On Fri, 2015-08-07 at 07:04 +1000, Dave Chinner wrote:
> On Thu, Aug 06, 2015 at 11:43:20AM -0600, Ross Zwisler wrote:
> > Update the DAX I/O path so that all operations that store data (I/O
> > writes, zeroing blocks, punching holes, etc.) properly synchronize the
> > stores to media using the PMEM API. This ensures that the data DAX is
> > writing is durable on media before the operation completes.
> >
> > Signed-off-by: Ross Zwisler <ross.zwis...@linux.intel.com>
> ....
> > + if (pgsz < PAGE_SIZE) {
> > memset(addr, 0, pgsz);
> > - else
> > + wb_cache_pmem((void __pmem *)addr, pgsz);
> > + } else {
> > clear_page(addr);
> > + wb_cache_pmem((void __pmem *)addr, PAGE_SIZE);
> > + }
>
> I'd much prefer to see these wrapped up in helper fuctions e.g.
> clear_page_pmem() rather than scatter them around randomly.
> Especially the barriers - the way they've been optimised is asking
> for people to get it wrong in the future. I'd much prefer to see
> the operations paired properly in a helper first (i.e. obviously
> correct) and then it can be optimised later if workloads start to
> show the barrier as a bottleneck...
>
> > +/*
> > + * This function's stores and flushes need to be synced to media by a
> > + * wmb_pmem() in the caller. We flush the data instead of writing it back
> > + * because we don't expect to read this newly zeroed data in the near
> > future.
> > + */
>
> That seems suboptimal. dax_new_buf() is called on newly allocated or
> unwritten buffers we are about to write to. Immediately after this
> we write the new data to the page, so we are effectively writting
> the whole page here.
>
> So why wouldn't we simply commit the whole page during the write and
> capture all this zeroing in the one flush/commit/barrier op?
>
> > static void dax_new_buf(void *addr, unsigned size, unsigned first, loff_t
> > pos,
> > loff_t end)
> > {
> > loff_t final = end - pos + first; /* The final byte of the buffer */
> >
> > - if (first > 0)
> > + if (first > 0) {
> > memset(addr, 0, first);
> > - if (final < size)
> > + flush_cache_pmem((void __pmem *)addr, first);
> > + }
> > + if (final < size) {
> > memset(addr + final, 0, size - final);
> > + flush_cache_pmem((void __pmem *)addr + final, size - final);
> > + }
> > }
> >
> > static bool buffer_written(struct buffer_head *bh)
> > @@ -108,6 +123,7 @@ static ssize_t dax_io(struct inode *inode, struct
> > iov_iter *iter,
> > loff_t bh_max = start;
> > void *addr;
> > bool hole = false;
> > + bool need_wmb = false;
> >
> > if (iov_iter_rw(iter) != WRITE)
> > end = min(end, i_size_read(inode));
> > @@ -145,18 +161,23 @@ static ssize_t dax_io(struct inode *inode, struct
> > iov_iter *iter,
> > retval = dax_get_addr(bh, &addr, blkbits);
> > if (retval < 0)
> > break;
> > - if (buffer_unwritten(bh) || buffer_new(bh))
> > + if (buffer_unwritten(bh) || buffer_new(bh)) {
> > dax_new_buf(addr, retval, first, pos,
> > end);
> > + need_wmb = true;
> > + }
> > addr += first;
> > size = retval - first;
> > }
> > max = min(pos + size, end);
> > }
> >
> > - if (iov_iter_rw(iter) == WRITE)
> > + if (iov_iter_rw(iter) == WRITE) {
> > len = copy_from_iter_nocache(addr, max - pos, iter);
> > - else if (!hole)
> > + if (!iter_is_iovec(iter))
> > + wb_cache_pmem((void __pmem *)addr, max - pos);
> > + need_wmb = true;
>
> Conditional pmem cache writeback after a "nocache" copy to the pmem?
> Comments, please.
>
> Cheers,
>
> Dave.
I agree with all your comments, and will address them in v2. Thank you for
the feedback.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
