Fix bugs in qla2xxx driver
Subsystems (3)
cover Fix bugs in qla2xxx driver
This series of patches contains 3 separate changes that fix some bugs in the qla2xxx driver.
Anastasia Kovaleva (3):
scsi: qla2xxx: Drop starvation counter on success
scsi: qla2xxx: make target send correct LOGO
scsi: qla2xxx: Remove incorrect trap
drivers/scsi/qla2xxx/qla_iocb.c | 11 +++++++++++ drivers/scsi/qla2xxx/qla_isr.c | 4 ++++ drivers/scsi/qla2xxx/qla_target.c | 21 ++++++++++++--------- 3 files changed, 27 insertions(+), 9 deletions(-)
Signature · 1 line
2.40.1
PATCH 1/3 scsi: qla2xxx: Drop starvation counter on success
Long-lived sessions under high load can accumulate a starvation counter, and the current implementation does not allow this counter to be reset during an active session.
If HBA sends correct ATIO IOCB, then it has enough resources to process commands and we should not call ISP recovery.
--- drivers/scsi/qla2xxx/qla_isr.c | 4 ++++ drivers/scsi/qla2xxx/qla_target.c | 11 +++++++++++ 2 files changed, 15 insertions(+)
drivers/scsi/qla2xxx/qla_isr.c +4 -0
@@ -1959,6 +1959,10 @@ qla2x00_async_event(scsi_qla_host_t *vha, struct rsp_que *rsp, uint16_t *mb)ql_dbg (ql_dbg_async, vha,0x5091 ,"Transceiver Removal\n" );break ; +case MBA_REJECTED_FCP_CMD : +ql_dbg (ql_dbg_async, vha,0x5092 ,"LS_RJT was sent. No resources to process the ELS request.\n" ); +break ; +default : ql_dbg(ql_dbg_async, vha, 0x5057, "Unknown AEN:%04x %04x %04x %04x\n",
drivers/scsi/qla2xxx/qla_target.c +11 -0
@@ -6801,6 +6801,7 @@ qlt_24xx_process_atio_queue(struct scsi_qla_host *vha, uint8_t ha_locked)struct qla_hw_data *ha = vha->hw;struct atio_from_isp *pkt;int cnt, i; +unsigned long flags =0 ;if (!ha->flags.fw_started)return ;
@@ -6826,6 +6827,16 @@ qlt_24xx_process_atio_queue(struct scsi_qla_host *vha, uint8_t ha_locked)qlt_send_term_exchange (ha->base_qpair,NULL , pkt, ha_locked,0 ); }else { +/* +* If we get correct ATIO, then HBA had enough memory +* to proceed without reset. +*/ +if (!ha_locked) +spin_lock_irqsave (&ha->hardware_lock, flags); + vha->hw->exch_starvation =0 ; +if (!ha_locked) +spin_unlock_irqrestore (&ha->hardware_lock, flags); +qlt_24xx_atio_pkt_all_vps (vha, (struct atio_from_isp *)pkt, ha_locked); }
Signature · 1 line
2.40.1
Re: [PATCH 1/3] scsi: qla2xxx: Drop starvation counter on success
Quoted 50 lines · Anastasia Kovaleva
Long-lived sessions under high load can accumulate a starvation counter, and the current implementation does not allow this counter to be reset during an active session.
If HBA sends correct ATIO IOCB, then it has enough resources to process commands and we should not call ISP recovery.
Fixes ead038556f64 ("qla2xxx: Add Dual mode support in the driver")Signed-off-by Anastasia Kovaleva <[email protected]>Reviewed-by Dmitry Bogdanov <[email protected]>--- drivers/scsi/qla2xxx/qla_isr.c | 4 ++++ drivers/scsi/qla2xxx/qla_target.c | 11 +++++++++++ 2 files changed, 15 insertions(+)drivers/scsi/qla2xxx/qla_isr.c +4
-0@@ -1959,6 +1959,10 @@ qla2x00_async_event(scsi_qla_host_t *vha, struct rsp_que *rsp, uint16_t *mb)ql_dbg (ql_dbg_async, vha,0x5091 ,"Transceiver Removal\n" );break ; +case MBA_REJECTED_FCP_CMD : +ql_dbg (ql_dbg_async, vha,0x5092 ,"LS_RJT was sent. No resources to process the ELS request.\n" ); +break ; +default : ql_dbg(ql_dbg_async, vha, 0x5057, "Unknown AEN:%04x %04x %04x %04x\n",drivers/scsi/qla2xxx/qla_target.c +4
-0@@ -6801,6 +6801,7 @@ qlt_24xx_process_atio_queue(struct scsi_qla_host *vha, uint8_t ha_locked)struct qla_hw_data *ha = vha->hw;struct atio_from_isp *pkt;int cnt, i; +unsigned long flags =0 ;if (!ha->flags.fw_started)return ;@@ -6826,6 +6827,16 @@ qlt_24xx_process_atio_queue(struct scsi_qla_host *vha, uint8_t ha_locked)qlt_send_term_exchange (ha->base_qpair,NULL , pkt, ha_locked,0 ); }else { +/* +* If we get correct ATIO, then HBA had enough memory +* to proceed without reset.
