| Message ID | 20220525081139.88165-1-hare@suse.de (mailing list archive) |
|---|---|
| State | Changes Requested |
| Headers | show |
| Series | scsi_dh_alua: mark port group as failed after ALUA transitioning timeout | expand |
On Wed, 2022-05-25 at 10:11 +0200, Hannes Reinecke wrote: > When ALUA transitioning timeout triggers the path group state must > be considered invalid. So add a new flag ALUA_PG_FAILED to indicate > that the path state isn't necessarily valid, and keep the existing > path state until we get a valid response from a RTPG. > > Cc: Brian Bunker <brian@purestorage.com> > Cc: Martin Wilck <mwilck@suse.de> > Signed-off-by: Hannes Reinecke <hare@suse.de> > --- > drivers/scsi/device_handler/scsi_dh_alua.c | 24 +++++++------------- > -- > 1 file changed, 7 insertions(+), 17 deletions(-) > > diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c > b/drivers/scsi/device_handler/scsi_dh_alua.c > index 1d9be771f3ee..6921490a5e65 100644 > --- a/drivers/scsi/device_handler/scsi_dh_alua.c > +++ b/drivers/scsi/device_handler/scsi_dh_alua.c > @@ -49,6 +49,7 @@ > #define ALUA_PG_RUN_RTPG 0x10 > #define ALUA_PG_RUN_STPG 0x20 > #define ALUA_PG_RUNNING 0x40 > +#define ALUA_PG_FAILED 0x80 > > static uint optimize_stpg; > module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR); > @@ -420,7 +421,7 @@ static enum scsi_disposition > alua_check_sense(struct scsi_device *sdev, > */ > rcu_read_lock(); > pg = rcu_dereference(h->pg); > - if (pg) > + if (pg && !(pg->flags & ALUA_PG_FAILED)) > pg->state = > SCSI_ACCESS_STATE_TRANSITIONING; > rcu_read_unlock(); > alua_check(sdev, false); You still return NEEDS_RETRY from alua_check_sense(), even if ALUA_PG_FAILED is set? > @@ -694,7 +695,7 @@ static int alua_rtpg(struct scsi_device *sdev, > struct alua_port_group *pg) > > skip_rtpg: > spin_lock_irqsave(&pg->lock, flags); > - if (transitioning_sense) > + if (transitioning_sense && !(pg->flags & ALUA_PG_FAILED)) > pg->state = SCSI_ACCESS_STATE_TRANSITIONING; > > if (group_id_old != pg->group_id || state_old != pg->state || > @@ -718,23 +719,10 @@ static int alua_rtpg(struct scsi_device *sdev, > struct alua_port_group *pg) > pg->interval = ALUA_RTPG_RETRY_DELAY; > err = SCSI_DH_RETRY; > } else { > - struct alua_dh_data *h; > - > - /* Transitioning time exceeded, set port to > standby */ > + /* Transitioning time exceeded, mark pg as > failed */ > err = SCSI_DH_IO; > - pg->state = SCSI_ACCESS_STATE_STANDBY; > + pg->flags |= ALUA_PG_FAILED; > pg->expiry = 0; > - rcu_read_lock(); > - list_for_each_entry_rcu(h, &pg->dh_list, > node) { > - if (!h->sdev) > - continue; > - h->sdev->access_state = > - (pg->state & > SCSI_ACCESS_STATE_MASK); > - if (pg->pref) > - h->sdev->access_state |= > - > SCSI_ACCESS_STATE_PREFE > RRED; > - } > - rcu_read_unlock(); > } > break; > case SCSI_ACCESS_STATE_OFFLINE: > @@ -746,6 +734,8 @@ static int alua_rtpg(struct scsi_device *sdev, > struct alua_port_group *pg) > /* Useable path if active */ > err = SCSI_DH_OK; > pg->expiry = 0; > + /* RTPG succeeded, clear ALUA_PG_FAILED */ > + pg->flags &= ~ALUA_PG_FAILED; Shouldn't this be done for any state except TRANSITIONING? Btw I've re-read SPC6 and found "The IMPLICIT TRANSITION TIME field indicates the