mmc: block: bypass the queue even if usage is present for hotplug

mmc: block: bypass the queue even if usage is present for hotplug

The commit 304419d8a7e9 ("mmc: core: Allocate per-request data using the
block layer core") refactored mechanism of queue handling caused
mmc_init_request() can be called just after mmc_cleanup_queue() caused null
pointer dereference.

Another commit bbdc74dc19e0 ("mmc: block: Prevent new req entering queue
after its cleanup") tried to fix the problem. However it actually miss one
corner case.

We could still reproduce the issue mentioned with these steps:
(1) insert a SD card and mount it
(2) hotplug it, so it will leave md->usage still be counted
(3) reboot the system which will sync data and umount the card

[Unable to handle kernel NULL pointer dereference at virtual address
00000000
[user pgtable: 4k pages, 48-bit VAs, pgd = ffff80007bab3000
[[0000000000000000] *pgd=000000007a828003, *pud=0000000078dce003,
*pmd=000000007aab6003, *pte=0000000000000000
[Internal error: Oops: 96000007 [#1] PREEMPT SMP
[Modules linked in:
[CPU: 3 PID: 3507 Comm: umount Tainted: G        W
4.13.0-rc1-next-20170720-00012-g9d9bf45 #33
[Hardware name: Firefly-RK3399 Board (DT)
[task: ffff80007a1de200 task.stack: ffff80007a01c000
[PC is at mmc_init_request+0x14/0xc4
[LR is at alloc_request_size+0x4c/0x74
[pc : [<ffff0000087d7150>] lr : [<ffff000008378fe0>] pstate: 600001c5
[sp : ffff80007a01f8f0

....

[[<ffff0000087d7150>] mmc_init_request+0x14/0xc4
[[<ffff000008378fe0>] alloc_request_size+0x4c/0x74
[[<ffff00000817ac28>] mempool_create_node+0xb8/0x17c
[[<ffff00000837aadc>] blk_init_rl+0x9c/0x120
[[<ffff000008396580>] blkg_alloc+0x110/0x234
[[<ffff000008396ac8>] blkg_create+0x424/0x468
[[<ffff00000839877c>] blkg_lookup_create+0xd8/0x14c
[[<ffff0000083796bc>] generic_make_request_checks+0x368/0x3b0
[[<ffff00000837b050>] generic_make_request+0x1c/0x240

So mmc_blk_put wouldn't calling blk_cleanup_queue which actually the
QUEUE_FLAG_DYING and QUEUE_FLAG_BYPASS should stay. Block core expect
blk_queue_bypass_{start, end} internally to bypass/drain the queue before
actually dying the queue, so it didn't expose API to set the queue bypass.
I think we should set QUEUE_FLAG_BYPASS whenever queue is removed, although
the md->usage is still counted, as no dispatch queue could be found then.

Fixes: 304419d8a7e9 ("mmc: core: Allocate per-request data using the block layer core")
Signed-off-by: Shawn Lin <shawn.lin@rock-chips.com>
Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>