On Thu, Feb 02, 2017 at 11:49:01PM -0800, Christoph Hellwig wrote: > On Thu, Feb 02, 2017 at 08:00:52AM -0500, Jeff Layton wrote: > > Yeah, that might work. You could kmalloc the buffer array according to > > the maxsize value. For small ones we could even consider using an on- > > stack buffer. > > For the block direct I/O code we defintively want to avoid any > allocations for small I/O a that shows up in the performance numbers. > And we'd like to reuse the on-stack bio_vec so that the defintion of > a small I/O can be as big as possible without blowing up the stack. Hmm... Reuse part is really nasty ;-/ OTOH, it might make sense to have a "fill bio_vec array" as separate primitive - having that sucker come from bio looks like an artificial restriction. OK, next question, seeing that you've dealt with O_DIRECT guts more than I have. When we have iov_iter_get_pages() fail on do_direct_IO() write with some blocks already allocated, we pick zero page as data source. So far, so good, but: * should we bother zeroing anything unless buffer_new() is true? * why, in case of more than a page worth of pending allocated blocks, do we bother with calling iov_iter_get_pages() again and again? We *do* take care not to allocate anything else after that point, but dio_get_page() will be calling iov_iter_get_pages() every time in that case - there's only one page in queue.
