On Thu, Aug 22, 2024 at 02:40:57AM GMT, Kamlesh Gurudasani wrote: > Kamlesh Gurudasani <kamlesh@xxxxxx> writes: > > > Daniel Jordan <daniel.m.jordan@xxxxxxxxxx> writes: > > > >> On Tue, Aug 20, 2024 at 12:06:47PM GMT, Herbert Xu wrote: > >>> On Mon, Aug 19, 2024 at 06:29:52PM -0400, Daniel Jordan wrote: > >>> > > >>> > The DIV_ROUND_UP approach reads a bit nicer to me, but I can imagine > >>> > oddball cases where rounding up is undesirable (say, near-zero values > >>> > for size, min_chunk, and align; padata_work_alloc_mt returns many fewer > >>> > works than requested; and a single unit of work is very expensive) so > >>> > that rounding up makes a bigger difference. So, the way it now is seems > >>> > ok. > >>> > >>> In that case let's do the max ahead of the align check: > >>> > >>> ps.chunk_size = max(ps.chunk_size, 1ul); > >>> ps.chunk_size = roundup(ps.chunk_size, job->align); > >>> > >>> If we do it after then it may come out unaligned (e.g., job->align = 8 > >>> and ps.chunk_size = 1). > >> > >> Sure, I think Kamlesh was the first to suggest max, so maybe Kamlesh > >> would like to make the change. I'll send a patch otherwise. > > Thanks for consideration, Daniel. I'll send a patch. > Sent. > > Just curious about one thing on line 495, > > nworks = max(job->size / max(job->min_chunk, job->align), 1ul); > > what happens if both min_chunk and align are 0. That's a fair point. It's another of those things that's not supposed to happen, but it's worth making padata robust to it.
