From: Christoph Hellwig <hch@xxxxxxxxxxxxx> Sent: Wednesday, November 22, 2023 11:32 PM > > On Thu, Nov 23, 2023 at 12:24:49AM +0000, Michael Kelley wrote: > > > I really do not want to expose vmap_pages_range. Please try to come up > > > with a good way to encapsulate your map at a certain address primitive > > > and implement it in vmalloc.c. > > > > To clarify, is your concern narrowly about vmap_pages_range() > > specifically? > > The prime concern is that it took a lot of effort to make > vmap_pages_range static and remove all the abuses. I absolutely > object to undoing that. OK, so I assume that means a new variant of vmap_pages_range(), such as one that always sets the page_shift parameter to PAGE_SIZE, is also disallowed because of the same potential for abuse. So the only way to map a system memory page to a vmalloc vaddr is via vmap() or some vmap() variant, which always creates a new vmalloc area via get_vm_area(). I've done the perf measurements, and that approach won't work for this use case. Independent of the alignment requirements, the churn in creating and removing a lot of vmalloc areas has too much perf impact. The use case needs to create a single vmalloc area, and then repeatedly map/unmap a page in that existing area. I'll have to handle the top-level problem in this patch set in a completely different way. Michael
