On Sun, Oct 25, 2020 at 04:44:38AM +0000, Matthew Wilcox wrote: > > On my laptop, I have about 31MB allocated to buffer_heads. > > buffer_head 182728 299910 104 39 1 : tunables 0 0 0 : slabdata 7690 7690 0 > > Reducing the size of the buffer_head by 8 bytes gets us to 96 bytes, > which means we get 42 per page instead of 39 and saves me 2MB of memory. > > I think b_end_io() is ripe for removal. It's only used while the I/O > is in progress, and it's always set to end_bio_bh_io_sync() which > may set the quiet bit, calls ->b_end_io and calls bio_put(). > > So how about this as an approach? Only another 40 or so call-sites > to take care of to eliminate b_end_io from the buffer_head. Yes, this > particular example should be entirely rewritten to do away with buffer > heads, but that's been true since 2006. I'm looking for an approach > which can be implemented quickly since the buffer_head does not appear > to be going away any time soon. I think this looks pretty reasonable.
