Re: [PATCH 2/2 v3] nbd: fix possible page fault for nbd disk

From: Xiubo Li
Date: Thu Aug 29 2019 - 20:58:51 EST


On 2019/8/30 7:49, Mike Christie wrote:
On 08/22/2019 02:59 AM, xiubli@xxxxxxxxxx wrote:
From: Xiubo Li <xiubli@xxxxxxxxxx>

When the NBD_CFLAG_DESTROY_ON_DISCONNECT flag is set and at the same
time when the socket is closed due to the server daemon is restarted,
just before the last DISCONNET is totally done if we start a new connection
by using the old nbd_index, there will be crashing randomly, like:

<3>[ 110.151949] block nbd1: Receive control failed (result -32)
<1>[ 110.152024] BUG: unable to handle page fault for address: 0000058000000840
<1>[ 110.152063] #PF: supervisor read access in kernel mode
<1>[ 110.152083] #PF: error_code(0x0000) - not-present page
<6>[ 110.152094] PGD 0 P4D 0
<4>[ 110.152106] Oops: 0000 [#1] SMP PTI
<4>[ 110.152120] CPU: 0 PID: 6698 Comm: kworker/u5:1 Kdump: loaded Not tainted 5.3.0-rc4+ #2
<4>[ 110.152136] Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
<4>[ 110.152166] Workqueue: knbd-recv recv_work [nbd]
<4>[ 110.152187] RIP: 0010:__dev_printk+0xd/0x67
<4>[ 110.152206] Code: 10 e8 c5 fd ff ff 48 8b 4c 24 18 65 48 33 0c 25 28 00 [...]
<4>[ 110.152244] RSP: 0018:ffffa41581f13d18 EFLAGS: 00010206
<4>[ 110.152256] RAX: ffffa41581f13d30 RBX: ffff96dd7374e900 RCX: 0000000000000000
<4>[ 110.152271] RDX: ffffa41581f13d20 RSI: 00000580000007f0 RDI: ffffffff970ec24f
<4>[ 110.152285] RBP: ffffa41581f13d80 R08: ffff96dd7fc17908 R09: 0000000000002e56
<4>[ 110.152299] R10: ffffffff970ec24f R11: 0000000000000003 R12: ffff96dd7374e900
<4>[ 110.152313] R13: 0000000000000000 R14: ffff96dd7374e9d8 R15: ffff96dd6e3b02c8
<4>[ 110.152329] FS: 0000000000000000(0000) GS:ffff96dd7fc00000(0000) knlGS:0000000000000000
<4>[ 110.152362] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
<4>[ 110.152383] CR2: 0000058000000840 CR3: 0000000067cc6002 CR4: 00000000001606f0
<4>[ 110.152401] Call Trace:
<4>[ 110.152422] _dev_err+0x6c/0x83
<4>[ 110.152435] nbd_read_stat.cold+0xda/0x578 [nbd]
<4>[ 110.152448] ? __switch_to_asm+0x34/0x70
<4>[ 110.152468] ? __switch_to_asm+0x40/0x70
<4>[ 110.152478] ? __switch_to_asm+0x34/0x70
<4>[ 110.152491] ? __switch_to_asm+0x40/0x70
<4>[ 110.152501] ? __switch_to_asm+0x34/0x70
<4>[ 110.152511] ? __switch_to_asm+0x40/0x70
<4>[ 110.152522] ? __switch_to_asm+0x34/0x70
<4>[ 110.152533] recv_work+0x35/0x9e [nbd]
<4>[ 110.152547] process_one_work+0x19d/0x340
<4>[ 110.152558] worker_thread+0x50/0x3b0
<4>[ 110.152568] kthread+0xfb/0x130
<4>[ 110.152577] ? process_one_work+0x340/0x340
<4>[ 110.152609] ? kthread_park+0x80/0x80
<4>[ 110.152637] ret_from_fork+0x35/0x40

This is very easy to reproduce by running the nbd-runner.

Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
---
drivers/block/nbd.c | 36 +++++++++++++++++++++++++++++++++++-
1 file changed, 35 insertions(+), 1 deletion(-)

diff --git a/drivers/block/nbd.c b/drivers/block/nbd.c
index 8c2f17b99224..a25b59725c6e 100644
--- a/drivers/block/nbd.c
+++ b/drivers/block/nbd.c
@@ -26,6 +26,7 @@
#include <linux/ioctl.h>
#include <linux/mutex.h>
#include <linux/compiler.h>
+#include <linux/completion.h>
#include <linux/err.h>
#include <linux/kernel.h>
#include <linux/slab.h>
@@ -80,6 +81,9 @@ struct link_dead_args {
#define NBD_RT_DESTROY_ON_DISCONNECT 6
#define NBD_RT_DISCONNECT_ON_CLOSE 7
+#define NBD_DESTROY_ON_DISCONNECT 0
+#define NBD_DISCONNECT_REQUESTED 1
+
struct nbd_config {
u32 flags;
unsigned long runtime_flags;
@@ -112,6 +116,9 @@ struct nbd_device {
struct list_head list;
struct task_struct *task_recv;
struct task_struct *task_setup;
+
+ struct completion destroy_complete;
+ unsigned long flags;
};
#define NBD_CMD_REQUEUED 1
@@ -222,6 +229,16 @@ static void nbd_dev_remove(struct nbd_device *nbd)
disk->private_data = NULL;
put_disk(disk);
}
+
+ /*
+ * Place this in the last just before the nbd is freed to
+ * make sure that the disk and the related kobject are also
+ * totally removed to avoid duplicate creation of the same
+ * one.
+ */
+ if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
+ complete(&nbd->destroy_complete);
+
kfree(nbd);
}
@@ -230,8 +247,8 @@ static void nbd_put(struct nbd_device *nbd)
if (refcount_dec_and_mutex_lock(&nbd->refs,
&nbd_index_mutex)) {
idr_remove(&nbd_index_idr, nbd->index);
- mutex_unlock(&nbd_index_mutex);
nbd_dev_remove(nbd);
+ mutex_unlock(&nbd_index_mutex);
}
}
@@ -1103,6 +1120,7 @@ static int nbd_disconnect(struct nbd_device *nbd)
dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
+ set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
send_disconnects(nbd);
return 0;
}
@@ -1596,6 +1614,7 @@ static int nbd_dev_add(int index)
nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
BLK_MQ_F_BLOCKING;
nbd->tag_set.driver_data = nbd;
+ init_completion(&nbd->destroy_complete);
err = blk_mq_alloc_tag_set(&nbd->tag_set);
if (err)
@@ -1761,6 +1780,16 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
mutex_unlock(&nbd_index_mutex);
return -EINVAL;
}
+
+ if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
Why does this have to be set? If this is not set would you end up
hitting the config_refs check:

if (refcount_read(&nbd->config_refs)) {

and possibly returning failure?

Yeah, this is a good question. Before I have also tried to fix it with this, but it still won't work for me.

From my test cases almost more than 50% times, the crash will be hit in the gap just after the nbd->config already been released, and before the nbd itself not yet, so the nbd->config_refs will be 0.



If you moved the complete() to nbd_config_put would it work if this bit
was set or not?

Tried it already, it still won't work.

There is one case that when disconnecting the nbd device, the userspace service will do the open()/release()  things, please see [1], and the sequence is not the same every time, if the NBD_CFLAG_DESTROY_ON_DISCONNECT bit is set the crash still exists.

So sometimes when the nbd_put() called from the nbd_config_put(), the &nbd->refs in nbd_put won't be 0, it could be 1. And it will be 0 just after the release() is triggered later.

