| Message ID | 20221213160104.198237-1-jlayton@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [nfs-utils] Don't allow junction tests to trigger automounts | expand |
> On Dec 13, 2022, at 11:01 AM, Jeff Layton <jlayton@kernel.org> wrote: > > JianHong reported some strange behavior with automounts on an nfs server > without an explicit pseudoroot. When clients issued a readdir in the > pseudoroot, automounted directories that were not yet mounted would show > up even if they weren't exported, though the clients wouldn't be able to > do anything with them. > > The issue was that triggering the automount on a directory would cause > the mountd upcall to time out, which would cause nfsd to include the > automounted dentry in the readdir response. Eventually, the automount > would work and report that it wasn't exported and subsequent attempts to > access the dentry would (properly) fail. > > We never want mountd to trigger an automount. The kernel should do that > if it wants to use it. Change the junction checks to do an O_PATH open > and use fstatat with AT_NO_AUTOMOUNT. > > Cc: Chuck Lever <chuck.lever@oracle.com> > Link: https://bugzilla.redhat.com/show_bug.cgi?id=2148353 And also https://bugzilla.kernel.org/show_bug.cgi?id=216777 ? > Reported-by: JianHong Yin <jiyin@redhat.com> > Signed-off-by: Jeff Layton <jlayton@kernel.org> > --- > support/junction/junction.c | 10 +++++----- > 1 file changed, 5 insertions(+), 5 deletions(-) > > diff --git a/support/junction/junction.c b/support/junction/junction.c > index 41cce261cb52..0628bb0ffffb 100644 > --- a/support/junction/junction.c > +++ b/support/junction/junction.c > @@ -63,7 +63,7 @@ junction_open_path(const char *pathname, int *fd) > if (pathname == NULL || fd == NULL) > return FEDFS_ERR_INVAL; > > - tmp = open(pathname, O_DIRECTORY); > + tmp = open(pathname, O_PATH|O_DIRECTORY); > if (tmp == -1) { > switch (errno) { > case EPERM: > @@ -93,7 +93,7 @@ junction_is_directory(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -121,7 +121,7 @@ junction_is_sticky_bit_set(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -155,7 +155,7 @@ junction_set_sticky_bit(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -393,7 +393,7 @@ junction_get_mode(const char *pathname, mode_t *mode) > if (retval != FEDFS_OK) > return retval; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, pathname); > (void)close(fd); > -- > 2.38.1 > -- Chuck Lever
On Tue, 2022-12-13 at 16:25 +0000, Chuck Lever III wrote: > > > On Dec 13, 2022, at 11:01 AM, Jeff Layton <jlayton@kernel.org> wrote: > > > > JianHong reported some strange behavior with automounts on an nfs server > > without an explicit pseudoroot. When clients issued a readdir in the > > pseudoroot, automounted directories that were not yet mounted would show > > up even if they weren't exported, though the clients wouldn't be able to > > do anything with them. > > > > The issue was that triggering the automount on a directory would cause > > the mountd upcall to time out, which would cause nfsd to include the > > automounted dentry in the readdir response. Eventually, the automount > > would work and report that it wasn't exported and subsequent attempts to > > access the dentry would (properly) fail. > > > > We never want mountd to trigger an automount. The kernel should do that > > if it wants to use it. Change the junction checks to do an O_PATH open > > and use fstatat with AT_NO_AUTOMOUNT. > > > > Cc: Chuck Lever <chuck.lever@oracle.com> > > Link: https://bugzilla.redhat.com/show_bug.cgi?id=2148353 > > And also https://bugzilla.kernel.org/show_bug.cgi?id=216777 ? > Yes indeed, thanks! Note too that I suspect there may also be a kernel error handling bug related to this. I know that nfsd_cross_mnt returned -ETIMEDOUT, but the READDIR response still contained the dentry. The follow-on stat call failed, but it seems like the readdir response shouldn't have included that dentry in the first place. I'm still looking at that bit, but we should probably aim to fix that too. > > > Reported-by: JianHong Yin <jiyin@redhat.com> > > Signed-off-by: Jeff Layton <jlayton@kernel.org> > > --- > > support/junction/junction.c | 10 +++++----- > > 1 file changed, 5 insertions(+), 5 deletions(-) > > > > diff --git a/support/junction/junction.c b/support/junction/junction.c > > index 41cce261cb52..0628bb0ffffb 100644 > > --- a/support/junction/junction.c > > +++ b/support/junction/junction.c > > @@ -63,7 +63,7 @@ junction_open_path(const char *pathname, int *fd) > > if (pathname == NULL || fd == NULL) > > return FEDFS_ERR_INVAL; > > > > - tmp = open(pathname, O_DIRECTORY); > > + tmp = open(pathname, O_PATH|O_DIRECTORY); > > if (tmp == -1) { > > switch (errno) { > > case EPERM: > > @@ -93,7 +93,7 @@ junction_is_directory(int fd, const char *path) > > { > > struct stat stb; > > > > - if (fstat(fd, &stb) == -1) { > > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > > xlog(D_GENERAL, "%s: failed to stat %s: %m", > > __func__, path); > > return FEDFS_ERR_ACCESS; > > @@ -121,7 +121,7 @@ junction_is_sticky_bit_set(int fd, const char *path) > > { > > struct stat stb; > > > > - if (fstat(fd, &stb) == -1) { > > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > > xlog(D_GENERAL, "%s: failed to stat %s: %m", > > __func__, path); > > return FEDFS_ERR_ACCESS; > > @@ -155,7 +155,7 @@ junction_set_sticky_bit(int fd, const char *path) > > { > > struct stat stb; > > > > - if (fstat(fd, &stb) == -1) { > > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > > xlog(D_GENERAL, "%s: failed to stat %s: %m", > > __func__, path); > > return FEDFS_ERR_ACCESS; > > @@ -393,7 +393,7 @@ junction_get_mode(const char *pathname, mode_t *mode) > > if (retval != FEDFS_OK) > > return retval; > > > > - if (fstat(fd, &stb) == -1) { > > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > > xlog(D_GENERAL, "%s: failed to stat %s: %m", > > __func__, pathname); > > (void)close(fd); > > -- > > 2.38.1 > > > > -- > Chuck Lever > > >
On Tue, 2022-12-13 at 11:01 -0500, Jeff Layton wrote: > JianHong reported some strange behavior with automounts on an nfs server > without an explicit pseudoroot. When clients issued a readdir in the > pseudoroot, automounted directories that were not yet mounted would show > up even if they weren't exported, though the clients wouldn't be able to > do anything with them. > > The issue was that triggering the automount on a directory would cause > the mountd upcall to time out, which would cause nfsd to include the > automounted dentry in the readdir response. Eventually, the automount > would work and report that it wasn't exported and subsequent attempts to > access the dentry would (properly) fail. > > We never want mountd to trigger an automount. The kernel should do that > if it wants to use it. Change the junction checks to do an O_PATH open > and use fstatat with AT_NO_AUTOMOUNT. > > Cc: Chuck Lever <chuck.lever@oracle.com> > Link: https://bugzilla.redhat.com/show_bug.cgi?id=2148353 > Reported-by: JianHong Yin <jiyin@redhat.com> > Signed-off-by: Jeff Layton <jlayton@kernel.org> > --- > support/junction/junction.c | 10 +++++----- > 1 file changed, 5 insertions(+), 5 deletions(-) > Ping? Steve, can you comment on (or merge) this patch? Thanks,
On 12/13/22 11:01 AM, Jeff Layton wrote: > JianHong reported some strange behavior with automounts on an nfs server > without an explicit pseudoroot. When clients issued a readdir in the > pseudoroot, automounted directories that were not yet mounted would show > up even if they weren't exported, though the clients wouldn't be able to > do anything with them. > > The issue was that triggering the automount on a directory would cause > the mountd upcall to time out, which would cause nfsd to include the > automounted dentry in the readdir response. Eventually, the automount > would work and report that it wasn't exported and subsequent attempts to > access the dentry would (properly) fail. > > We never want mountd to trigger an automount. The kernel should do that > if it wants to use it. Change the junction checks to do an O_PATH open > and use fstatat with AT_NO_AUTOMOUNT. > > Cc: Chuck