Re: [PATCH V2] Btrfs: Full direct I/O and AIO read implementation.

[Josef Bacik <josef@redhat.com>]

On Tue, Feb 16, 2010 at 02:28:00PM -0500, jim owens wrote:
> Josef Bacik wrote:
>> On Wed, Feb 10, 2010 at 01:53:50PM -0500, jim owens wrote:
>>> +
>>> +static int btrfs_dio_hole_read(struct btrfs_diocb *diocb, u64 hole_len)
>>> +{
>>> +	int err = 0;
>>> +	diocb->umc.todo = hole_len;
>>> +	while (diocb->umc.todo) {
>>> +		struct bio_vec uv;
>>> +		char *out;
>>> +
>>> +		err = btrfs_dio_get_user_bvec(&uv, &diocb->umc);
>>> +		if (err)
>>> +			goto fail;
>>> +		diocb->start += uv.bv_len;
>>> +		out = kmap_atomic(uv.bv_page, KM_USER0);
>>> +		memset(out + uv.bv_offset, 0, uv.bv_len);
>>> +		kunmap_atomic(out, KM_USER0);
>>> +
>>
>> /me hands jim a zero_user_page()
>>
>>> +		btrfs_dio_done_with_out(&uv, NULL);
>>> +	}
>
> As promised... why I'm keeping my code :)
>
> cscope says there is no zero_user_page() so I think
> you mean the same one I looked at zero_user(), which
> calls the far more complicated zero_user_segments().
>
> I did not use it for 3 reasons:
>
> 1) More generated code (but maybe gcc reduces it).
>
> 2) My btrfs_dio_done_with_out() must do (so repeat) the
> flush_dcache_page which is not a noop on every arch.
>
> 3) Everywhere else in btrfs does the same as my code.
>

Hmm I swore it was zero_user_page but now I can't find it.  Yeah the extra
flush_dcace_page is reason enough to not use it, seems reasonable enough to me.
Thanks,

Josef