Am 20.04.21 um 18:52 schrieb Vladimir Sementsov-Ogievskiy: > 20.04.2021 18:04, Kevin Wolf wrote: >> Am 20.04.2021 um 16:31 hat Vladimir Sementsov-Ogievskiy geschrieben: >>> 15.04.2021 18:22, Kevin Wolf wrote: >>>> In order to avoid RMW cycles, is_allocated_sectors() treats zeroed areas >>>> like non-zero data if the end of the checked area isn't aligned. This >>>> can improve the efficiency of the conversion and was introduced in >>>> commit 8dcd3c9b91a. >>>> >>>> However, it comes with a correctness problem: qemu-img convert is >>>> supposed to sparsify areas that contain only zeros, which it doesn't do >>>> any more. It turns out that this even happens when not only the >>>> unaligned area is zeroed, but also the blocks before and after it. In >>>> the bug report, conversion of a fragmented 10G image containing only >>>> zeros resulted in an image consuming 2.82 GiB even though the expected >>>> size is only 4 KiB. >>>> >>>> As a tradeoff between both, let's ignore zeroed sectors only after >>>> non-zero data to fix the alignment, but if we're only looking at zeros, >>>> keep them as such, even if it may mean additional RMW cycles. >>>> >>> >>> Hmm.. If I understand correctly, we are going to do unaligned >>> write-zero. And that helps. >> >> This can happen (mostly raw images on block devices, I think?), but >> usually it just means skipping the write because we know that the target >> image is already zeroed. >> >> What it does mean is that if the next part is data, we'll have an >> unaligned data write. >> >>> Doesn't that mean that alignment is wrongly detected? >> >> The problem is that you can have bdrv_block_status_above() return the >> same allocation status multiple times in a row, but *pnum can be >> unaligned for the conversion. >> >> We only look at a single range returned by it when detecting the >> alignment, so it could be that we have zero buffers for both 0-11 and >> 12-16 and detect two misaligned ranges, when both together are a >> perfectly aligned zeroed range. >> >> In theory we could try to do some lookahead and merge ranges where >> possible, which should give us the perfect result, but it would make the >> code considerably more complicated. (Whether we want to merge them >> doesn't only depend on the block status, but possibly also on the >> content of a DATA range.) >> >> Kevin >> > > Oh, I understand now the problem, thanks for explanation. > > Hmm, yes that means, that if the whole buf is zero, is_allocated_sectors must > not align it down, to be possibly "merged" with next chunk if it is zero too. > > But it's still good to align zeroes down, if data starts somewhere inside the > buf, isn't it? > > what about something like this: > > diff --git a/qemu-img.c b/qemu-img.c > index babb5573ab..d1704584a0 100644 > --- a/qemu-img.c > +++ b/qemu-img.c > @@ -1167,19 +1167,39 @@ static int is_allocated_sectors(const uint8_t *buf, > int n, int *pnum, > } > } > > + if (i == n) { > + /* > + * The whole buf is the same. > + * > + * if it's data, just return it. It's the old behavior. > + * > + * if it's zero, just return too. It will work good if target is > alredy > + * zeroed. And if next chunk is zero too we'll have no RMW and no > reason > + * to write data. > + */ > + *pnum = i; > + return !is_zero; > + } > + > tail = (sector_num + i) & (alignment - 1); > if (tail) { > if (is_zero && i <= tail) { > - /* treat unallocated areas which only consist > - * of a small tail as allocated. */ > + /* > + * For sure next sector after i is data, and it will rewrite this > + * tail anyway due to RMW. So, let's just write data now. > + */ > is_zero = false; > } > if (!is_zero) { > - /* align up end offset of allocated areas. */ > + /* If possible, align up end offset of allocated areas. */ > i += alignment - tail; > i = MIN(i, n); > } else { > - /* align down end offset of zero areas. */ > + /* > + * For sure next sector after i is data, and it will rewrite this > + * tail anyway due to RMW. Better is avoid RMW and write zeroes > up > + * to aligned bound. > + */ > i -= tail; > } > } > >
I think we forgot to follow up on this. Has anyone tested this suggestion? Otherwise, I would try to rerun the tests I did with the my old and Kevins suggestion. Peter