[ adding Konstantin ] On Wed, Aug 24, 2016 at 7:40 AM, Kani, Toshimitsu <toshi.k...@hpe.com> wrote: > On Tue, 2016-08-23 at 21:48 -0700, Dan Williams wrote: >> On Tue, Aug 23, 2016 at 9:28 PM, Kani, Toshimitsu <toshi.k...@hpe.com >> > >> > BTW, why does devm_memremap_pages() put a whole range to >> > pgmap_radix as device memory, but only initialize page->pgmap for >> > its data range? Is there particular reason for this inconsistency? >> >> The radix tree is indexed by section number, but we don't always >> initialize a full section. The cases when we don't use a full >> section is when it overlaps device metadata, or if a platform >> multiplexes the device memory range with another resource within the >> same section. > > I see, but I still feel odd about making get_dev_pagemap() to work for > metadata, but get_page() -> get_zone_device_page() to crash like this. > follow_devmap_pmd() assumes get_page() to work when get_dev_pagemap() > returns a valid pgmap... >
We could leave the unmapped portions of a section out of the radix, but I'm also worried about keeping the get_dev_pagemap() lookup cheap. I saw that Konstantin has some proposed changes to the radix api to make it easier to fill a range [1]. I might switch to radix_tree_fill_range() when it becomes available. [1]: https://github.com/koct9i/linux/commits/radix-tree
