This RFC is posted to collect feedback and to be used as a reference implementation for the upcoming LSF/MM discussion. The code is not of production quality and needs more testing. Guaranteed CMA (GCMA) is designed to improve utilization of reserved memory carveouts without compromising their advantages of: 1. Guaranteed success of allocation (as long as total allocation size is below the size of the reservation. 2. Low allocation latency. The idea is that carved out memory when not used for its primary purpose can be donated and used as an extension of the pagecache and any donated page can be taken back at any moment with minimal latency and guaranteed success. To achieve this, GCMA needs to use memory that is not addressable by the kernel (can't be pinned) and that contains content that can be discarded. To provide such memory we reintroduce cleancache idea [1] with two major changes. New implementation: 1. Avoids intrusive hooks into filesystem code, limiting them to two hooks for filesystem mount/unmount events and a hook for bdev invalidation. 2. Manages fs->inode->page tracking and handles pools of donated pages inside cleancache itself, freeing backends of this burden. Cleancache provides a simple interface to its backends which lets them donate a bunch of pages to cleancache, take a page back for own use and return the page back to cleancache when not needed. With cleancache in place, GCMA becomes a thin layer linking CMA allocator to cleancache which allows existing CMA API to be used for continuous memory allocations with additional guarantees listed above. The limitation of GCMA is that its donated memory can be used only to extend file-backed pagecache. Note that both CMA and GCMA can be used at the same time. The patchset borrows some ideas and code from previous implementations of the cleancache and GCMA [2] as well as Android's reference patchset [3] implemented by Minchan Kim and used by many Android vendors. [1] https://elixir.bootlin.com/linux/v5.16.20/source/Documentation/vm/cleancache.rst [2] https://lore.kernel.org/lkml/1424721263-25314-1-git-send-email-sj38.park@xxxxxxxxx/ [3] https://android-review.googlesource.com/q/topic:%22gcma_6.12%22 Minchan Kim (1): mm: introduce GCMA Suren Baghdasaryan (2): mm: implement cleancache mm: integrate GCMA with CMA using dt-bindings arch/powerpc/kernel/fadump.c | 2 +- block/bdev.c | 8 + fs/super.c | 3 + include/linux/cleancache.h | 88 ++++ include/linux/cma.h | 2 +- include/linux/fs.h | 7 + include/linux/gcma.h | 12 + kernel/dma/contiguous.c | 11 +- mm/Kconfig | 32 ++ mm/Makefile | 2 + mm/cleancache.c | 926 +++++++++++++++++++++++++++++++++++ mm/cma.c | 33 +- mm/cma.h | 1 + mm/cma_sysfs.c | 10 + mm/filemap.c | 63 ++- mm/gcma.c | 155 ++++++ mm/truncate.c | 21 +- 17 files changed, 1356 insertions(+), 20 deletions(-) create mode 100644 include/linux/cleancache.h create mode 100644 include/linux/gcma.h create mode 100644 mm/cleancache.c create mode 100644 mm/gcma.c base-commit: 40caf747267c18b6206e26a37d6ea6b695236c11 -- 2.49.0.rc1.451.g8f38331e32-goog
