"qemu-ndb -d" will call ioctl 'NBD_DISCONNECT' first, however, following
message was found:
block nbd0: Send disconnect failed -32
Which indicate that something is wrong with the server. Then,
"qemu-nbd -d" will call ioctl 'NBD_CLEAR_SOCK', however ioctl can't clear
requests after commit 2516ab1543fd("nbd: only clear the queue on device
teardown"). And in the meantime, request can't complete through timeout
because nbd_xmit_timeout() will always return 'BLK_EH_RESET_TIMER', which
means such request will never be completed in this situation.
Now that the flag 'NBD_CMD_INFLIGHT' can make sure requests won't
complete multiple times, switch back to call nbd_clear_sock() in
nbd_clear_sock_ioctl(), so that inflight requests can be cleared.
commit cddce0116058 ("nbd: Aovid double completion of a request")
try to fix that nbd_clear_que() and recv_work() can complete a
request concurrently. However, the problem still exists:
t1 t2 t3
nbd_disconnect_and_put
flush_workqueue recv_work blk_mq_complete_request blk_mq_complete_request_remote -> this is true WRITE_ONCE(rq->state, MQ_RQ_COMPLETE) blk_mq_raise_softirq blk_done_softirq blk_complete_reqs nbd_complete_rq blk_mq_end_request blk_mq_free_request WRITE_ONCE(rq->state, MQ_RQ_IDLE)
nbd_clear_que
blk_mq_tagset_busy_iter
nbd_clear_req __blk_mq_free_request blk_mq_put_tag
blk_mq_complete_request -> complete again
There are three places where request can be completed in nbd:
recv_work(), nbd_clear_que() and nbd_xmit_timeout(). Since they
all hold cmd->lock before completing the request, it's easy to
avoid the problem by setting and checking a cmd flag.
Signed-off-by: Yu Kuai <email address hidden>
Reviewed-by: Ming Lei <email address hidden>
Reviewed-by: Josef Bacik <email address hidden>
Link: https://<email address hidden>
Signed-off-by: Jens Axboe <email address hidden>
(cherry picked from 07175cb1baf4c51051b1fbd391097e349f9a02a9)
Signed-off-by: Matthew Ruffell <email address hidden>
Acked-by: Tim Gardner <email address hidden>
Acked-by: Bartlomiej Zolnierkiewicz <email address hidden>
Signed-off-by: Stefan Bader <email address hidden>
While handling a response message from server, nbd_read_stat() will
try to get request by tag, and then complete the request. However,
this is problematic if nbd haven't sent a corresponding request
message:
PCIE_EDR is enabling support to handle events generated when a PCIE port
disconnects to handle errors. From the comments given in the commit which
adds this option and its help, it sounds like if the OS enables DPC (down-
stream port control) which allows to control PCIE ports in parallel to the
firmware, it should also enable EDR.
Signed-off-by: Michael Reed <email address hidden>
Acked-by: Stefan Bader <email address hidden>
[Added annotation enforcement and bug reference, and adjust annotation
to force arm64 to the same setting as it was before]
Acked-by: Kleber Sacilotto de Souza <email address hidden>
Signed-off-by: Stefan Bader <email address hidden>
PCI: pciehp: Ignore Link Down/Up caused by error-induced Hot Reset
Stuart Hayes reports that an error handled by DPC at a Root Port results
in pciehp gratuitously bringing down a subordinate hotplug port:
RP -- UP -- DP -- UP -- DP (hotplug) -- EP
pciehp brings the slot down because the Link to the Endpoint goes down.
That is caused by a Hot Reset being propagated as a result of DPC.
Per PCIe Base Spec 5.0, section 6.6.1 "Conventional Reset":
For a Switch, the following must cause a hot reset to be sent on all
Downstream Ports: [...]
* The Data Link Layer of the Upstream Port reporting DL_Down status.
In Switches that support Link speeds greater than 5.0 GT/s, the
Upstream Port must direct the LTSSM of each Downstream Port to the
Hot Reset state, but not hold the LTSSMs in that state. This permits
each Downstream Port to begin Link training immediately after its
hot reset completes. This behavior is recommended for all Switches.
* Receiving a hot reset on the Upstream Port.
Once DPC recovers, pcie_do_recovery() walks down the hierarchy and
invokes pcie_portdrv_slot_reset() to restore each port's config space.
At that point, a hotplug interrupt is signaled per PCIe Base Spec r5.0,
section 6.7.3.4 "Software Notification of Hot-Plug Events":
If the Port is enabled for edge-triggered interrupt signaling using
MSI or MSI-X, an interrupt message must be sent every time the logical
AND of the following conditions transitions from FALSE to TRUE: [...]
* The Hot-Plug Interrupt Enable bit in the Slot Control register is
set to 1b.
* At least one hot-plug event status bit in the Slot Status register
and its associated enable bit in the Slot Control register are both
set to 1b.
Prevent pciehp from gratuitously bringing down the slot by clearing the
error-induced Data Link Layer State Changed event before restoring
config space. Afterwards, check whether the link has unexpectedly
failed to retrain and synthesize a DLLSC event if so.
Allow each pcie_port_service_driver (one of them being pciehp) to define
a slot_reset callback and re-use the existing pm_iter() function to
iterate over the callbacks.
Thereby, the Endpoint driver remains bound throughout error recovery and
may restore the device to working state.
Surprise removal during error recovery is detected through a Presence
Detect Changed event. The hotplug port is expected to not signal that
event as a result of a Hot Reset.
The issue isn't DPC-specific, it also occurs when an error is handled by
AER through aer_root_reset(). So while the issue was noticed only now,
it's been around since 2006 when AER support was first introduced.