So I just place the complete() before "free(nbd);", or there will be another Call trace will be seen very often:

   2489 Aug 20 18:10:04 lxbfd2 kernel: sysfs: cannot create duplicate filename '/devices/virtual/block/nbd0'
   2490 Aug 20 18:10:04 lxbfd2 kernel: CPU: 0 PID: 8635 Comm: nbd-clid Kdump: loaded Tainted: G      D 5.1.18-300.fc30.x86_64 #1
   2491 Aug 20 18:10:04 lxbfd2 kernel: Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
   2492 Aug 20 18:10:04 lxbfd2 kernel: Call Trace:
   2493 Aug 20 18:10:04 lxbfd2 kernel: dump_stack+0x5c/0x80
   2494 Aug 20 18:10:04 lxbfd2 kernel: sysfs_warn_dup.cold+0x17/0x2d
   2495 Aug 20 18:10:04 lxbfd2 kernel: sysfs_create_dir_ns+0xb6/0xd0
   2496 Aug 20 18:10:04 lxbfd2 kernel: kobject_add_internal+0xb7/0x280
   2497 Aug 20 18:10:04 lxbfd2 kernel: kobject_add+0x7e/0xb0
   2498 Aug 20 18:10:04 lxbfd2 kernel: ? _cond_resched+0x15/0x30
   2499 Aug 20 18:10:04 lxbfd2 kernel: device_add+0x12b/0x690
   2500 Aug 20 18:10:04 lxbfd2 kernel: __device_add_disk+0x1b5/0x470
   2501 Aug 20 18:10:04 lxbfd2 kernel: nbd_dev_add+0x21d/0x2b0 [nbd]
   2502 Aug 20 18:10:04 lxbfd2 kernel: nbd_genl_connect+0x16e/0x630 [nbd]
   2503 Aug 20 18:10:04 lxbfd2 kernel: genl_family_rcv_msg+0x1a9/0x3b0
   2504 Aug 20 18:10:04 lxbfd2 kernel: ? __switch_to_asm+0x35/0x70
   2505 Aug 20 18:10:04 lxbfd2 kernel: ? __switch_to_asm+0x41/0x70
   2506 Aug 20 18:10:04 lxbfd2 kernel: ? __switch_to+0x11f/0x4c0
   2507 Aug 20 18:10:04 lxbfd2 kernel: ? __switch_to_asm+0x41/0x70
   [...]


[1] https://lkml.org/lkml/2019/5/29/125



+ test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
+ mutex_unlock(&nbd_index_mutex);
+
+ /* Wait untill the recv_work exit */
If that is all you need we could do a flush_workqueue(nbd->recv_workq)
(you would need Jens's for next branch which has some work queue changes
in nbd).

Yeah, this makes sense.

I think that might be too messy with how we do the puts for the
nbd_device and config and the locking though.


+ wait_for_completion(&nbd->destroy_complete);
The completion is allocated as part of the nbd device struct. Right
after the other thread does a complete() we will free the nbd device
struct, and we could still access the destroy_complete completion in
this thread in wait_for_completion.

You would want to do DECLARE_COMPLETION_ONSTACK in this function, make
destroy_complete a pointer and set the destroy_complete pointer to the
completion declared in this function.

In nbd_dev_remove you would then just check if destroy_complete is set
to a non-NULL pointer.

Hmm, good catch.

Will fix it.

Thanks Mike.
BRs
Xiubo





+ goto again;
+ }
+
if (!refcount_inc_not_zero(&nbd->refs)) {
mutex_unlock(&nbd_index_mutex);
if (index == -1)
@@ -1817,7 +1846,10 @@ static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
&config->runtime_flags);
+ set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
put_dev = true;
+ } else {
+ clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
}
if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
@@ -1987,10 +2019,12 @@ static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
&config->runtime_flags))
put_dev = true;
+ set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
} else {
if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
&config->runtime_flags))
refcount_inc(&nbd->refs);
+ clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
}
if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {