| Message ID | 20220715023928.2701166-1-libaokun1@huawei.com |
|---|---|
| State | Not Applicable |
| Headers | show |
| Series | [4.19] ext4: fix race condition between ext4_ioctl_setflags and ext4_fiemap | expand |
On Fri, Jul 15, 2022 at 10:39:28AM +0800, Baokun Li wrote: > This patch and problem analysis is based on v4.19 LTS. > The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch > is incorporated in v5.7-rc1. This patch avoids this problem by switching > to iomap in ext4_fiemap. > > Hulk Robot reported a BUG on stable 4.19.252: > ================================================================== > kernel BUG at fs/ext4/extents_status.c:762! > invalid opcode: 0000 [#1] SMP KASAN PTI > CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 > RIP: 0010:ext4_es_cache_extent+0x30e/0x370 > [...] > Call Trace: > ext4_cache_extents+0x238/0x2f0 > ext4_find_extent+0x785/0xa40 > ext4_fiemap+0x36d/0xe90 > do_vfs_ioctl+0x6af/0x1200 > [...] > ================================================================== > > Above issue may happen as follows: > ------------------------------------- > cpu1 cpu2 > _____________________|_____________________ > do_vfs_ioctl > ext4_ioctl > ext4_ioctl_setflags > ext4_ind_migrate > do_vfs_ioctl > ioctl_fiemap > ext4_fiemap > ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) > ext4_fill_fiemap_extents > down_write(&EXT4_I(inode)->i_data_sem); > ext4_ext_check_inode > ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) > memset(ei->i_data, 0, sizeof(ei->i_data)) > up_write(&EXT4_I(inode)->i_data_sem); > down_read(&EXT4_I(inode)->i_data_sem); > ext4_find_extent > ext4_cache_extents > ext4_es_cache_extent > BUG_ON(end < lblk) > > We can easily reproduce this problem with the syzkaller testcase: > ``` > 02:37:07 executing program 3: > r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) > ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) > mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) > r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) > ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) > ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) > ``` > > To solve this issue, we use __generic_block_fiemap() instead of > generic_block_fiemap() and add inode_lock_shared to avoid race condition. > > Reported-by: Hulk Robot <hulkci@huawei.com> > Signed-off-by: Baokun Li <libaokun1@huawei.com> > --- > fs/ext4/extents.c | 15 +++++++++++---- > 1 file changed, 11 insertions(+), 4 deletions(-) What is the git commit id of this change in Linus's tree? If it is not in Linus's tree, why not? confused, greg k-h
在 2022/7/15 22:10, Greg KH 写道: > On Fri, Jul 15, 2022 at 10:39:28AM +0800, Baokun Li wrote: >> This patch and problem analysis is based on v4.19 LTS. >> The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch >> is incorporated in v5.7-rc1. This patch avoids this problem by switching >> to iomap in ext4_fiemap. >> >> Hulk Robot reported a BUG on stable 4.19.252: >> ================================================================== >> kernel BUG at fs/ext4/extents_status.c:762! >> invalid opcode: 0000 [#1] SMP KASAN PTI >> CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 >> RIP: 0010:ext4_es_cache_extent+0x30e/0x370 >> [...] >> Call Trace: >> ext4_cache_extents+0x238/0x2f0 >> ext4_find_extent+0x785/0xa40 >> ext4_fiemap+0x36d/0xe90 >> do_vfs_ioctl+0x6af/0x1200 >> [...] >> ================================================================== >> >> Above issue may happen as follows: >> ------------------------------------- >> cpu1 cpu2 >> _____________________|_____________________ >> do_vfs_ioctl >> ext4_ioctl >> ext4_ioctl_setflags >> ext4_ind_migrate >> do_vfs_ioctl >> ioctl_fiemap >> ext4_fiemap >> ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) >> ext4_fill_fiemap_extents >> down_write(&EXT4_I(inode)->i_data_sem); >> ext4_ext_check_inode >> ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) >> memset(ei->i_data, 