minimum amount of time in seconds the application client should wait prior to timing out an implicit state transition (see 5.18.2.3)". This is unclear to me. "minimum amount of time" suggests that the host could decide to wait longer until it times out the transition. Worse, the spec doesn't clearly define when the transition is supposed to have started. From the host's PoV it makes sense to assume the start time is when it first encounters either TRANSITIONING from an RTPG, or NOT READY / 04 / 0a from a regular I/O, which is what we currently do. But the spec is remarkably unclear about it. Finally, the spec explicitly says that the timeout applies only to implicit transitions, without saying what to do in the case of an explicit transition... Regards, Martin
On 5/25/22 13:20, Martin Wilck wrote: > On Wed, 2022-05-25 at 10:11 +0200, Hannes Reinecke wrote: >> When ALUA transitioning timeout triggers the path group state must >> be considered invalid. So add a new flag ALUA_PG_FAILED to indicate >> that the path state isn't necessarily valid, and keep the existing >> path state until we get a valid response from a RTPG. >> >> Cc: Brian Bunker <brian@purestorage.com> >> Cc: Martin Wilck <mwilck@suse.de> >> Signed-off-by: Hannes Reinecke <hare@suse.de> >> --- >> drivers/scsi/device_handler/scsi_dh_alua.c | 24 +++++++------------- >> -- >> 1 file changed, 7 insertions(+), 17 deletions(-) >> >> diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c >> b/drivers/scsi/device_handler/scsi_dh_alua.c >> index 1d9be771f3ee..6921490a5e65 100644 >> --- a/drivers/scsi/device_handler/scsi_dh_alua.c >> +++ b/drivers/scsi/device_handler/scsi_dh_alua.c >> @@ -49,6 +49,7 @@ >> #define ALUA_PG_RUN_RTPG 0x10 >> #define ALUA_PG_RUN_STPG 0x20 >> #define ALUA_PG_RUNNING 0x40 >> +#define ALUA_PG_FAILED 0x80 >> >> static uint optimize_stpg; >> module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR); >> @@ -420,7 +421,7 @@ static enum scsi_disposition >> alua_check_sense(struct scsi_device *sdev, >> */ >> rcu_read_lock(); >> pg = rcu_dereference(h->pg); >> - if (pg) >> + if (pg && !(pg->flags & ALUA_PG_FAILED)) >> pg->state = >> SCSI_ACCESS_STATE_TRANSITIONING; >> rcu_read_unlock(); >> alua_check(sdev, false); > > You still return NEEDS_RETRY from alua_check_sense(), even if > ALUA_PG_FAILED is set? > >> @@ -694,7 +695,7 @@ static int alua_rtpg(struct scsi_device *sdev, >> struct alua_port_group *pg) >> >> skip_rtpg: >> spin_lock_irqsave(&pg->lock, flags); >> - if (transitioning_sense) >> + if (transitioning_sense && !(pg->flags & ALUA_PG_FAILED)) >> pg->state = SCSI_ACCESS_STATE_TRANSITIONING; >> >> if (group_id_old != pg->group_id || state_old != pg->state || >> @@ -718,23 +719,10 @@ static int alua_rtpg(struct scsi_device *sdev, >> struct alua_port_group *pg) >> pg->interval = ALUA_RTPG_RETRY_DELAY; >> err = SCSI_DH_RETRY; >> } else { >> - struct alua_dh_data *h; >> - >> - /* Transitioning time exceeded, set port to >> standby */ >> + /* Transitioning time exceeded, mark pg as >> failed */ >> err = SCSI_DH_IO; >> - pg->state = SCSI_ACCESS_STATE_STANDBY; >> + pg->flags |= ALUA_PG_FAILED; >> pg->expiry = 0; >> - rcu_read_lock(); >> - list_for_each_entry_rcu(h, &pg->dh_list, >> node) { >> - if (!h->sdev) >> - continue; >> - h->sdev->access_state = >> - (pg->state & >> SCSI_ACCESS_STATE_MASK); >> - if (pg->pref) >> - h->sdev->access_state |= >> - >> SCSI_ACCESS_STATE_PREFE >> RRED; >> - } >> - rcu_read_unlock(); >> } >> break; >> case SCSI_ACCESS_STATE_OFFLINE: >> @@ -746,6 +734,8 @@ static int alua_rtpg(struct scsi_device *sdev, >> struct alua_port_group *pg) >> /* Useable path if active */ >> err = SCSI_DH_OK; >> pg->expiry = 0; >> + /* RTPG succeeded, clear ALUA_PG_FAILED */ >> + pg->flags &= ~ALUA_PG_FAILED; > > Shouldn't this be done for any state except TRANSITIONING? > Why, but it does. We're only entering this block if the state is not TRANSITIONING. (It's part of a 'switch' statement) > Btw I've re-read SPC6 and found "The IMPLICIT TRANSITION TIME field > indicates the minimum amount of time in seconds the application client > should wait prior to timing out an implicit state transition (see > 5.18.2.3)". > > This is unclear to me. "minimum amount of time" suggests that the host > could decide to wait longer until it times out the transition. Worse, > the spec doesn't clearly define when the transition is supposed to have > started. From the host's PoV it makes sense to assume the start time is > when it first encounters either TRANSITIONING from an RTPG, or NOT > READY / 04 / 0a from a regular I/O, which is what we currently do. But > the spec is remarkably unclear about it. Finally, the spec explicitly > says that the timeout applies only to implicit transitions, without > saying what to do in the case of an explicit transition... > Why, but it does; "5.15.2.5 Transitions between target port asymmetric access states" speaks at length what should happen when transitions fail. Problem is, a timeout expiring is something outside the spec, so one can't be sure that the array does what's expected here. But the one key takeaway is that any port in transitioning is allowed to ignore TMFs, so any SCSI EH is pointless anyway. Cheers, Hannes
On Wed, 2022-05-25 at 14:07 +0200, Hannes Reinecke wrote: > On 5/25/22 13:20, Martin Wilck wrote: > > On Wed, 2022-05-25 at 10:11 +0200, Hannes Reinecke wrote: > > > When ALUA transitioning timeout triggers the path group state > > > must > > > be considered invalid. So add a new flag ALUA_PG_FAILED to > > > indicate > > > that the path state isn't necessarily valid, and keep the > > > existing > > > path state until we get a valid response from a RTPG. > > > > > > Cc: Brian Bunker <brian@purestorage.com> > > > Cc: Martin Wilck <mwilck@suse.de> > > > Signed-off-by: Hannes Reinecke <hare@suse.de> > > > --- > > > drivers/scsi/device_handler/scsi_dh_alua.c | 24 +++++++-------- > > > ----- > > > -- > > > 1 file changed, 7 insertions(+), 17 deletions(-) > > > > > > diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c > > > b/drivers/scsi/device_handler/scsi_dh_alua.c > > > index 1d9be771f3ee..6921490a5e65 100644 > > > --- a/drivers/scsi/device_handler/scsi_dh_alua.c > > > +++ b/drivers/scsi/device_handler/scsi_dh_alua.c > > > @@ -49,6 +49,7 @@ > > > #define ALUA_PG_RUN_RTPG 0x10 > > > #define ALUA_PG_RUN_STPG 0x20 > > > #define ALUA_PG_RUNNING 0x40 > > > +#define ALUA_PG_FAILED 0x80 > > > > > > static uint optimize_stpg; > > > module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR); > > > @@ -420,7 +421,7 @@ static enum scsi_disposition > > > alua_check_sense(struct scsi_device *sdev, > > > */ > > > rcu_read_lock(); > > > pg = rcu_dereference(h->pg); > > > - if (pg) > > > + if (pg && !