drivers/scsi/qla2xxx/qla_target.c qlt_24xx_process_atio_queue(struct scsi_qla_host *vha, uint8_t ha_locked) · line 6842
Quoted 10 lines
drivers/scsi/qla2xxx/qla_target.c +7
-0+*/ +if (!ha_locked) +spin_lock_irqsave (&ha->hardware_lock, flags); + vha->hw->exch_starvation =0 ; +if (!ha_locked) +spin_unlock_irqrestore (&ha->hardware_lock, flags); +qlt_24xx_atio_pkt_all_vps (vha, (struct atio_from_isp *)pkt, ha_locked); }
Why not just 'WRITE_ONCE()' and drop the spinlock?
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH 2/3 scsi: qla2xxx: make target send correct LOGO
In reply to Anastasia KovalevaUpon removing the ACL from the target, it sends a LOGO command to the initiator to break the connection. But HBA fills port_name and port_id of the LOGO command with all zeroes, which is not valid. The initiator sends a reject for this command, but it is not being processed on the target, since it assumes LOGO can never fail. This leaves a system in a state where the initiator thinks it is still logged in to the target and can send commands to it, but the target ignores all incoming commands from this initiator.
If, in such a situation, the initiator sends some command (e.g. during a scan), after not receiving a response for a timeout duration, it sends ABORT for the command. After a timeout on receiving an ABORT response, the initiator sends LOGO to the target. Only after that, the initiator can successfully relogin to the target and restore the connection. In the end, this whole situation hangs the system for approximately a minute.
By default, the driver sends a LOGO command to HBA filling only port_id, expecting HBA to match port_id with the correct port_name from it's internal table. HBA doesn't do that, instead filling these fields with all zeroes.
This patch makes the driver send a LOGO command to HBA with port_name and port_id in the I/O PARMETER fields. HBA then copies these values to corresponding fields in the LOGO command frame.
--- drivers/scsi/qla2xxx/qla_iocb.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
drivers/scsi/qla2xxx/qla_iocb.c +11 -0
@@ -2486,6 +2486,17 @@ qla24xx_logout_iocb(srb_t *sp, struct logio_entry_24xx *logio) logio->port_id[1 ] = sp->fcport->d_id.b.area; logio->port_id[2 ] = sp->fcport->d_id.b.domain; logio->vp_index = sp->vha->vp_idx; + logio->io_parameter[0 ] =cpu_to_le32 (sp->vha->d_id.b.al_pa | + sp->vha->d_id.b.area <<8 | + sp->vha->d_id.b.domain <<16 ); + logio->io_parameter[1 ] =cpu_to_le32 (sp->vha->port_name[3 ] | + sp->vha->port_name[2 ] <<8 | + sp->vha->port_name[1 ] <<16 | + sp->vha->port_name[0 ] <<24 ); + logio->io_parameter[2 ] =cpu_to_le32 (sp->vha->port_name[7 ] | + sp->vha->port_name[6 ] <<8 | + sp->vha->port_name[5 ] <<16 | + sp->vha->port_name[4 ] <<24 ); } static void
Signature · 1 line
2.40.1
Re: [PATCH 2/3] scsi: qla2xxx: make target send correct LOGO
Quoted 44 lines · Anastasia Kovaleva
Upon removing the ACL from the target, it sends a LOGO command to the initiator to break the connection. But HBA fills
port_nameandport_idof the LOGO command with all zeroes, which is not valid. The initiator sends a reject for this command, but it is not being processed on the target, since it assumes LOGO can never fail. This leaves a system in a state where the initiator thinks it is still logged in to the target and can send commands to it, but the target ignores all incoming commands from this initiator.If, in such a situation, the initiator sends some command (e.g. during a scan), after not receiving a response for a timeout duration, it sends ABORT for the command. After a timeout on receiving an ABORT response, the initiator sends LOGO to the target. Only after that, the initiator can successfully relogin to the target and restore the connection. In the end, this whole situation hangs the system for approximately a minute.