0, sizeof(ei->i_data)) >> up_write(&EXT4_I(inode)->i_data_sem); >> down_read(&EXT4_I(inode)->i_data_sem); >> ext4_find_extent >> ext4_cache_extents >> ext4_es_cache_extent >> BUG_ON(end < lblk) >> >> We can easily reproduce this problem with the syzkaller testcase: >> ``` >> 02:37:07 executing program 3: >> r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) >> ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) >> mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) >> r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) >> ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) >> ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) >> ``` >> >> To solve this issue, we use __generic_block_fiemap() instead of >> generic_block_fiemap() and add inode_lock_shared to avoid race condition. >> >> Reported-by: Hulk Robot <hulkci@huawei.com> >> Signed-off-by: Baokun Li <libaokun1@huawei.com> >> --- >> fs/ext4/extents.c | 15 +++++++++++---- >> 1 file changed, 11 insertions(+), 4 deletions(-) > What is the git commit id of this change in Linus's tree? > > If it is not in Linus's tree, why not? > > confused, > > greg k-h > . This patch does not exist in the Linus' tree. This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") is incorporated in v5.7-rc1. However, this problem can be found by asserting after patch 4068664e3cd2 ("ext4: fix extent_status fragmentation for plain files") is incorporated into v5.6-rc1. If someone don't want to convert ext4_fiemap to iomap, this patch may help. I also CC linux-ext4 in hopes of getting some advice. Feel free to improve it if something wrong. Thanks a lot!
On Sat, Jul 16, 2022 at 10:33:30AM +0800, Baokun Li wrote: > 在 2022/7/15 22:10, Greg KH 写道: > > On Fri, Jul 15, 2022 at 10:39:28AM +0800, Baokun Li wrote: > > > This patch and problem analysis is based on v4.19 LTS. > > > The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch > > > is incorporated in v5.7-rc1. This patch avoids this problem by switching > > > to iomap in ext4_fiemap. > > > > > > Hulk Robot reported a BUG on stable 4.19.252: > > > ================================================================== > > > kernel BUG at fs/ext4/extents_status.c:762! > > > invalid opcode: 0000 [#1] SMP KASAN PTI > > > CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 > > > RIP: 0010:ext4_es_cache_extent+0x30e/0x370 > > > [...] > > > Call Trace: > > > ext4_cache_extents+0x238/0x2f0 > > > ext4_find_extent+0x785/0xa40 > > > ext4_fiemap+0x36d/0xe90 > > > do_vfs_ioctl+0x6af/0x1200 > > > [...] > > > ================================================================== > > > > > > Above issue may happen as follows: > > > ------------------------------------- > > > cpu1 cpu2 > > > _____________________|_____________________ > > > do_vfs_ioctl > > > ext4_ioctl > > > ext4_ioctl_setflags > > > ext4_ind_migrate > > > do_vfs_ioctl > > > ioctl_fiemap > > > ext4_fiemap > > > ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) > > > ext4_fill_fiemap_extents > > > down_write(&EXT4_I(inode)->i_data_sem); > > > ext4_ext_check_inode > > > ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) > > > memset(ei->i_data, 0, sizeof(ei->i_data)) > > > up_write(&EXT4_I(inode)->i_data_sem); > > > down_read(&EXT4_I(inode)->i_data_sem); > > > ext4_find_extent > > > ext4_cache_extents > > > ext4_es_cache_extent > > > BUG_ON(end < lblk) > > > > > > We can easily reproduce this problem with the syzkaller testcase: > > > ``` > > > 02:37:07 executing program 3: > > > r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) > > > ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) > > > mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) > > > r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) > > > ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) > > > ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) > > > ``` > > > > > > To solve this