* Xiao Yang (yangx...@cn.fujitsu.com) wrote:
> On 2019/12/13 0:37, Dr. David Alan Gilbert (git) wrote:
> > - res = fuse_buf_copy(&pipe_buf, buf,
> > - FUSE_BUF_FORCE_SPLICE | FUSE_BUF_SPLICE_NONBLOCK);
> > - if (res< 0) {
> > - if (res == -EAGAIN || res == -EINVAL) {
> > - /*
> > - * Should only get EAGAIN on kernels with
> > - * broken SPLICE_F_NONBLOCK support (<=
> > - * 2.6.35) where this error or a short read is
> > - * returned even if the pipe itself is not
> > - * full
> > - *
> > - * EINVAL might mean that splice can't handle
> > - * this combination of input and output.
> > - */
> > - if (res == -EAGAIN)
> > - se->broken_splice_nonblock = 1;
> > -
> > - pthread_setspecific(se->pipe_key, NULL);
> > - fuse_ll_pipe_free(llp);
> > - goto fallback;
> > - }
> > - res = -res;
> > - goto clear_pipe;
> > - }
> > -
> > - if (res != 0&& res< len) {
> > - struct fuse_bufvec mem_buf = FUSE_BUFVEC_INIT(len);
> > - void *mbuf;
> > - size_t now_len = res;
> > - /*
> > - * For regular files a short count is either
> > - * 1) due to EOF, or
> > - * 2) because of broken SPLICE_F_NONBLOCK (see above)
> > - *
> > - * For other inputs it's possible that we overflowed
> > - * the pipe because of small buffer fragments.
> > - */
> > -
> > - res = posix_memalign(&mbuf, pagesize, len);
> > - if (res != 0)
> > - goto clear_pipe;
> > -
> > - mem_buf.buf[0].mem = mbuf;
> > - mem_buf.off = now_len;
> > - res = fuse_buf_copy(&mem_buf, buf, 0);
> > - if (res> 0) {
> > - char *tmpbuf;
> > - size_t extra_len = res;
> > - /*
> > - * Trickiest case: got more data. Need to get
> > - * back the data from the pipe and then fall
> > - * back to regular write.
> > - */
> > - tmpbuf = malloc(headerlen);
> > - if (tmpbuf == NULL) {
> > - free(mbuf);
> > - res = ENOMEM;
> > - goto clear_pipe;
> > - }
> > - res = read_back(llp->pipe[0], tmpbuf, headerlen);
> > - free(tmpbuf);
> > - if (res != 0) {
> > - free(mbuf);
> > - goto clear_pipe;
> > - }
> > - res = read_back(llp->pipe[0], mbuf, now_len);
> > - if (res != 0) {
> > - free(mbuf);
> > - goto clear_pipe;
> > - }
> > - len = now_len + extra_len;
> > - iov[iov_count].iov_base = mbuf;
> > - iov[iov_count].iov_len = len;
> > - iov_count++;
> > - res = fuse_send_msg(se, ch, iov, iov_count);
> > - free(mbuf);
> > - return res;
> > - }
> > - free(mbuf);
> > - res = now_len;
> > - }
> > - len = res;
> > - out->len = headerlen + len;
> > -
> > - if (se->debug) {
> > - fuse_log(FUSE_LOG_DEBUG,
> > - " unique: %llu, success, outsize: %i (splice)\n",
> > - (unsigned long long) out->unique, out->len);
> > - }
> > -
> > - splice_flags = 0;
> > - if ((flags& FUSE_BUF_SPLICE_MOVE)&&
> > - (se->conn.want& FUSE_CAP_SPLICE_MOVE))
> > - splice_flags |= SPLICE_F_MOVE;
> > -
> > - res = splice(llp->pipe[0], NULL, ch ? ch->fd : se->fd,
> > - NULL, out->len, splice_flags);
> Hi,
>
> 1) In buffer.c, fuse_buf_splice() uses splice(2) to copy/move data in some
> cases if the syscall is supported.
> 2) One pipe needs to be passed to splice(2) so splice(2) without one pipe
> fails and then go back to use other ways(e.g. use fuse_buf_fd_to_fd()).
> 3) fuse_buf_copy() calls fuse_buf_splice() indirectly and this patch has
> removed all pipes used by fuse_buf_copy().
>
> Is it necessary to leave the code related to splice(2)? Is it going to be
> used in future?
> We have to use splice(2) by the correct CONFIG_SPLICE macro If necessary.
Yes, I think we never set HAVE_SPLICE; so that code can go.
I'll change this patch to remove that as well.
Dave
> Best Regards,
> Xiao Yang
>
>
--
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK
