On 4/18/24 13:04, Dave Chinner wrote:
On Thu, Apr 18, 2024 at 11:39:25AM +0800, Xiubo Li wrote:Hi all BTW, is this a known issue and has it been fixed already ? I can reproduce this always with my VMs: <4>[ 9009.171195] <4>[ 9009.171205] ====================================================== <4>[ 9009.171208] WARNING: possible circular locking dependency detected <4>[ 9009.171211] 6.9.0-rc3+ #49 Not tainted <4>[ 9009.171214] ------------------------------------------------------ <4>[ 9009.171216] kswapd0/149 is trying to acquire lock: <4>[ 9009.171219] ffff88811346a920 (&xfs_nondir_ilock_class){++++}-{4:4}, at: xfs_reclaim_inode+0x3ac/0x590 [xfs] <4>[ 9009.171580] <4>[ 9009.171580] but task is already holding lock: <4>[ 9009.171583] ffffffff8bb33100 (fs_reclaim){+.+.}-{0:0}, at: balance_pgdat+0x5d9/0xad0 <4>[ 9009.171593] <4>[ 9009.171593] which lock already depends on the new lock. <4>[ 9009.171593] <4>[ 9009.171595] <4>[ 9009.171595] the existing dependency chain (in reverse order) is: <4>[ 9009.171597] <4>[ 9009.171597] -> #1 (fs_reclaim){+.+.}-{0:0}: <4>[ 9009.171603] __lock_acquire+0x7da/0x1030 <4>[ 9009.171610] lock_acquire+0x15d/0x400 <4>[ 9009.171614] fs_reclaim_acquire+0xb5/0x100 <4>[ 9009.171618] prepare_alloc_pages.constprop.0+0xc5/0x230 <4>[ 9009.171622] __alloc_pages+0x12a/0x3f0 <4>[ 9009.171625] alloc_pages_mpol+0x175/0x340 <4>[ 9009.171630] stack_depot_save_flags+0x4c5/0x510 <4>[ 9009.171635] kasan_save_stack+0x30/0x40 <4>[ 9009.171640] kasan_save_track+0x10/0x30 <4>[ 9009.171643] __kasan_slab_alloc+0x83/0x90 <4>[ 9009.171646] kmem_cache_alloc+0x15e/0x4a0 <4>[ 9009.171652] __alloc_object+0x35/0x370 <4>[ 9009.171659] __create_object+0x22/0x90 <4>[ 9009.171665] __kmalloc_node_track_caller+0x477/0x5b0 <4>[ 9009.171672] krealloc+0x5f/0x110 <4>[ 9009.171679] xfs_iext_insert_raw+0x4b2/0x6e0 [xfs] <4>[ 9009.172172] xfs_iext_insert+0x2e/0x130 [xfs]The only krealloc() in this path is: new = krealloc(ifp->if_data, new_size, GFP_KERNEL | __GFP_NOLOCKDEP | __GFP_NOFAIL); And it explicitly uses __GFP_NOLOCKDEP to tell lockdep not to warn about this allocation because of this false positive situation. Oh. I've seen this before. This is a KASAN bug, and I'm pretty sure I've posted a patch to fix it a fair while back that nobody seemed to care about enough to review or merge it. That is: kasan_save_stack() is doing a fixed GFP_KERNEL allocation in an context where GFP_KERNEL allocations are known to generate lockdep false positives. This occurs depsite the XFS and general memory allocation code doing exactly the right thing to avoid the lockdep false positives (i.e. using and obeying __GFP_NOLOCKDEP). The kasan code ends up in stack_depot_save_flags(), which does a GFP_KERNEL allocation but filters out __GFP_NOLOCKDEP and does not add it back. Hence kasan generates the false positive lockdep warnings, not the code doing the original allocation. kasan and/or stack_depot_save_flags() needs fixing here.
Hi Dave, Thanks very much for your feedback. BTW, do you have the link of your patch ? I can help test it. Thanks - Xiubo
-Dave.