issue, we use __generic_block_fiemap() instead of > > > generic_block_fiemap() and add inode_lock_shared to avoid race condition. > > > > > > Reported-by: Hulk Robot <hulkci@huawei.com> > > > Signed-off-by: Baokun Li <libaokun1@huawei.com> > > > --- > > > fs/ext4/extents.c | 15 +++++++++++---- > > > 1 file changed, 11 insertions(+), 4 deletions(-) > > What is the git commit id of this change in Linus's tree? > > > > If it is not in Linus's tree, why not? > > > > confused, > > > > greg k-h > > . > > This patch does not exist in the Linus' tree. > > This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap > to use iomap framework") is incorporated in v5.7-rc1. Then why not ask for that change to be added instead? thanks, greg k-h
在 2022/7/19 19:26, Greg KH 写道: > On Sat, Jul 16, 2022 at 10:33:30AM +0800, Baokun Li wrote: >> 在 2022/7/15 22:10, Greg KH 写道: >>> On Fri, Jul 15, 2022 at 10:39:28AM +0800, Baokun Li wrote: >>>> This patch and problem analysis is based on v4.19 LTS. >>>> The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch >>>> is incorporated in v5.7-rc1. This patch avoids this problem by switching >>>> to iomap in ext4_fiemap. >>>> >>>> Hulk Robot reported a BUG on stable 4.19.252: >>>> ================================================================== >>>> kernel BUG at fs/ext4/extents_status.c:762! >>>> invalid opcode: 0000 [#1] SMP KASAN PTI >>>> CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 >>>> RIP: 0010:ext4_es_cache_extent+0x30e/0x370 >>>> [...] >>>> Call Trace: >>>> ext4_cache_extents+0x238/0x2f0 >>>> ext4_find_extent+0x785/0xa40 >>>> ext4_fiemap+0x36d/0xe90 >>>> do_vfs_ioctl+0x6af/0x1200 >>>> [...] >>>> ================================================================== >>>> >>>> Above issue may happen as follows: >>>> ------------------------------------- >>>> cpu1 cpu2 >>>> _____________________|_____________________ >>>> do_vfs_ioctl >>>> ext4_ioctl >>>> ext4_ioctl_setflags >>>> ext4_ind_migrate >>>> do_vfs_ioctl >>>> ioctl_fiemap >>>> ext4_fiemap >>>> ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) >>>> ext4_fill_fiemap_extents >>>> down_write(&EXT4_I(inode)->i_data_sem); >>>> ext4_ext_check_inode >>>> ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) >>>> memset(ei->i_data, 0, sizeof(ei->i_data)) >>>> up_write(&EXT4_I(inode)->i_data_sem); >>>> down_read(&EXT4_I(inode)->i_data_sem); >>>> ext4_find_extent >>>> ext4_cache_extents >>>> ext4_es_cache_extent >>>> BUG_ON(end < lblk) >>>> >>>> We can easily reproduce this problem with the syzkaller testcase: >>>> ``` >>>> 02:37:07 executing program 3: >>>> r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) >>>> ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) >>>> mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) >>>> r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) >>>> ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) >>>> ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) >>>> ``` >>>> >>>> To solve this issue, we use __generic_block_fiemap() instead of >>>> generic_block_fiemap() and add inode_lock_shared to avoid race condition. >>>> >>>> Reported-by: Hulk Robot <hulkci@huawei.com> >>>> Signed-off-by: Baokun Li <libaokun1@huawei.com> >>>> --- >>>> fs/ext4/extents.c | 15 +++++++++++---- >>>> 1 file changed, 11 insertions(+), 4 deletions(-) >>> What is the git commit id of this change in Linus's tree? >>> >>> If it is not in Linus's tree, why not? >>> >>> confused, >>> >>> greg k-h >>> . >> This patch does not exist in the Linus' tree. >> >> This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap >> to use iomap framework") is incorporated in v5.7-rc1. > Then why not ask for that change to be added instead? > > thanks, > > greg k-h > . If we want to switch to the iomap framework, we need to analyze and integrate about 60 patches. The workload may be greater than that of solving this problem alone. Thank you! Thanks a lot!
On Tue, Jul 19, 2022 at 08:15:13PM +0800, Baokun Li wrote: > 在 2022/7/19 19:26, Greg KH 写道: > > On Sat, Jul 16, 2022 at 10:33:30AM +0800, Baokun Li wrote: > > > 在 2022/7/15 22:10, Greg KH 写道: > > > > On Fri, Jul 15, 2022 at 10:39:28AM +0800, Baokun Li wrote: > > > > > This patch and problem analysis is based on v4.19 LTS. > > > > > The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch > > > > > is incorporated in v5.7-rc1. This patch avoids this problem by switching > > > > > to iomap in ext4_fiemap. > > > > > > > > > > Hulk Robot reported a BUG on stable 4.19.252: > > > > > ================================================================== > > > > > kernel BUG at fs/ext4/extents_status.c:762! > > > > > invalid opcode: 0000 [#1] SMP KASAN PTI > > > > > CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 > > > > > RIP: 0010:ext4_es_cache_extent+0x30e/0x370 > > > > > [...] > > > > > Call Trace: > > > > > ext4_cache_extents+0x238/0x2f0 > > > > > ext4_find_extent+0x785/0xa40 > > > > > ext4_fiemap+0x36d/0xe90 > > > > > do_vfs_ioctl+0x6af/0x1200 > > > > > [...] > > > > > ================================================================== > > > > > > > > > > Above issue may happen as follows: > > > > > ------------------------------------- > > > > > cpu1 cpu2 > > > > > _____________________|_____________________ > > > > > do_vfs_ioctl > > > > > ext4_ioctl > > > > > ext4_ioctl_setflags > > > > > ext4_ind_migrate > > > > > do_vfs_ioctl > > > > > ioctl_fiemap > > > > > ext4_fiemap > > > > > ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) > > > > > ext4_fill_fiemap_extents > > > > > down_write(&EXT4_I(inode)->i_data_sem); > > > > > ext4_ext_check_inode > > > > > ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) > > > > > memset(ei->i_data, 0, sizeof(ei->i_data)) > > > > > up_write(&EXT4_I(inode)->i_data_sem); > > > > > down_read(&EXT4_I(inode)->i_data_sem); > > > > > ext4_find_extent > > > > > ext4_cache_extents > > > > > ext4_es_cache_extent > > > > > BUG_ON(end < lblk) > > > > > > > > > > We can easily reproduce this problem with the syzkaller testcase: > > > > > ``` > > > > > 02:37:07 executing program 3: > > > > > r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) > > > > > ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) > > > > > mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) > > > > > r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) > > > > > ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) > > > > > ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) > > > > > ``` > > > > > > > > > > To solve this issue, we use __generic_block_fiemap() instead of > > > > > generic_block_fiemap() and add inode_lock_shared to avoid race condition. > > > > > > > > > > Reported-by: Hulk Robot <hulkci@huawei.com> > > > > > Signed-off-by: Baokun Li <libaokun1@huawei.com> > > > > > --- > > > > > fs/ext4/extents.c | 15 +++++++++++---- > > > > > 1 file changed, 11 insertions(+), 4 deletions(-) > > > > What is the git commit id of this change in Linus's tree? > > > > > > > > If it is not in Linus's tree, why not? > > > > > > > > confused, > > > > > > > > greg k-h > > > > . > > > This patch does not exist in the Linus' tree. > > > > > > This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap > > > to use iomap framework") is incorporated in v5.7-rc1. > > Then why not ask for that change to be added instead? > > > > thanks, > > > > greg k-h > > . > > If we want to switch to the iomap framework, we need to analyze and > integrate about 60 patches. > > The workload may be greater than that of solving this problem alone. 95% of the time we take a patch that is not in Linus's tree, it is buggy and causes problems in the long run. See what those 60 patches really require and if this issue really does need all of that. Or better yet, take the effort here and move off of 4.19 to a newer kernel without this problem in it. What is preventing you from doing that today? 4.19 is not going to be around for forever, and will probably not even be getting fixes for stuff like RETBLEED, so are you _SURE_ you want to keep using it? thanks, greg k-h
