentacles of the Linux block subsystem have found their way into the the SCSI midlevel. The st and sg drivers are proof of why this is bad design as they are char devices.
If the generic_unplug_device() call is removed then the sg driver will periodically have its commands suspended on the SCSI mid level queue until the block subsystem decides to send something to the device in question. This can be seconds (which isn't a pleasant thing to do to a cdwriter).
Doug Gilbert
Oops output:
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
Unable to handle kernel paging request at virtual address 0a080294 c01b7688 *pde = 00000000 Oops: 0000 CPU: 0 EIP: 0010:[<c01b7688>]
Using defaults from ksymoops -t elf32-i386 -a i386
EFLAGS: 00210002
eax: 0a08021c ebx: 00200202 ecx: 00000010 edx: 0a08021c
esi: e6634064 edi: e6634040 ebp: c1bc8940 esp: d55bfefc
ds: 0018 es: 0018 ss: 0018
Process cdparanoia (pid: 3380, stackpage=d55bf000)
Stack: ea8bfc80 c01f8fe7 0a08021c ea8bfc80 d55bff6c dd2f0000 00007770 c01f9bc0
00001770 00000001 d55bff6c e6634000 e6634040 08058b64 c01f8c17 e6634000
e6634040 d55bff6c 00001770 00000001 e3cae540 ffffffea 00000000 000077a0
Call Trace: [<c01f8fe7>] [<c01f9bc0>] [<c01f8c17>] [<c012ebe6>] [<c0106c2b>]
Code: 80 7a 78 00 74 15 c6 42 78 00 8d 42 28 39 42 28 74 09 52 8bQuoted 1 line
Quoted 1 line
EIP;
c01b7688<generic_unplug_device+8/30> <=====
Trace; c01f8fe6 <sg_common_write+1d6/1f0>
Trace; c01f9bc0 <sg_cmd_done_bh+0/280>
Trace; c01f8c16 <sg_write+256/280>
Trace; c012ebe6 <sys_write+96/d0>
Trace; c0106c2a <system_call+32/38>
Code; c01b7688 <generic_unplug_device+8/30>
00000000 <_EIP>:
Code; c01b7688 <generic_unplug_device+8/30> <=====
0: 80 7a 78 00 cmpb $0x0,0x78(%edx) <=====
Code; c01b768c <generic_unplug_device+c/30>
4: 74 15 je 1b <_EIP+0x1b> c01b76a2
<generic_unplug_device+22/30>
Code; c01b768e <generic_unplug_device+e/30>
6: c6 42 78 00 movb $0x0,0x78(%edx)
Code; c01b7692 <generic_unplug_device+12/30>
a: 8d 42 28 lea 0x28(%edx),%eax
Code; c01b7694 <generic_unplug_device+14/30>
d: 39 42 28 cmp %eax,0x28(%edx)
Code; c01b7698 <generic_unplug_device+18/30>
10: 74 09 je 1b <_EIP+0x1b> c01b76a2
<generic_unplug_device+22/30>
Code; c01b769a <generic_unplug_device+1a/30>
12: 52 push %edx
Code; c01b769a <generic_unplug_device+1a/30>
13: 8b 00 mov (%eax),%eax
Quoted 7 lines · Douglas Gilbert
Quoted 3 lines · [email protected] <<nameless>> wrote:
The ksymoops output. Let me know if there is anything else I can offer to help? This happened when ripping a cd with cdparanoia on an IDE drive with the ide-scsi stuff.
....
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
Quoted 1 line
EIP;
c01b7688<generic_unplug_device+8/30> <=====Trace;
c01f8fe6<sg_common_write+1d6/1f0>
Trace;c01f9bc0<sg_cmd_done_bh+0/280> #### bizarre
Trace;c01f8c16<sg_write+256/280>
generic_unplug_device()was an addition into the sg driver by Jens Axboe. Under heavy stress testing I have also received an oops from this function.
It looks like a race in that sg_cmd_done_bh can be completed before generic_unplug_device is called (and thus on a free'd scsi request). We then pass an invalid queue to generic_unplug_device.
Quoted 4 lines
It is there because the tentacles of the Linux block subsystem have found their way into the the SCSI midlevel. The st and sg drivers are proof of why this is bad design as they are char devices.
Not so. The SCSI mid level should just force this unplug, or have a special do_req that forces imediate completion (or start of execution, rather). The unplug in sg is needed because of that.
Quoted 6 lines
If the
generic_unplug_device()call is removed then the sg driver will periodically have its commands suspended on the SCSI mid level queue until the block subsystem decides to send something to the device in question. This can be seconds (which isn't a pleasant thing to do to a cdwriter).
