| Message ID | 20220118165803.3667947-1-maier@linux.ibm.com (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
| Series | zfcp: fix failed recovery on gone remote port with non-NPIV FCP devices | expand |
Steffen, > Suppose we have an environment with a number of non-NPIV FCP devices > (virtual HBAs / FCP devices / zfcp "adapter"s) sharing the same > physical FCP channel (HBA port) and its I_T nexus. [...] Applied to 5.17/scsi-staging, thanks!
On Tue, 18 Jan 2022 17:58:03 +0100, Steffen Maier wrote: > Suppose we have an environment with a number of non-NPIV FCP devices > (virtual HBAs / FCP devices / zfcp "adapter"s) sharing the same physical > FCP channel (HBA port) and its I_T nexus. Plus a number of storage target > ports zoned to such shared channel. Now one target port logs out of the > fabric causing an RSCN. Zfcp reacts with an ADISC ELS and subsequent port > recovery depending on the ADISC result. This happens on all such FCP > devices (in different Linux images) concurrently as they all receive a copy > of this RSCN. In the following we look at one of those FCP devices. > > [...] Applied to 5.17/scsi-fixes, thanks! [1/1] zfcp: fix failed recovery on gone remote port with non-NPIV FCP devices https://git.kernel.org/mkp/scsi/c/8c9db6679be4
diff --git a/drivers/s390/scsi/zfcp_fc.c b/drivers/s390/scsi/zfcp_fc.c index d24cafe02708..511bf8e0a436 100644 --- a/drivers/s390/scsi/zfcp_fc.c +++ b/drivers/s390/scsi/zfcp_fc.c @@ -521,6 +521,8 @@ static void zfcp_fc_adisc_handler(void *data) goto out; } + /* re-init to undo drop from zfcp_fc_adisc() */ + port->d_id = ntoh24(adisc_resp->adisc_port_id); /* port is good, unblock rport without going through erp */ zfcp_scsi_schedule_rport_register(port); out: @@ -534,6 +536,7 @@ static int zfcp_fc_adisc(struct zfcp_port *port) struct zfcp_fc_req *fc_req; struct zfcp_adapter *adapter = port->adapter; struct Scsi_Host *shost = adapter->scsi_host; + u32 d_id; int ret; fc_req = kmem_cache_zalloc(zfcp_fc_req_cache, GFP_ATOMIC); @@ -558,7 +561,15 @@ static int zfcp_fc_adisc(struct zfcp_port *port) fc_req->u.adisc.req.adisc_cmd = ELS_ADISC; hton24(fc_req->u.adisc.req.adisc_port_id, fc_host_port_id(shost)); - ret = zfcp_fsf_send_els(adapter, port->d_id, &fc_req->ct_els, + d_id = port->d_id; /* remember as destination for send els below */ + /* + * Force fresh GID_PN lookup on next port recovery. + * Must happen after request setup and before sending request, + * to prevent race with port->d_id re-init in zfcp_fc_adisc_handler(). + */ + port->d_id = 0; + + ret = zfcp_fsf_send_els(adapter, d_id, &fc_req->ct_els, ZFCP_FC_CTELS_TMO); if (ret) kmem_cache_free(zfcp_fc_req_cache, fc_req);
Suppose we have an environment with a number of non-NPIV FCP devices (virtual HBAs / FCP devices / zfcp "adapter"s) sharing the same physical FCP channel (HBA port) and its I_T nexus. Plus a number of storage target ports zoned to such shared channel. Now one target port logs out of the fabric causing an RSCN. Zfcp reacts with an ADISC ELS and subsequent port recovery depending on the ADISC result. This happens on all such FCP devices (in different Linux images) concurrently as they all receive a copy of this RSCN. In the following we look at one of those FCP devices. Requests other than FSF_QTCB_FCP_CMND can be slow until they get a response. Depending on which requests are affected by slow responses, there are different recovery outcomes. Here we want to fix failed recoveries on port or adapter level by avoiding recovery requests that can be slow. We need the cached N_Port_ID for the remote port "link" test with ADISC. Just before sending the ADISC, we now intentionally forget the old cached N_Port_ID. The idea is that on receiving an RSCN for a port, we have to assume that any cached information about this port is stale. This forces a fresh new GID_PN [FC-GS] nameserver lookup on any subsequent recovery for the same port. Since we typically can still communicate with the nameserver efficiently, we now reach steady state quicker: Either the nameserver still does not know about the port so we stop recovery, or the nameserver already knows the port potentially with a new N_Port_ID and we can successfully and quickly perform open port recovery. For the one case, where ADISC returns successfully, we re-initialize port->d_id because that case does not involve any port recovery. This also solves a problem if the storage WWPN quickly logs into the fabric again but with a different N_Port_ID. Such as on virtual WWPN takeover during target NPIV failover. [https://www.redbooks.ibm.com/abstracts/redp5477.html] In that case the RSCN from the storage FDISC was ignored by zfcp and we could not successfully recover the failover. On some later failback on the storage, we could have been lucky if the virtual WWPN got the same old N_Port_ID from the SAN switch as we still had cached. Then the related RSCN triggered a successful port reopen recovery. However, there is no guarantee to get the same N_Port_ID on NPIV FDISC. Even though NPIV-enabled FCP devices are not affected by this problem, this code change optimizes recovery time for gone remote ports as a side effect. The timely drop of cached N_Port_IDs prevents unnecessary slow open port attempts. While the problem might have been in code before v2.6.32 commit 799b76d09aee ("[SCSI] zfcp: Decouple gid_pn requests from erp") this fix depends on the gid_pn_work introduced with that commit, so we mark it as culprit to satisfy fix dependencies. Note: Point-to-point remote port is already handled separately and gets its N_Port_ID from the cached peer_d_id. So resetting port->d_id in general does not affect PtP. Suggested-by: Benjamin Block <bblock@linux.ibm.com> Fixes: 799b76d09aee ("[SCSI] zfcp: Decouple gid_pn requests from erp") Cc: <stable@vger.kernel.org> #2.6.32+ Reviewed-by: Benjamin Block <bblock@linux.ibm.com> Signed-off-by: Steffen Maier <maier@linux.ibm.com> --- drivers/s390/scsi/zfcp_fc.c | 13 ++++++++++++- 1 file changed, 12 insertions(+), 1 deletion(-) base-commit: 2576e153cd982d540b212e989458edc42ad1b390