A cancel completes a fetched io command and sets io->cmd to NULL, but the io still counts as ready. Only ubq->canceling stops ublk_queue_rq() from using the NULL io->cmd. Two paths clear that flag too early: partial FETCH round recovery, two queues ------------------- -------------------- task A: FETCH tags 0-2 q0 ready: q0->canceling = false task A exits: q0 task exits: start_cancel(): mark queues start_cancel(): ub->canceling tags 0-2: io->cmd = NULL is still set -> q0 not marked task B: FETCH tag 3 q0: io->cmd = NULL queue ready: canceling = false q1 ready START_DEV END_USER_RECOVERY read -> ublk_queue_cmd(NULL) read on q0 -> ublk_queue_cmd(NULL) A FETCH round runs from the open of /dev/ublkcN until ublk_reset_ch_dev() resets the queues. A canceled command can't be fetched again in the same round. So the ready check only needs one fact: did this round see a cancel? ub->canceling already records it: ublk_start_cancel() sets it under cancel_mutex before any command is taken. It is only cleared too early, when all queues are ready. Fix: - clear ub->canceling only in ublk_reset_ch_dev() - when a queue gets ready, clear ubq->canceling only if ub->canceling is not set, and check it under cancel_mutex The check and the marking are ordered by cancel_mutex: check first: queue cleared -> start_cancel() marks it again -> command taken cancel first: check sees ub->canceling -> queue stays canceling While ub->canceling is set, no queue clears its flag. So the "all queues marked" shortcut in ublk_start_cancel() is correct again. UBLK_F_BATCH_IO follows the same rule. It does not dispatch through io->cmd, so it never hit the oops, but a round which lost a fetch command is treated the same way: it has to start over. Fixes: 728cbac5fe21 ("ublk: move device reset into ublk_ch_release()") Fixes: 3f3850785594 ("ublk: fix batch I/O recovery -ENODEV error") Cc: stable@vger.kernel.org # v6.17+: needs cancel_mutex Reported-by: Josef Bacik Closes: https://lore.kernel.org/linux-block/20260928-b4-ublk-cancel-stop-v1-0-4a4360232a46@toxicpanda.com/ Tested-by: Josef Bacik Signed-off-by: Ming Lei --- drivers/block/ublk_drv.c | 54 ++++++++++++++++++++++++---------------- 1 file changed, 33 insertions(+), 21 deletions(-) diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c index b32872421951..b202d97f0dc2 100644 --- a/drivers/block/ublk_drv.c +++ b/drivers/block/ublk_drv.c @@ -334,6 +334,11 @@ struct ublk_device { u16 nr_queue_ready; bool unprivileged_daemons; struct mutex cancel_mutex; + /* + * A cancel started in this FETCH round. Set by ublk_set_canceling(), + * cleared only by ublk_reset_ch_dev() when a new round starts. While + * it is set, no queue clears its ->canceling. + */ bool canceling; pid_t ublksrv_tgid; struct delayed_work exit_work; @@ -2415,6 +2420,11 @@ static void ublk_reset_ch_dev(struct ublk_device *ub) spin_unlock(&ubq->cancel_lock); } + /* a new FETCH round starts, the queues stay canceling until ready */ + mutex_lock(&ub->cancel_mutex); + ub->canceling = false; + mutex_unlock(&ub->cancel_mutex); + /* set to NULL, otherwise new tasks cannot mmap io_cmd_buf */ ub->mm = NULL; ub->nr_queue_ready = 0; @@ -3024,11 +3034,19 @@ static void ublk_reset_io_flags(struct ublk_queue *ubq, struct ublk_io *io) } /* reset per-queue io flags */ -static void ublk_queue_reset_io_flags(struct ublk_queue *ubq) +static void ublk_queue_reset_io_flags(struct ublk_device *ub, + struct ublk_queue *ubq) { - spin_lock(&ubq->cancel_lock); - ubq->canceling = false; - spin_unlock(&ubq->cancel_lock); + /* + * A cancel in this FETCH round took a command which still counts as + * ready, so the queue has to stay canceling. ub->canceling is set + * under cancel_mutex before any command is taken: either we see it + * here, or the cancel marks this queue again later. + */ + mutex_lock(&ub->cancel_mutex); + if (!ub->canceling) + ubq->canceling = false; + mutex_unlock(&ub->cancel_mutex); ubq->fail_io = false; } @@ -3050,24 +3068,17 @@ static void ublk_mark_io_ready(struct ublk_device *ub, u16 q_id, ub->nr_queue_ready++; /* - * Reset queue flags as soon as this queue is ready. - * This clears the canceling flag, allowing batch FETCH commands - * to succeed during recovery without waiting for all queues. + * Reset queue flags as soon as this queue is ready. Unless + * this round saw a cancel, this clears the canceling flag, + * allowing batch FETCH commands to succeed during recovery + * without waiting for all queues. */ - ublk_queue_reset_io_flags(ubq); + ublk_queue_reset_io_flags(ub, ubq); } - /* Check if all queues are ready */ - if (ublk_dev_ready(ub)) { - /* - * All queues ready - clear device-level canceling flag - * and wake ublk_dev_ready() waiters. - */ - mutex_lock(&ub->cancel_mutex); - ub->canceling = false; - mutex_unlock(&ub->cancel_mutex); + /* All queues ready - wake ublk_dev_ready() waiters */ + if (ublk_dev_ready(ub)) wake_up_var(&ub->nr_queue_ready); - } } static inline int ublk_check_cmd_op(u32 cmd_op) @@ -4433,9 +4444,10 @@ static bool ublk_validate_user_pid(struct ublk_device *ub, pid_t ublksrv_pid) /* * Wait until all queues have fetched their I/O commands, and return with - * ub->mutex held and readiness guaranteed: then every queue's ->canceling - * is cleared. Ready may regress between wakeup and mutex_lock() (F_BATCH - * UNPREP, daemon death), so re-check it under the mutex and wait again. + * ub->mutex held and readiness guaranteed. The queues stay canceling if + * this round saw a cancel, see ublk_queue_reset_io_flags(). Ready may + * regress between wakeup and mutex_lock() (F_BATCH UNPREP, daemon death), + * so re-check it under the mutex and wait again. */ static int ublk_wait_dev_ready_and_lock(struct ublk_device *ub) { -- 2.55.0