Fix a deadlock in the UFS driver
Subsystems (9)
Bart Van Assche
reviewerv3 cover Fix a deadlock in the UFS driver
Hi Martin,
This patch series fixes a deadlock in the UFS driver between the suspend/resume code and the SCSI error handler. Please consider this patch series for the next merge window.
Thanks,
Bart.
Changes compared to v2:
* Replaced the custom error handling code in ufshcd_eh_timed_out() with a call to ufshcd_link_recovery().
Changes compared to v1:
* Added support in the SCSI core for failing SCSI commands quickly during host recovery.
* Removed the patch that splits the ufshcd_err_handler() function.
* Fixed the code in ufshcd_eh_timed_out() for handling command timeouts.
* Removed the power management notifier again.
Bart Van Assche (8):
scsi: core: Fix a race between scsi_done() and scsi_timeout()
scsi: core: Change the return type of .eh_timed_out()
scsi: core: Support failing requests while recovering
scsi: ufs: Remove an outdated comment
scsi: ufs: Use 'else' in ufshcd_set_dev_pwr_mode()
scsi: ufs: Try harder to change the power mode
scsi: ufs: Track system suspend / resume activity
scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Documentation/scsi/scsi_eh.rst | 7 ++-- drivers/message/fusion/mptsas.c | 8 ++--- drivers/scsi/libiscsi.c | 26 +++++++------- drivers/scsi/megaraid/megaraid_sas_base.c | 7 ++-- drivers/scsi/mvumi.c | 4 +-- drivers/scsi/qla4xxx/ql4_os.c | 8 ++--- drivers/scsi/scsi_error.c | 41 +++++++++++------------ drivers/scsi/scsi_lib.c | 8 +++-- drivers/scsi/scsi_transport_fc.c | 6 ++-- drivers/scsi/scsi_transport_srp.c | 8 ++--- drivers/scsi/storvsc_drv.c | 4 +-- drivers/scsi/virtio_scsi.c | 4 +-- drivers/ufs/core/ufshcd.c | 37 +++++++++++++++----- include/scsi/libiscsi.h | 2 +- include/scsi/scsi_host.h | 14 +++++++- include/scsi/scsi_transport_fc.h | 2 +- include/scsi/scsi_transport_srp.h | 2 +- include/ufs/ufshcd.h | 5 ++- 18 files changed, 115 insertions(+), 78 deletions(-)
Bart Van Assche
reviewerPATCH v3 1/8 scsi: core: Fix a race between scsi_done() and scsi_timeout()
If there is a race between scsi_done() and scsi_timeout() and if scsi_timeout() loses the race, scsi_timeout() should not reset the request timer. Hence change the return value for this case from BLK_EH_RESET_TIMER into BLK_EH_DONE.
Although the block layer holds a reference on a request (req->ref) while calling a timeout handler, restarting the timer (blk_add_timer()) while a request is being completed is racy.
--- drivers/scsi/scsi_error.c | 14 +++----------- 1 file changed, 3 insertions(+), 11 deletions(-)
drivers/scsi/scsi_error.c +3 -11
@@ -343,19 +343,11 @@ enum blk_eh_timer_return scsi_timeout(struct request *req)if (rtn ==BLK_EH_DONE ) {/* -* Set the command to complete first in order to prevent a real -* completion from releasing the command while error handling -* is using it. If the command was already completed, then the -* lower level driver beat the timeout handler, and it is safe -* to return without escalating error recovery. -* -* If timeout handling lost the race to a real completion, the -* block layer may ignore that due to a fake timeout injection, -* so return RESET_TIMER to allow error handling another shot -* at this command. +* If scsi_done() has already set SCMD_STATE_COMPLETE, do not +* modify *scmd. */ if (test_and_set_bit (SCMD_STATE_COMPLETE , &scmd->state)) -return BLK_EH_RESET_TIMER ; +return BLK_EH_DONE ;if (scsi_abort_command (scmd) !=SUCCESS ) {set_host_byte (scmd,DID_TIME_OUT );scsi_eh_scmd_add (scmd);
Re: [PATCH v3 1/8] scsi: core: Fix a race between scsi_done() and scsi_timeout()
Quoted 29 lines · Bart Van Assche
If there is a race between
scsi_done()andscsi_timeout()and ifscsi_timeout()loses the race,scsi_timeout()should not reset the request timer. Hence change the return value for this case fromBLK_EH_RESET_TIMERintoBLK_EH_DONE.Although the block layer holds a reference on a request (
req->ref) while calling a timeout handler, restarting the timer (blk_add_timer()) while a request is being completed is racy.Cc Keith Busch <[email protected]>Cc Christoph Hellwig <[email protected]>Cc Ming Lei <[email protected]>Cc John Garry <[email protected]>Cc Mike Christie <[email protected]>Cc Hannes Reinecke <[email protected]>Reported-by Adrian Hunter <[email protected]>Fixes 065990bd198e ("scsi: set timed out out mq requests to complete")Signed-off-by Bart Van Assche <[email protected]>--- drivers/scsi/scsi_error.c | 14 +++----------- 1 file changed, 3 insertions(+), 11 deletions(-)drivers/scsi/scsi_error.c +0
-0@@ -343,19 +343,11 @@ enum blk_eh_timer_return scsi_timeout(struct request *req)if (rtn ==BLK_EH_DONE ) {
drivers/scsi/scsi_error.c enum blk_eh_timer_return scsi_timeout(struct request *req) · line 354
Quoted 10 lines
drivers/scsi/scsi_error.c +0
-9/* -* Set the command to complete first in order to prevent a real -* completion from releasing the command while error handling -* is using it. If the command was already completed, then the -* lower level driver beat the timeout handler, and it is safe -* to return without escalating error recovery. -* -* If timeout handling lost the race to a real completion, the -* block layer may ignore that due to a fake timeout injection, -* so return RESET_TIMER to allow error handling another shot
I've been wondering about this code too.
I think the patch is correct for the normal cases, but I didn't understand the old fake timeout comment case. From the comment it seemed like that was the reason we did the RESET_TIMER. Does that not exist anymore or was it just bogus?
The commit you referenced actually was returning BLK_EH_DONE like we want. This commit:
commit f1342709d18af97b0e71449d5696b8873d1a456c
Author: Keith Busch <[email protected]> Date: Mon Nov 26 09:54:29 2018 -0700
scsi: Do not rely on blk-mq for double completions
changed it to BLK_EH_RESET_TIMER and changed the above comment to mention the fake timeout case. However, the commit message mentioned the patch was done because we didn't want scsi digging the block layer.
If the fake injection thingy is bogus, then it seems ok to me.
Quoted 10 lines
drivers/scsi/scsi_error.c +3
-2-* at this command. +* If scsi_done() has already set SCMD_STATE_COMPLETE, do not +* modify *scmd. */ if (test_and_set_bit (SCMD_STATE_COMPLETE , &scmd->state)) -return BLK_EH_RESET_TIMER ; +return BLK_EH_DONE ;if (scsi_abort_command (scmd) !=SUCCESS ) {set_host_byte (scmd,DID_TIME_OUT );scsi_eh_scmd_add (scmd);
Bart Van Assche
reviewerRe: [PATCH v3 1/8] scsi: core: Fix a race between scsi_done() and scsi_timeout()
Quoted 15 lines · Mike Christie
drivers/scsi/scsi_error.c enum blk_eh_timer_return scsi_timeout(struct request *req) · line 354
Quoted 11 lines · Bart Van Assche
if (rtn == BLK_EH_DONE) { /*drivers/scsi/scsi_error.c +0
-9-* Set the command to complete first in order to prevent a real -* completion from releasing the command while error handling -* is using it. If the command was already completed, then the -* lower level driver beat the timeout handler, and it is safe -* to return without escalating error recovery. -* -* If timeout handling lost the race to a real completion, the -* block layer may ignore that due to a fake timeout injection, -* so return RESET_TIMER to allow error handling another shot I've been wondering about this code too.
