On Mon, Sep 29, 2014 at 11:49:54PM +0200, David Sterba wrote:
> On Mon, Sep 29, 2014 at 12:45:12PM -0700, Omar Sandoval wrote:
> > > > @@ -639,8 +640,7 @@ static int send_header(struct send_ctx *sctx)
> > > > - return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
> > > > - &sctx->send_off);
> > > > + return write_buf(sctx->send_filp, &hdr, sizeof(hdr),
> > > > &sctx->send_off);
> > >
> > > > ret = write_buf(sctx->send_filp, sctx->send_buf,
> > > > sctx->send_size,
> > > > - &sctx->send_off);
> > > > + &sctx->send_off);
> > >
> > > Please do not fold unrelated changes.
> >
> > My metric for "related" here was that these were call sites of a function I
> > directly modified.
>
> The changes are only in the whitespace, that's not necessary. It's
> usually ok to fix style issues in the code you modify directly.
>
> > Is the preferred form to just split style fixes that we encounter into
> > a separate patch in the series?
>
> Well, I may only express my point of view. Yes, split the style-only
> changes into another patch and don't send it :)
>
> The problem with patches that do not effectively change anything is that
> they pollute git history and just add extra step when one has to look
> for a patch that broke something, or eg. change context of following
> patches and make backporting a bit more tedious. Code cleanups are fine,
> but there's usually a point of making the code more readable, compact,
> etc.
>
> The coding style should be perfect from the beginning. Nobody will
> probably point out minor style violations during review, because it just
> pointless for a patch that fixes a real bug.
Thank you for the perspective :)
On Tue, Sep 30, 2014 at 12:27:43PM -0700, Zach Brown wrote:
> On Sun, Sep 28, 2014 at 03:26:04PM -0700, Omar Sandoval wrote:
> > On Sun, Sep 28, 2014 at 01:48:11AM -0700, Omar Sandoval wrote:
> > > diff --git a/fs/btrfs/send.c b/fs/btrfs/send.c
> > > index 6528aa6..e0be577 100644
> > > --- a/fs/btrfs/send.c
> > > +++ b/fs/btrfs/send.c
> > > @@ -515,7 +515,8 @@ static int write_buf(struct file *filp, const void
> > > *buf, u32 len, loff_t *off)
> > > set_fs(KERNEL_DS);
> > >
> > > while (pos < len) {
> > > - ret = vfs_write(filp, (char *)buf + pos, len - pos, off);
> > > + ret = vfs_write(filp, (__force const char __user *)buf + pos,
> > > + len - pos, off);
> > > /* TODO handle that correctly */
> > > /*if (ret == -ERESTARTSYS) {
> > > continue;
> >
> > Actually, looking at this now, it looks like this is just an open-coded
> > kernel_write. I think this could be made a bit cleaner by using that
> > instead;
>
> Agreed, but notice that you'll want to be careful to update
> write_buf()'s *off because passing a dereferenced copy to kernel_write()
> will lose the pos update that vfs_write() is currently taking care of.
>
> A carefully placed "*off += ret" in write_buf() will be fine. (As fine
> as having a magical private file position in the send context ever was.)
>
> > the tradeoff is that each call to kernel_write will do the address
> > space flip-flop, so if the write gets split up into many calls,
> > there'd be some slight overhead. That's probably a microoptimization,
> > but
>
> Yeah, I don't think that overhead is going to be significant given all
> of the work that's going on.
>
> > I think it's worth looking
> > into making kernel_read and kernel_write handle the retry logic.
>
> I disagree. I wouldn't broaden the scope to add retrying on behalf of
> all kernel_write() callers and write methods (it's exported to modules,
> too). I'd leave the looping in btrfs and just call kernel_write() to
> get rid of the segment juggling.
>
> - z
That sounds fair. I'll submit a v2 that replaces vfs_write with kernel_write.
--
Omar
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
