JeffleXu
2021-Aug-20 05:03 UTC
[Virtio-fs] [virtiofsd PATCH v4 4/4] virtiofsd: support per-file DAX in FUSE_LOOKUP
On 8/19/21 9:08 PM, Dr. David Alan Gilbert wrote:> * JeffleXu (jefflexu at linux.alibaba.com) wrote: >> >> >> On 8/18/21 3:00 AM, Dr. David Alan Gilbert wrote: >>> * Jeffle Xu (jefflexu at linux.alibaba.com) wrote: >>>> For passthrough, when the corresponding virtiofs in guest is mounted >>>> with '-o dax=inode', advertise that the file is capable of per-file >>>> DAX if the inode in the backend fs is marked with FS_DAX_FL flag. >>>> >>>> Signed-off-by: Jeffle Xu <jefflexu at linux.alibaba.com> >>>> --- >>>> tools/virtiofsd/passthrough_ll.c | 43 ++++++++++++++++++++++++++++++++ >>>> 1 file changed, 43 insertions(+) >>>> >>>> diff --git a/tools/virtiofsd/passthrough_ll.c b/tools/virtiofsd/passthrough_ll.c >>>> index 5b6228210f..4cbd904248 100644 >>>> --- a/tools/virtiofsd/passthrough_ll.c >>>> +++ b/tools/virtiofsd/passthrough_ll.c >>>> @@ -171,6 +171,7 @@ struct lo_data { >>>> int allow_direct_io; >>>> int announce_submounts; >>>> int perfile_dax_cap; /* capability of backend fs */ >>>> + bool perfile_dax; /* enable per-file DAX or not */ >>>> bool use_statx; >>>> struct lo_inode root; >>>> GHashTable *inodes; /* protected by lo->mutex */ >>>> @@ -716,6 +717,10 @@ static void lo_init(void *userdata, struct fuse_conn_info *conn) >>>> >>>> if (conn->capable & FUSE_CAP_PERFILE_DAX && lo->perfile_dax_cap ) { >>>> conn->want |= FUSE_CAP_PERFILE_DAX; >>>> + lo->perfile_dax = 1; >>>> + } >>>> + else { >>>> + lo->perfile_dax = 0; >>>> } >>>> } >>>> >>>> @@ -983,6 +988,41 @@ static int do_statx(struct lo_data *lo, int dirfd, const char *pathname, >>>> return 0; >>>> } >>>> >>>> +/* >>>> + * If the file is marked with FS_DAX_FL or FS_XFLAG_DAX, then DAX should be >>>> + * enabled for this file. >>>> + */ >>>> +static bool lo_should_enable_dax(struct lo_data *lo, struct lo_inode *dir, >>>> + const char *name) >>>> +{ >>>> + int res, fd; >>>> + int ret = false;; >>>> + unsigned int attr; >>>> + struct fsxattr xattr; >>>> + >>>> + if (!lo->perfile_dax) >>>> + return false; >>>> + >>>> + /* Open file without O_PATH, so that ioctl can be called. */ >>>> + fd = openat(dir->fd, name, O_NOFOLLOW); >>>> + if (fd == -1) >>>> + return false; >>> >>> Doesn't that defeat the whole benefit of using O_PATH - i.e. that we >>> might stumble into a /dev node or something else we're not allowed to >>> open? >> >> As far as I know, virtiofsd will pivot_root/chroot to the source >> directory, and can only access files inside the source directory >> specified by "-o source=". Then where do these unexpected files come >> from? Besides, fd opened without O_PATH here is temporary and used for >> FS_IOC_GETFLAGS/FS_IOC_FSGETXATTR ioctl only. It's closed when the >> function returns. > > The guest is still allowed to mknod. > See: > https://lists.gnu.org/archive/html/qemu-devel/2021-01/msg05461.html > > also it's legal to expose a root filesystem for a guest; the virtiofsd > should *never* open a device other than O_PATH - and it's really tricky > to do a check to see if it is a device in a race-free way. >Fine. Got it. However the returned fd (opened without O_PATH) is only used for FS_IOC_GETFLAGS/FS_IOC_FSGETXATTR ioctl, while in most cases for special device files, these two ioctls should return -ENOTTY. If it's really a security issue, then lo_inode_open() could be used to get a temporary fd, i.e., check if it's a special file before opening. After all, FUSE_OPEN also handles in this way. Besides, I can't understand what "race-free way" means. -- Thanks, Jeffle
