Messages in this thread | | | Date | Tue, 9 Jun 2015 17:58:40 +0800 | Subject | Re: [PATCH v4 1/5] fs: kiocb: introduce IOCB_DONT_DIRTY_PAGE flag for direct IO | From | Ming Lei <> |
| |
On Tue, Jun 9, 2015 at 2:29 PM, Christoph Hellwig <hch@infradead.org> wrote: > On Sat, Jun 06, 2015 at 08:42:33AM +0800, Ming Lei wrote: >> Both ITER_KVEC and ITER_BVEC doesn't mean the pages are kernel >> page, for example of loop and swap. That is why this patch is more flexiable, >> and won't cause regression since the users may have different dirtying >> rules as you mentioned last time. >> >> http://marc.info/?t=143093223200001&r=1&w=2 > > Again, we never dirty pages for the caller when doing general purpose > kernel I/O. For loop we either don't need it (totally in-kernel block > I/O) or the caller takes care of it (direct I/O on the loop device). > > The swap code currently only uses ->direct_IO and thus isn't affected > by this flag. If we use it for write the low-level I/O code should not > dirty the page either. > > So to repeat myself: for the current state of affairs adding a flag > that every sensible user has to set is a horrible interface. If for
There shouldn't be lots of such uses, and the flag still has the document benifit, that means the caller should think and check the current dirtying usage.
> some unforseen reason we'll need the flag later on it should have > reverse polarity, and only be added when needed.
OK, I will remove the flag in v5 since loop dio is the 1st read ->direct_IO via ITER_KVEC/ITER_BVEC inside kernel.
But I am wondering it is good to decide dirtying by ITER_KVEC and ITER_BVEC, and dirtying pages should better be checked case by case, at least for loop, the usage is a bit special(fs over loop is taking care of that)
Thanks, Ming
| |