On 12 Nov 2024, at 9:06, Sebastian Feld wrote: > On Tue, Nov 12, 2024 at 2:56 PM Benjamin Coddington <bcodding@xxxxxxxxxx> wrote: >> >> On 12 Nov 2024, at 6:27, Sebastian Feld wrote: >> >>> On Mon, Nov 11, 2024 at 8:25 AM Seiichi Ikarashi (Fujitsu) >>> <s.ikarashi@xxxxxxxxxxx> wrote: >>>> >>>> The rsize/wsize values are not multiples of 1024 but multiples of PAGE_SIZE >>>> or powers of 2 if < PAGE_SIZE as defined in fs/nfs/internal.h:nfs_io_size(). >>> >>> *facepalm* >>> >>> How should this work at all in a heterogeneous environment where >>> pagesizes can be 4k or 64k (ARM)? >>> >>> IMHO this is a BIG, rsize and wsize should count in 1024 bytes, and >>> warn if there is no exact match to a page size. Otherwise non-portable >>> chaos rules. >> >> >> I'm not following you - is "BIG" an acronym? > > I hit the wrong key. I wanted to write "BUG" > >> >> Can you explain what you mean by non-portable chaos? I'm having trouble >> seeing the problem. > > x86-only-world-view: There are other platforms like PowerPC or ARM > which can have other page sizes, and even the default page size for a > platform can vary. ARM can do 4k, 64k defaults, servers default to > 64k, IOT machines to 4k. > So this is NOT a documentation bug, it's a bug in the code which > should do what the doc says. Not the other way around. What's the bug in the code again? I'm still not seeing the bug. Why should the code set the max io read/write size to a multiple of 1024 instead of a multiple of the page size? Ben