Lever <chuck.lever@oracle.com> > Link: https://bugzilla.redhat.com/show_bug.cgi?id=2148353 > Reported-by: JianHong Yin <jiyin@redhat.com> > Signed-off-by: Jeff Layton <jlayton@kernel.org> Committed... (tag: nfs-utils-2-6-3-rc6) steved > --- > support/junction/junction.c | 10 +++++----- > 1 file changed, 5 insertions(+), 5 deletions(-) > > diff --git a/support/junction/junction.c b/support/junction/junction.c > index 41cce261cb52..0628bb0ffffb 100644 > --- a/support/junction/junction.c > +++ b/support/junction/junction.c > @@ -63,7 +63,7 @@ junction_open_path(const char *pathname, int *fd) > if (pathname == NULL || fd == NULL) > return FEDFS_ERR_INVAL; > > - tmp = open(pathname, O_DIRECTORY); > + tmp = open(pathname, O_PATH|O_DIRECTORY); > if (tmp == -1) { > switch (errno) { > case EPERM: > @@ -93,7 +93,7 @@ junction_is_directory(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -121,7 +121,7 @@ junction_is_sticky_bit_set(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -155,7 +155,7 @@ junction_set_sticky_bit(int fd, const char *path) > { > struct stat stb; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, path); > return FEDFS_ERR_ACCESS; > @@ -393,7 +393,7 @@ junction_get_mode(const char *pathname, mode_t *mode) > if (retval != FEDFS_OK) > return retval; > > - if (fstat(fd, &stb) == -1) { > + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { > xlog(D_GENERAL, "%s: failed to stat %s: %m", > __func__, pathname); > (void)close(fd);
diff --git a/support/junction/junction.c b/support/junction/junction.c index 41cce261cb52..0628bb0ffffb 100644 --- a/support/junction/junction.c +++ b/support/junction/junction.c @@ -63,7 +63,7 @@ junction_open_path(const char *pathname, int *fd) if (pathname == NULL || fd == NULL) return FEDFS_ERR_INVAL; - tmp = open(pathname, O_DIRECTORY); + tmp = open(pathname, O_PATH|O_DIRECTORY); if (tmp == -1) { switch (errno) { case EPERM: @@ -93,7 +93,7 @@ junction_is_directory(int fd, const char *path) { struct stat stb; - if (fstat(fd, &stb) == -1) { + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { xlog(D_GENERAL, "%s: failed to stat %s: %m", __func__, path); return FEDFS_ERR_ACCESS; @@ -121,7 +121,7 @@ junction_is_sticky_bit_set(int fd, const char *path) { struct stat stb; - if (fstat(fd, &stb) == -1) { + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { xlog(D_GENERAL, "%s: failed to stat %s: %m", __func__, path); return FEDFS_ERR_ACCESS; @@ -155,7 +155,7 @@ junction_set_sticky_bit(int fd, const char *path) { struct stat stb; - if (fstat(fd, &stb) == -1) { + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { xlog(D_GENERAL, "%s: failed to stat %s: %m", __func__, path); return FEDFS_ERR_ACCESS; @@ -393,7 +393,7 @@ junction_get_mode(const char *pathname, mode_t *mode) if (retval != FEDFS_OK) return retval; - if (fstat(fd, &stb) == -1) { + if (fstatat(fd, "", &stb, AT_NO_AUTOMOUNT|AT_EMPTY_PATH) == -1) { xlog(D_GENERAL, "%s: failed to stat %s: %m", __func__, pathname); (void)close(fd);
JianHong reported some strange behavior with automounts on an nfs server without an explicit pseudoroot. When clients issued a readdir in the pseudoroot, automounted directories that were not yet mounted would show up even if they weren't exported, though the clients wouldn't be able to do anything with them. The issue was that triggering the automount on a directory would cause the mountd upcall to time out, which would cause nfsd to include the automounted dentry in the readdir response. Eventually, the automount would work and report that it wasn't exported and subsequent attempts to access the dentry would (properly) fail. We never want mountd to trigger an automount. The kernel should do that if it wants to use it. Change the junction checks to do an O_PATH open and use fstatat with AT_NO_AUTOMOUNT. Cc: Chuck Lever <chuck.lever@oracle.com> Link: https://bugzilla.redhat.com/show_bug.cgi?id=2148353 Reported-by: JianHong Yin <jiyin@redhat.com> Signed-off-by: Jeff Layton <jlayton@kernel.org> --- support/junction/junction.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)