On Tue, Feb 16, 2021 at 06:48:24PM -0800, ira.weiny@xxxxxxxxx wrote: > From: Ira Weiny <ira.weiny@xxxxxxxxx> > > These kmaps are thread local and don't need to be atomic. So they can use > the more efficient kmap_local_page(). However, the mapping of pages in > the stripes and the additional parity and qstripe pages are a bit > trickier because the unmapping must occur in the opposite order from the > mapping. Furthermore, the pointer array in __raid_recover_end_io() may > get reordered. > > Convert these calls to kmap_local_page() taking care to reverse the > unmappings of any page arrays as well as being careful with the mappings > of any special pages such as the parity and qstripe pages. Though there's one more allocation for the additional array, I don't see a simpler way to avoid it and track the array reordering at a lower memory cost. At least it's not in a performance critical code and the array size is reasonably small.