On Tue, Jul 19, 2022 at 01:26:57PM +0200, Greg KH wrote: > On Sat, Jul 16, 2022 at 10:33:30AM +0800, Baokun Li wrote: > > This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap > > to use iomap framework") is incorporated in v5.7-rc1. > > Then why not ask for that change to be added instead? Switching over to use the iomap framework is a quite invasive change, which is fraught with danager and potential performance regressions. So it's really not something that would be considered safe for an LTS kernel. As an upstream developer I'd ask why are people trying to use a kernel as old as 4.19, but RHEL has done more insane things than that. Also, I know what the answer is, and it's just too depressing for a nice summer day like this. :-) - Ted
On Tue, Jul 19, 2022 at 02:25:18PM +0200, Greg KH wrote: > > 95% of the time we take a patch that is not in Linus's tree, it is buggy > and causes problems in the long run. So if we really want a 4.19 LTS specific patch, I'd be OK with signing off on it from an ext4 perspective.... IF AND ONLY IF someone is willing to tell me that they ran "kvm-xfstests -c ext4/all -g auto" or the equivalent before and after applying the patch, and is willing to certify that there are no test regressions. Helpful links: * https://thunk.org/gce-xfstests * https://thunk.org/android-xfstests * Documentation links from https://github.com/tytso/xfstests-bld * https://github.com/tytso/xfstests-bld/blob/master/Documentation/kvm-quickstart.md * https://github.com/tytso/xfstests-bld/blob/master/Documentation/gce-xfstests.md (Note that running "-c ext4/all -g auto" will take some 12+ hours if the tests are run serially, which is why using gce-xfstests's lightweight test manager to run the file system test configurations in parallel is a big win.) > Or better yet, take the effort here and move off of 4.19 to a newer > kernel without this problem in it. What is preventing you from doing > that today? 4.19 is not going to be around for forever, and will > probably not even be getting fixes for stuff like RETBLEED, so are you > _SURE_ you want to keep using it? Or yeah, maybe it's better/cheaper/time for you to move off of 4.19. :-) - Ted P.S. If we go down this path, Greg K-H may also insist on getting the bug fix to the 5.4 LTS kernel, so that a bug isn't just fixed in 4.19 LTS but not 5.4 LTS. In which case, the same requirement of running "-c ext4/all -g auto" and showing that there are no test regressions is going to be a requirement for 5.4 LTS as well.
On Tue, Jul 19, 2022 at 02:57:18PM -0400, Theodore Ts'o wrote: > P.S. If we go down this path, Greg K-H may also insist on getting the > bug fix to the 5.4 LTS kernel, so that a bug isn't just fixed in 4.19 > LTS but not 5.4 LTS. In which case, the same requirement of running > "-c ext4/all -g auto" and showing that there are no test regressions > is going to be a requirement for 5.4 LTS as well. Yes, if this issue is also in 5.4 or any newer kernels, it HAS to be fixed there at the same time, or before, as we can not have anyone upgrading from an older kernel to a newer one and having a regression. thanks, greg k-h
在 2022/7/19 22:06, Theodore Ts'o 写道: > On Tue, Jul 19, 2022 at 01:26:57PM +0200, Greg KH wrote: >> On Sat, Jul 16, 2022 at 10:33:30AM +0800, Baokun Li wrote: >>> This problem persists until the patch d3b6f23f7167("ext4: move ext4_fiemap >>> to use iomap framework") is incorporated in v5.7-rc1. >> Then why not ask for that change to be added instead? > Switching over to use the iomap framework is a quite invasive change, > which is fraught with danager and potential performance regressions. > So it's really not something that would be considered safe for an LTS > kernel. > > As an upstream developer I'd ask why are people trying to use a kernel > as old as 4.19, but RHEL has done more insane things than that. Also, > I know what the answer is, and it's just too depressing for a nice > summer day like this. :-) > > - Ted > . Indeed.