By default, the driver sends a LOGO command to HBA filling only
port_id, expecting HBA to matchport_idwith the correctport_namefrom it's internal table. HBA doesn't do that, instead filling these fields with all zeroes.This patch makes the driver send a LOGO command to HBA with
port_nameandport_idin the I/O PARMETER fields. HBA then copies these values to corresponding fields in the LOGO command frame.Signed-off-by Anastasia Kovaleva <[email protected]>Reviewed-by Dmitry Bogdanov <[email protected]>--- drivers/scsi/qla2xxx/qla_iocb.c | 11 +++++++++++ 1 file changed, 11 insertions(+)drivers/scsi/qla2xxx/qla_iocb.c +4
-0@@ -2486,6 +2486,17 @@ qla24xx_logout_iocb(srb_t *sp, struct logio_entry_24xx *logio) logio->port_id[1 ] = sp->fcport->d_id.b.area; logio->port_id[2 ] = sp->fcport->d_id.b.domain; logio->vp_index = sp->vha->vp_idx; + logio->io_parameter[0 ] =cpu_to_le32 (sp->vha->d_id.b.al_pa | + sp->vha->d_id.b.area <<8 | + sp->vha->d_id.b.domain <<16 ); + logio->io_parameter[1] = cpu_to_le32(sp->vha->port_name[3] |
drivers/scsi/qla2xxx/qla_iocb.c qla24xx_logout_iocb(srb_t *sp, struct logio_entry_24xx *logio) · line 2502
Quoted 10 lines
drivers/scsi/qla2xxx/qla_iocb.c +7
-0+ sp->vha->port_name[2 ] <<8 | + sp->vha->port_name[1 ] <<16 | + sp->vha->port_name[0] << 24); + logio->io_parameter[2 ] =cpu_to_le32 (sp->vha->port_name[7 ] | + sp->vha->port_name[6 ] <<8 | + sp->vha->port_name[5 ] <<16 | + sp->vha->port_name[4 ] <<24 ); } static void
Now that looks like serious debugging. Well done.
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
PATCH 3/3 scsi: qla2xxx: Remove incorrect trap
In reply to Anastasia KovalevaThis BUG_ON() is triggered when there is no fc_port with a certain loop ID in the scsi host vp_fcports list, but there is one in lport_loopid_map. As these two data structures do not change simultaneously and atomically, such a trap is invalid.
--- drivers/scsi/qla2xxx/qla_target.c | 10 +--------- 1 file changed, 1 insertion(+), 9 deletions(-)
drivers/scsi/qla2xxx/qla_target.c +1 -9
@@ -5190,15 +5190,7 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,ql_dbg (ql_dbg_disc, vha,0x20fc ,"%s: logo %llx res %d sess %p " , __func__, wwn, res, sess); -if (res ==0 ) { -/* -* cmd went upper layer, look for qlt_xmit_tm_rsp() -* for LOGO_ACK & sess delete -*/ -BUG_ON (!sess); - res =0 ; - }else { -/* cmd did not go to upper layer. */ +if (res) {if (sess) {qlt_schedule_sess_for_deletion (sess); res =0 ;
Signature · 1 line
2.40.1
Re: [PATCH 3/3] scsi: qla2xxx: Remove incorrect trap
Quoted 24 lines · Anastasia Kovaleva
This
BUG_ON()is triggered when there is nofc_portwith a certain loop ID in the scsi hostvp_fcportslist, but there is one inlport_loopid_map. As these two data structures do not change simultaneously and atomically, such a trap is invalid.Fixes 726b85487067 ("qla2xxx: Add framework for async fabric discovery")Signed-off-by Anastasia Kovaleva <[email protected]>Reviewed-by Dmitry Bogdanov <[email protected]>--- drivers/scsi/qla2xxx/qla_target.c | 10 +--------- 1 file changed, 1 insertion(+), 9 deletions(-)drivers/scsi/qla2xxx/qla_target.c +0
-3@@ -5190,15 +5190,7 @@ static int qlt_24xx_handle_els(struct scsi_qla_host *vha,ql_dbg (ql_dbg_disc, vha,0x20fc ,"%s: logo %llx res %d sess %p " , __func__, wwn, res, sess); -if (res ==0 ) { -/* -* cmd went upper layer, look for qlt_xmit_tm_rsp()
drivers/scsi/qla2xxx/qla_target.c static int qlt_24xx_handle_els(struct scsi_qla_host *vha, · line 5196
Quoted 10 lines
drivers/scsi/qla2xxx/qla_target.c +1
-6-* for LOGO_ACK & sess delete -*/ -BUG_ON (!sess); - res =0 ; - }else { -/* cmd did not go to upper layer. */ +if (res) {if (sess) {qlt_schedule_sess_for_deletion (sess); res =0 ;
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich