| Message ID | 20200623140242.98864-1-maier@linux.ibm.com (mailing list archive) |
|---|---|
| State | Mainlined |
| Commit | 936e6b85da0476dd2edac7c51c68072da9fb4ba2 |
| Headers | show |
| Series | zfcp: fix panic on ERP timeout for previously dismissed ERP action | expand |
On Tue, 23 Jun 2020 16:02:42 +0200, Steffen Maier wrote: > Suppose that, for unrelated reasons, FSF requests on behalf of recovery > are very slow and can run into the ERP timeout. > > In the case at hand, we did adapter recovery to a large degree. > However due to the slowness a LUN open is pending > so the corresponding fc_rport remains blocked. > After fast_io_fail_tmo we trigger close physical port recovery > for the port under which the LUN should have been opened. > The new higher order port recovery > dismisses the pending LUN open ERP action and > dismisses the pending LUN open FSF request. > Such dismissal decouples the ERP action from the pending corresponding > FSF request by setting zfcp_fsf_req->erp_action to NULL > (among other things) [zfcp_erp_strategy_check_fsfreq()]. > > [...] Applied to 5.8/scsi-fixes, thanks! [1/1] scsi: zfcp: Fix panic on ERP timeout for previously dismissed ERP action https://git.kernel.org/mkp/scsi/c/936e6b85da04
diff --git a/drivers/s390/scsi/zfcp_erp.c b/drivers/s390/scsi/zfcp_erp.c index db320dab1fee..79f6e8fb03ca 100644 --- a/drivers/s390/scsi/zfcp_erp.c +++ b/drivers/s390/scsi/zfcp_erp.c @@ -577,7 +577,10 @@ static void zfcp_erp_strategy_check_fsfreq(struct zfcp_erp_action *act) ZFCP_STATUS_ERP_TIMEDOUT)) { req->status |= ZFCP_STATUS_FSFREQ_DISMISSED; zfcp_dbf_rec_run("erscf_1", act); - req->erp_action = NULL; + /* lock-free concurrent access with + * zfcp_erp_timeout_handler() + */ + WRITE_ONCE(req->erp_action, NULL); } if (act->status & ZFCP_STATUS_ERP_TIMEDOUT) zfcp_dbf_rec_run("erscf_2", act); @@ -613,8 +616,14 @@ void zfcp_erp_notify(struct zfcp_erp_action *erp_action, unsigned long set_mask) void zfcp_erp_timeout_handler(struct timer_list *t) { struct zfcp_fsf_req *fsf_req = from_timer(fsf_req, t, timer); - struct zfcp_erp_action *act = fsf_req->erp_action; + struct zfcp_erp_action *act; + if (fsf_req->status & ZFCP_STATUS_FSFREQ_DISMISSED) + return; + /* lock-free concurrent access with zfcp_erp_strategy_check_fsfreq() */ + act = READ_ONCE(fsf_req->erp_action); + if (!act) + return; zfcp_erp_notify(act, ZFCP_STATUS_ERP_TIMEDOUT); }
Suppose that, for unrelated reasons, FSF requests on behalf of recovery are very slow and can run into the ERP timeout. In the case at hand, we did adapter recovery to a large degree. However due to the slowness a LUN open is pending so the corresponding fc_rport remains blocked. After fast_io_fail_tmo we trigger close physical port recovery for the port under which the LUN should have been opened. The new higher order port recovery dismisses the pending LUN open ERP action and dismisses the pending LUN open FSF request. Such dismissal decouples the ERP action from the pending corresponding FSF request by setting zfcp_fsf_req->erp_action to NULL (among other things) [zfcp_erp_strategy_check_fsfreq()]. If now the ERP timeout for the pending open LUN request runs out, we must not use zfcp_fsf_req->erp_action in the ERP timeout handler. This is a problem since v4.15 commit 75492a51568b ("s390/scsi: Convert timers to use timer_setup()"). Before that we intentionally only passed zfcp_erp_action as context argument to zfcp_erp_timeout_handler(). Note: The lifetime of the corresponding zfcp_fsf_req object continues until a (late) response or an (unrelated) adapter recovery. Just like the regular response path ignores dismissed requests [zfcp_fsf_req_complete() => zfcp_fsf_protstatus_eval() => return early] the ERP timeout handler now needs to ignore dismissed requests. So simply return early in the ERP timeout handler if the FSF request is marked as dismissed in its status flags. To protect against the race where zfcp_erp_strategy_check_fsfreq() dismisses and sets zfcp_fsf_req->erp_action to NULL after our previous status flag check, return early if zfcp_fsf_req->erp_action is NULL. After all, the former ERP action does not need to be woken up as that was already done as part of the dismissal above [zfcp_erp_action_dismiss()]. This fixes the following panic due to kernel page fault in IRQ context: Unable to handle kernel pointer dereference in virtual kernel address space Failing address: 0000000000000000 TEID: 0000000000000483 Fault in home space mode while using kernel ASCE. AS:000009859238c00b R2:00000e3e7ffd000b R3:00000e3e7ffcc007 S:00000e3e7ffd7000 P:000000000000013d Oops: 0004 ilc:2 [#1] SMP Modules linked in: ... CPU: 82 PID: 311273 Comm: stress Kdump: loaded Tainted: G E X ... Hardware name: IBM 8561 T01 701 (LPAR) Krnl PSW : 0404c00180000000 001fffff80549be0 (zfcp_erp_notify+0x40/0xc0 [zfcp]) R:0 T:1 IO:0 EX:0 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3 Krnl GPRS: 0000000000000080 00000e3d00000000 00000000000000f0 0000000000030000 000000010028e700 000000000400a39c 000000010028e700 00000e3e7cf87e02 0000000010000000 0700098591cb67f0 0000000000000000 0000000000000000 0000033840e9a000 0000000000000000 001fffe008d6bc18 001fffe008d6bbc8 Krnl Code: 001fffff80549bd4: a7180000 lhi %r1,0 001fffff80549bd8: 4120a0f0 la %r2,240(%r10) #001fffff80549bdc: a53e0003 llilh %r3,3 >001fffff80549be0: ba132000 cs %r1,%r3,0(%r2) 001fffff80549be4: a7740037 brc 7,1fffff80549c52 001fffff80549be8: e320b0180004 lg %r2,24(%r11) 001fffff80549bee: e31020e00004 lg %r1,224(%r2) 001fffff80549bf4: 412020e0 la %r2,224(%r2) Call Trace: [<001fffff80549be0>] zfcp_erp_notify+0x40/0xc0 [zfcp] [<00000985915e26f0>] call_timer_fn+0x38/0x190 [<00000985915e2944>] expire_timers+0xfc/0x190 [<00000985915e2ac4>] run_timer_softirq+0xec/0x218 [<0000098591ca7c4c>] __do_softirq+0x144/0x398 [<00000985915110aa>] do_softirq_own_stack+0x72/0x88 [<0000098591551b58>] irq_exit+0xb0/0xb8 [<0000098591510c6a>] do_IRQ+0x82/0xb0 [<0000098591ca7140>] ext_int_handler+0x128/0x12c [<0000098591722d98>] clear_subpage.constprop.13+0x38/0x60 ([<000009859172ae4c>] clear_huge_page+0xec/0x250) [<000009859177e7a2>] do_huge_pmd_anonymous_page+0x32a/0x768 [<000009859172a712>] __handle_mm_fault+0x88a/0x900 [<000009859172a860>] handle_mm_fault+0xd8/0x1b0 [<0000098591529ef6>] do_dat_exception+0x136/0x3e8 [<0000098591ca6d34>] pgm_check_handler+0x1c8/0x220 Last Breaking-Event-Address: [<001fffff80549c88>] zfcp_erp_timeout_handler+0x10/0x18 [zfcp] Kernel panic - not syncing: Fatal exception in interrupt Signed-off-by: Steffen Maier <maier@linux.ibm.com> Fixes: 75492a51568b ("s390/scsi: Convert timers to use timer_setup()") Cc: <stable@vger.kernel.org> #4.15+ Reviewed-by: Julian Wiedmann <jwi@linux.ibm.com> --- Martin, James, this zfcp fix for a seldom panic would be something for v5.8-rcX and applies to 5.8/scsi-fixes. drivers/s390/scsi/zfcp_erp.c | 13 +++++++++++-- 1 file changed, 11 insertions(+), 2 deletions(-)