On Tue, Feb 06, 2024 at 05:14:42PM -0600, Steve French wrote:
> The code in question is a little hard to follow, and may eventually
> get rewritten by later folio/netfs patches from David Howells but the
> problem is in
> cifs_write_back_from_locked_folio() and cifs_writepages_region() where
> after the write (of maximum write size) completes, the next write
> skips to the beginning of the next page (leaving the tail end of the
> previous page unwritten). This is not an issue with typical servers
> and typical wsize values because those will almost always be a
> multiple of 4096, but in the bug report the server in question was old
> and had sent a value for maximum write size that was not a multiple of
> 4096.
>
> This can be a temporary fix, that can be removed as netfs/folios
> implementation improves here - but in the short term the easiest way
> to fix this seems to be to round the negotiated maximum_write_size
> down if not a multiple of 4096, to be a multiple of 4096 (this can be
> removed in the future when the folios code is found which caused
> this), and also warn the user if they pick a wsize that is not
> recommended, not a multiple of 4096.
Seems like a sensible stopgap, but probably the patch should use
PAGE_SIZE rather than plain 4096 (what about
Alpha/Sparc/powerpc-64k/arm64-{16,64}k?)
Also, what if the server says its max-write-size is 2048 bytes?
Also, does the code work well if the max-write-size is, say, 20480
bytes? (ie an odd multiple of PAGE_SIZE is fine; it doesn't need to be
a power-of-two?)
