On Wed, 2017-06-14 at 18:45 -0600, Toshi Kani wrote: > On Fri, 2017-06-09 at 13:23 -0700, Dan Williams wrote: > > Allow device-mapper to route copy_from_iter operations to the > > per-target implementation. In order for the device stacking to work > > we need a dax_dev and a pgoff relative to that device. This gives > > each layer of the stack the information it needs to look up the > > operation pointer for the next level. > > > > This conceptually allows for an array of mixed device drivers with > > varying copy_from_iter implementations. > > > > Cc: Toshi Kani <toshi.kani@xxxxxxx> > > Reviewed-by: Mike Snitzer <snitzer@xxxxxxxxxx> > > Signed-off-by: Dan Williams <dan.j.williams@xxxxxxxxx> > > I was worried about possible overhead with additional stub calls, but > it looks fine with a single thread fio write test with direct=1. > > 92.62% [kernel.kallsyms] [k] __copy_user_nocache > 0.04% [kernel.kallsyms] [k] entry_SYSCALL_64_fastpath > 0.08% libpthread-2.22.so [.] __GI___libc_write > 0.01% [kernel.kallsyms] [k] sys_write > 0.02% [kernel.kallsyms] [k] vfs_write > 0.02% [kernel.kallsyms] [k] __vfs_write > 0.02% [kernel.kallsyms] [k] ext4_file_write_iter > 0.02% [kernel.kallsyms] [k] dax_iomap_rw > 0.03% [kernel.kallsyms] [k] iomap_apply > 0.04% [kernel.kallsyms] [k] dax_iomap_actor > 0.01% [kernel.kallsyms] [k] dax_copy_from_iter > 0.01% [kernel.kallsyms] [k] dm_dax_copy_from_iter > 0.01% [kernel.kallsyms] [k] linear_dax_copy_from_iter > 0.03% [kernel.kallsyms] [k] copy_from_iter_flushcache > 0.00% [kernel.kallsyms] [k] pmem_copy_from_iter I had bs=256k, which was too big for this test. bs=4k result is not this pretty at all, only 23% in __copy_user_nocache. This change accounts for approx. 1% with 4k. Given we have larger overheads in many other functions in the path, the change looks acceptable (I keep my review-by). I'd prefer to reduce code in the path, though. Thanks, -Toshi
