> > > dm-thinp (MB/s) loopback (MB/s) loop faster by
> > >
> > > ==============+======================================================
> > > random jobs=4 | 18496.0 33522.0 1.68x
> > > zeros jobs=4 | 8119.2 9767.2 1.20x
> > > ==============+======================================================
> > > random jobs=1 | 7330.5 12330.0 1.81x
> > > zeros jobs=1 | 4965.2 6799.9 1.11x
> >
> > This looks more reasonable in terms of throughput.
> >
> > One major worry here is that checking every write is blowing your cache,
> > so you could have a major impact on performance in general. Even for
> > O_DIRECT writes, you are now accessing the memory. Have you looked into
> > doing non-temporal memory compares instead? I think that would be the
> > way to go.
>
> So I found your patch in the thread. For each vector, use memcmp() instead and
> hope it does the right thing. You can compare with empty_zero_page. That
> should drastically cut down on the amount of hand rolled code you have in
> bio_is_zero_filled() at the moment.
I ran these against dm-thinp within provision_block() when checking
memcmp(page_address(ZERO_PAGE()), data, bv.bv_len) != 0
and got these numbers that show memcmp() based testing is rather slow:
zeroed jobs=1 1268.1 MB/s
zeroed jobs=4 1445.6 MB/s
random jobs=1 7554.1 MB/s
random jobs=4 17677.0 MB/s
I updated the checking code to use __uint128_t, and performance got even
better than with 64-bit integers (note that gcc is emulating 128bit).
These are the numbers w/ __uint128_t:
zeroed jobs=1 6424 MB/s
zeroed jobs=4 10738 MB/s MB/s
random jobs=1 7928 MB/s
random jobs=4 19435 MB/s MB/s
Jens says to try non-temporal memory compares to deal with cache
invalidation, but that sounds like arch-specific assembler unless I am
mistaken. Does GCC have something I can use to flag the buffer as not to
be cached? If not, I'm happy to poke some asm code into the check
routine, but I'm a bit rusty there.
Can someone suggest the inline asm that I would use to rep until a
non-zero value appears? Assume that we are 8 or 16-byte aligned, probably
the bigger the word sized checks are the better, performance wise.
My latest version is attached, and includes moving bvec_kunmap_irq out of
the for loops in hopes of getting gcc to unroll loops.
-Eric
---
block/bio.c | 74 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 files changed, 74 insertions(+), 0 deletions(-)
diff --git a/block/bio.c b/block/bio.c
index 8c2e55e..262f190 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -511,6 +511,80 @@ void zero_fill_bio(struct bio *bio)
}
EXPORT_SYMBOL(zero_fill_bio);
+bool bio_is_zero_filled(struct bio *bio)
+{
+ unsigned i, count, left;
+ unsigned long flags;
+ struct bio_vec bv;
+ struct bvec_iter iter;
+ char *data = NULL, *p;
+ __uint128_t *parch;
+
+ bio_for_each_segment(bv, bio, iter) {
+ data = bvec_kmap_irq(&bv, &flags);
+ p = data;
+ left = bv.bv_len;
+
+ if (unlikely( data == NULL ))
+ continue;
+
+
+ /* check unaligned bytes at the beginning of p */
+ if (unlikely( ( (uintptr_t)p & (sizeof(__uint128_t)-1) ) != 0 )) {
+ count = sizeof(__uint128_t) - ( (uintptr_t)p & (sizeof(__uint128_t)-1) );
+ for (i = 0; i < count; i++) {
+ if (*p)
+ break;
+ p++;
+ }
+ if (i < count)
+ goto out_false;
+ left -= count;
+ }
+
+ /* we should be word aligned now */
+ BUG_ON(unlikely( ((uintptr_t)p & (sizeof(__uint128_t)-1) ) != 0 ));
+
+ /* now check in word-sized chunks */
+ parch = (__uint128_t*)p;
+ count = left >> ilog2(sizeof(__uint128_t)); /* count = left / sizeof(__uint128_t) */;
+ for (i = 0; i < count; i++) {
+ if (*parch)
+ break;
+ parch++;
+ }
+ if (i < count)
+ goto out_false;
+ left -= count << ilog2(sizeof(__uint128_t)); /* left -= count*sizeof(__uint128_t) */
+
+ /* check remaining unaligned values at the end */
+ p = (char*)parch;
+ if (unlikely(left > 0))
+ {
+ for (i = 0; i < left; i++) {
+ if (*p)
+ break;
+ p++;
+ }
+ if (i < count)
+ goto out_false;
+ left = 0;
+ }
+
+ bvec_kunmap_irq(data, &flags);
+ BUG_ON(unlikely( left > 0 ));
+ BUG_ON(unlikely( data+bv.bv_len != p ));
+ }
+
+ return true;
+
+out_false:
+ if (data != NULL)
+ bvec_kunmap_irq(data, &flags);
+ return false;
+}
+EXPORT_SYMBOL(bio_is_zero_filled);
+
/**
* bio_put - release a reference to a bio
* @bio: bio to release reference to
--
1.7.1
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
