Re: [PATCH] cache: Tolerate short writes in print_slot

[Hristo Venev <hristo@venev.name>]

[-- Attachment #1: Type: text/plain, Size: 4359 bytes --]

On Thu, 2021-10-07 at 10:35 +0100, John Keeping wrote:
> Have you seen these errors in practice, or is this just theoretical?
>
> In recent (since v2.6.33) versions of Linux, all files should support
> sendfile(), especially since we expect out_fd to be a socket or pipe.

Even though I haven't seen any errors from sendfile, I'm not sure if it
always works with fuse filesystems and with third-party filesystems
like zfs. The sendfile(2) man page says this:

"Applications may wish to fall back to read(2)/write(2) in the case
where sendfile() fails with EINVAL or ENOSYS."

> > On the read/write path, use write_in_full which deals with short
> > writes.
> > 
> > Signed-off-by: Hristo Venev <hristo@venev.name>
> > ---
> >  cache.c | 46 ++++++++++++++++++++++++----------------------
> >  1 file changed, 24 insertions(+), 22 deletions(-)
> > 
> > diff --git a/cache.c b/cache.c
> > index 55199e8..85cfbd9 100644
> > --- a/cache.c
> > +++ b/cache.c
> > @@ -85,40 +85,42 @@ static int close_slot(struct cache_slot *slot)
> >  /* Print the content of the active cache slot (but skip the key). */
> >  static int print_slot(struct cache_slot *slot)
> >  {
> > -#ifdef HAVE_LINUX_SENDFILE
> > -       off_t start_off;
> > -       int ret;
> > +       off_t off;
> > +       ssize_t i;
> > +
> > +       off = slot->keylen + 1;
> >  
> > -       start_off = slot->keylen + 1;
> > +#ifdef HAVE_LINUX_SENDFILE
> > +       off_t size;
> 
> decl-after-stmt if HAVE_LINUX_SENDFILE is set.

I didn't know that compilers that don't support that exist outside
museums. I will fix it in v2.

> > +       size = slot->cache_st.st_size;
> >  
> >         do {
> > -               ret = sendfile(STDOUT_FILENO, slot->cache_fd,
> > &start_off,
> > -                               slot->cache_st.st_size - start_off);
> > -               if (ret < 0) {
> > +               i = sendfile(STDOUT_FILENO, slot->cache_fd, &off,
> > size - off);
> 
> Why is ret renamed?  i is normally a loop index variable, using it for
> the return value here is strange, please stick with "ret".

I will fix this in v2.

> > +               if (i < 0) {
> >                         if (errno == EAGAIN || errno == EINTR)
> >                                 continue;
> > +                       /* Fall back to read/write on EINVAL */
> > +                       if (errno == EINVAL)
> > +                               break;
> >                         return errno;
> >                 }
> > -               return 0;
> > +               if (off == size)
> > +                       return 0;
> >         } while (1);
> > -#else
> > -       ssize_t i, j;
> > +#endif
> >  
> > -       i = lseek(slot->cache_fd, slot->keylen + 1, SEEK_SET);
> > -       if (i != slot->keylen + 1)
> > +       if (lseek(slot->cache_fd, off, SEEK_SET) != off)
> >                 return errno;
> >  
> >         do {
> > -               i = j = xread(slot->cache_fd, slot->buf, sizeof(slot-
> > >buf));
> > -               if (i > 0)
> > -                       j = xwrite(STDOUT_FILENO, slot->buf, i);
> > -       } while (i > 0 && j == i);
> > -
> > -       if (i < 0 || j != i)
> > -               return errno;
> > -       else
> > -               return 0;
> > -#endif
> > +               i = xread(slot->cache_fd, slot->buf, sizeof(slot-
> > >buf));
> > +               if (i < 0)
> > +                       return errno;
> > +               if (i == 0)
> > +                       return 0;
> > +               if (write_in_full(STDOUT_FILENO, slot->buf, i) < 0)
> > +                       return errno;
> > +       } while (1);
> >  }
> >  
> >  /* Check if the slot has expired */
> > -- 
> > 2.31.1
> > 


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 858 bytes --]