Quoted 3 lines
I think the patch is correct for the normal cases, but I didn't understand the old fake timeout comment case. From the comment it seemed like that was the reason we did the
RESET_TIMER. Does that not exist anymore or was it just bogus?
Before commit 15f73f5b3e59 ("blk-mq: move failure injection out of blk_mq_complete_request") the scsi_mq_done() function cleared the SCMD_STATE_COMPLETE bit in case of fake timeout injection. I think that commit made the above comment incorrect.
Quoted 7 lines
The commit you referenced actually was returning
BLK_EH_DONElike we want. This commit:commit
f1342709d18af97b0e71449d5696b8873d1a456cAuthor: Keith Busch <[email protected]> Date: Mon Nov 26 09:54:29 2018 -0700
Quoted 8 lines
scsi: Do not rely on blk-mq for double completions
changed it to
BLK_EH_RESET_TIMERand changed the above comment to mention the fake timeout case. However, the commit message mentioned the patch was done because we didn't want scsi digging the block layer.If the fake injection thingy is bogus, then it seems ok to me.
Hmm ... I probably should modify the Fixes tag.
Thanks,
Bart.
Bart Van Assche
reviewerPATCH v3 2/8 scsi: core: Change the return type of .eh_timed_out()
In reply to Bart Van AsscheCommit 6600593cbd93 ("block: rename BLK_EH_NOT_HANDLED to BLK_EH_DONE") made it impossible for .eh_timed_out() implementations to call scsi_done() without causing a crash. Restore support for SCSI timeout handlers to call scsi_done() as follows:
* Change all .eh_timed_out() handlers as follows:
- Change the return type into enum scsi_timeout_action.
- Change BLK_EH_RESET_TIMER into SCSI_EH_RESET_TIMER.
- Change BLK_EH_DONE into SCSI_EH_NOT_HANDLED.
* In scsi_timeout(), convert the SCSI_EH_* values into BLK_EH_* values.
--- Documentation/scsi/scsi_eh.rst | 7 +++-- drivers/message/fusion/mptsas.c | 8 +++--- drivers/scsi/libiscsi.c | 26 +++++++++--------- drivers/scsi/megaraid/megaraid_sas_base.c | 7 +++-- drivers/scsi/mvumi.c | 4 +-- drivers/scsi/qla4xxx/ql4_os.c | 8 +++--- drivers/scsi/scsi_error.c | 33 +++++++++++++---------- drivers/scsi/scsi_transport_fc.c | 6 ++--- drivers/scsi/scsi_transport_srp.c | 8 +++--- drivers/scsi/storvsc_drv.c | 4 +-- drivers/scsi/virtio_scsi.c | 4 +-- include/scsi/libiscsi.h | 2 +- include/scsi/scsi_host.h | 14 +++++++++- include/scsi/scsi_transport_fc.h | 2 +- include/scsi/scsi_transport_srp.h | 2 +- 15 files changed, 77 insertions(+), 58 deletions(-)
Documentation/scsi/scsi_eh.rst +5 -2
@@ -92,14 +92,17 @@ The timeout handler is scsi_timeout(). When a timeout occurs, this function 1. invokes optional hostt->eh_timed_out() callback. Return value can be one of - - BLK_EH_RESET_TIMER + - SCSI_EH_RESET_TIMER This indicates that more time is required to finish the command. Timer is restarted. - - BLK_EH_DONE + - SCSI_EH_NOT_HANDLED eh_timed_out() callback did not handle the command. Step #2 is taken. + - SCSI_EH_DONE + eh_timed_out() completed the command. + 2. scsi_abort_command() is invoked to schedule an asynchronous abort which may issue a retry scmd->allowed + 1 times. Asynchronous aborts are not invoked for commands for which the SCSI_EH_ABORT_SCHEDULED flag is set (this
drivers/message/fusion/mptsas.c +4 -4
@@ -1952,12 +1952,12 @@ mptsas_qcmd(struct Scsi_Host *shost, struct scsi_cmnd *SCpnt)* @sc: scsi command that the midlayer is about to time out * **/ -static enum blk_eh_timer_return mptsas_eh_timed_out (struct scsi_cmnd *sc) +static enum scsi_timeout_action mptsas_eh_timed_out (struct scsi_cmnd *sc) {MPT_SCSI_HOST *hd;MPT_ADAPTER *ioc;VirtDevice *vdevice; -enum blk_eh_timer_return rc =BLK_EH_DONE ; +enum scsi_timeout_action rc =SCSI_EH_NOT_HANDLED ; hd =shost_priv (sc->device->host);if (hd ==NULL ) {
@@ -1980,7 +1980,7 @@ static enum blk_eh_timer_return mptsas_eh_timed_out(struct scsi_cmnd *sc)dtmprintk (ioc,printk (MYIOC_s_WARN_FMT": %s: ioc is in reset," "SML need to reset the timer (sc=%p)\n" , ioc->name, __func__, sc)); - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; } vdevice = sc->device->hostdata; if (vdevice && vdevice->vtarget && (vdevice->vtarget->inDMD
@@ -1988,7 +1988,7 @@ static enum blk_eh_timer_return mptsas_eh_timed_out(struct scsi_cmnd *sc)dtmprintk (ioc,printk (MYIOC_s_WARN_FMT": %s: target removed " "or in device removal delay (sc=%p)\n" , ioc->name, __func__, sc)); - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; }
drivers/scsi/libiscsi.c +13 -13
@@ -2071,9 +2071,9 @@ static int iscsi_has_ping_timed_out(struct iscsi_conn *conn)return 0 ; } -enum blk_eh_timer_return iscsi_eh_cmd_timed_out (struct scsi_cmnd *sc) +enum scsi_timeout_action iscsi_eh_cmd_timed_out (struct scsi_cmnd *sc) { -enum blk_eh_timer_return rc =BLK_EH_DONE ; +enum scsi_timeout_action rc =SCSI_EH_NOT_HANDLED ;struct iscsi_task *task =NULL , *running_task;struct iscsi_cls_session *cls_session;struct iscsi_session *session;
@@ -2093,7 +2093,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)* Raced with completion. Blk layer has taken ownership * so let timeout code complete it now. */ - rc =BLK_EH_DONE ; + rc =SCSI_EH_NOT_HANDLED ;spin_unlock (&session->back_lock); goto done; }
@@ -2102,7 +2102,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)* Racing with the completion path right now, so give it more * time so that path can complete it like normal. */ - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; task =NULL ;spin_unlock (&session->back_lock); goto done;
@@ -2120,21 +2120,21 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)if (unlikely (system_state !=SYSTEM_RUNNING )) { sc->result =DID_NO_CONNECT <<16 ;ISCSI_DBG_EH (session,"sc on shutdown, handled\n" ); - rc =BLK_EH_DONE ; + rc =SCSI_EH_NOT_HANDLED ; goto done; }/* * We are probably in the middle of iscsi recovery so let * that complete and handle the error. */ - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; } conn = session->leadconn;if (!conn) {/* In the middle of shuting down */ - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; }
