| Message ID | 20200121050153.13290-1-ghe@suse.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | ocfs2: fix the oops problem when write cloned file | expand |
Hello Guys, Happy Lunar New Year. If you are free, please help to review this patch. Since this is a regression issue, broke writing cloned files. The root cause is inode buffer head is NULL when calling ocfs2_refcount_cow. Thanks Gang > -----Original Message----- > From: Gang He > Sent: 2020年1月21日 13:02 > To: mark@fasheh.com; jlbec@evilplan.org; joseph.qi@linux.alibaba.com > Cc: Gang He <GHe@suse.com>; linux-kernel@vger.kernel.org; > ocfs2-devel@oss.oracle.com; akpm@linux-foundation.org > Subject: [PATCH] ocfs2: fix the oops problem when write cloned file > > Writing a cloned file triggers a kernel oops and the user-space command > process is also killed by the system. > The bug can be reproduced stably, e.g. > 1)create a file under ocfs2 file system directory. > journalctl -b > aa.txt > 2)create a cloned file for this file. > reflink aa.txt bb.txt > 3)write the cloned file with dd command. > dd if=/dev/zero of=bb.txt bs=512 count=1 conv=notrunc > > The dd command is killed by the kernel, then you can see the oops message > via dmesg command. > > [ 463.875404] BUG: kernel NULL pointer dereference, address: > 0000000000000028 [ 463.875413] #PF: supervisor read access in kernel > mode [ 463.875416] #PF: error_code(0x0000) - not-present page > [ 463.875418] PGD 0 P4D 0 [ 463.875425] Oops: 0000 [#1] SMP PTI > [ 463.875431] CPU: 1 PID: 2291 Comm: dd Tainted: G OE > 5.3.16-2-default > [ 463.875433] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), > BIOS Bochs 01/01/2011 [ 463.875500] RIP: > 0010:ocfs2_refcount_cow+0xa4/0x5d0 [ocfs2] [ 463.875505] Code: 06 89 > 6c 24 38 89 eb f6 44 24 3c 02 74 be 49 8b 47 28 [ 463.875508] RSP: > 0018:ffffa2cb409dfce8 EFLAGS: 00010202 [ 463.875512] RAX: > ffff8b1ebdca8000 RBX: 0000000000000001 RCX: ffff8b1eb73a9df0 > [ 463.875515] RDX: 0000000000056a01 RSI: 0000000000000000 RDI: > 0000000000000000 [ 463.875517] RBP: 0000000000000001 R08: > ffff8b1eb73a9de0 R09: 0000000000000000 [ 463.875520] R10: > 0000000000000001 R11: 0000000000000000 R12: 0000000000000000 > [ 463.875522] R13: ffff8b1eb922f048 R14: 0000000000000000 R15: > ffff8b1eb922f048 [ 463.875526] FS: 00007f8f44d15540(0000) > GS:ffff8b1ebeb00000(0000) knlGS:0000000000000000 [ 463.875529] CS: > 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 463.875532] CR2: > 0000000000000028 CR3: 000000003c17a000 CR4: 00000000000006e0 > [ 463.875546] Call Trace: > [ 463.875596] ? ocfs2_inode_lock_full_nested+0x18b/0x960 [ocfs2] > [ 463.875648] ocfs2_file_write_iter+0xaf8/0xc70 [ocfs2] [ 463.875672] > new_sync_write+0x12d/0x1d0 [ 463.875688] vfs_write+0xad/0x1a0 > [ 463.875697] ksys_write+0xa1/0xe0 [ 463.875710] > do_syscall_64+0x60/0x1f0 [ 463.875743] > entry_SYSCALL_64_after_hwframe+0x49/0xbe > [ 463.875758] RIP: 0033:0x7f8f4482ed44 > [ 463.875762] Code: 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f 80 00 > 00 00 [ 463.875765] RSP: 002b:00007fff300a79d8 EFLAGS: 00000246 > ORIG_RAX: 0000000000000001 [ 463.875769] RAX: ffffffffffffffda RBX: > 0000000000000000 RCX: 00007f8f4482ed44 [ 463.875771] RDX: > 0000000000000200 RSI: 000055f771b5c000 RDI: 0000000000000001 > [ 463.875774] RBP: 0000000000000200 R08: 00007f8f44af9c78 R09: > 0000000000000003 [ 463.875776] R10: 000000000000089f R11: > 0000000000000246 R12: 000055f771b5c000 [ 463.875779] R13: > 0000000000000200 R14: 0000000000000000 R15: 000055f771b5c000 > > This regression problem was introduced by commit e74540b28556 ("ocfs2: > protect extent tree in ocfs2_prepare_inode_for_write()"). > > Fixes: e74540b28556 ("ocfs2: protect extent tree in > ocfs2_prepare_inode_for_write()"). > Signed-off-by: Gang He <ghe@suse.com> > --- > fs/ocfs2/file.c | 14 ++++++-------- > 1 file changed, 6 insertions(+), 8 deletions(-) > > diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c index > 9876db52913a..6cd5e4924e4d 100644 > --- a/fs/ocfs2/file.c > +++ b/fs/ocfs2/file.c > @@ -2101,17 +2101,15 @@ static int ocfs2_is_io_unaligned(struct inode > *inode, size_t count, loff_t pos) static int > ocfs2_inode_lock_for_extent_tree(struct inode *inode, > struct buffer_head **di_bh, > int meta_level, > - int overwrite_io, > int write_sem, > int wait) > { > int ret = 0; > > if (wait) > - ret = ocfs2_inode_lock(inode, NULL, meta_level); > + ret = ocfs2_inode_lock(inode, di_bh, meta_level); > else > - ret = ocfs2_try_inode_lock(inode, > - overwrite_io ? NULL : di_bh, meta_level); > + ret = ocfs2_try_inode_lock(inode, di_bh, meta_level); > if (ret < 0) > goto out; > > @@ -2136,6 +2134,7 @@ static int ocfs2_inode_lock_for_extent_tree(struct > inode *inode, > > out_unlock: > brelse(*di_bh); > + *di_bh = NULL; > ocfs2_inode_unlock(inode, meta_level); > out: > return ret; > @@ -2177,7 +2176,6 @@ static int ocfs2_prepare_inode_for_write(struct file > *file, > ret = ocfs2_inode_lock_for_extent_tree(inode, > &di_bh, > meta_level, > - overwrite_io, > write_sem, > wait); > if (ret < 0) { > @@ -2233,13 +2231,13 @@ static int ocfs2_prepare_inode_for_write(struct > file *file, > &di_bh, > meta_level, > write_sem); > + meta_level = 1; > + write_sem = 1; > ret = ocfs2_inode_lock_for_extent_tree(inode, > &di_bh, > meta_level, > - overwrite_io, > - 1, > + write_sem, > wait); > - write_sem = 1; > if (ret < 0) { > if (ret != -EAGAIN) > mlog_errno(ret); > -- > 2.12.3
Hello Joseph, Changwei, Sunny and all, Could you help to review this patch? This patch will fix the oops problem caused by write ocfs2 clone files. The root cause is inode buffer head is NULL when calling ocfs2_refcount_cow. Secondly, we should use EX meta lock when calling ocfs2_refcount_cow. Thanks a lot. Gang
On 20/2/3 10:17, Gang He wrote: > Hello Joseph, Changwei, Sunny and all, > > Could you help to review this patch? > This patch will fix the oops problem caused by write ocfs2 clone files. > The root cause is inode buffer head is NULL when calling ocfs2_refcount_cow. > Secondly, we should use EX meta lock when calling ocfs2_refcount_cow. > Before commit e74540b28556, we may also use NULL buffer head in case of overwrite, so why there is no such issue before? Thanks, Joseph
Sure. I will try to allocate some time for this change. - Changwei On 2/3/20 10:17 AM, Gang He wrote: > Hello Joseph, Changwei, Sunny and all, > > Could you help to review this patch? > This patch will fix the oops problem caused by write ocfs2 clone files. > The root cause is inode buffer head is NULL when calling ocfs2_refcount_cow. > Secondly, we should use EX meta lock when calling ocfs2_refcount_cow. > > Thanks a lot. > Gang > > ________________________________________ > From: Gang He <GHe@suse.com> > Sent: Tuesday, January 21, 2020 1:02 PM > To: mark@fasheh.com; jlbec@evilplan.org; joseph.qi@linux.alibaba.com > Cc: Gang He; linux-kernel@vger.kernel.org; ocfs2-devel@oss.oracle.com; akpm@linux-foundation.org > Subject: [PATCH] ocfs2: fix the oops problem when write cloned file > > Writing a cloned file triggers a kernel oops and the user-space > command process is also killed by the system. > The bug can be reproduced stably, e.g. > 1)create a file under ocfs2 