(pg->flags & ALUA_PG_FAILED)) > > > pg->state = > > > SCSI_ACCESS_STATE_TRANSITIONING; > > > rcu_read_unlock(); > > > alua_check(sdev, false); > > > > You still return NEEDS_RETRY from alua_check_sense(), even if > > ALUA_PG_FAILED is set? > > > > > @@ -694,7 +695,7 @@ static int alua_rtpg(struct scsi_device > > > *sdev, > > > struct alua_port_group *pg) > > > > > > skip_rtpg: > > > spin_lock_irqsave(&pg->lock, flags); > > > - if (transitioning_sense) > > > + if (transitioning_sense && !(pg->flags & ALUA_PG_FAILED)) > > > pg->state = SCSI_ACCESS_STATE_TRANSITIONING; > > > > > > if (group_id_old != pg->group_id || state_old != pg- > > > >state || > > > @@ -718,23 +719,10 @@ static int alua_rtpg(struct scsi_device > > > *sdev, > > > struct alua_port_group *pg) > > > pg->interval = ALUA_RTPG_RETRY_DELAY; > > > err = SCSI_DH_RETRY; > > > } else { > > > - struct alua_dh_data *h; > > > - > > > - /* Transitioning time exceeded, set port > > > to > > > standby */ > > > + /* Transitioning time exceeded, mark pg > > > as > > > failed */ > > > err = SCSI_DH_IO; > > > - pg->state = SCSI_ACCESS_STATE_STANDBY; > > > + pg->flags |= ALUA_PG_FAILED; > > > pg->expiry = 0; > > > - rcu_read_lock(); > > > - list_for_each_entry_rcu(h, &pg->dh_list, > > > node) { > > > - if (!h->sdev) > > > - continue; > > > - h->sdev->access_state = > > > - (pg->state & > > > SCSI_ACCESS_STATE_MASK); > > > - if (pg->pref) > > > - h->sdev->access_state |= > > > - > > > SCSI_ACCESS_STATE > > > _PREFE > > > RRED; > > > - } > > > - rcu_read_unlock(); > > > } > > > break; > > > case SCSI_ACCESS_STATE_OFFLINE: > > > @@ -746,6 +734,8 @@ static int alua_rtpg(struct scsi_device > > > *sdev, > > > struct alua_port_group *pg) > > > /* Useable path if active */ > > > err = SCSI_DH_OK; > > > pg->expiry = 0; > > > + /* RTPG succeeded, clear ALUA_PG_FAILED */ > > > + pg->flags &= ~ALUA_PG_FAILED; > > > > Shouldn't this be done for any state except TRANSITIONING? > > > Why, but it does. > We're only entering this block if the state is not TRANSITIONING. > (It's part of a 'switch' statement) Right, the only exception is SCSI_ACCESS_STATE_OFFLINE, for which the ALUA_PG_FAILED state should of cause persist. Sorry for bothering. Martin
On Wed, 2022-05-25 at 21:33 +0200, Martin Wilck wrote: > On Wed, 2022-05-25 at 14:07 +0200, Hannes Reinecke wrote: > > On 5/25/22 13:20, Martin Wilck wrote: > > > On Wed, 2022-05-25 at 10:11 +0200, Hannes Reinecke wrote: > > > > When ALUA transitioning timeout triggers the path group state > > > > must > > > > be considered invalid. So add a new flag ALUA_PG_FAILED to > > > > indicate > > > > that the path state isn't necessarily valid, and keep the > > > > existing > > > > path state until we get a valid response from a RTPG. > > > > > > > > Cc: Brian Bunker <brian@purestorage.com> > > > > Cc: Martin Wilck <mwilck@suse.de> > > > > Signed-off-by: Hannes Reinecke <hare@suse.de> > > > > --- > > > > drivers/scsi/device_handler/scsi_dh_alua.c | 24 +++++++------ > > > > -- > > > > ----- > > > > -- > > > > 1 file changed, 7 insertions(+), 17 deletions(-) > > > > > > > > diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c > > > > b/drivers/scsi/device_handler/scsi_dh_alua.c > > > > index 1d9be771f3ee..6921490a5e65 100644 > > > > --- a/drivers/scsi/device_handler/scsi_dh_alua.c > > > > +++ b/drivers/scsi/device_handler/scsi_dh_alua.c > > > > @@ -49,6 +49,7 @@ > > > > #define ALUA_PG_RUN_RTPG 0x10 > > > > #define ALUA_PG_RUN_STPG 0x20 > > > > #define ALUA_PG_RUNNING 0x40 > > > > +#define ALUA_PG_FAILED 0x80 > > > > > > > > static uint optimize_stpg; > > > > module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR); > > > > @@ -420,7 +421,7 @@ static enum scsi_disposition > > > > alua_check_sense(struct scsi_device *sdev, > > > > */ > > > > rcu_read_lock(); > > > > pg = rcu_dereference(h->pg); > > > > - if (pg) > > > > + if (pg && !(pg->flags & > > > > ALUA_PG_FAILED)) > > > > pg->state = > > > > SCSI_ACCESS_STATE_TRANSITIONING; > > > > rcu_read_unlock(); > > > > alua_check(sdev, false); > > > > > > You still return NEEDS_RETRY from alua_check_sense(), even if > > > ALUA_PG_FAILED is set? > > > > > > > @@ -694,7 +695,7 @@ static int alua_rtpg(struct scsi_device > > > > *sdev, > > > > struct alua_port_group *pg) > > > > > > > > skip_rtpg: > > > > spin_lock_irqsave(&pg->lock, flags); > > > > - if (transitioning_sense) > > > > + if (transitioning_sense && !(pg->flags & > > > > ALUA_PG_FAILED)) > > > > pg->state = SCSI_ACCESS_STATE_TRANSITIONING; > > > > > > > > if (group_id_old != pg->group_id || state_old != pg- > > > > > state || > > > > @@ -718,23 +719,10 @@ static int alua_rtpg(struct scsi_device > > > > *sdev, > > > > struct alua_port_group *pg) > > > > pg->interval = ALUA_RTPG_RETRY_DELAY; > > > > err = SCSI_DH_RETRY; > > > > } else { > > > > - struct alua_dh_data *h; > > > > - > > > > - /* Transitioning time exceeded, set > > > > port > > > > to > > > > standby */ > > > > + /* Transitioning time exceeded, mark pg > > > > as > > > > failed */ > > > > err = SCSI_DH_IO; > > > > - pg->state = SCSI_ACCESS_STATE_STANDBY; > > > > + pg->flags |= ALUA_PG_FAILED; > > > > pg->expiry = 0; > > > > - rcu_read_lock(); > > > > - list_for_each_entry_rcu(h, &pg- > > > > >dh_list, > > > > node) { > > > > - if (!h->sdev) > > > > - continue; > > > > - h->sdev->access_state = > > > > - (pg->state & > > > > SCSI_ACCESS_STATE_MASK); > > > > - if (pg->pref) > > > > - h->sdev->access_state > > > > |= > > > > - > > > > SCSI_ACCESS_STA > > > > TE > > > > _PREFE > > > > RRED; > > > > - } > > > > - rcu_read_unlock(); > > > > } > > > > break; > > > > case SCSI_ACCESS_STATE_OFFLINE: > > > > @@ -746,6 +734,8 @@ static int alua_rtpg(struct scsi_device > > > > *sdev, > > > > struct alua_port_group *pg) > > > > /* Useable path if active */ > > > > err = SCSI_DH_OK; > > > > pg->expiry = 0; > > > > + /* RTPG succeeded, clear ALUA_PG_FAILED */ > > > > + pg->flags &= ~ALUA_PG_FAILED; > > > > > > Shouldn't this be done for any state except TRANSITIONING? > > > > > Why, but it does. > > We're only entering this block if the state is not TRANSITIONING. > > (It's part of a 'switch' statement) > > Right, the only exception is