在 2022/7/20 3:14, Greg KH 写道: > On Tue, Jul 19, 2022 at 02:57:18PM -0400, Theodore Ts'o wrote: >> P.S. If we go down this path, Greg K-H may also insist on getting the >> bug fix to the 5.4 LTS kernel, so that a bug isn't just fixed in 4.19 >> LTS but not 5.4 LTS. In which case, the same requirement of running >> "-c ext4/all -g auto" and showing that there are no test regressions >> is going to be a requirement for 5.4 LTS as well. > Yes, if this issue is also in 5.4 or any newer kernels, it HAS to be > fixed there at the same time, or before, as we can not have anyone > upgrading from an older kernel to a newer one and having a regression. > > thanks, > > greg k-h > . Indeed.
diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c index 6c492fca60c4..38aaf48e94cb 100644 --- a/fs/ext4/extents.c +++ b/fs/ext4/extents.c @@ -5198,13 +5198,18 @@ int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, return error; } + inode_lock_shared(inode); /* fallback to generic here if not in extents fmt */ - if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) - return generic_block_fiemap(inode, fieinfo, start, len, + if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) { + error = __generic_block_fiemap(inode, fieinfo, start, len, ext4_get_block); + goto out_unlock; + } - if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS)) - return -EBADR; + if (fiemap_check_flags(fieinfo, EXT4_FIEMAP_FLAGS)) { + error = -EBADR; + goto out_unlock; + } if (fieinfo->fi_flags & FIEMAP_FLAG_XATTR) { error = ext4_xattr_fiemap(inode, fieinfo); @@ -5225,6 +5230,8 @@ int ext4_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo, error = ext4_fill_fiemap_extents(inode, start_blk, len_blks, fieinfo); } +out_unlock: + inode_unlock_shared(inode); return error; }
This patch and problem analysis is based on v4.19 LTS. The d3b6f23f7167("ext4: move ext4_fiemap to use iomap framework") patch is incorporated in v5.7-rc1. This patch avoids this problem by switching to iomap in ext4_fiemap. Hulk Robot reported a BUG on stable 4.19.252: ================================================================== kernel BUG at fs/ext4/extents_status.c:762! invalid opcode: 0000 [#1] SMP KASAN PTI CPU: 7 PID: 2845 Comm: syz-executor Not tainted 4.19.252 #46 RIP: 0010:ext4_es_cache_extent+0x30e/0x370 [...] Call Trace: ext4_cache_extents+0x238/0x2f0 ext4_find_extent+0x785/0xa40 ext4_fiemap+0x36d/0xe90 do_vfs_ioctl+0x6af/0x1200 [...] ================================================================== Above issue may happen as follows: ------------------------------------- cpu1 cpu2 _____________________|_____________________ do_vfs_ioctl ext4_ioctl ext4_ioctl_setflags ext4_ind_migrate do_vfs_ioctl ioctl_fiemap ext4_fiemap ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS) ext4_fill_fiemap_extents down_write(&EXT4_I(inode)->i_data_sem); ext4_ext_check_inode ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS) memset(ei->i_data, 0, sizeof(ei->i_data)) up_write(&EXT4_I(inode)->i_data_sem); down_read(&EXT4_I(inode)->i_data_sem); ext4_find_extent ext4_cache_extents ext4_es_cache_extent BUG_ON(end < lblk) We can easily reproduce this problem with the syzkaller testcase: ``` 02:37:07 executing program 3: r0 = openat(0xffffffffffffff9c, &(0x7f0000000040)='./file0\x00', 0x26e1, 0x0) ioctl$FS_IOC_FSSETXATTR(r0, 0x40086602, &(0x7f0000000080)={0x17e}) mkdirat(0xffffffffffffff9c, &(0x7f00000000c0)='./file1\x00', 0x1ff) r1 = openat(0xffffffffffffff9c, &(0x7f0000000100)='./file1\x00', 0x0, 0x0) ioctl$FS_IOC_FIEMAP(r1, 0xc020660b, &(0x7f0000000180)={0x0, 0x1, 0x0, 0xef3, 0x6, []}) (async, rerun: 32) ioctl$FS_IOC_FSSETXATTR(r1, 0x40086602, &(0x7f0000000140)={0x17e}) (rerun: 32) ``` To solve this issue, we use __generic_block_fiemap() instead of generic_block_fiemap() and add inode_lock_shared to avoid race condition. Reported-by: Hulk Robot <hulkci@huawei.com> Signed-off-by: Baokun Li <libaokun1@huawei.com> --- fs/ext4/extents.c | 15 +++++++++++---- 1 file changed, 11 insertions(+), 4 deletions(-)