On 17.01.20 10:55, Kevin Wolf wrote: > Am 17.01.2020 um 10:12 hat Max Reitz geschrieben: >> On 17.01.20 00:26, Alberto Garcia wrote: >>> On Tue 14 Jan 2020 03:15:48 PM CET, Max Reitz wrote: >>>>> @@ -219,7 +219,7 @@ static int l2_load(BlockDriverState *bs, uint64_t >>>>> offset, >>>>> * Writes one sector of the L1 table to the disk (can't update single >>>>> entries >>>>> * and we really don't want bdrv_pread to perform a read-modify-write) >>>>> */ >>>>> -#define L1_ENTRIES_PER_SECTOR (512 / 8) >>>>> +#define L1_ENTRIES_PER_SECTOR (BDRV_SECTOR_SIZE / 8) >>>>> int qcow2_write_l1_entry(BlockDriverState *bs, int l1_index) >>>> >>>> Here it’s because the comment is wrong: “Can’t update single entries” – >>>> yes, we can. We’d just have to do a bdrv_pwrite() to a single entry. >>> >>> What's the point of qcow2_write_l1_entry() then? >> >> I think the point was that we couldn’t, for a long time, because the >> block layer only provided sector-granularity access. This function >> simply was never changed when the block layer gained the ability to do >> byte-granularity I/O. >> >> (We’d still need this function, but only for the endian swap, I think.) > > We still can't do byte-granularity writes with O_DIRECT, because that's > a kernel requirement.
Ah, yes. But that makes BDRV_SECTOR_SIZE the wrong choice.
> The comment explains that we don't want to do a RMW cycle to write a
> single entry because that would be slower than just writing a whole
> sector. I think this is still accurate. Maybe we should change the
> comment to say "can't necessarily update". (The part that looks really
> wrong in the comment is "bdrv_pread", that should be "bdrv_pwrite"...)
Hm. But we wouldn’t do an RMW cycle without O_DIRECT, would we?
> Now, what's wrong about the logic to avoid the RMW is that it assumes
> a fixed required alignment of 512. What it should do is looking at
> bs->file->bl.request_alignment and rounding accordingly.
Yes.
>>>>> @@ -3836,7 +3837,7 @@ qcow2_co_copy_range_from(BlockDriverState *bs,
>>>>> case QCOW2_CLUSTER_NORMAL:
>>>>> child = s->data_file;
>>>>> copy_offset += offset_into_cluster(s, src_offset);
>>>>> - if ((copy_offset & 511) != 0) {
>>>>> + if (!QEMU_IS_ALIGNED(copy_offset, BDRV_SECTOR_SIZE)) {
>>>>
>>>> Hm. I don’t get this one.
>>>
>>> Checking the code (e.g. block_copy_do_copy()) it seems that the whole
>>> chunk must be cluster aligned so I don't get this one either.
>>
>> Hm, how did you get to block_copy_do_copy()? That’s part of the
>> block-copy infrastructure that’s only used for the backup job, as far as
>> I’m aware. It’s different from copy_range.
>>
>> I don’t see any limitation for copy_range. I suppose maybe it doesn’t
>> work for anything that isn’t aligned to physical sectors? But the qcow2
>> driver shouldn’t care about that.
>>
>> On thing’s for sure, the raw driver doesn’t care about it.
>
> I don't understand this one either.
Good. :-)
Max
signature.asc
Description: OpenPGP digital signature
