Changes since v3 [1]:
* Kill the i_daxdma_lock, and do not impose any new locking constraints
on filesystem implementations (Dave)
* Reuse the existing i_mmap_lock for synchronizing against
get_user_pages() by unmapping and causing punch-hole/truncate to
re-fault the page before get_user_pages() can elevate the page reference
count (Jan)
* Create a dax-specifc address_space_operations instance for each
filesystem. This allows page->mapping to be set for dax pages. (Jan).
* Change the ext4 and ext2 policy of 'mount -o dax' vs a device that
does not support dax. This converts any environments that may have
been using 'page-less' dax back to using page cache.
* Rename wait_on_devmap_idle() to wait_on_atomic_one(), a generic
facility for waiting for an atomic counter to reach a value of '1'.
[1]: https://lwn.net/Articles/737273/
---
Background:
get_user_pages() pins file backed memory pages for access by dma
devices. However, it only pins the memory pages not the page-to-file
offset association. If a file is truncated the pages are mapped out of
the file and dma may continue indefinitely into a page that is owned by
a device driver. This breaks coherency of the file vs dma, but the
assumption is that if userspace wants the file-space truncated it does
not matter what data is inbound from the device, it is not relevant
anymore. The only expectation is that dma can safely continue while the
filesystem reallocates the block(s).
Problem:
This expectation that dma can safely continue while the filesystem
changes the block map is broken by dax. With dax the target dma page
*is* the filesystem block. The model of leaving the page pinned for dma,
but truncating the file block out of the file, means that the filesytem
is free to reallocate a block under active dma to another file and now
the expected data-incoherency situation has turned into active
data-corruption.
Solution:
Defer all filesystem operations (fallocate(), truncate()) on a dax mode
file while any page/block in the file is under active dma. This solution
assumes that dma is transient. Cases where dma operations are known to
not be transient, like RDMA, have been explicitly disabled via
commits like 5f1d43de5416 "IB/core: disable memory registration of
filesystem-dax vmas".
The dax_flush_dma() routine is called by filesystems with a lock held
against mm faults (i_mmap_lock). It then invalidates all mappings to
trigger any subsequent get_user_pages() to block on i_mmap_lock. Finally
it scans/rescans all pages in the mapping until it observes all pages
idle.
So far this solution only targets xfs since it already implements
xfs_break_layouts in all the locations that would need this
synchronization. It applies on top of the vmem_altmap / dev_pagemap
reworks from Christoph.
---
Dan Williams (18):
mm, dax: introduce pfn_t_special()
ext4: auto disable dax instead of failing mount
ext2: auto disable dax instead of failing mount
dax: require 'struct page' by default for filesystem dax
dax: stop using VM_MIXEDMAP for dax
dax: stop using VM_HUGEPAGE for dax
dax: store pfns in the radix
tools/testing/nvdimm: add 'bio_delay' mechanism
mm, dax: enable filesystems to trigger dev_pagemap ->page_free callbacks
mm, dev_pagemap: introduce CONFIG_DEV_PAGEMAP_OPS
fs, dax: introduce DEFINE_FSDAX_AOPS
xfs: use DEFINE_FSDAX_AOPS
ext4: use DEFINE_FSDAX_AOPS
ext2: use DEFINE_FSDAX_AOPS
mm, fs, dax: use page->mapping to warn if dma collides with truncate
wait_bit: introduce {wait_on,wake_up}_atomic_one
mm, fs, dax: dax_flush_dma, handle dma vs block-map-change collisions
xfs, dax: wire up dax_flush_dma support via a new xfs_sync_dma helper
arch/powerpc/platforms/Kconfig | 1
arch/powerpc/sysdev/axonram.c | 2
drivers/dax/device.c | 1
drivers/dax/super.c | 100 ++++++++++-
drivers/nvdimm/pmem.c | 3
drivers/s390/block/Kconfig | 1
drivers/s390/block/dcssblk.c | 3
fs/Kconfig | 8 +
fs/dax.c | 295 ++++++++++++++++++++++++++++-----
fs/ext2/ext2.h | 1
fs/ext2/file.c | 1
fs/ext2/inode.c | 23 ++-
fs/ext2/namei.c | 18 --
fs/ext2/super.c | 13 +
fs/ext4/file.c | 1
fs/ext4/inode.c | 6 +
fs/ext4/super.c | 15 +-
fs/xfs/Makefile | 3
fs/xfs/xfs_aops.c | 2
fs/xfs/xfs_aops.h | 1
fs/xfs/xfs_dma.c | 81 +++++++++
fs/xfs/xfs_dma.h | 24 +++
fs/xfs/xfs_file.c | 8 -
fs/xfs/xfs_ioctl.c | 7 -
fs/xfs/xfs_iops.c | 12 +
fs/xfs/xfs_super.c | 20 +-
include/linux/dax.h | 70 +++++++-
include/linux/memremap.h | 28 +--
include/linux/mm.h | 62 +++++--
include/linux/pfn_t.h | 13 +
include/linux/vma.h | 23 +++
include/linux/wait_bit.h | 13 +
kernel/memremap.c | 30 +++
kernel/sched/wait_bit.c | 59 ++++++-
mm/Kconfig | 5 +
mm/gup.c | 5 +
mm/hmm.c | 13 -
mm/huge_memory.c | 6 -
mm/ksm.c | 3
mm/madvise.c | 2
mm/memory.c | 22 ++
mm/migrate.c | 3
mm/mlock.c | 5 -
mm/mmap.c | 8 -
mm/swap.c | 3
tools/testing/nvdimm/Kbuild | 1
tools/testing/nvdimm/test/iomap.c | 62 +++++++
tools/testing/nvdimm/test/nfit.c | 34 ++++
tools/testing/nvdimm/test/nfit_test.h | 1
49 files changed, 918 insertions(+), 203 deletions(-)
create mode 100644 fs/xfs/xfs_dma.c
create mode 100644 fs/xfs/xfs_dma.h
create mode 100644 include/linux/vma.h
> So far this solution only targets xfs since it already implements
> xfs_break_layouts in all the locations that would need this
> synchronization. It applies on top of the vmem_altmap / dev_pagemap
> reworks from Christoph.
Those got a rebase since your posting of this series, btw.