Message ID | 1499726857-7875-2-git-send-email-mauricfo@linux.vnet.ibm.com (mailing list archive) |
---|---|
State | Changes Requested, archived |
Headers | show |
On 07/11/2017 12:47 AM, Mauricio Faria de Oliveira wrote: > According to SPC-4 (5.15.2.4.5 Unavailable state), the unavailable > state may (or may not) transition to other states (e.g., microcode > downloading or hardware error, which may be temporary or permanent). > > But, scsi_dh_alua currently fails I/O requests early on once that > state occurs (in alua_prep_fn()) preventing path checkers in such > function path to actually check if I/O still fails or now works. > > And that prevents a path activation (alua_activate()) which could > update the PG state if it eventually recovered to an active state, > thus resume I/O. (This is also the case with the standby state.) > > This might cause device-mapper multipath to fail all paths to some > storage system that moves the controllers to the unavailable state > for firmware upgrades, and never recover regardless of the storage > system doing upgrades one controller at a time and get them online. > > Then I/O requests are blocked indefinitely due to queue_if_no_path > but the underlying individual paths are fully operational, and can > be verified as such through other function paths (e.g., SG_IO): > > # multipath -l > mpatha (360050764008100dac000000000000100) dm-0 IBM,2145 > size=40G features='2 queue_if_no_path retain_attached_hw_handler' > hwhandler='1 alua' wp=rw > |-+- policy='service-time 0' prio=0 status=enabled > | |- 1:0:1:0 sdf 8:80 failed undef running > | `- 2:0:1:0 sdn 8:208 failed undef running > `-+- policy='service-time 0' prio=0 status=enabled > |- 1:0:0:0 sdb 8:16 failed undef running > `- 2:0:0:0 sdj 8:144 failed undef running > > # strace -e read \ > sg_dd blk_sgio=0 \ > if=/dev/sdj of=/dev/null bs=512 count=1 iflag=direct \ > 2>&1 | grep 512 > read(3, 0x3fff7ba80000, 512) = -1 EIO (Input/output error) > > # strace -e ioctl \ > sg_dd blk_sgio=1 \ > if=/dev/sdj of=/dev/null bs=512 count=1 iflag=direct \ > 2>&1 | grep 512 > ioctl(3, SG_IO, {'S', SG_DXFER_FROM_DEV, cmd[10]=[28, 00, 00, 00, > 00, 00, 00, 00, 01, 00], <...>) = 0 > > So, allow I/O to paths of PGs in unavailable/standby state, so path > checkers can actually check them. > > Also, schedule a recheck when unavailable/standby state is detected > (in alua_check_sense()) to update pg->state, and quiet further SCSI > error messages (in alua_prep_fn()). > > Once a path checker eventually detects a working/active state again, > the PG state is normally updated on path activation (alua_activate(), > as it schedules a recheck), thus I/O requests are no longer failed. > > Signed-off-by: Mauricio Faria de Oliveira <mauricfo@linux.vnet.ibm.com> > Reported-by: Naresh Bannoth <nbannoth@in.ibm.com> > > --- > v2: > - also add support for standby state to alua_check_sense(), alua_prep_fn() > (Bart Van Assche <Bart.VanAssche@sandisk.com>) > > drivers/scsi/device_handler/scsi_dh_alua.c | 25 +++++++++++++++++++++++++ > 1 file changed, 25 insertions(+) > > diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c > index c01b47e5b55a..a1cf3d6aa853 100644 > --- a/drivers/scsi/device_handler/scsi_dh_alua.c > +++ b/drivers/scsi/device_handler/scsi_dh_alua.c > @@ -431,6 +431,26 @@ static int alua_check_sense(struct scsi_device *sdev, > alua_check(sdev, false); > return NEEDS_RETRY; > } > + if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0b) { > + /* > + * LUN Not Accessible - target port in standby state. > + * > + * Do not retry, so failover to another target port occur. > + * Schedule a recheck to update state for other functions. > + */ > + alua_check(sdev, true); > + return SUCCESS; > + } > + if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0c) { > + /* > + * LUN Not Accessible - target port in unavailable state. > + * > + * Do not retry, so failover to another target port occur. > + * Schedule a recheck to update state for other functions. > + */ > + alua_check(sdev, true); > + return SUCCESS; > + } > break; > case UNIT_ATTENTION: > if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) { > @@ -1057,6 +1077,8 @@ static void alua_check(struct scsi_device *sdev, bool force) > * > * Fail I/O to all paths not in state > * active/optimized or active/non-optimized. > + * Allow I/O to paths in state unavailable/standby > + * so path checkers can actually check them. > */ > static int alua_prep_fn(struct scsi_device *sdev, struct request *req) > { > @@ -1072,6 +1094,9 @@ static int alua_prep_fn(struct scsi_device *sdev, struct request *req) > rcu_read_unlock(); > if (state == SCSI_ACCESS_STATE_TRANSITIONING) > ret = BLKPREP_DEFER; > + else if (state == SCSI_ACCESS_STATE_UNAVAILABLE || > + state == SCSI_ACCESS_STATE_STANDBY) > + req->rq_flags |= RQF_QUIET; > else if (state != SCSI_ACCESS_STATE_OPTIMAL && > state != SCSI_ACCESS_STATE_ACTIVE && > state != SCSI_ACCESS_STATE_LBA) { > NACK. The whole _point_ of having device handlers is to _avoid_ I/O errors during booting. And the ALUA checker is prepared to handle this situation properly. The directio checker of course doesn't know about this, but then no-one expected the directio checker to work with ALUA. Cheers, Hannes
On 07/11/2017 06:18 AM, Hannes Reinecke wrote: > NACK. > > The whole_point_ of having device handlers is to_avoid_ I/O errors > during booting. > > And the ALUA checker is prepared to handle this situation properly. > The directio checker of course doesn't know about this, but then no-one > expected the directio checker to work with ALUA. I lacked that more holistic understanding. Thanks for explaining. Now, for the sake of logging/debugging... Any problem with patches 2 and 4? Also, it seems the Unavailable/Standby states would not be logged without a recheck from alua_check_sense(), since the only callers of alua_rtpg_queue() are alua_activate() and alua_check[_sense]() [the call from alua_check_vpd() is only in the initialization path]. Isn't there a point in scheduling a recheck once those conditions are found in alua_check_sense() to get them logged? - since valid path checkers won't go through that function. (and it occurred to me that the state-change check of patch 3 can be done there, simpler.) cheers,
On 07/11/2017 12:32 PM, Mauricio Faria de Oliveira wrote: > Also, it seems the Unavailable/Standby states would not be logged > without a recheck from alua_check_sense(), since the only callers > of alua_rtpg_queue() are alua_activate() and alua_check[_sense]() Well, actually it does get logged if when activating/switching path groups but shouldn't be the case with only a single path group.
diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c index c01b47e5b55a..a1cf3d6aa853 100644 --- a/drivers/scsi/device_handler/scsi_dh_alua.c +++ b/drivers/scsi/device_handler/scsi_dh_alua.c @@ -431,6 +431,26 @@ static int alua_check_sense(struct scsi_device *sdev, alua_check(sdev, false); return NEEDS_RETRY; } + if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0b) { + /* + * LUN Not Accessible - target port in standby state. + * + * Do not retry, so failover to another target port occur. + * Schedule a recheck to update state for other functions. + */ + alua_check(sdev, true); + return SUCCESS; + } + if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0c) { + /* + * LUN Not Accessible - target port in unavailable state. + * + * Do not retry, so failover to another target port occur. + * Schedule a recheck to update state for other functions. + */ + alua_check(sdev, true); + return SUCCESS; + } break; case UNIT_ATTENTION: if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) { @@ -1057,6 +1077,8 @@ static void alua_check(struct scsi_device *sdev, bool force) * * Fail I/O to all paths not in state * active/optimized or active/non-optimized. + * Allow I/O to paths in state unavailable/standby + * so path checkers can actually check them. */ static int alua_prep_fn(struct scsi_device *sdev, struct request *req) { @@ -1072,6 +1094,9 @@ static int alua_prep_fn(struct scsi_device *sdev, struct request *req) rcu_read_unlock(); if (state == SCSI_ACCESS_STATE_TRANSITIONING) ret = BLKPREP_DEFER; + else if (state == SCSI_ACCESS_STATE_UNAVAILABLE || + state == SCSI_ACCESS_STATE_STANDBY) + req->rq_flags |= RQF_QUIET; else if (state != SCSI_ACCESS_STATE_OPTIMAL && state != SCSI_ACCESS_STATE_ACTIVE && state != SCSI_ACCESS_STATE_LBA) {
According to SPC-4 (5.15.2.4.5 Unavailable state), the unavailable state may (or may not) transition to other states (e.g., microcode downloading or hardware error, which may be temporary or permanent). But, scsi_dh_alua currently fails I/O requests early on once that state occurs (in alua_prep_fn()) preventing path checkers in such function path to actually check if I/O still fails or now works. And that prevents a path activation (alua_activate()) which could update the PG state if it eventually recovered to an active state, thus resume I/O. (This is also the case with the standby state.) This might cause device-mapper multipath to fail all paths to some storage system that moves the controllers to the unavailable state for firmware upgrades, and never recover regardless of the storage system doing upgrades one controller at a time and get them online. Then I/O requests are blocked indefinitely due to queue_if_no_path but the underlying individual paths are fully operational, and can be verified as such through other function paths (e.g., SG_IO): # multipath -l mpatha (360050764008100dac000000000000100) dm-0 IBM,2145 size=40G features='2 queue_if_no_path retain_attached_hw_handler' hwhandler='1 alua' wp=rw |-+- policy='service-time 0' prio=0 status=enabled | |- 1:0:1:0 sdf 8:80 failed undef running | `- 2:0:1:0 sdn 8:208 failed undef running `-+- policy='service-time 0' prio=0 status=enabled |- 1:0:0:0 sdb 8:16 failed undef running `- 2:0:0:0 sdj 8:144 failed undef running # strace -e read \ sg_dd blk_sgio=0 \ if=/dev/sdj of=/dev/null bs=512 count=1 iflag=direct \ 2>&1 | grep 512 read(3, 0x3fff7ba80000, 512) = -1 EIO (Input/output error) # strace -e ioctl \ sg_dd blk_sgio=1 \ if=/dev/sdj of=/dev/null bs=512 count=1 iflag=direct \ 2>&1 | grep 512 ioctl(3, SG_IO, {'S', SG_DXFER_FROM_DEV, cmd[10]=[28, 00, 00, 00, 00, 00, 00, 00, 01, 00], <...>) = 0 So, allow I/O to paths of PGs in unavailable/standby state, so path checkers can actually check them. Also, schedule a recheck when unavailable/standby state is detected (in alua_check_sense()) to update pg->state, and quiet further SCSI error messages (in alua_prep_fn()). Once a path checker eventually detects a working/active state again, the PG state is normally updated on path activation (alua_activate(), as it schedules a recheck), thus I/O requests are no longer failed. Signed-off-by: Mauricio Faria de Oliveira <mauricfo@linux.vnet.ibm.com> Reported-by: Naresh Bannoth <nbannoth@in.ibm.com> --- v2: - also add support for standby state to alua_check_sense(), alua_prep_fn() (Bart Van Assche <Bart.VanAssche@sandisk.com>) drivers/scsi/device_handler/scsi_dh_alua.c | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+)