Because the scsi_request_fn can quit with commands on the queue. That's a SCSI internal issue. This is not a block layer bug.
scsi/sg.c renamed from scsi/sg.c~ +3 -1
@@ -645,6 +645,7 @@Scsi_Request * SRpnt;Sg_device * sdp = sfp->parentdp;sg_io_hdr_t * hp = &srp->header; +request_queue_t * q; srp->data.cmd_opcode = cmnd[0 ];/* hold opcode of command */ hp->status =0 ;
@@ -673,6 +674,7 @@return -ENODEV ; } SRpnt =scsi_allocate_request (sdp->device); + q = &SRpnt->sr_device->request_queue;if (SRpnt ==NULL ) {SCSI_LOG_TIMEOUT (1 ,printk ("sg_write: no mem\n" ));sg_finish_rem_req (srp);
@@ -715,7 +717,7 @@ (void *)SRpnt->sr_buffer, hp->dxfer_len, sg_cmd_done_bh, timeout, SG_DEFAULT_RETRIES);/* dxfer_len overwrites SRpnt->sr_bufflen, hence need for b_malloc_len */ -generic_unplug_device (&SRpnt->sr_device->request_queue); +generic_unplug_device (q);return 0 ; }
Signature · 1 line
Jens Axboe
Quoted 3 lines · Jens Axboe
It looks like a race in that
sg_cmd_done_bhcan be completed beforegeneric_unplug_deviceis called (and thus on a free'd scsi request). We then pass an invalid queue togeneric_unplug_device.
(corrected version, scsi_allocate_request can of course fail)
scsi/sg.c renamed from scsi/sg.c~ +3 -1
@@ -645,6 +645,7 @@Scsi_Request * SRpnt;Sg_device * sdp = sfp->parentdp;sg_io_hdr_t * hp = &srp->header; +request_queue_t * q; srp->data.cmd_opcode = cmnd[0 ];/* hold opcode of command */ hp->status =0 ;
@@ -680,6 +681,7 @@ } srp->my_cmdp = SRpnt; + q = &SRpnt->sr_device->request_queue; SRpnt->sr_request.rq_dev = sdp->i_rdev; SRpnt->sr_request.rq_status =RQ_ACTIVE ; SRpnt->sr_sense_buffer[0 ] =0 ;
@@ -715,7 +717,7 @@ (void *)SRpnt->sr_buffer, hp->dxfer_len, sg_cmd_done_bh, timeout, SG_DEFAULT_RETRIES);/* dxfer_len overwrites SRpnt->sr_bufflen, hence need for b_malloc_len */ -generic_unplug_device (&SRpnt->sr_device->request_queue); +generic_unplug_device (q);return 0 ; }
Signature · 1 line
Jens Axboe
Quoted 7 lines · Jens Axboe
Quoted 3 lines · Jens Axboe
It looks like a race in that
sg_cmd_done_bhcan be completed beforegeneric_unplug_deviceis called (and thus on a free'd scsi request). We then pass an invalid queue togeneric_unplug_device.(corrected version,
scsi_allocate_requestcan of course fail)
Jens,
Prior to this patch (actually the first one you posted today) sg_dd would frequently crash in generic_unplug_device when tested against the scsi_debug adapter driver. [I have hacked up that driver to simulate a large number of (ram) disks to test Richard Gooch's 2000+ scsi disk patch.]
The way scsi_debug handles all its commands, the bottom half handler in sg will be called before scsi_do_req() completes. With this patch the problem goes away.
Doug Gilbert
Quoted 16 lines · Douglas Gilbert
Quoted 7 lines · Jens Axboe
Quoted 3 lines · Jens Axboe
It looks like a race in that
sg_cmd_done_bhcan be completed beforegeneric_unplug_deviceis called (and thus on a free'd scsi request). We then pass an invalid queue togeneric_unplug_device.(corrected version,
scsi_allocate_requestcan of course fail)Jens,
Prior to this patch (actually the first one you posted today)sg_ddwould frequently crash ingeneric_unplug_devicewhen tested against thescsi_debugadapter driver. [I have hacked up that driver to simulate a large number of (ram) disks to test Richard Gooch's 2000+ scsi disk patch.]
Quoted 3 lines
The way
scsi_debughandles all its commands, the bottom half handler in sg will be called beforescsi_do_req()completes. With this patch the problem goes away.
Good, so at least that long standing race has been closed now.
Signature · 1 line
Jens Axboe