file system directory. > journalctl -b > aa.txt > 2)create a cloned file for this file. > reflink aa.txt bb.txt > 3)write the cloned file with dd command. > dd if=/dev/zero of=bb.txt bs=512 count=1 conv=notrunc > > The dd command is killed by the kernel, then you can see the oops > message via dmesg command. > > [ 463.875404] BUG: kernel NULL pointer dereference, address: 0000000000000028 > [ 463.875413] #PF: supervisor read access in kernel mode > [ 463.875416] #PF: error_code(0x0000) - not-present page > [ 463.875418] PGD 0 P4D 0 > [ 463.875425] Oops: 0000 [#1] SMP PTI > [ 463.875431] CPU: 1 PID: 2291 Comm: dd Tainted: G OE 5.3.16-2-default > [ 463.875433] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 > [ 463.875500] RIP: 0010:ocfs2_refcount_cow+0xa4/0x5d0 [ocfs2] > [ 463.875505] Code: 06 89 6c 24 38 89 eb f6 44 24 3c 02 74 be 49 8b 47 28 > [ 463.875508] RSP: 0018:ffffa2cb409dfce8 EFLAGS: 00010202 > [ 463.875512] RAX: ffff8b1ebdca8000 RBX: 0000000000000001 RCX: ffff8b1eb73a9df0 > [ 463.875515] RDX: 0000000000056a01 RSI: 0000000000000000 RDI: 0000000000000000 > [ 463.875517] RBP: 0000000000000001 R08: ffff8b1eb73a9de0 R09: 0000000000000000 > [ 463.875520] R10: 0000000000000001 R11: 0000000000000000 R12: 0000000000000000 > [ 463.875522] R13: ffff8b1eb922f048 R14: 0000000000000000 R15: ffff8b1eb922f048 > [ 463.875526] FS: 00007f8f44d15540(0000) GS:ffff8b1ebeb00000(0000) knlGS:0000000000000000 > [ 463.875529] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > [ 463.875532] CR2: 0000000000000028 CR3: 000000003c17a000 CR4: 00000000000006e0 > [ 463.875546] Call Trace: > [ 463.875596] ? ocfs2_inode_lock_full_nested+0x18b/0x960 [ocfs2] > [ 463.875648] ocfs2_file_write_iter+0xaf8/0xc70 [ocfs2] > [ 463.875672] new_sync_write+0x12d/0x1d0 > [ 463.875688] vfs_write+0xad/0x1a0 > [ 463.875697] ksys_write+0xa1/0xe0 > [ 463.875710] do_syscall_64+0x60/0x1f0 > [ 463.875743] entry_SYSCALL_64_after_hwframe+0x49/0xbe > [ 463.875758] RIP: 0033:0x7f8f4482ed44 > [ 463.875762] Code: 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f 80 00 00 00 > [ 463.875765] RSP: 002b:00007fff300a79d8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 > [ 463.875769] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f8f4482ed44 > [ 463.875771] RDX: 0000000000000200 RSI: 000055f771b5c000 RDI: 0000000000000001 > [ 463.875774] RBP: 0000000000000200 R08: 00007f8f44af9c78 R09: 0000000000000003 > [ 463.875776] R10: 000000000000089f R11: 0000000000000246 R12: 000055f771b5c000 > [ 463.875779] R13: 0000000000000200 R14: 0000000000000000 R15: 000055f771b5c000 > > This regression problem was introduced by commit e74540b28556 ("ocfs2: > protect extent tree in ocfs2_prepare_inode_for_write()"). > > Fixes: e74540b28556 ("ocfs2: protect extent tree in ocfs2_prepare_inode_for_write()"). > Signed-off-by: Gang He <ghe@suse.com> > --- > fs/ocfs2/file.c | 14 ++++++-------- > 1 file changed, 6 insertions(+), 8 deletions(-) > > diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c > index 9876db52913a..6cd5e4924e4d 100644 > --- a/fs/ocfs2/file.c > +++ b/fs/ocfs2/file.c > @@ -2101,17 +2101,15 @@ static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos) > static int ocfs2_inode_lock_for_extent_tree(struct inode *inode, > struct buffer_head **di_bh, > int meta_level, > - int overwrite_io, > int write_sem, > int wait) > { > int ret = 0; > > if (wait) > - ret = ocfs2_inode_lock(inode, NULL, meta_level); > + ret = ocfs2_inode_lock(inode, di_bh, meta_level); > else > - ret = ocfs2_try_inode_lock(inode, > - overwrite_io ? NULL : di_bh, meta_level); > + ret = ocfs2_try_inode_lock(inode, di_bh, meta_level); > if (ret < 0) > goto out; > > @@ -2136,6 +2134,7 @@ static int ocfs2_inode_lock_for_extent_tree(struct inode *inode, > > out_unlock: > brelse(*di_bh); > + *di_bh = NULL; > ocfs2_inode_unlock(inode, meta_level); > out: > return ret; > @@ -2177,7 +2176,6 @@ static int ocfs2_prepare_inode_for_write(struct file *file, > ret = ocfs2_inode_lock_for_extent_tree(inode, > &di_bh, > meta_level, > - overwrite_io, > write_sem, > wait); > if (ret < 0) { > @@ -2233,13 +2231,13 @@ static int ocfs2_prepare_inode_for_write(struct file *file, > &di_bh, > meta_level, > write_sem); > + meta_level = 1; > + write_sem = 1; > ret = ocfs2_inode_lock_for_extent_tree(inode, > &di_bh, > meta_level, > - overwrite_io, > - 1, > + write_sem, > wait); > - write_sem = 1; > if (ret < 0) { > if (ret != -EAGAIN) > mlog_errno(ret); > -- > 2.12.3 >
Hi Joseph, before calling ocfs2_refcount_cow() in the function ocfs2_prepare_inode_for_write(), we do not use inode buffer_head. So, we can let buffer_head is NULL. But, when we invoke ocfs2_refcount_cow() function, we have to pass inode buffer_head without NULL pointer. That is why writing clone file will cause oops and kill the user-space process. Thanks Gang
On 20/2/3 13:32, Gang He wrote: > Hi Joseph, > > before calling ocfs2_refcount_cow() in the function ocfs2_prepare_inode_for_write(), we do not use inode buffer_head. > So, we can let buffer_head is NULL. > But, when we invoke ocfs2_refcount_cow() function, we have to pass inode buffer_head without NULL pointer. > That is why writing clone file will cause oops and kill the user-space process. > Okay, so before commit e74540b28556, we will always get a valid buffer head in ocfs2_prepare_inode_for_refcount(). You can feel free to add: Reviewed-by: Joseph Qi <joseph.qi@linux.alibaba.com> BTW, you'd better resend the patch in a single thread, for the convenience of merging by akpm. > > ________________________________________ > From: Joseph Qi <joseph.qi@linux.alibaba.com> > Sent: Monday, February 3, 2020 1:15 PM > To: Gang He; mark@fasheh.com; jlbec@evilplan.org; gechangwei@live.cn; Shuning Zhang > Cc: linux-kernel@vger.kernel.org; ocfs2-devel@oss.oracle.com; akpm@linux-foundation.org > Subject: Re: [PATCH] ocfs2: fix the oops problem when write cloned file > > > > On 20/2/3 10:17, Gang He wrote: >> Hello Joseph, Changwei, Sunny and all, >> >> Could you help to review this patch? >> This patch will fix the oops problem caused by write ocfs2 clone files. >> The root cause is inode buffer head is NULL when calling ocfs2_refcount_cow. >> Secondly, we should use EX meta lock when calling ocfs2_refcount_cow. >> > Before commit e74540b28556, we may also use NULL buffer head in case of > overwrite, so why there is no such issue before? > > Thanks, > Joseph >
diff --git a/fs/ocfs2/file.c b/fs/ocfs2/file.c index 9876db52913a..6cd5e4924e4d 100644 --- a/fs/ocfs2/file.c +++ b/fs/ocfs2/file.c @@ -2101,17 +2101,15 @@ static int ocfs2_is_io_unaligned(struct inode *inode, size_t count, loff_t pos) static int ocfs2_inode_lock_for_extent_tree(struct inode *inode, struct buffer_head **di_bh, int meta_level, - int overwrite_io, int write_sem, int wait) { int ret = 0; if (wait) - ret = ocfs2_inode_lock(inode, NULL, meta_level); + ret = ocfs2_inode_lock(inode, di_bh, meta_level); else - ret = ocfs2_try_inode_lock(inode, - overwrite_io ? NULL : di_bh, meta_level); + ret = ocfs2_try_inode_lock(inode, di_bh, meta_level); if (ret < 0) goto out; @@ -2136,6 +2134,7 @@ static int ocfs2_inode_lock_for_extent_tree(struct inode *inode, out_unlock: brelse(*di_bh); + *di_bh = NULL; ocfs2_inode_unlock(inode, meta_level); out: return ret; @@ -2177,7 +2176,6 @@ static int ocfs2_prepare_inode_for_write(struct file *file, ret = ocfs2_inode_lock_for_extent_tree(inode, &di_bh, meta_level, - overwrite_io, write_sem, wait); if (ret < 0) { @@ -2233,13 +2231,13 @@ static int ocfs2_prepare_inode_for_write(struct file *file, &di_bh, meta_level, write_sem); + meta_level = 1; + write_sem = 1; ret = ocfs2_inode_lock_for_extent_tree(inode, &di_bh, meta_level, - overwrite_io, - 1, + write_sem, wait); - write_sem = 1; if (ret < 0) { if (ret != -EAGAIN) mlog_errno(ret);
Writing a cloned file triggers a kernel oops and the user-space command process is also killed by the system. The bug can be reproduced stably, e.g. 1)create a file under ocfs2 file system directory. journalctl -b > aa.txt 2)create a cloned file for this file. reflink aa.txt bb.txt 3)write the cloned file with dd command. dd if=/dev/zero of=bb.txt bs=512 count=1 conv=notrunc The dd command is killed by the kernel, then you can see the oops message via dmesg command. [ 463.875404] BUG: kernel NULL pointer dereference, address: 0000000000000028 [ 463.875413] #PF: supervisor read access in kernel mode [ 463.875416] #PF: error_code(0x0000) - not-present page [ 463.875418] PGD 0 P4D 0 [ 463.875425] Oops: 0000 [#1] SMP PTI [ 463.875431] CPU: 1 PID: 2291 Comm: dd Tainted: G OE 5.3.16-2-default [ 463.875433] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011 [ 463.875500] RIP: 0010:ocfs2_refcount_cow+0xa4/0x5d0 [ocfs2] [ 463.875505] Code: 06 89 6c 24 38 89 eb f6 44 24 3c 02 74 be 49 8b 47 28 [ 463.875508] RSP: 0018:ffffa2cb409dfce8 EFLAGS: 00010202 [ 463.875512] RAX: ffff8b1ebdca8000 RBX: 0000000000000001 RCX: ffff8b1eb73a9df0 [ 463.875515] RDX: 0000000000056a01 RSI: 0000000000000000 RDI: 0000000000000000 [ 463.875517] RBP: 0000000000000001 R08: ffff8b1eb73a9de0 R09: 0000000000000000 [ 463.875520] R10: 0000000000000001 R11: 0000000000000000 R12: 0000000000000000 [ 463.875522] R13: ffff8b1eb922f048 R14: 0000000000000000 R15: ffff8b1eb922f048 [ 463.875526] FS: 00007f8f44d15540(0000) GS:ffff8b1ebeb00000(0000) knlGS:0000000000000000 [ 463.875529] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 463.875532] CR2: 0000000000000028 CR3: 000000003c17a000 CR4: 00000000000006e0 [ 463.875546] Call Trace: [ 463.875596] ? ocfs2_inode_lock_full_nested+0x18b/0x960 [ocfs2] [ 463.875648] ocfs2_file_write_iter+0xaf8/0xc70 [ocfs2] [ 463.875672] new_sync_write+0x12d/0x1d0 [ 463.875688] vfs_write+0xad/0x1a0 [ 463.875697] ksys_write+0xa1/0xe0 [ 463.875710] do_syscall_64+0x60/0x1f0 [ 463.875743] entry_SYSCALL_64_after_hwframe+0x49/0xbe [ 463.875758] RIP: 0033:0x7f8f4482ed44 [ 463.875762] Code: 00 f7 d8 64 89 02 48 c7 c0 ff ff ff ff eb b7 0f 1f 80 00 00 00 [ 463.875765] RSP: 002b:00007fff300a79d8 EFLAGS: 00000246 ORIG_RAX: 0000000000000001 [ 463.875769] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f8f4482ed44 [ 463.875771] RDX: 0000000000000200 RSI: 000055f771b5c000 RDI: 0000000000000001 [ 463.875774] RBP: 0000000000000200 R08: 00007f8f44af9c78 R09: 0000000000000003 [ 463.875776] R10: 000000000000089f R11: 0000000000000246 R12: 000055f771b5c000 [ 463.875779] R13: 0000000000000200 R14: 0000000000000000 R15: 000055f771b5c000 This regression problem was introduced by commit e74540b28556 ("ocfs2: protect extent tree in ocfs2_prepare_inode_for_write()"). Fixes: e74540b28556 ("ocfs2: protect extent tree in ocfs2_prepare_inode_for_write()"). Signed-off-by: Gang He <ghe@suse.com> --- fs/ocfs2/file.c | 14 ++++++-------- 1 file changed, 6 insertions(+), 8 deletions(-)