@@ -2151,7 +2151,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)"Last data xfer at %lu. Last timeout was at " "%lu\n.", task->last_xfer, task->last_timeout); task->have_checked_conn = false; - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; }
@@ -2162,7 +2162,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)* and can let the iscsi eh handle it */ if (iscsi_has_ping_timed_out (conn)) { - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; }
@@ -2200,7 +2200,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc) task->last_xfer, running_task->last_xfer, task->last_timeout);spin_unlock (&session->back_lock); - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; } }
@@ -2216,14 +2216,14 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc)*/ if (READ_ONCE (conn->ping_task)) { task->have_checked_conn = true; - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; goto done; }/* Make sure there is a transport check done */ iscsi_send_nopout (conn,NULL ); task->have_checked_conn = true; - rc =BLK_EH_RESET_TIMER ; + rc =SCSI_EH_RESET_TIMER ; done:spin_unlock_bh (&session->frwd_lock);
@@ -2232,7 +2232,7 @@ enum blk_eh_timer_return iscsi_eh_cmd_timed_out(struct scsi_cmnd *sc) task->last_timeout = jiffies;iscsi_put_task (task); } -ISCSI_DBG_EH (session,"return %s\n" , rc ==BLK_EH_RESET_TIMER ? +ISCSI_DBG_EH (session,"return %s\n" , rc ==SCSI_EH_RESET_TIMER ?"timer reset" :"shutdown or nh" );return rc; }
drivers/scsi/megaraid/megaraid_sas_base.c +3 -4
@@ -2927,15 +2927,14 @@ static int megasas_generic_reset(struct scsi_cmnd *scmd)* Sets the FW busy flag and reduces the host->can_queue if the * cmd has not been completed within the timeout period. */ -static enum -blk_eh_timer_return megasas_reset_timer (struct scsi_cmnd *scmd) +static enum scsi_timeout_action megasas_reset_timer (struct scsi_cmnd *scmd) {struct megasas_instance *instance;unsigned long flags;if (time_after (jiffies, scmd->jiffies_at_alloc + (scmd_timeout *2 ) *HZ )) { -return BLK_EH_DONE ; +return SCSI_EH_NOT_HANDLED ; } instance = (struct megasas_instance *)scmd->device->host->hostdata;
@@ -2949,7 +2948,7 @@ blk_eh_timer_return megasas_reset_timer(struct scsi_cmnd *scmd)spin_unlock_irqrestore (instance->host->host_lock, flags); } -return BLK_EH_RESET_TIMER ; +return SCSI_EH_RESET_TIMER ; }/**
drivers/scsi/mvumi.c +2 -2
@@ -2109,7 +2109,7 @@ static int mvumi_queue_command(struct Scsi_Host *shost,return 0 ; } -static enum blk_eh_timer_return mvumi_timed_out (struct scsi_cmnd *scmd) +static enum scsi_timeout_action mvumi_timed_out (struct scsi_cmnd *scmd) {struct mvumi_cmd *cmd =mvumi_priv (scmd)->cmd_priv;struct Scsi_Host *host = scmd->device->host;
@@ -2137,7 +2137,7 @@ static enum blk_eh_timer_return mvumi_timed_out(struct scsi_cmnd *scmd)mvumi_return_cmd (mhba, cmd);spin_unlock_irqrestore (mhba->shost->host_lock, flags); -return BLK_EH_DONE ; +return SCSI_EH_NOT_HANDLED ; } static int
drivers/scsi/qla4xxx/ql4_os.c +4 -4
@@ -116,7 +116,7 @@ static int qla4xxx_iface_set_param(struct Scsi_Host *shost, void *data,static int qla4xxx_get_iface_param (struct iscsi_iface *iface,enum iscsi_param_type param_type,int param,char *buf); -static enum blk_eh_timer_return qla4xxx_eh_cmd_timed_out (struct scsi_cmnd *sc); +static enum scsi_timeout_action qla4xxx_eh_cmd_timed_out (struct scsi_cmnd *sc);static struct iscsi_endpoint *qla4xxx_ep_connect (struct Scsi_Host *shost,struct sockaddr *dst_addr,int non_blocking);
@@ -1871,17 +1871,17 @@ static void qla4xxx_conn_get_stats(struct iscsi_cls_conn *cls_conn,return ; } -static enum blk_eh_timer_return qla4xxx_eh_cmd_timed_out (struct scsi_cmnd *sc) +static enum scsi_timeout_action qla4xxx_eh_cmd_timed_out (struct scsi_cmnd *sc) {struct iscsi_cls_session *session;unsigned long flags; -enum blk_eh_timer_return ret =BLK_EH_DONE ; +enum scsi_timeout_action ret =SCSI_EH_NOT_HANDLED ; session =starget_to_session (scsi_target (sc->device));spin_lock_irqsave (&session->lock, flags);if (session->state ==ISCSI_SESSION_FAILED ) - ret =BLK_EH_RESET_TIMER ; + ret =SCSI_EH_RESET_TIMER ;spin_unlock_irqrestore (&session->lock, flags);return ret;
drivers/scsi/scsi_error.c +19 -14
@@ -328,7 +328,6 @@ void scsi_eh_scmd_add(struct scsi_cmnd *scmd)enum blk_eh_timer_return scsi_timeout (struct request *req) {struct scsi_cmnd *scmd =blk_mq_rq_to_pdu (req); -enum blk_eh_timer_return rtn =BLK_EH_DONE ;struct Scsi_Host *host = scmd->device->host;trace_scsi_dispatch_cmd_timeout (scmd);
@@ -338,23 +337,29 @@ enum blk_eh_timer_return scsi_timeout(struct request *req)if (host->eh_deadline !=-1 && !host->last_reset) host->last_reset = jiffies; -if (host->hostt->eh_timed_out) - rtn = host->hostt->eh_timed_out (scmd); - -if (rtn ==BLK_EH_DONE ) { -/* -* If scsi_done() has already set SCMD_STATE_COMPLETE, do not -* modify *scmd. -*/ -if (test_and_set_bit (SCMD_STATE_COMPLETE , &scmd->state)) +if (host->hostt->eh_timed_out) { +switch (host->hostt->eh_timed_out (scmd)) { +case SCSI_EH_DONE :return BLK_EH_DONE ; -if (scsi_abort_command (scmd) !=SUCCESS ) { -set_host_byte (scmd,DID_TIME_OUT ); -scsi_eh_scmd_add (scmd); +case SCSI_EH_RESET_TIMER : +return BLK_EH_RESET_TIMER ; +case SCSI_EH_NOT_HANDLED : +break ; } } -return rtn; +/* +* If scsi_done() has already set SCMD_STATE_COMPLETE, do not modify +* *scmd. +*/ +if (test_and_set_bit (SCMD_STATE_COMPLETE , &scmd->state)) +return BLK_EH_DONE ; +if (scsi_abort_command (scmd) !=SUCCESS ) { +set_host_byte (scmd,DID_TIME_OUT ); +scsi_eh_scmd_add (scmd); + } + +return BLK_EH_DONE ; }/**
