Re: [PATCH v2] splice: sendfile() at once fails for big files

[leroy christophe <christophe.leroy@c-s.fr>]

Le 06/05/2015 16:23, Jens Axboe a écrit :
> On 05/05/2015 09:41 PM, Linus Torvalds wrote:
>> Jens, ping?
>>
>> The test results should make this a no-brainer, but I hate how random
>> these flag ops.
>
> Missed the original, apparently. I too am confused how this is a
> correctness fix and not just an optimization.
>
> +               if (read_len < len)
> +                       sd->flags |= SPLICE_F_MORE;
> +               else if (!more)
> +                       sd->flags &= ~SPLICE_F_MORE;
>
> Should that check be for 'more', not '!more'?
>
>
@@ -1204,6 +1204,7 @@ ssize_t splice_direct_to_actor(struct file *in, 
struct splice_desc *sd,
       * Don't block on output, we have to drain the direct pipe.
       */
      sd->flags &= ~SPLICE_F_NONBLOCK;
+    more = sd->flags & SPLICE_F_MORE;

      while (len) {
          size_t read_len;
@@ -1216,6 +1217,10 @@ ssize_t splice_direct_to_actor(struct file *in, 
struct splice_desc *sd,
          read_len = ret;
          sd->total_len = read_len;

+        if (read_len < len)
+            sd->flags |= SPLICE_F_MORE;
+        else if (!more)
+            sd->flags &= ~SPLICE_F_MORE;



'more' contains whether sendfile() has been called with SPLICE_F_MORE or 
not.
Until all bytes are processed, we have to force SPLICE_F_MORE regardless 
of how sendfile() was called.
Once all bytes have been read, we have to reset the flags according to 
how sendfile() was called, so if 'more' is NOT set, we have to clear 
SPLICE_F_MORE from sd->flags (which was unconditionaly set for 
processing the first bytes)

Christophe