On Fri, Mar 8, 2024 at 7:04 AM Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > > On Thu, Mar 07, 2024 at 07:49:16PM -0800, Alexei Starovoitov wrote: > > Ok. I think I figured it out. > > Please try the attached patch. > > I don't think this is the right thing. The probem is that > the PCI code shouldn't really be using ioremap_page_range if it is > not an ioremap area, but instead directly call into > vmap_range_noflush (or an added back vmap_range to avoid all the > duplication) similar to the vunmap case in vunmap_range. vmap_range_noflush() is static in mm/vmalloc.c There is vmap_pages_range_noflush() that is in mm/internal.h, but it needs pages instead of phys_addr_t. Newly introduced vm_area_map_pages() needs struct vm_struct *area and struct page **pages. In this PCI case there is no vm_struct and no pages. ioremap_page_range() is the only api that fits. afaict.