Greg Kurz
2021-Aug-24 10:15 UTC
[Virtio-fs] [virtiofsd PATCH v4 4/4] virtiofsd: support per-file DAX in FUSE_LOOKUP
On Fri, 20 Aug 2021 13:03:23 +0800 JeffleXu <jefflexu at linux.alibaba.com> wrote:> > > On 8/19/21 9:08 PM, Dr. David Alan Gilbert wrote: > > * JeffleXu (jefflexu at linux.alibaba.com) wrote: > >> > >> > >> On 8/18/21 3:00 AM, Dr. David Alan Gilbert wrote: > >>> * Jeffle Xu (jefflexu at linux.alibaba.com) wrote: > >>>> For passthrough, when the corresponding virtiofs in guest is mounted > >>>> with '-o dax=inode', advertise that the file is capable of per-file > >>>> DAX if the inode in the backend fs is marked with FS_DAX_FL flag. > >>>> > >>>> Signed-off-by: Jeffle Xu <jefflexu at linux.alibaba.com> > >>>> --- > >>>> tools/virtiofsd/passthrough_ll.c | 43 ++++++++++++++++++++++++++++++++ > >>>> 1 file changed, 43 insertions(+) > >>>> > >>>> diff --git a/tools/virtiofsd/passthrough_ll.c b/tools/virtiofsd/passthrough_ll.c > >>>> index 5b6228210f..4cbd904248 100644 > >>>> --- a/tools/virtiofsd/passthrough_ll.c > >>>> +++ b/tools/virtiofsd/passthrough_ll.c > >>>> @@ -171,6 +171,7 @@ struct lo_data { > >>>> int allow_direct_io; > >>>> int announce_submounts; > >>>> int perfile_dax_cap; /* capability of backend fs */ > >>>> + bool perfile_dax; /* enable per-file DAX or not */ > >>>> bool use_statx; > >>>> struct lo_inode root; > >>>> GHashTable *inodes; /* protected by lo->mutex */ > >>>> @@ -716,6 +717,10 @@ static void lo_init(void *userdata, struct fuse_conn_info *conn) > >>>> > >>>> if (conn->capable & FUSE_CAP_PERFILE_DAX && lo->perfile_dax_cap ) { > >>>> conn->want |= FUSE_CAP_PERFILE_DAX; > >>>> + lo->perfile_dax = 1; > >>>> + } > >>>> + else { > >>>> + lo->perfile_dax = 0; > >>>> } > >>>> } > >>>> > >>>> @@ -983,6 +988,41 @@ static int do_statx(struct lo_data *lo, int dirfd, const char *pathname, > >>>> return 0; > >>>> } > >>>> > >>>> +/* > >>>> + * If the file is marked with FS_DAX_FL or FS_XFLAG_DAX, then DAX should be > >>>> + * enabled for this file. > >>>> + */ > >>>> +static bool lo_should_enable_dax(struct lo_data *lo, struct lo_inode *dir, > >>>> + const char *name) > >>>> +{ > >>>> + int res, fd; > >>>> + int ret = false;; > >>>> + unsigned int attr; > >>>> + struct fsxattr xattr; > >>>> + > >>>> + if (!lo->perfile_dax) > >>>> + return false; > >>>> + > >>>> + /* Open file without O_PATH, so that ioctl can be called. */ > >>>> + fd = openat(dir->fd, name, O_NOFOLLOW); > >>>> + if (fd == -1) > >>>> + return false; > >>> > >>> Doesn't that defeat the whole benefit of using O_PATH - i.e. that we > >>> might stumble into a /dev node or something else we're not allowed to > >>> open? > >> > >> As far as I know, virtiofsd will pivot_root/chroot to the source > >> directory, and can only access files inside the source directory > >> specified by "-o source=". Then where do these unexpected files come > >> from? Besides, fd opened without O_PATH here is temporary and used for > >> FS_IOC_GETFLAGS/FS_IOC_FSGETXATTR ioctl only. It's closed when the > >> function returns. > > > > The guest is still allowed to mknod. > > See: > > https://lists.gnu.org/archive/html/qemu-devel/2021-01/msg05461.html > > > > also it's legal to expose a root filesystem for a guest; the virtiofsd > > should *never* open a device other than O_PATH - and it's really tricky > > to do a check to see if it is a device in a race-free way. > > > > Fine. Got it. However the returned fd (opened without O_PATH) is only > used for FS_IOC_GETFLAGS/FS_IOC_FSGETXATTR ioctl, while in most cases > for special device files, these two ioctls should return -ENOTTY. >The actual problem is that a FIFO will cause openat() to block until the other end of the FIFO is open for writing...> If it's really a security issue, then lo_inode_open() could be used to... and cause a DoS on virtiofsd. So yes, this is a security issue and lo_inode_open() was introduced specifically to handle this.> get a temporary fd, i.e., check if it's a special file before opening. > After all, FUSE_OPEN also handles in this way. Besides, I can't > understand what "race-free way" means. >"race-free way" means a way that guarantees that file type cannot change between the time you check it and the time you open it (TOCTOU error). For example, doing a plain stat(), checking st_mode and proceeding to open() is wrong : nothing prevents the file to be unlinked and replaced by something else between stat() and open(). We avoid that by keeping O_PATH fds around and using lo_inode_open() instead of openat(). In your case, it seems that you should do the checking after you have an actual lo_inode for the target file, and pass that to lo_should_enable_dax() instead of the parent lo_inode and target name. Cheers, -- Greg
JeffleXu
2021-Sep-08 10:34 UTC
[Virtio-fs] [virtiofsd PATCH v4 4/4] virtiofsd: support per-file DAX in FUSE_LOOKUP
On 8/24/21 6:15 PM, Greg Kurz wrote:> On Fri, 20 Aug 2021 13:03:23 +0800 > JeffleXu <jefflexu at linux.alibaba.com> wrote: >> >> Fine. Got it. However the returned fd (opened without O_PATH) is only >> used for FS_IOC_GETFLAGS/FS_IOC_FSGETXATTR ioctl, while in most cases >> for special device files, these two ioctls should return -ENOTTY. >> > > The actual problem is that a FIFO will cause openat() to block until > the other end of the FIFO is open for writing...Got it.> >> If it's really a security issue, then lo_inode_open() could be used to > > ... and cause a DoS on virtiofsd. So yes, this is a security issue and > lo_inode_open() was introduced specifically to handle this. > >> get a temporary fd, i.e., check if it's a special file before opening. >> After all, FUSE_OPEN also handles in this way. Besides, I can't >> understand what "race-free way" means. >> > > "race-free way" means a way that guarantees that file type > cannot change between the time you check it and the time > you open it (TOCTOU error). For example, doing a plain stat(), > checking st_mode and proceeding to open() is wrong : nothing > prevents the file to be unlinked and replaced by something > else between stat() and open(). > > We avoid that by keeping O_PATH fds around and using > lo_inode_open() instead of openat().Thanks for the detailed explanation. Got it.> > In your case, it seems that you should do the checking after > you have an actual lo_inode for the target file, and pass > that to lo_should_enable_dax() instead of the parent lo_inode > and target name. >Yes, that will be more reasonable. Thanks. -- Thanks, Jeffle