SCSI_ACCESS_STATE_OFFLINE, for which the > ALUA_PG_FAILED state should of cause persist. Sorry for bothering. > What I'm still missing here is how we avoid sending further I/O down the path after the timeout has expired. alua_check_sense() will return NEEDS_RETRY, even if ALUA_PG_FAILED is set. Without this patch, alua_prep_fn() would return an error because it would see STANDBY state. With this patch, the state will stay TRANSITIONING, and (after 6056a92ceb2a ("scsi: scsi_dh_alua: Properly handle the ALUA transitioning state")) alua_prep_fn() will return BLK_STS_OK. IMO alua_check_sense() should return SUCCESS if ALUA_PG_FAILED is set, and alua_prep_fn() should return an error. Regards Martin
diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c b/drivers/scsi/device_handler/scsi_dh_alua.c index 1d9be771f3ee..6921490a5e65 100644 --- a/drivers/scsi/device_handler/scsi_dh_alua.c +++ b/drivers/scsi/device_handler/scsi_dh_alua.c @@ -49,6 +49,7 @@ #define ALUA_PG_RUN_RTPG 0x10 #define ALUA_PG_RUN_STPG 0x20 #define ALUA_PG_RUNNING 0x40 +#define ALUA_PG_FAILED 0x80 static uint optimize_stpg; module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR); @@ -420,7 +421,7 @@ static enum scsi_disposition alua_check_sense(struct scsi_device *sdev, */ rcu_read_lock(); pg = rcu_dereference(h->pg); - if (pg) + if (pg && !(pg->flags & ALUA_PG_FAILED)) pg->state = SCSI_ACCESS_STATE_TRANSITIONING; rcu_read_unlock(); alua_check(sdev, false); @@ -694,7 +695,7 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg) skip_rtpg: spin_lock_irqsave(&pg->lock, flags); - if (transitioning_sense) + if (transitioning_sense && !(pg->flags & ALUA_PG_FAILED)) pg->state = SCSI_ACCESS_STATE_TRANSITIONING; if (group_id_old != pg->group_id || state_old != pg->state || @@ -718,23 +719,10 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg) pg->interval = ALUA_RTPG_RETRY_DELAY; err = SCSI_DH_RETRY; } else { - struct alua_dh_data *h; - - /* Transitioning time exceeded, set port to standby */ + /* Transitioning time exceeded, mark pg as failed */ err = SCSI_DH_IO; - pg->state = SCSI_ACCESS_STATE_STANDBY; + pg->flags |= ALUA_PG_FAILED; pg->expiry = 0; - rcu_read_lock(); - list_for_each_entry_rcu(h, &pg->dh_list, node) { - if (!h->sdev) - continue; - h->sdev->access_state = - (pg->state & SCSI_ACCESS_STATE_MASK); - if (pg->pref) - h->sdev->access_state |= - SCSI_ACCESS_STATE_PREFERRED; - } - rcu_read_unlock(); } break; case SCSI_ACCESS_STATE_OFFLINE: @@ -746,6 +734,8 @@ static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg) /* Useable path if active */ err = SCSI_DH_OK; pg->expiry = 0; + /* RTPG succeeded, clear ALUA_PG_FAILED */ + pg->flags &= ~ALUA_PG_FAILED; break; } spin_unlock_irqrestore(&pg->lock, flags);
When ALUA transitioning timeout triggers the path group state must be considered invalid. So add a new flag ALUA_PG_FAILED to indicate that the path state isn't necessarily valid, and keep the existing path state until we get a valid response from a RTPG. Cc: Brian Bunker <brian@purestorage.com> Cc: Martin Wilck <mwilck@suse.de> Signed-off-by: Hannes Reinecke <hare@suse.de> --- drivers/scsi/device_handler/scsi_dh_alua.c | 24 +++++++--------------- 1 file changed, 7 insertions(+), 17 deletions(-)