From mboxrd@z Thu Jan 1 00:00:00 1970 From: Roberto Sassu Subject: Re: [syzbot] [reiserfs?] possible deadlock in open_xa_dir Date: Fri, 02 Jun 2023 18:46:27 +0200 Message-ID: References: <0000000000007bedb605f119ed9f@google.com> <00000000000000964605faf87416@google.com> <1020d006-c698-aacc-bcc3-92e5b237ef91@huaweicloud.com> <29fcea18-d720-d5df-0e00-eb448e6bbfcf@suse.com> <07c825a21fb4c57f4290158e529d32f4e0e0fbf0.camel@huaweicloud.com> Mime-Version: 1.0 Content-Transfer-Encoding: quoted-printable Return-path: In-Reply-To: <07c825a21fb4c57f4290158e529d32f4e0e0fbf0.camel@huaweicloud.com> List-ID: Content-Type: text/plain; charset="windows-1252" To: Jeff Mahoney , Paul Moore , syzbot Cc: hdanton@sina.com, linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org, reiserfs-devel@vger.kernel.org, roberto.sassu@huawei.com, syzkaller-bugs@googlegroups.com, peterz@infradead.org, mingo@redhat.com, will@kernel.org, Jan Kara On Fri, 2023-06-02 at 09:20 +0200, Roberto Sassu wrote: > On Thu, 2023-06-01 at 17:22 -0400, Jeff Mahoney wrote: > > On 5/31/23 05:49, Roberto Sassu wrote: > > > On 5/5/2023 11:36 PM, Paul Moore wrote: > > > > On Fri, May 5, 2023 at 4:51=E2=80=AFPM syzbot > > > > wrote: > > > > > syzbot has bisected this issue to: > > > > >=20 > > > > > commit d82dcd9e21b77d338dc4875f3d4111f0db314a7c > > > > > Author: Roberto Sassu > > > > > Date: Fri Mar 31 12:32:18 2023 +0000 > > > > >=20 > > > > > reiserfs: Add security prefix to xattr name in=20 > > > > > reiserfs_security_write() > > > > >=20 > > > > > bisection log: =20 > > > > > https://syzkaller.appspot.com/x/bisect.txt?x=3D14403182280000 > > > > > start commit: 3c4aa4434377 Merge tag 'ceph-for-6.4-rc1' of=20 > > > > > https://githu.. > > > > > git tree: upstream > > > > > final oops: =20 > > > > > https://syzkaller.appspot.com/x/report.txt?x=3D16403182280000 > > > > > console output: https://syzkaller.appspot.com/x/log.txt?x=3D12403= 182280000 > > > > > kernel config: =20 > > > > > https://syzkaller.appspot.com/x/.config?x=3D73a06f6ef2d5b492 > > > > > dashboard link:=20 > > > > > https://syzkaller.appspot.com/bug?extid=3D8fb64a61fdd96b50f3b8 > > > > > syz repro: =20 > > > > > https://syzkaller.appspot.com/x/repro.syz?x=3D12442414280000 > > > > > C reproducer: https://syzkaller.appspot.com/x/repro.c?x=3D176a7= 318280000 > > > > >=20 > > > > > Reported-by: syzbot+8fb64a61fdd96b50f3b8@syzkaller.appspotmail.com > > > > > Fixes: d82dcd9e21b7 ("reiserfs: Add security prefix to xattr name= in=20 > > > > > reiserfs_security_write()") > > > > >=20 > > > > > For information about bisection process see:=20 > > > > > https://goo.gl/tpsmEJ#bisection > > > >=20 > > > > I don't think Roberto's patch identified above is the actual root > > > > cause of this problem as reiserfs_xattr_set_handle() is called in > > > > reiserfs_security_write() both before and after the patch. However, > > > > due to some bad logic in reiserfs_security_write() which Roberto > > > > corrected, I'm thinking that it is possible this code is being > > > > exercised for the first time and syzbot is starting to trigger a > > > > locking issue in the reiserfs code ... ? > > >=20 > > > + Jan, Jeff (which basically restructured the lock) > > >=20 > > > + Petr, Ingo, Will >=20 > Peter, clearly (sorry!) >=20 > > I involve the lockdep experts, to get a bit of help on this. > >=20 > > Yep, looks like that's been broken since it was added in 2009. Since=20 > > there can't be any users of it, it'd make sense to drop the security=20 > > xattr support from reiserfs entirely. >=20 > Thanks, Jeff. Will make a patch to implement your suggestion. Ok, I tried first to disable security xattr initialization and keep the xattr handler. Setting the security xattr manually triggers a lockdep warning. Even worse, setting a trusted xattr manually triggers that too. So, not sure how we should proceed. Have you looked at: https://lore.kernel.org/linux-kernel/8a48ede1-3a45-7c3c-39e9-36001ac09283@h= uaweicloud.com/ That silences the lockdep warning, but I'm far from saying that it won't have any side effect... Thanks Roberto > Meanwhile, I learned how to read lockdep a bit better. The following > format could have helped me to understand it more quickly. The proposal > is simply to change #n to CPU#n at the top of the trace, define labels > L#n for the locks, and add them where effectively are held. >=20 > [ 77.746561][ T5418] -> CPU1 (&sbi->lock){+.+.}-{3:3}: > [ 77.753772][ T5418] lock_acquire+0x23e/0x630 > [ 77.758792][ T5418] __mutex_lock_common+0x1d8/0x2530 > [ 77.764504][ T5418] (L3) mutex_lock_nested+0x1b/0x20 > [ 77.769868][ T5418] reiserfs_write_lock+0x70/0xc0 > [ 77.775321][ T5418] reiserfs_mkdir+0x321/0x870 > [ 77.780509][ T5418] (L2) open_xa_dir+0x259/0x540 > [ 77.785440][ T5418] xattr_lookup+0x17/0x210 > [ 77.790378][ T5418] reiserfs_xattr_set_handle+0xda/0xc80 > [ 77.796448][ T5418] reiserfs_security_write+0x134/0x190 > [ 77.802416][ T5418] reiserfs_new_inode+0x13bf/0x1a90 > [ 77.808124][ T5418] reiserfs_create+0x3b1/0x680 > [ 77.813399][ T5418] path_openat+0xf1e/0x2c10 > [ 77.818415][ T5418] do_filp_open+0x22a/0x440 > [ 77.823433][ T5418] do_sys_openat2+0x10f/0x430 > [ 77.828624][ T5418] __x64_sys_creat+0x11e/0x160 > [ 77.833905][ T5418] do_syscall_64+0x41/0xc0 > [ 77.838926][ T5418] entry_SYSCALL_64_after_hwframe+0x63/0xcd > [ 77.845514][ T5418]=20 > [ 77.845514][ T5418] -> CPU0 (&type->i_mutex_dir_key#8/3){+.+.}-{3:3}: > [ 77.854118][ T5418] validate_chain+0x166b/0x58e0 > [ 77.859488][ T5418] __lock_acquire+0x125b/0x1f80 > [ 77.864853][ T5418] lock_acquire+0x23e/0x630 > [ 77.869909][ T5418] (L4) down_write_nested+0x3d/0x50 > [ 77.875186][ T5418] open_xa_dir+0x134/0x540 > [ 77.880117][ T5418] xattr_lookup+0x17/0x210 > [ 77.885050][ T5418] reiserfs_xattr_get+0xe1/0x4a0 > [ 77.890501][ T5418] __vfs_getxattr+0x2fe/0x350 > [ 77.895802][ T5418] smk_fetch+0x98/0xf0 > [ 77.900382][ T5418] smack_d_instantiate+0x5d5/0xa20 > [ 77.906018][ T5418] security_d_instantiate+0x6b/0xb0 > [ 77.911736][ T5418] d_instantiate_new+0x5e/0xe0 > [ 77.917013][ T5418] (L1) reiserfs_create+0x5ee/0x680 > [ 77.922293][ T5418] path_openat+0xf1e/0x2c10 > [ 77.927308][ T5418] do_filp_open+0x22a/0x440 > [ 77.932330][ T5418] do_sys_openat2+0x10f/0x430 > [ 77.937515][ T5418] __x64_sys_creat+0x11e/0x160 > [ 77.942874][ T5418] do_syscall_64+0x41/0xc0 > [ 77.947796][ T5418] entry_SYSCALL_64_after_hwframe+0x63/0xcd > [ 77.954200][ T5418]=20 > [ 77.954200][ T5418] other info that might help us debug this: > [ 77.954200][ T5418]=20 > [ 77.964508][ T5418] Possible unsafe locking scenario: > [ 77.964508][ T5418]=20 > [ 77.972034][ T5418] CPU0 CPU1 > [ 77.977394][ T5418] ---- ---- > [ 77.982748][ T5418] L1: lock(&sbi->lock); > [ 77.986726][ T5418] L2: lock(&type->i_m= utex_dir_key#8/3); > [ 77.994618][ T5418] L3: lock(&sbi->lock= ); > [ 78.001118][ T5418] L4: lock(&type->i_mutex_dir_key#8/3); >=20 > Thanks >=20 > Roberto >=20 > > > First of all, the lockdep warning is trivial to reproduce: > > >=20 > > > # dd if=3D/dev/zero of=3Dreiserfs.img bs=3D1M count=3D100 > > > # losetup -f --show reiserfs.img > > > /dev/loop0 > > > # mkfs.reiserfs /dev/loop0 > > > # mount /dev/loop0 /mnt/ > > > # touch file0 > > >=20 > > > In the testing system, Smack is the major LSM. > > >=20 > > > Ok, so the warning here is clear: > > >=20 > > > https://syzkaller.appspot.com/x/log.txt?x=3D12403182280000 > > >=20 > > > However, I was looking if that can really happen. From this: > > >=20 > > > [ 77.746561][ T5418] -> #1 (&sbi->lock){+.+.}-{3:3}: > > > [ 77.753772][ T5418] lock_acquire+0x23e/0x630 > > > [ 77.758792][ T5418] __mutex_lock_common+0x1d8/0x2530 > > > [ 77.764504][ T5418] mutex_lock_nested+0x1b/0x20 > > > [ 77.769868][ T5418] reiserfs_write_lock+0x70/0xc0 > > > [ 77.775321][ T5418] reiserfs_mkdir+0x321/0x870 > > >=20 > > > I see that the lock is taken in reiserfs_write_lock(), while lockdep = says: > > >=20 > > > [ 77.710227][ T5418] but task is already holding lock: > > > [ 77.717587][ T5418] ffff88807568d090 (&sbi->lock){+.+.}-{3:3}, at:= =20 > > > reiserfs_write_lock_nested+0x4a/0xb0 > > >=20 > > > which is in a different place, I believe here: > > >=20 > > > int reiserfs_paste_into_item(struct reiserfs_transaction_handle *th, > > > /* Path to the pasted item. */ > > > [...] > > >=20 > > > depth =3D reiserfs_write_unlock_nested(sb); > > > dquot_free_space_nodirty(inode, pasted_size); > > > reiserfs_write_lock_nested(sb, depth); > > > return retval; > > > } > > >=20 > > > This is called by reiserfs_add_entry(), which is called by=20 > > > reiserfs_create() (it is in the lockdep trace). After returning to=20 > > > reiserfs_create(), d_instantiate_new() is called. > > >=20 > > > I don't know exactly, I take the part that the lock is held. But if i= t=20 > > > is held, how d_instantiate_new() can be executed in another task? > > >=20 > > > static int reiserfs_create(struct mnt_idmap *idmap, struct inode *dir, > > > struct dentry *dentry, umode_t mode, bool ex= cl) > > > { > > >=20 > > > [...] > > >=20 > > > reiserfs_write_lock(dir->i_sb); > > >=20 > > > retval =3D journal_begin(&th, dir->i_sb, jbegin_count); > > >=20 > > > [...] > > >=20 > > > d_instantiate_new(dentry, inode); > > > retval =3D journal_end(&th); > > >=20 > > > out_failed: > > > reiserfs_write_unlock(dir->i_sb); > > >=20 > > > If the lock is held, the scenario lockdep describes cannot happen. An= y=20 > > > thoughts? > >=20 > > It's important to understand that the reiserfs write lock was added as = a=20 > > subsystem-specific replacement for the BKL. Given that reiserfs was=20 > > dying already back then, it made more sense from a time management=20 > > perspective to emulate that behavior internally rather than use new=20 > > locking when practically nobody cared anymore. > >=20 > > See reiserfs_write_unlock_nested and reiserfs_write_lock_nested paired = > > throughout the code. It drops the lock when it passes a point where=20 > > it's likely to schedule, just like the BKL would have. > >=20 > > Yes, it's a mess. Just let it die quietly. > >=20 > > -Jeff > >=20