| Message ID | 20211013095101.641329-1-chenxiaosong2@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [4.19,v2] VFS: Fix fuseblk memory leak caused by mount concurrency | expand |
On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: > If two processes mount same superblock, memory leak occurs: > > CPU0 | CPU1 > do_new_mount | do_new_mount > fs_set_subtype | fs_set_subtype > kstrdup | > | kstrdup > memrory leak | > > Fix this by adding a write lock while calling fs_set_subtype. > > Linus's tree already have refactoring patchset [1], one of them can fix this bug: > c30da2e981a7 (fuse: convert to use the new mount API) > > Since we did not merge the refactoring patchset in this branch, I create this patch. > > [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ > > Fixes: 79c0b2df79eb (add filesystem subtype support) > Cc: David Howells <dhowells@redhat.com> > Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> > --- > v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk memory leak caused by mount concurrency) > v2: Use write lock while writing superblock > > fs/namespace.c | 9 ++++++--- > 1 file changed, 6 insertions(+), 3 deletions(-) As you are referring to a fuse-only patch above, why are you trying to resolve this issue in the core namespace code instead? How does fuse have anything to do with this? confused, greg k-h
在 2021/10/13 18:11, Greg KH 写道: > On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: >> If two processes mount same superblock, memory leak occurs: >> >> CPU0 | CPU1 >> do_new_mount | do_new_mount >> fs_set_subtype | fs_set_subtype >> kstrdup | >> | kstrdup >> memrory leak | >> >> Fix this by adding a write lock while calling fs_set_subtype. >> >> Linus's tree already have refactoring patchset [1], one of them can fix this bug: >> c30da2e981a7 (fuse: convert to use the new mount API) >> >> Since we did not merge the refactoring patchset in this branch, I create this patch. >> >> [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ >> >> Fixes: 79c0b2df79eb (add filesystem subtype support) >> Cc: David Howells <dhowells@redhat.com> >> Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> >> --- >> v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk memory leak caused by mount concurrency) >> v2: Use write lock while writing superblock >> >> fs/namespace.c | 9 ++++++--- >> 1 file changed, 6 insertions(+), 3 deletions(-) > > As you are referring to a fuse-only patch above, why are you trying to > resolve this issue in the core namespace code instead? > > How does fuse have anything to do with this? > > confused, > > greg k-h > . > Now, only `fuse_fs_type` and `fuseblk_fs_type` has `FS_HAS_SUBTYPE` flag in kernel code, but maybe there is a filesystem module(`struct file_system_type` has `FS_HAS_SUBTYPE` flag). And only mounting fuseblk filesystem(e.g. ntfs) will occur memory leak now.
在 2021/10/13 18:38, chenxiaosong (A) 写道: > 在 2021/10/13 18:11, Greg KH 写道: >> On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: >>> If two processes mount same superblock, memory leak occurs: >>> >>> CPU0 | CPU1 >>> do_new_mount | do_new_mount >>> fs_set_subtype | fs_set_subtype >>> kstrdup | >>> | kstrdup >>> memrory leak | >>> >>> Fix this by adding a write lock while calling fs_set_subtype. >>> >>> Linus's tree already have refactoring patchset [1], one of them can >>> fix this bug: >>> c30da2e981a7 (fuse: convert to use the new mount API) >>> >>> Since we did not merge the refactoring patchset in this branch, I >>> create this patch. >>> >>> [1] >>> https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ >>> >>> >>> Fixes: 79c0b2df79eb (add filesystem subtype support) >>> Cc: David Howells <dhowells@redhat.com> >>> Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> >>> --- >>> v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk memory >>> leak caused by mount concurrency) >>> v2: Use write lock while writing superblock >>> >>> fs/namespace.c | 9 ++++++--- >>> 1 file changed, 6 insertions(+), 3 deletions(-) >> >> As you are referring to a fuse-only patch above, why are you trying to >> resolve this issue in the core namespace code instead? >> >> How does fuse have anything to do with this? >> >> confused, >> >> greg k-h >> . >> > > Now, only `fuse_fs_type` and `fuseblk_fs_type` has `FS_HAS_SUBTYPE` flag > in kernel code, but maybe there is a filesystem module(`struct > file_system_type` has `FS_HAS_SUBTYPE` flag). And only mounting fuseblk > filesystem(e.g. ntfs) will occur memory leak now. How about updating the subject as: VFS: Fix memory leak caused by mounting fs with subtype concurrency?
