On 08.08.20 17:05, Alan Stern wrote:
> On Sat, Aug 08, 2020 at 08:59:09AM +0200, Martin Kepplinger wrote:
>> On 07.08.20 16:30, Alan Stern wrote:
>>> On Fri, Aug 07, 2020 at 11:51:21AM +0200, Martin Kepplinger wrote:
>>>> it's really strange: below is the change I'm trying. Of course that's
>>>> only for testing the functionality, nothing how a patch could look like.
>>>>
>>>> While I remember it had worked, now (weirdly since I tried that mounting
>>>> via fstab) it doesn't anymore!
>>>>
>>>> What I understand (not much): I handle the error with "retry" via the
>>>> new flag, but scsi_decide_disposition() returns SUCCESS because of "no
>>>> more retries"; but it's the first and only time it's called.
>>>
>>> Are you saying that scmd->allowed is set to 0? Or is scsi_notry_cmd()
>>> returning a nonzero value? Whichever is true, why does it happen that
>>> way?
>>
>> scsi_notry_cmd() is returning 1. (it's retry 1 of 5 allowed).
>>
>> why is it returning 1? REQ_FAILFAST_DEV is set. It's DID_OK, then "if
>> (status_byte(scmd->result) != CHECK_CONDITION)" appearently is not true,
>> then at the end it returns 1 because of REQ_FAILFAST_DEV.
>>
>> that seems to come from the block layer. why and when? could I change
>> that so that the scsi error handling stays in control?
>
> The only place I see where that flag might get set is in
> blk_mq_bio_to_request() in block/blk-mq.c, which does:
>
> if (bio->bi_opf & REQ_RAHEAD)
> rq->cmd_flags |= REQ_FAILFAST_MASK;
>
> So apparently read-ahead reads are supposed to fail fast (i.e., without
> retries), presumably because they are optional after all.
>
>>> What is the failing command? Is it a READ(10)?
>>
>> Not sure how I'd answer that, but here's the test to trigger the error:
>>
>> mount /dev/sda1 /mnt
>> cd /mnt
>> ls
>> cp file ~/ (if ls "works" and doesn't yet trigger the error)
>>
>> and that's the (familiar looking) logs when doing so. again: despite the
>> mentioned workaround in scsi_error and the new expected_media_change
>> flag *is* set and gets cleared, as it should be. REQ_FAILFAST_DEV seems
>> to override what I want to do is scsi_error:
>>
>> [ 55.557629] sd 0:0:0:0: [sda] tag#0 UNKNOWN(0x2003) Result:
>> hostbyte=0x00 driverbyte=0x08 cmd_age=0s
>> [ 55.557639] sd 0:0:0:0: [sda] tag#0 Sense Key : 0x6 [current]
>> [ 55.557646] sd 0:0:0:0: [sda] tag#0 ASC=0x28 ASCQ=0x0
>> [ 55.557657] sd 0:0:0:0: [sda] tag#0 CDB: opcode=0x28 28 00 00 08 fc
>> e0 00 00 01 00
>
> Yes, 0x28 is READ(10). Likely this is a read-ahead request, although I
> don't know how we can tell for sure.
>
>> [ 55.557666] blk_update_request: I/O error, dev sda, sector 589024 op
>> 0x0:(READ) flags 0x80700 phys_seg 1 prio class 0
>> [ 55.568899] sd 0:0:0:0: [sda] tag#0 device offline or changed
>> [ 55.574691] blk_update_request: I/O error, dev sda, sector 589025 op
>> 0x0:(READ) flags 0x80700 phys_seg 1 prio class 0
>> [ 55.585756] sd 0:0:0:0: [sda] tag#0 device offline or changed
>> [ 55.591562] blk_update_request: I/O error, dev sda, sector 589026 op
>> 0x0:(READ) flags 0x80700 phys_seg 1 prio class 0
>> [ 55.602274] sd 0:0:0:0: [sda] tag#0 device offline or changed
>> (... goes on with the same)
>
> Is such a drastic response really appropriate for the failure of a
> read-ahead request? It seems like a more logical response would be to
> let the request fail but keep the device online.
>
> Of course, that would only solve part of your problem -- your log would
> still get filled with those "I/O error" messages even though they
> wouldn't be fatal. Probably a better approach would be to make the new
> expecting_media_change flag override scsi_no_retry_cmd().
>
> But this is not my area of expertise. Maybe someone else will have more
> to say.
>
> Alan Stern
>
Hey Alan, I'm really glad for that, I suspected some of this but I have
little experience in scsi/block layers, so that is super helpful.
I'd appreciate an opinion on the below workaround that *seems* to work
now (let's see, I had thought so before :)
Whether or not this helps to find a real solution, let's see. But
integration of such a flag in the error handling paths is what's
interesting for now:
--- a/drivers/scsi/scsi_error.c
+++ b/drivers/scsi/scsi_error.c
@@ -565,6 +565,17 @@ int scsi_check_sense(struct scsi_cmnd *scmd)
return NEEDS_RETRY;
}
}
+ if (scmd->device->expecting_media_change) {
+ if (sshdr.asc == 0x28 && sshdr.ascq == 0x00) {
+ /* clear expecting_media_change in
+ * scsi_noretry_cmd() because we need
+ * to override possible "failfast" overrides
+ * that block readahead can cause.
+ */
+ return NEEDS_RETRY;
+ }
+ }
+
/*
* we might also expect a cc/ua if another LUN on the target
* reported a UA with an ASC/ASCQ of 3F 0E -
@@ -1744,6 +1755,15 @@ int scsi_noretry_cmd(struct scsi_cmnd *scmd)
return (scmd->request->cmd_flags & REQ_FAILFAST_DRIVER);
}
+ /*
+ * We need to have retries when expecting_media_change is set.
+ * So we need to return success and clear the flag here.
+ */
+ if (scmd->device->expecting_media_change) {
+ scmd->device->expecting_media_change = 0;
+ return 0;
+ }
+
if (status_byte(scmd->result) != CHECK_CONDITION)
return 0;
diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
index d90fefffe31b..bb583e403b81 100644
--- a/drivers/scsi/sd.c
+++ b/drivers/scsi/sd.c
@@ -3642,6 +3642,8 @@ static int sd_resume(struct device *dev)
if (!sdkp) /* E.g.: runtime resume at the start of sd_probe() */
return 0;
+ sdkp->device->expecting_media_change = 1;
+
if (!sdkp->device->manage_start_stop)
return 0;
diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
index bc5909033d13..f5fc1af68e00 100644
--- a/include/scsi/scsi_device.h
+++ b/include/scsi/scsi_device.h
@@ -169,6 +169,7 @@ struct scsi_device {
* this device */
unsigned expecting_cc_ua:1; /* Expecting a CHECK_CONDITION/UNIT_ATTN
* because we did a bus reset. */
+ unsigned expecting_media_change:1;
unsigned use_10_for_rw:1; /* first try 10-byte read / write */
unsigned use_10_for_ms:1; /* first try 10-byte mode sense/select */
unsigned set_dbd_for_ms:1; /* Set "DBD" field in mode sense */
--
