> > > Yes, but it's gains very little ... architectures that don't want it can
So show a specific device where the virtual merge accounting is useful.
(1) The device that is often used in alpha or pa-risc environments ---
because the accounting is not used on other archs.
(2) The device that is performance-sensitive --- not something outdated or
unusual.
(3) And the device that has limited sg-list size, so that generic I/O
requests made by the kernel hit this limit. (if the sg-list is so big that
nr_phys_segments of most requests fits into it, you don't need to count
nr_hw_segments --- because nr_hw_segments < nr_phys_segments and
nr_phys_segments already fits).
[ the device that traverses its sg-list slowly doesn't fall into category
(3), beacuse virtual merging would happen with or without nr_hw_segments
accounting ]
Mikulas
--