On Wed, Oct 13, 2021 at 06:49:06PM +0800, chenxiaosong (A) wrote: > 在 2021/10/13 18:38, chenxiaosong (A) 写道: > > 在 2021/10/13 18:11, Greg KH 写道: > > > On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: > > > > If two processes mount same superblock, memory leak occurs: > > > > > > > > CPU0 | CPU1 > > > > do_new_mount | do_new_mount > > > > fs_set_subtype | fs_set_subtype > > > > kstrdup | > > > > | kstrdup > > > > memrory leak | > > > > > > > > Fix this by adding a write lock while calling fs_set_subtype. > > > > > > > > Linus's tree already have refactoring patchset [1], one of them > > > > can fix this bug: > > > > c30da2e981a7 (fuse: convert to use the new mount API) > > > > > > > > Since we did not merge the refactoring patchset in this branch, > > > > I create this patch. > > > > > > > > [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ > > > > > > > > > > > > Fixes: 79c0b2df79eb (add filesystem subtype support) > > > > Cc: David Howells <dhowells@redhat.com> > > > > Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> > > > > --- > > > > v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk > > > > memory leak caused by mount concurrency) > > > > v2: Use write lock while writing superblock > > > > > > > > fs/namespace.c | 9 ++++++--- > > > > 1 file changed, 6 insertions(+), 3 deletions(-) > > > > > > As you are referring to a fuse-only patch above, why are you trying to > > > resolve this issue in the core namespace code instead? > > > > > > How does fuse have anything to do with this? > > > > > > confused, > > > > > > greg k-h > > > . > > > > > > > Now, only `fuse_fs_type` and `fuseblk_fs_type` has `FS_HAS_SUBTYPE` flag > > in kernel code, but maybe there is a filesystem module(`struct > > file_system_type` has `FS_HAS_SUBTYPE` flag). And only mounting fuseblk > > filesystem(e.g. ntfs) will occur memory leak now. > > How about updating the subject as: VFS: Fix memory leak caused by mounting > fs with subtype concurrency? That would be a better idea, but still, this is not obvious that this is the correct fix at all...
在 2021/10/20 20:30, Greg KH 写道: > On Wed, Oct 13, 2021 at 06:49:06PM +0800, chenxiaosong (A) wrote: >> 在 2021/10/13 18:38, chenxiaosong (A) 写道: >>> 在 2021/10/13 18:11, Greg KH 写道: >>>> On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: >>>>> If two processes mount same superblock, memory leak occurs: >>>>> >>>>> CPU0 | CPU1 >>>>> do_new_mount | do_new_mount >>>>> fs_set_subtype | fs_set_subtype >>>>> kstrdup | >>>>> | kstrdup >>>>> memrory leak | >>>>> >>>>> Fix this by adding a write lock while calling fs_set_subtype. >>>>> >>>>> Linus's tree already have refactoring patchset [1], one of them >>>>> can fix this bug: >>>>> c30da2e981a7 (fuse: convert to use the new mount API) >>>>> >>>>> Since we did not merge the refactoring patchset in this branch, >>>>> I create this patch. >>>>> >>>>> [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ >>>>> >>>>> >>>>> Fixes: 79c0b2df79eb (add filesystem subtype support) >>>>> Cc: David Howells <dhowells@redhat.com> >>>>> Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> >>>>> --- >>>>> v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk >>>>> memory leak caused by mount concurrency) >>>>> v2: Use write lock while writing superblock >>>>> >>>>> fs/namespace.c | 9 ++++++--- >>>>> 1 file changed, 6 insertions(+), 3 deletions(-) >>>> >>>> As you are referring to a fuse-only patch above, why are you trying to >>>> resolve this issue in the core namespace code instead? >>>> >>>> How does fuse have anything to do with this? >>>> >>>> confused, >>>> >>>> greg k-h >>>> . >>>> >>> >>> Now, only `fuse_fs_type` and `fuseblk_fs_type` has `FS_HAS_SUBTYPE` flag >>> in kernel code, but maybe there is a filesystem module(`struct >>> file_system_type` has `FS_HAS_SUBTYPE` flag). And only mounting fuseblk >>> filesystem(e.g. ntfs) will occur memory leak now. >> >> How about updating the subject as: VFS: Fix memory leak caused by mounting >> fs with subtype concurrency? > > That would be a better idea, but still, this is not obvious that this is > the correct fix at all... > . > Why is this patch not correct? Can you tell me more about it? Thanks.