drivers/scsi/scsi_transport_fc.c +3 -3
@@ -2530,15 +2530,15 @@ static int fc_vport_match(struct attribute_container *cont,* Notes: * This routine assumes no locks are held on entry. */ -enum blk_eh_timer_return +enum scsi_timeout_action fc_eh_timed_out (struct scsi_cmnd *scmd) {struct fc_rport *rport =starget_to_rport (scsi_target (scmd->device));if (rport->port_state ==FC_PORTSTATE_BLOCKED ) -return BLK_EH_RESET_TIMER ; +return SCSI_EH_RESET_TIMER ; -return BLK_EH_DONE ; +return SCSI_EH_NOT_HANDLED ; }EXPORT_SYMBOL (fc_eh_timed_out);
drivers/scsi/scsi_transport_srp.c +4 -4
@@ -594,13 +594,13 @@ EXPORT_SYMBOL(srp_reconnect_rport);* @scmd: SCSI command. * * If a timeout occurs while an rport is in the blocked state, ask the SCSI -* EH to continue waiting (BLK_EH_RESET_TIMER). Otherwise let the SCSI core -* handle the timeout (BLK_EH_DONE). +* EH to continue waiting (SCSI_EH_RESET_TIMER). Otherwise let the SCSI core +* handle the timeout (SCSI_EH_NOT_HANDLED). * * Note: This function is called from soft-IRQ context and with the request * queue lock held. */ -enum blk_eh_timer_return srp_timed_out (struct scsi_cmnd *scmd) +enum scsi_timeout_action srp_timed_out (struct scsi_cmnd *scmd) {struct scsi_device *sdev = scmd->device;struct Scsi_Host *shost = sdev->host;
@@ -611,7 +611,7 @@ enum blk_eh_timer_return srp_timed_out(struct scsi_cmnd *scmd)return rport && rport->fast_io_fail_tmo <0 && rport->dev_loss_tmo <0 && i->f->reset_timer_if_blocked &&scsi_device_blocked (sdev) ? -BLK_EH_RESET_TIMER :BLK_EH_DONE ; +SCSI_EH_RESET_TIMER :SCSI_EH_NOT_HANDLED ; }EXPORT_SYMBOL (srp_timed_out);
drivers/scsi/storvsc_drv.c +2 -2
@@ -1643,13 +1643,13 @@ static int storvsc_host_reset_handler(struct scsi_cmnd *scmnd)* be unbounded on Azure. Reset the timer unconditionally to give the host a * chance to perform EH. */ -static enum blk_eh_timer_return storvsc_eh_timed_out (struct scsi_cmnd *scmnd) +static enum scsi_timeout_action storvsc_eh_timed_out (struct scsi_cmnd *scmnd) {#if IS_ENABLED (CONFIG_SCSI_FC_ATTRS )if (scmnd->device->host->transportt == fc_transport_template)return fc_eh_timed_out (scmnd);#endif -return BLK_EH_RESET_TIMER ; +return SCSI_EH_RESET_TIMER ; }static bool storvsc_scsi_cmd_ok (struct scsi_cmnd *scmnd)
drivers/scsi/virtio_scsi.c +2 -2
@@ -731,9 +731,9 @@ static void virtscsi_commit_rqs(struct Scsi_Host *shost, u16 hwq)* latencies might be higher than on bare metal. Reset the timer * unconditionally to give the host a chance to perform EH. */ -static enum blk_eh_timer_return virtscsi_eh_timed_out (struct scsi_cmnd *scmnd) +static enum scsi_timeout_action virtscsi_eh_timed_out (struct scsi_cmnd *scmnd) { -return BLK_EH_RESET_TIMER ; +return SCSI_EH_RESET_TIMER ; }static struct scsi_host_template virtscsi_host_template = {
include/scsi/libiscsi.h +1 -1
@@ -393,7 +393,7 @@ extern int iscsi_eh_recover_target(struct scsi_cmnd *sc);extern int iscsi_eh_session_reset (struct scsi_cmnd *sc);extern int iscsi_eh_device_reset (struct scsi_cmnd *sc);extern int iscsi_queuecommand (struct Scsi_Host *host,struct scsi_cmnd *sc); -extern enum blk_eh_timer_return iscsi_eh_cmd_timed_out (struct scsi_cmnd *sc); +extern enum scsi_timeout_action iscsi_eh_cmd_timed_out (struct scsi_cmnd *sc);/* * iSCSI host helpers.
include/scsi/scsi_host.h +13 -1
@@ -27,6 +27,18 @@ struct scsi_transport_template;#define MODE_INITIATOR 0x01#define MODE_TARGET 0x02 +/** +* enum scsi_timeout_action - How to handle a command that timed out. +* @SCSI_EH_DONE: The command has already been completed. +* @SCSI_EH_RESET_TIMER: Reset the timer and continue waiting for completion. +* @SCSI_EH_NOT_HANDLED: The command has not yet finished. Abort the command. +*/ +enum scsi_timeout_action { +SCSI_EH_DONE , +SCSI_EH_RESET_TIMER , +SCSI_EH_NOT_HANDLED , +}; +struct scsi_host_template {/* * Put fields referenced in IO submission path together in
@@ -331,7 +343,7 @@ struct scsi_host_template {* * Status: OPTIONAL */ -enum blk_eh_timer_return (*eh_timed_out)(struct scsi_cmnd *); +enum scsi_timeout_action (*eh_timed_out)(struct scsi_cmnd *);/* * Optional routine that allows the transport to decide if a cmd * is retryable. Return true if the transport is in a state the
include/scsi/scsi_transport_fc.h +1 -1
@@ -862,7 +862,7 @@ struct fc_vport *fc_vport_create(struct Scsi_Host *shost, int channel,int fc_vport_terminate (struct fc_vport *vport);int fc_block_rport (struct fc_rport *rport);int fc_block_scsi_eh (struct scsi_cmnd *cmnd); -enum blk_eh_timer_return fc_eh_timed_out (struct scsi_cmnd *scmd); +enum scsi_timeout_action fc_eh_timed_out (struct scsi_cmnd *scmd);bool fc_eh_should_retry_cmd (struct scsi_cmnd *scmd);static inline struct Scsi_Host *fc_bsg_to_shost (struct bsg_job *job)
include/scsi/scsi_transport_srp.h +1 -1
@@ -118,7 +118,7 @@ extern int srp_reconnect_rport(struct srp_rport *rport);extern void srp_start_tl_fail_timers (struct srp_rport *rport);extern void srp_remove_host (struct Scsi_Host *);extern void srp_stop_rport_timers (struct srp_rport *rport); -enum blk_eh_timer_return srp_timed_out (struct scsi_cmnd *scmd); +enum scsi_timeout_action srp_timed_out (struct scsi_cmnd *scmd);/** * srp_chkready() - evaluate the transport layer state before I/O
Bart Van Assche
reviewerPATCH v3 3/8 scsi: core: Support failing requests while recovering
In reply to Bart Van AsscheThe current behavior for SCSI commands submitted while error recovery is ongoing is to retry command submission after error recovery has finished. See also the scsi_host_in_recovery() check in scsi_host_queue_ready(). Add support for failing SCSI commands while host recovery is in progress. This functionality will be used to fix a deadlock in the UFS driver.
--- drivers/scsi/scsi_lib.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/scsi/scsi_lib.c +5 -3
@@ -1337,9 +1337,6 @@ static inline int scsi_host_queue_ready(struct request_queue *q, struct scsi_device *sdev, struct scsi_cmnd *cmd) { - if (scsi_host_in_recovery(shost)) - return 0; -if (atomic_read (&shost->host_blocked) >0 ) {if (scsi_host_busy (shost) >0 ) goto starved;
@@ -1727,6 +1724,11 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, ret =BLK_STS_RESOURCE ;if (!scsi_target_queue_ready (shost, sdev)) goto out_put_budget; +if (unlikely (scsi_host_in_recovery (shost))) { +if (req->cmd_flags &REQ_FAILFAST_MASK ) + ret =BLK_STS_OFFLINE ; + goto out_dec_target_busy; + }if (!scsi_host_queue_ready (q, shost, sdev, cmd)) goto out_dec_target_busy;
Mike Christie
maintainerRe: [PATCH v3 3/8] scsi: core: Support failing requests while recovering
Quoted 33 lines · Bart Van Assche
The current behavior for SCSI commands submitted while error recovery is ongoing is to retry command submission after error recovery has finished. See also the
scsi_host_in_recovery()check inscsi_host_queue_ready(). Add support for failing SCSI commands while host recovery is in progress. This functionality will be used to fix a deadlock in the UFS driver.Cc Christoph Hellwig <[email protected]>Cc Ming Lei <[email protected]>Cc John Garry <[email protected]>Cc Mike Christie <[email protected]>Cc Hannes Reinecke <[email protected]>Signed-off-by Bart Van Assche <[email protected]>--- drivers/scsi/scsi_lib.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/scsi/scsi_lib.c +0
-3@@ -1337,9 +1337,6 @@ static inline int scsi_host_queue_ready(struct request_queue *q, struct scsi_device *sdev, struct scsi_cmnd *cmd) { - if (scsi_host_in_recovery(shost)) - return 0; -if (atomic_read (&shost->host_blocked) >0 ) {if (scsi_host_busy (shost) >0 ) goto starved;@@ -1727,6 +1724,11 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, ret =BLK_STS_RESOURCE ;
drivers/scsi/scsi_lib.c static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, · line 1734
Quoted 10 lines
drivers/scsi/scsi_lib.c +5
-0if (!scsi_target_queue_ready (shost, sdev)) goto out_put_budget; +if (unlikely (scsi_host_in_recovery (shost))) { +if (req->cmd_flags &REQ_FAILFAST_MASK ) + ret =BLK_STS_OFFLINE ; + goto out_dec_target_busy; + }if (!scsi_host_queue_ready (q, shost, sdev, cmd)) goto out_dec_target_busy;
This might add a regression to dm-multipath or it at least makes the behavior difficult to understand for users and support people.
If there is a transport issue and the cmd times out and the abort does as well or fails, then we would start the scsi eh recovery. When the driver/transport class figures out it's a transport issue we will block the scsi_device. So before this patch we would requeue the cmd then we would wait until the fast_io_fail (FC) or replacement_timer (iscsi) to fire before failing commands upwards and failing paths.
With this patch we can end up doing 2 or 3 things depending on timing:
1. If the cmd hits the code above we will fail the command and cause a path failure.
2. If driver/transport blocks the scsi_device first then we would hit the scsi_device state check in scsi_queue_rq and not fail the path like before.
3. With or without your patch, if dm_mq_queue_rq calls the busy callback first (this does blk_lld_busy -> scsi_mq_lld_busy -> scsi_host_in_recovery) then it might see all the paths in recovery. It considers this a temp condition and will requeue cmds. So in this case we will not fail the path.
I'm not sure what the correct fix is. Maybe not fast fail REQ_FAILFAST_TRANSPORT above, or maybe add a new FAILFAST type that you could use?
Bart Van Assche
reviewerRe: [PATCH v3 3/8] scsi: core: Support failing requests while recovering
Quoted 62 lines · Mike Christie
Quoted 33 lines · Bart Van Assche
The current behavior for SCSI commands submitted while error recovery is ongoing is to retry command submission after error recovery has finished. See also the
scsi_host_in_recovery()check inscsi_host_queue_ready(). Add support for failing SCSI commands while host recovery is in progress. This functionality will be used to fix a deadlock in the UFS driver.Cc Christoph Hellwig <[email protected]>Cc Ming Lei <[email protected]>Cc John Garry <[email protected]>Cc Mike Christie <[email protected]>Cc Hannes Reinecke <[email protected]>Signed-off-by Bart Van Assche <[email protected]>--- drivers/scsi/scsi_lib.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/scsi/scsi_lib.c +0
-3@@ -1337,9 +1337,6 @@ static inline int scsi_host_queue_ready(struct request_queue *q, struct scsi_device *sdev, struct scsi_cmnd *cmd) { - if (scsi_host_in_recovery(shost)) - return 0; -if (atomic_read (&shost->host_blocked) >0 ) {if (scsi_host_busy (shost) >0 ) goto starved;@@ -1727,6 +1724,11 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, ret =BLK_STS_RESOURCE ;drivers/scsi/scsi_lib.c static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, · line 1734
Quoted 10 lines
drivers/scsi/scsi_lib.c +5
-0if (!scsi_target_queue_ready (shost, sdev)) goto out_put_budget; +if (unlikely (scsi_host_in_recovery (shost))) { +if (req->cmd_flags &REQ_FAILFAST_MASK ) + ret =BLK_STS_OFFLINE ; + goto out_dec_target_busy; + }if (!scsi_host_queue_ready (q, shost, sdev, cmd)) goto out_dec_target_busy;This might add a regression to dm-multipath or it at least makes the behavior difficult to understand for users and support people.
If there is a transport issue and the cmd times out and the abort does as well or fails, then we would start the scsi eh recovery. When the driver/transport class figures out it's a transport issue we will block the
scsi_device. So before this patch we would requeue the cmd then we would wait until thefast_io_fail(FC) orreplacement_timer(iscsi) to fire before failing commands upwards and failing paths.With this patch we can end up doing 2 or 3 things depending on timing:
1. If the cmd hits the code above we will fail the command and cause a path failure.2. If driver/transport blocks the
scsi_devicefirst then we would hit thescsi_devicestate check inscsi_queue_rqand not fail the path like before.
Quoted 7 lines
3. With or without your patch, if
dm_mq_queue_rqcalls the busy callback first (this doesblk_lld_busy->scsi_mq_lld_busy->scsi_host_in_recovery) then it might see all the paths in recovery. It considers this a temp condition and will requeue cmds. So in this case we will not fail the path.I'm not sure what the correct fix is. Maybe not fast fail
REQ_FAILFAST_TRANSPORTabove, or maybe add a new FAILFAST type that you could use?
Hi Mike,
I will look into introducing a new request flag for this purpose.
Thanks,
Bart.
Bart Van Assche
reviewerPATCH v3 4/8 scsi: ufs: Remove an outdated comment
In reply to Bart Van AsscheAlthough the host lock had to be held by ufshcd_clk_scaling_start_busy() callers when that function was introduced, that is no longer the case today. Hence remove the comment that claims that callers of this function must hold the host lock.
--- drivers/ufs/core/ufshcd.c | 1 - 1 file changed, 1 deletion(-)
drivers/ufs/core/ufshcd.c +0 -1
@@ -2013,7 +2013,6 @@ static void ufshcd_exit_clk_gating(struct ufs_hba *hba)destroy_workqueue (hba->clk_gating.clk_gating_workq); } -/* Must be called with host lock acquired */ static void ufshcd_clk_scaling_start_busy (struct ufs_hba *hba) {bool queue_resume_work = false;
Re: [PATCH v3 4/8] scsi: ufs: Remove an outdated comment
Quoted 8 lines · Bart Van Assche
Although the host lock had to be held by
ufshcd_clk_scaling_start_busy()
callers when that function was introduced, that is no longer the case today. Hence remove the comment that claims that callers of this function
must hold the host lock.Signed-off-by Bart Van Assche <[email protected]>
Re: [PATCH v3 4/8] scsi: ufs: Remove an outdated comment
In reply to Bart Van AsscheQuoted 6 lines · Bart Van Assche
Although the host lock had to be held by
ufshcd_clk_scaling_start_busy()callers when that function was introduced, that is no longer the case today. Hence remove the comment that claims that callers of this function must hold the host lock.Signed-off-by Bart Van Assche <[email protected]>
Quoted 16 lines
--- drivers/ufs/core/ufshcd.c | 1 - 1 file changed, 1 deletion(-)drivers/ufs/core/ufshcd.c +0
-1@@ -2013,7 +2013,6 @@ static void ufshcd_exit_clk_gating(struct ufs_hba *hba)destroy_workqueue (hba->clk_gating.clk_gating_workq); } -/* Must be called with host lock acquired */ static void ufshcd_clk_scaling_start_busy (struct ufs_hba *hba) {bool queue_resume_work = false;
Bart Van Assche
reviewerPATCH v3 5/8 scsi: ufs: Use 'else' in ufshcd_set_dev_pwr_mode()
In reply to Bart Van AsscheConvert if (ret) { ... } if (!ret) { ... } into
if (ret) { ... } else { ... }.
