At the moment we allocate and register the Scsi_Host object corresponding
to a zfcp adapter (FCP device) very early in the life cycle of the adapter
- even before we fully discover and initialize the underlying
firmware/hardware. This had the advantage that we could already use the
Scsi_Host object, and fill in all its information during said discover and
initialize.
Due to commit 737eb78e82 ("block: Delay default elevator initialization")
(first released in v5.4), we noticed a regression that would prevent us
from using any storage volume if zfcp is configured with support for DIF or
DIX (zfcp.dif=1 || zfcp.dix=1). Doing so would result in an illegal memory
access as soon as the first request is sent with such an configuration. As
example for a crash resulting from this:
scsi host0: scsi_eh_0: sleeping
scsi host0: zfcp
qdio: 0.0.1900 ZFCP on SC 4bd using AI:1 QEBSM:0 PRI:1 TDD:1 SIGA: W AP
scsi 0:0:0:0: scsi scan: INQUIRY pass 1 length 36
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:0000000035c7c007 R3:00000001effcc007 S:00000001effd1000 P:000000000000003d
Oops: 0004 ilc:3 [#1] PREEMPT SMP DEBUG_PAGEALLOC
Modules linked in: ...
CPU: 1 PID: 783 Comm: kworker/u760:5 Kdump: loaded Not tainted 5.6.0-rc2-bb-next+ #1
Hardware name: ...
Workqueue: scsi_wq_0 fc_scsi_scan_rport [scsi_transport_fc]
Krnl PSW : 0704e00180000000 000003ff801fcdae (scsi_queue_rq+0x436/0x740 [scsi_mod])
R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:2 PM:0 RI:0 EA:3
Krnl GPRS: 0fffffffffffffff 0000000000000000 0000000187150120 0000000000000000
000003ff80223d20 000000000000018e 000000018adc6400 0000000187711000
000003e0062337e8 00000001ae719000 0000000187711000 0000000187150000
00000001ab808100 0000000187150120 000003ff801fcd74 000003e0062336a0
Krnl Code: 000003ff801fcd9e: e310a35c0012 lt %r1,860(%r10)
000003ff801fcda4: a7840010 brc 8,000003ff801fcdc4
#000003ff801fcda8: e310b2900004 lg %r1,656(%r11)
>000003ff801fcdae: d71710001000 xc 0(24,%r1),0(%r1)
000003ff801fcdb4: e310b2900004 lg %r1,656(%r11)
000003ff801fcdba: 41201018 la %r2,24(%r1)
000003ff801fcdbe: e32010000024 stg %r2,0(%r1)
000003ff801fcdc4: b904002b lgr %r2,%r11
Call Trace:
[<000003ff801fcdae>] scsi_queue_rq+0x436/0x740 [scsi_mod]
([<000003ff801fcd74>] scsi_queue_rq+0x3fc/0x740 [scsi_mod])
[<00000000349c9970>] blk_mq_dispatch_rq_list+0x390/0x680
[<00000000349d1596>] blk_mq_sched_dispatch_requests+0x196/0x1a8
[<00000000349c7a04>] __blk_mq_run_hw_queue+0x144/0x160
[<00000000349c7ab6>] __blk_mq_delay_run_hw_queue+0x96/0x228
[<00000000349c7d5a>] blk_mq_run_hw_queue+0xd2/0xe0
[<00000000349d194a>] blk_mq_sched_insert_request+0x192/0x1d8
[<00000000349c17b8>] blk_execute_rq_nowait+0x80/0x90
[<00000000349c1856>] blk_execute_rq+0x6e/0xb0
[<000003ff801f8ac2>] __scsi_execute+0xe2/0x1f0 [scsi_mod]
[<000003ff801fef98>] scsi_probe_and_add_lun+0x358/0x840 [scsi_mod]
[<000003ff8020001c>] __scsi_scan_target+0xc4/0x228 [scsi_mod]
[<000003ff80200254>] scsi_scan_target+0xd4/0x100 [scsi_mod]
[<000003ff802d8b96>] fc_scsi_scan_rport+0x96/0xc0 [scsi_transport_fc]
[<0000000034245ce8>] process_one_work+0x458/0x7d0
[<00000000342462a2>] worker_thread+0x242/0x448
[<0000000034250994>] kthread+0x15c/0x170
[<0000000034e1979c>] ret_from_fork+0x30/0x38
INFO: lockdep is turned off.
