On Fri, Jul 27, 2018 at 04:23:10PM -0500, Eric Blake wrote:
> On 07/27/2018 04:03 PM, Nir Soffer wrote:
> >When using block device on RHEL 7.5, file plugin fails to zero with this
> >error (copied from strace):
> >
> >[pid 39551] fallocate(8, FALLOC_FL_ZERO_RANGE, 1536, 64000) = -1 ENODEV (No such device)
> >
> >This is expected error according to the manual:
> >
> >ENODEV fd does not refer to a regular file or a directory.  (If fd is a
> >pipe or FIFO, a different error results.)
>
> The man page is out-of-date; newer kernels support
> FALLOC_FL_ZERO_RANGE on block devices, in a manner that leaves the
> pages marked allocated (that is, no discard happens). The kernel
> also supports FALLOC_FL_PUNCH_HOLE to guarantee a read-zeroes result
> (using discard, if that works), and
> FALLOC_FL_PUNCH_HOLE|FALLOC_FL_NO_HIDE_STALE (which only discards,
> and no longer guarantees a read-zeroes result).  But you are also
> right that in all cases, the errno returned when the operation
> cannot be supported is not always EOPNOSUPP.
>
> >
> >Treat this error as EOPNOSUPP.
> >
> >Tested only on Fedora 28; I don't know how to build nbdkit on RHEL, but
> >the change is pretty trivial.
> >---
> >  plugins/file/file.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> >
> >diff --git a/plugins/file/file.c b/plugins/file/file.c
> >index b6e33de..a7c07fb 100644
> >--- a/plugins/file/file.c
> >+++ b/plugins/file/file.c
> >@@ -243,7 +243,7 @@ file_zero (void *handle, uint32_t count, uint64_t offset, int may_trim)
> >    if (may_trim) {
> >      r = fallocate (h->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
> >                offset, count);
> >-    if (r == -1 && errno != EOPNOTSUPP) {
> >+    if (r == -1 && errno != EOPNOTSUPP && errno != ENODEV) {
> >        nbdkit_error ("zero: %m");
> >      }
> >      /* PUNCH_HOLE is older; if it is not supported, it is likely that
>
> Given the recent thread on qemu about the different fallocate flags
> in relation to block devices, we may need to revisit the logic in
> this function to more closely follow the flags (use
> FALLOC_FL_ZERO_RANGE when may_trim is false, and
> FALLOC_FL_PUNCH_HOLE when it is true; as well as tweak the
> .discard() callback to use
> FALLOC_FL_PUNCH_HOLE|FALLOC_FL_NO_HIDE_STALE).
>
> https://lists.gnu.org/archive/html/qemu-devel/2018-07/msg05135.html
>
> >@@ -254,7 +254,7 @@ file_zero (void *handle, uint32_t count, uint64_t offset, int may_trim)
> >  #ifdef FALLOC_FL_ZERO_RANGE
> >    r = fallocate (h->fd, FALLOC_FL_ZERO_RANGE, offset, count);
> >-  if (r == -1 && errno != EOPNOTSUPP) {
> >+  if (r == -1 && errno != EOPNOTSUPP && errno != ENODEV) {
> >      nbdkit_error ("zero: %m");
> >    }
> >  #else
>
> Are we sure that the caller still sees the correct EOPNOTSUPP errno
> value that it uses in deciding whether to attempt the manual
> fallbacks? You may need to explicitly set errno.
>
> >@@ -288,11 +288,11 @@ file_trim (void *handle, uint32_t count, uint64_t offset)
> >    struct handle *h = handle;
> >    /* Trim is advisory; we don't care if it fails for anything other
> >-   * than EIO or EPERM. */
> >+   * than EIO, EPERM, or ENODEV (kernel 3.10) */
> >    r = fallocate (h->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
> >                   offset, count);
> >    if (r < 0) {
> >-    if (errno != EPERM && errno != EIO) {
> >+    if (errno != EPERM && errno != EIO && errno != ENODEV) {
>
> This last hunk looks wrong.  We want to ignore ENODEV errors in this
> code path (the same way we are already ignoring EOPNOTSUPP errors).

I should have read the mailing list first ...

I pushed Nir's initial patch.  Let me know if additional
changes are required.