--- drivers/ufs/core/ufshcd.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)
drivers/ufs/core/ufshcd.c +2 -3
@@ -8798,10 +8798,9 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,scsi_print_sense_hdr (sdp,NULL , &sshdr); ret = -EIO ; } - } - -if (!ret) + }else { hba->curr_dev_pwr_mode = pwr_mode; + }scsi_device_put (sdp); hba->host->eh_noresume =0 ;
Re: [PATCH v3 5/8] scsi: ufs: Use 'else' in ufshcd_set_dev_pwr_mode()
Quoted 4 lines · Bart Van Assche
Convert if (ret) { ... } if (!ret) { ... } into
if (ret) { ... } else { ... }.Signed-off-by Bart Van Assche <[email protected]>
Re: [PATCH v3 5/8] scsi: ufs: Use 'else' in ufshcd_set_dev_pwr_mode()
In reply to Bart Van AsscheQuoted 4 lines · Bart Van Assche
Convert if (ret) { ... } if (!ret) { ... } into
if (ret) { ... } else { ... }.Signed-off-by Bart Van Assche <[email protected]>
Quoted 21 lines
--- drivers/ufs/core/ufshcd.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-)drivers/ufs/core/ufshcd.c +2
-3@@ -8798,10 +8798,9 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,scsi_print_sense_hdr (sdp,NULL , &sshdr); ret = -EIO ; } - } - -if (!ret) + }else { hba->curr_dev_pwr_mode = pwr_mode; + }scsi_device_put (sdp); hba->host->eh_noresume =0 ;
Bart Van Assche
reviewerPATCH v3 6/8 scsi: ufs: Try harder to change the power mode
In reply to Bart Van AsscheInstead of only retrying the START STOP UNIT command if a unit attention is reported, repeat it if any SCSI error is reported by the device or if the command timed out.
--- drivers/ufs/core/ufshcd.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/ufs/core/ufshcd.c +3 -3
@@ -8784,9 +8784,9 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,for (retries =3 ; retries >0 ; --retries) { ret =scsi_execute (sdp, cmd,DMA_NONE ,NULL ,0 ,NULL , &sshdr,START_STOP_TIMEOUT ,0 ,0 ,RQF_PM ,NULL ); -if (!scsi_status_is_check_condition (ret) || - !scsi_sense_valid (&sshdr) || - sshdr.sense_key !=UNIT_ATTENTION ) +if (ret <0 ) +break ; +if (host_byte (ret) ==0 &&scsi_status_is_good (ret))break ; }if (ret) {
Re: [PATCH v3 6/8] scsi: ufs: Try harder to change the power mode
Quoted 7 lines · Bart Van Assche
Instead of only retrying the START STOP UNIT command if a unit attention
is reported, repeat it if any SCSI error is reported by the device or if
the command timed out.Signed-off-by Bart Van Assche <[email protected]>
Re: [PATCH v3 6/8] scsi: ufs: Try harder to change the power mode
In reply to Bart Van AsscheQuoted 14 lines · Bart Van Assche
Instead of only retrying the START STOP UNIT command if a unit attention is reported, repeat it if any SCSI error is reported by the device or if the command timed out.
Signed-off-by Bart Van Assche <[email protected]>--- drivers/ufs/core/ufshcd.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)drivers/ufs/core/ufshcd.c +0
-0
drivers/ufs/core/ufshcd.c static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba, · line 8788
Quoted 8 lines
drivers/ufs/core/ufshcd.c +2
-3@@ -8784,9 +8784,9 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,for (retries =3 ; retries >0 ; --retries) { ret =scsi_execute (sdp, cmd,DMA_NONE ,NULL ,0 ,NULL , &sshdr,START_STOP_TIMEOUT ,0 ,0 ,RQF_PM ,NULL ); -if (!scsi_status_is_check_condition (ret) || - !scsi_sense_valid (&sshdr) || - sshdr.sense_key !=UNIT_ATTENTION ) +if (ret <0 ) +break ;
Could use a comment to answer: why not retry if ret < 0 ?
drivers/ufs/core/ufshcd.c static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba, · line 8790
Quoted 2 lines
drivers/ufs/core/ufshcd.c +1
-0+if (host_byte (ret) ==0 &&scsi_status_is_good (ret))break ;
Also could use a comment: why not just "if (!ret)" ?
Quoted 2 lines
drivers/ufs/core/ufshcd.c +0
-0}if (ret) {
Bart Van Assche
reviewerRe: [PATCH v3 6/8] scsi: ufs: Try harder to change the power mode
Quoted 10 lines · Adrian Hunter
Quoted 9 lines · Bart Van Assche
Instead of only retrying the START STOP UNIT command if a unit attention is reported, repeat it if any SCSI error is reported by the device or if the command timed out.
Signed-off-by Bart Van Assche <[email protected]>--- drivers/ufs/core/ufshcd.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
Quoted 15 lines
drivers/ufs/core/ufshcd.c static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba, · line 8788
Quoted 13 lines
drivers/ufs/core/ufshcd.c +2
-3@@ -8784,9 +8784,9 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,for (retries =3 ; retries >0 ; --retries) { ret =scsi_execute (sdp, cmd,DMA_NONE ,NULL ,0 ,NULL , &sshdr,START_STOP_TIMEOUT ,0 ,0 ,RQF_PM ,NULL ); -if (!scsi_status_is_check_condition (ret) || - !scsi_sense_valid (&sshdr) || - sshdr.sense_key !=UNIT_ATTENTION ) +if (ret <0 ) +break ;Could use a comment to answer: why not retry if ret < 0 ?
OK, I will add a comment above that if-test.
Quoted 4 lines
drivers/ufs/core/ufshcd.c static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba, · line 8790
Quoted 2 lines
drivers/ufs/core/ufshcd.c +1
-0+if (host_byte (ret) ==0 &&scsi_status_is_good (ret))break ;Also could use a comment: why not just "if (!ret)" ?
The UFS spec says "CONDITION MET - Not used for UFS" so changing the above test into "if (!ret)" would be valid. If nobody objects I will change the above test into "if (ret == 0)" instead of adding a comment.
Thanks,
Bart.
Bart Van Assche
reviewerPATCH v3 7/8 scsi: ufs: Track system suspend / resume activity
In reply to Bart Van AsscheAdd a new boolean variable that tracks whether the system is suspending, suspended or resuming. This information will be used in a later patch to fix a deadlock between the SCSI error handler and the suspend code.
--- drivers/ufs/core/ufshcd.c | 2 ++ include/ufs/ufshcd.h | 5 ++++- 2 files changed, 6 insertions(+), 1 deletion(-)
drivers/ufs/core/ufshcd.c +2 -0
@@ -9253,6 +9253,7 @@ static int ufshcd_wl_suspend(struct device *dev) hba =shost_priv (sdev->host);down (&hba->host_sem); + hba->system_suspending = true;if (pm_runtime_suspended (dev)) goto out;
@@ -9294,6 +9295,7 @@ static int ufshcd_wl_resume(struct device *dev) hba->curr_dev_pwr_mode, hba->uic_link_state);if (!ret) hba->is_sys_suspended = false; + hba->system_suspending = false;up (&hba->host_sem);return ret; }
include/ufs/ufshcd.h +4 -1
@@ -802,7 +802,9 @@ struct ufs_hba_monitor {* @caps: bitmask with information about UFS controller capabilities * @devfreq: frequency scaling information owned by the devfreq core * @clk_scaling: frequency scaling information owned by the UFS driver -* @is_sys_suspended: whether or not the entire system has been suspended +* @system_suspending: system suspend has been started and system resume has +* not yet finished. +* @is_sys_suspended: UFS device has been suspended because of system suspend * @urgent_bkops_lvl: keeps track of urgent bkops level for device * @is_urgent_bkops_lvl_checked: keeps track if the urgent bkops level for * device is known or not.
@@ -943,6 +945,7 @@ struct ufs_hba {struct devfreq *devfreq;struct ufs_clk_scaling clk_scaling; +bool system_suspending;bool is_sys_suspended;enum bkops_status urgent_bkops_lvl;
Re: [PATCH v3 7/8] scsi: ufs: Track system suspend / resume activity
Quoted 5 lines · Bart Van Assche
Add a new boolean variable that tracks whether the system is suspending, suspended or resuming. This information will be used in a later patch to fix a deadlock between the SCSI error handler and the suspend code.
Signed-off-by Bart Van Assche <[email protected]>
Quoted 48 lines
--- drivers/ufs/core/ufshcd.c | 2 ++ include/ufs/ufshcd.h | 5 ++++- 2 files changed, 6 insertions(+), 1 deletion(-)drivers/ufs/core/ufshcd.c +2
-0@@ -9253,6 +9253,7 @@ static int ufshcd_wl_suspend(struct device *dev) hba =shost_priv (sdev->host);down (&hba->host_sem); + hba->system_suspending = true;if (pm_runtime_suspended (dev)) goto out;@@ -9294,6 +9295,7 @@ static int ufshcd_wl_resume(struct device *dev) hba->curr_dev_pwr_mode, hba->uic_link_state);if (!ret) hba->is_sys_suspended = false; + hba->system_suspending = false;up (&hba->host_sem);return ret; }include/ufs/ufshcd.h +4
-1@@ -802,7 +802,9 @@ struct ufs_hba_monitor {* @caps: bitmask with information about UFS controller capabilities * @devfreq: frequency scaling information owned by the devfreq core * @clk_scaling: frequency scaling information owned by the UFS driver -* @is_sys_suspended: whether or not the entire system has been suspended +* @system_suspending: system suspend has been started and system resume has +* not yet finished. +* @is_sys_suspended: UFS device has been suspended because of system suspend * @urgent_bkops_lvl: keeps track of urgent bkops level for device * @is_urgent_bkops_lvl_checked: keeps track if the urgent bkops level for * device is known or not. @@ -943,6 +945,7 @@ struct ufs_hba {struct devfreq *devfreq;struct ufs_clk_scaling clk_scaling; +bool system_suspending;bool is_sys_suspended;enum bkops_status urgent_bkops_lvl;
Bart Van Assche
reviewerPATCH v3 8/8 scsi: ufs: Fix a deadlock between PM and the SCSI error handler
In reply to Bart Van AsscheThe following deadlock has been observed on multiple test setups:
* ufshcd_wl_suspend() is waiting for blk_execute_rq() to complete while it holds host_sem.
* ufshcd_eh_host_reset_handler() invokes ufshcd_err_handler() and the latter function tries to obtain host_sem.
This is a deadlock because blk_execute_rq() can't execute SCSI commands while the host is in the SHOST_RECOVERY state and because the error handler cannot make progress either.
Fix this deadlock as follows:
* Fail attempts to suspend the system while the SCSI error handler is in progress.
* If the system is suspending and a START STOP UNIT command times out, handle the SCSI command timeout from inside the context of the SCSI timeout handler instead of activating the SCSI error handler.
* Reduce the START STOP UNIT command timeout to one second since on Android devices a kernel panic is triggered if an attempt to suspend the system takes more than 20 seconds. One second should be enough for the START STOP UNIT command since this command completes in less than a millisecond for the UFS devices I have access to.
The runtime power management code is not affected by this deadlock since hba->host_sem is not touched by the runtime power management functions in the UFS driver.
--- drivers/ufs/core/ufshcd.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)
drivers/ufs/core/ufshcd.c +22 -1
@@ -8295,6 +8295,26 @@ static void ufshcd_async_scan(void *data, async_cookie_t cookie) } } +static enum scsi_timeout_action ufshcd_eh_timed_out (struct scsi_cmnd *scmd) +{ +struct ufs_hba *hba =shost_priv (scmd->device->host); + +if (!hba->system_suspending) { +/* Activate the error handler in the SCSI core. */ +return SCSI_EH_NOT_HANDLED ; + } + +/* +* Handle errors directly to prevent a deadlock between +* ufshcd_set_dev_pwr_mode() and ufshcd_err_handler(). +*/ +ufshcd_link_recovery (hba); +dev_info (hba->dev,"%s() finished; outstanding_tasks = %#lx.\n" , + __func__, hba->outstanding_tasks); + +return hba->outstanding_tasks ?SCSI_EH_RESET_TIMER :SCSI_EH_DONE ; +} +static const struct attribute_group *ufshcd_driver_groups[] = { &ufs_sysfs_unit_descriptor_group, &ufs_sysfs_lun_attributes_group,
@@ -8329,6 +8349,7 @@ static struct scsi_host_template ufshcd_driver_template = { .eh_abort_handler = ufshcd_abort, .eh_device_reset_handler = ufshcd_eh_device_reset_handler, .eh_host_reset_handler = ufshcd_eh_host_reset_handler, + .eh_timed_out = ufshcd_eh_timed_out, .this_id = -1, .sg_tablesize = SG_ALL, .cmd_per_lun = UFSHCD_CMD_PER_LUN,
@@ -8783,7 +8804,7 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,*/ for (retries =3 ; retries >0 ; --retries) { ret =scsi_execute (sdp, cmd,DMA_NONE ,NULL ,0 ,NULL , &sshdr, -START_STOP_TIMEOUT ,0 ,0 ,RQF_PM ,NULL ); +1 *HZ ,0 ,REQ_FAILFAST_DEV ,RQF_PM ,NULL );if (ret <0 )break ;if (host_byte (ret) ==0 &&scsi_status_is_good (ret))
Re: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Quoted 7 lines · Bart Van Assche
+ ufshcd_link_recovery(hba);
+ dev_info(hba->dev, "%s() finished;outstanding_tasks= %#lx.\n",
+__func__,hba->outstanding_tasks); +
+ returnhba->outstanding_tasks?SCSI_EH_RESET_TIMER:SCSI_EH_DONE;
Bart,
you have reset the device and host, in the case, there are pending TMs, Should be cleared locally, just like ufshcd_err_handler() does?
Kind regards,
Bean
Bart Van Assche
reviewerRe: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Quoted 9 lines · Bean Huo
Quoted 7 lines · Bart Van Assche
+ ufshcd_link_recovery(hba);
+ dev_info(hba->dev, "%s() finished;outstanding_tasks= %#lx.\n",
+__func__,hba->outstanding_tasks); +
+ returnhba->outstanding_tasks?SCSI_EH_RESET_TIMER:SCSI_EH_DONE;
Quoted 4 lines
Bart,
you have reset the device and host, in the case, there are pending TMs, Should be cleared locally, just like
ufshcd_err_handler()does?
Hi Bean,
Do you agree with the following?
* SCSI task management functions are only submitted by the SCSI error handler.
* The ufshcd_link_recovery() call added by this patch can only be invoked during system suspend.
* System suspend only happens after processes and kernel threads have been frozen and after sync() finished. Hence, no I/O should be in progress when ufshcd_wl_suspend() is called and the SCSI error handler should not be running either.
* Hence, no SCSI commands other than START STOP UNIT should be in progress when the ufshcd_link_recovery() call added by this patch is invoked. No TMFs should be in progress either.
Thanks,
Bart.
Re: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Quoted 30 lines · Bart Van Assche
Quoted 8 lines
Quoted 7 lines
+ ufshcd_link_recovery(hba);
+ dev_info(hba->dev, "%s() finished;outstanding_tasks= %#lx.\n",
+__func__,hba->outstanding_tasks); +
+ returnhba->outstanding_tasks?SCSI_EH_RESET_TIMER:SCSI_EH_DONE;Quoted 5 lines
Bart,
you have reset the device and host, in the case, there are pending TMs, Should be cleared locally, just like
ufshcd_err_handler()does?Hi Bean,
Do you agree with the following?
* SCSI task management functions are only submitted by the SCSI error handler.
* Theufshcd_link_recovery()call added by this patch can only be invoked during system suspend.
* System suspend only happens after processes and kernel threads have been frozen and aftersync()finished. Hence, no I/O should be in progress whenufshcd_wl_suspend()is called and the SCSI error handler
should not be running either.
* Hence, no SCSI commands other than START STOP UNIT should be in progress when theufshcd_link_recovery()call added by this patch is invoked. No TMFs should be in progress either.
Quoted 3 lines
Thanks,
Bart.
Hi Bart,
Yes, install the whole series of changes, ufshcd_link_recovery() will be called during the system is in suspending in the case of SSU timeouts. in this case, outstanding_tasks should be 0?
Kind regards,
Bean
Bart Van Assche
reviewerRe: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Quoted 3 lines · Bean Huo
Yes, install the whole series of changes,
ufshcd_link_recovery()will be called during the system is in suspending in the case of SSU timeouts. in this case,outstanding_tasksshould be 0?
Hi Bean,
'outstanding_tasks' needs to be changed into 'outstanding_reqs' in patch 8/8. My understanding is that ufshcd_link_recovery() calls ufshcd_complete_requests() indirectly. That causes the bits corresponding to completed requests to be cleared from outstanding_reqs but does not guarantee that outstanding_reqs will be zero.
Thanks,
Bart.
Re: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
In reply to Bart Van AsscheQuoted 21 lines · Bart Van Assche
The following deadlock has been observed on multiple test setups:
*ufshcd_wl_suspend()is waiting forblk_execute_rq()to complete while it holdshost_sem.
*ufshcd_eh_host_reset_handler()invokesufshcd_err_handler()and the latter function tries to obtainhost_sem.
This is a deadlock becauseblk_execute_rq()can't execute SCSI commands while the host is in theSHOST_RECOVERYstate and because the error handler cannot make progress either.Fix this deadlock as follows:
* Fail attempts to suspend the system while the SCSI error handler is in progress.
* If the system is suspending and a START STOP UNIT command times out, handle the SCSI command timeout from inside the context of the SCSI timeout handler instead of activating the SCSI error handler.
* Reduce the START STOP UNIT command timeout to one second since on Android devices a kernel panic is triggered if an attempt to suspend the system takes more than 20 seconds. One second should be enough for the START STOP UNIT command since this command completes in less than a millisecond for the UFS devices I have access to.
Quoted 5 lines
The runtime power management code is not affected by this deadlock since
hba->host_semis not touched by the runtime power management functions in the UFS driver.Signed-off-by Bart Van Assche <[email protected]>
Comment below. Nevertheless,
Quoted 16 lines
--- drivers/ufs/core/ufshcd.c | 23 ++++++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-)drivers/ufs/core/ufshcd.c +4
-0@@ -8295,6 +8295,26 @@ static void ufshcd_async_scan(void *data, async_cookie_t cookie) } } +static enum scsi_timeout_action ufshcd_eh_timed_out (struct scsi_cmnd *scmd) +{ +struct ufs_hba *hba =shost_priv (scmd->device->host); +
drivers/ufs/core/ufshcd.c static void ufshcd_async_scan(void *data, async_cookie_t cookie) · line 8311
Quoted 10 lines
drivers/ufs/core/ufshcd.c +10
-0+if (!hba->system_suspending) { +/* Activate the error handler in the SCSI core. */ +return SCSI_EH_NOT_HANDLED ; + } + +/* +* Handle errors directly to prevent a deadlock between +* ufshcd_set_dev_pwr_mode() and ufshcd_err_handler(). +*/ +ufshcd_link_recovery (hba);
This is OK for now, but ufshcd_link_recovery() doesn't check for errors *after* reset, and retry error handling - in which case it might need also the error handling that ufshcd_err_handler() does.
It might be better to have only 1 error handling function.
Quoted 26 lines
drivers/ufs/core/ufshcd.c +8
-1+dev_info (hba->dev,"%s() finished; outstanding_tasks = %#lx.\n" , + __func__, hba->outstanding_tasks); + +return hba->outstanding_tasks ?SCSI_EH_RESET_TIMER :SCSI_EH_DONE ; +} +static const struct attribute_group *ufshcd_driver_groups[] = { &ufs_sysfs_unit_descriptor_group, &ufs_sysfs_lun_attributes_group,@@ -8329,6 +8349,7 @@ static struct scsi_host_template ufshcd_driver_template = { .eh_abort_handler = ufshcd_abort, .eh_device_reset_handler = ufshcd_eh_device_reset_handler, .eh_host_reset_handler = ufshcd_eh_host_reset_handler, + .eh_timed_out = ufshcd_eh_timed_out, .this_id = -1, .sg_tablesize = SG_ALL, .cmd_per_lun = UFSHCD_CMD_PER_LUN,@@ -8783,7 +8804,7 @@ static int ufshcd_set_dev_pwr_mode(struct ufs_hba *hba,*/ for (retries =3 ; retries >0 ; --retries) { ret =scsi_execute (sdp, cmd,DMA_NONE ,NULL ,0 ,NULL , &sshdr, -START_STOP_TIMEOUT ,0 ,0 ,RQF_PM ,NULL ); +1 *HZ ,0 ,REQ_FAILFAST_DEV ,RQF_PM ,NULL );if (ret <0 )break ;if (host_byte (ret) ==0 &&scsi_status_is_good (ret))
Bart Van Assche
reviewerRe: [PATCH v3 8/8] scsi: ufs: Fix a deadlock between PM and the SCSI error handler
Quoted 5 lines · Adrian Hunter
This is OK for now, but
ufshcd_link_recovery()doesn't check for errors *after* reset, and retry error handling - in which case it might need also the error handling thatufshcd_err_handler()does.It might be better to have only 1 error handling function.
Hi Adrian,
I'm fine with sharing code between error handling contexts but I'm not sure a single error handling function is the best solution. Having a single error handling function and calling it from the timeout function would require to add so many if-tests into ufshcd_err_handler() that it would make that function unreadable.
Thanks,
Bart.