Last Breaking-Event-Address:
[<000003ff801fbc36>] scsi_add_cmd_to_list+0x9e/0xa8 [scsi_mod]
Kernel panic - not syncing: Fatal exception: panic_on_oops
While this issue is exposed by the commit named above, this is only by
accident. The real issue exists for longer already - basically since it's
possible to use blk-mq via scsi-mq, and blk-mq pre-allocates all requests
for a tag-set during initialization of the same. For a given Scsi_Host
object this is done when adding the object to the midlayer
(`scsi_add_host()` and such). In `scsi_mq_setup_tags()` the midlayer
calculates how much memory is required for a single scsi_cmnd, and its
additional data, which also might include space for additional protection
data - depending on whether the Scsi_Host has any form of protection
capabilities (`scsi_host_get_prot()`).
The problem is now thus, because zfcp does this step before we actually
know whether the firmware/hardware has these capabilities, we don't set any
protection capabilities in the Scsi_Host object. And so, no space is
allocated for additional protection data for requests in the Scsi_Host
tag-set.
Once we go through discover and initialize the FCP device firmware/hardware
fully (this is done via the firmware commands "Exchange Config Data" and
"Exchange Port Data") we find out whether it actually supports DIF and DIX,
and we set the corresponding capabilities in the Scsi_Host object (in
`zfcp_scsi_set_prot()`). Now the Scsi_Host potentially has protection
capabilities, but the already allocated requests in the tag-set don't have
any space allocated for that.
When we then trigger target scanning or add scsi_devices manually, the
midlayer will use requests from that tag-set, and before sending most
requests, it will also call `scsi_mq_prep_fn()`. To prepare the scsi_cmnd
this function will check again whether the used Scsi_Host has any
protection capabilities - and now it potentially has - and if so, it will
try to initialize the assumed to be preallocated structures and thus it
causes the crash, like shown above.
Before delaying the default elevator initialization with the commit named
above, we always would also allocate an elevator for any scsi_device before
ever sending any requests - in contrast to now, where we do it after
device-probing. That elevator in turn would have its own tag-set, and that
is initialized after we went through discovery and initialization of the
underlying firmware/hardware. So requests from that tag-set can be
allocated properly, and if used - unless the user changes/disabled the
default elevator - this would hide the underlying issue.
To fix this for any configuration - with or without an elevator - we move
the allocation and registration of the Scsi_Host object for a given FCP
device to after the first complete discovery and initialization of the
underlying firmware/hardware. By doing that we can make all basic
properties of the Scsi_Host known to the midlayer by the time we call
`scsi_add_host()`, including whether we have any protection capabilities.
To do that we have to delay all the accesses that we would have done in the
past during discovery and initialization, and do them instead once we are
finished with it. The previous patches ramp up to this by fencing and
factoring out all these accesses, and make it possible to re-do them later
on. In addition we make also use of the diagnostic buffers we recently
added with
commit 92953c6e0a ("scsi: zfcp: signal incomplete or error for sync exchange config/port data")
commit 7e418833e6 ("scsi: zfcp: diagnostics buffer caching and use for exchange port data")
commit 088210233e ("scsi: zfcp: add diagnostics buffer for exchange config data")
(first released in v5.5), because these already cache all the information
we need for that "re-do operation" - the information cached are always
updated during xconf or xport data, so it won't be stale.
In addition to the move and re-do, this patch also updates the
function-documentation of `zfcp_scsi_adapter_register()` and changes how it
reports if a Scsi_Host object already exists. In that case future
recovery-operations can skip this step completely and behave much like they
would do in the past - zfcp does not release a once allocated Scsi_Host
object unless the corresponding FCP device is deconstructed completely.
Link: https://lore.kernel.org/r/030dd6da318bbb529f0b5268ec65cebcd20fc0a3.1588956679.git.bblock@linux.ibm.com
Reviewed-by: Steffen Maier <maier@linux.ibm.com>
Signed-off-by: Benjamin Block <bblock@linux.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
d0dff2ac98