On Tue, Oct 26, 2021 at 10:18:11AM +0800, chenxiaosong (A) wrote: > > > 在 2021/10/20 20:30, Greg KH 写道: > > On Wed, Oct 13, 2021 at 06:49:06PM +0800, chenxiaosong (A) wrote: > > > 在 2021/10/13 18:38, chenxiaosong (A) 写道: > > > > 在 2021/10/13 18:11, Greg KH 写道: > > > > > On Wed, Oct 13, 2021 at 05:51:01PM +0800, ChenXiaoSong wrote: > > > > > > If two processes mount same superblock, memory leak occurs: > > > > > > > > > > > > CPU0 | CPU1 > > > > > > do_new_mount | do_new_mount > > > > > > fs_set_subtype | fs_set_subtype > > > > > > kstrdup | > > > > > > | kstrdup > > > > > > memrory leak | > > > > > > > > > > > > Fix this by adding a write lock while calling fs_set_subtype. > > > > > > > > > > > > Linus's tree already have refactoring patchset [1], one of them > > > > > > can fix this bug: > > > > > > c30da2e981a7 (fuse: convert to use the new mount API) > > > > > > > > > > > > Since we did not merge the refactoring patchset in this branch, > > > > > > I create this patch. > > > > > > > > > > > > [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ > > > > > > > > > > > > > > > > > > Fixes: 79c0b2df79eb (add filesystem subtype support) > > > > > > Cc: David Howells <dhowells@redhat.com> > > > > > > Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> > > > > > > --- > > > > > > v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk > > > > > > memory leak caused by mount concurrency) > > > > > > v2: Use write lock while writing superblock > > > > > > > > > > > > fs/namespace.c | 9 ++++++--- > > > > > > 1 file changed, 6 insertions(+), 3 deletions(-) > > > > > > > > > > As you are referring to a fuse-only patch above, why are you trying to > > > > > resolve this issue in the core namespace code instead? > > > > > > > > > > How does fuse have anything to do with this? > > > > > > > > > > confused, > > > > > > > > > > greg k-h > > > > > . > > > > > > > > > > > > > Now, only `fuse_fs_type` and `fuseblk_fs_type` has `FS_HAS_SUBTYPE` flag > > > > in kernel code, but maybe there is a filesystem module(`struct > > > > file_system_type` has `FS_HAS_SUBTYPE` flag). And only mounting fuseblk > > > > filesystem(e.g. ntfs) will occur memory leak now. > > > > > > How about updating the subject as: VFS: Fix memory leak caused by mounting > > > fs with subtype concurrency? > > > > That would be a better idea, but still, this is not obvious that this is > > the correct fix at all... > > . > > > Why is this patch not correct? Can you tell me more about it? Thanks. You need to prove that it is correct, and you need to get maintainers to approve it. thanks, greg k-h
diff --git a/fs/namespace.c b/fs/namespace.c index 2f3c6a0350a8..396ff1bcfdad 100644 --- a/fs/namespace.c +++ b/fs/namespace.c @@ -2490,9 +2490,12 @@ static int do_new_mount(struct path *path, const char *fstype, int sb_flags, return -ENODEV; mnt = vfs_kern_mount(type, sb_flags, name, data); - if (!IS_ERR(mnt) && (type->fs_flags & FS_HAS_SUBTYPE) && - !mnt->mnt_sb->s_subtype) - mnt = fs_set_subtype(mnt, fstype); + if (!IS_ERR(mnt) && (type->fs_flags & FS_HAS_SUBTYPE)) { + down_write(&mnt->mnt_sb->s_umount); + if (!mnt->mnt_sb->s_subtype) + mnt = fs_set_subtype(mnt, fstype); + up_write(&mnt->mnt_sb->s_umount); + } put_filesystem(type); if (IS_ERR(mnt))
If two processes mount same superblock, memory leak occurs: CPU0 | CPU1 do_new_mount | do_new_mount fs_set_subtype | fs_set_subtype kstrdup | | kstrdup memrory leak | Fix this by adding a write lock while calling fs_set_subtype. Linus's tree already have refactoring patchset [1], one of them can fix this bug: c30da2e981a7 (fuse: convert to use the new mount API) Since we did not merge the refactoring patchset in this branch, I create this patch. [1] https://patchwork.kernel.org/project/linux-fsdevel/patch/20190903113640.7984-3-mszeredi@redhat.com/ Fixes: 79c0b2df79eb (add filesystem subtype support) Cc: David Howells <dhowells@redhat.com> Signed-off-by: ChenXiaoSong <chenxiaosong2@huawei.com> --- v1: Can not mount sshfs ([PATCH linux-4.19.y] VFS: Fix fuseblk memory leak caused by mount concurrency) v2: Use write lock while writing superblock fs/namespace.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)