From: John Hubbard <jhubbard@xxxxxxxxxx>
Hi,
I'm calling this RFC v2, even though with all the discussion it actually
feels
like about v7 or so. But now that the dust has settled, it's time to show a
surprisingly small, cleaner approach. Jan and Jerome came up with a scheme
(discussed in more detail in "track gup-pinned pages" commit description)
that
does not require any additional struct page fields. This approach has the
additional advantage of being very lightweight and therefore fast, because
a) it mostly just does atomics,
b) unlike previous approaches, there is no need to remove and re-add to
LRUs.
c) it uses the same lock-free algorithms that get_user_pages already
relies upon.
This RFC shows the following:
1) A patch to get the call site conversion started:
mm: introduce put_user_page*(), placeholder versions
2) A sample call site conversion:
infiniband/mm: convert put_page() to put_user_page*()
...NOT shown: all of the other 100+ gup call site conversions. Again,
those are in various states of progress and disrepair, at [1].
3) Tracking, instrumentation, and documentation patches, once all the call
sites have been converted.
4) A small refactoring patch that I'm also going to submit separately, for
the page_cache_add_speculative() routine.
This seems to be working pretty well here. I've converted enough call sites
(there is git repo [1] with that, which gets rebased madly, but it's there if
you really want to try some early testing) to run things such as fio.
Performance: here is an fio run on an NVMe drive, using this for the fio
configuration file:
[reader]
direct=1
ioengine=libaio
blocksize=4096
size=1g
numjobs=1
rw=read
iodepth=64
reader: (g=0): rw=read, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=libaio, iodepth=64
fio-3.3
Starting 1 process
Jobs: 1 (f=1)
reader: (groupid=0, jobs=1): err= 0: pid=7011: Sun Feb 3 20:36:51 2019
read: IOPS=190k, BW=741MiB/s (778MB/s)(1024MiB/1381msec)
slat (nsec): min=2716, max=57255, avg=4048.14, stdev=1084.10
clat (usec): min=20, max=12485, avg=332.63, stdev=191.77
lat (usec): min=22, max=12498, avg=336.72, stdev=192.07
clat percentiles (usec):
| 1.00th=[ 322], 5.00th=[ 322], 10.00th=[ 322], 20.00th=[ 326],
| 30.00th=[ 326], 40.00th=[ 326], 50.00th=[ 326], 60.00th=[ 326],
| 70.00th=[ 326], 80.00th=[ 330], 90.00th=[ 330], 95.00th=[ 330],
| 99.00th=[ 478], 99.50th=[ 717], 99.90th=[ 1074], 99.95th=[ 1090],
| 99.99th=[12256]
bw ( KiB/s): min=730152, max=776512, per=99.22%, avg=753332.00, stdev=32781.47, samples=2
iops : min=182538, max=194128, avg=188333.00, stdev=8195.37, samples=2
lat (usec) : 50=0.01%, 100=0.01%, 250=0.07%, 500=99.26%, 750=0.38%
lat (usec) : 1000=0.02%
lat (msec) : 2=0.24%, 20=0.02%
cpu : usr=15.07%, sys=84.13%, ctx=10, majf=0, minf=74
IO depths : 1=0.1%, 2=0.1%, 4=0.1%, 8=0.1%, 16=0.1%, 32=0.1%, >=64=100.0%
submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.1%, >=64=0.0%
issued rwts: total=262144,0,0,0 short=0,0,0,0 dropped=0,0,0,0
latency : target=0, window=0, percentile=100.00%, depth=64
Run status group 0 (all jobs):
READ: bw=741MiB/s (778MB/s), 741MiB/s-741MiB/s (778MB/s-778MB/s), io=1024MiB (1074MB), run=1381-1381msec
Disk stats (read/write):
nvme0n1: ios=216966/0, merge=0/0, ticks=6112/0, in_queue=704, util=91.34%
A write up of the larger problem follows (co-written with Jérôme Glisse):
Overview
========
Some kernel components (file systems, device drivers) need to access
memory that is specified via process virtual address. For a long time, the
API to achieve that was get_user_pages ("GUP") and its variations. However,
GUP has critical limitations that have been overlooked; in particular, GUP
does not interact correctly with filesystems in all situations. That means
that file-backed memory + GUP is a recipe for potential problems, some of
which have already occurred in the field.
GUP was first introduced for Direct IO (O_DIRECT), allowing filesystem code
to get the struct page behind a virtual address and to let storage hardware
perform a direct copy to or from that page. This is a short-lived access
pattern, and as such, the window for a concurrent writeback of GUP'd page
was small enough that there were not (we think) any reported problems.
Also, userspace was expected to understand and accept that Direct IO was
not synchronized with memory-mapped access to that data, nor with any
process address space changes such as munmap(), mremap(), etc.
Over the years, more GUP uses have appeared (virtualization, device
drivers, RDMA) that can keep the pages they get via GUP for a long period
of time (seconds, minutes, hours, days, ...). This long-term pinning makes
an underlying design problem more obvious.
In fact, there are a number of key problems inherent to GUP:
Interactions with file systems
==============================
File systems expect to be able to write back data, both to reclaim pages,
and for data integrity. Allowing other hardware (NICs, GPUs, etc) to gain
write access to the file memory pages means that such hardware can dirty the
pages, without the filesystem being aware. This can, in some cases
(depending on filesystem, filesystem options, block device, block device
options, and other variables), lead to data corruption, and also to kernel
bugs of the form:
kernel BUG at /build/linux-fQ94TU/linux-4.4.0/fs/ext4/inode.c:1899!
backtrace:
ext4_writepage
__writepage
write_cache_pages
ext4_writepages
do_writepages
__writeback_single_inode
writeback_sb_inodes
__writeback_inodes_wb
wb_writeback
wb_workfn
process_one_work
worker_thread
kthread
ret_from_fork
...which is due to the file system asserting that there are still buffer
heads attached:
({ \
BUG_ON(!PagePrivate(page)); \
((struct buffer_head *)page_private(page)); \
})
Dave Chinner's description of this is very clear:
"The fundamental issue is that ->page_mkwrite must be called on every
write access to a clean file backed page, not just the first one.
How long the GUP reference lasts is irrelevant, if the page is clean
and you need to dirty it, you must call ->page_mkwrite before it is
marked writeable and dirtied. Every. Time."
This is just one symptom of the larger design problem: filesystems do not
actually support get_user_pages() being called on their pages, and letting
hardware write directly to those pages--even though that pattern has been
going on since about 2005 or so.
Long term GUP
=============
Long term GUP is an issue when FOLL_WRITE is specified to GUP (so, a
writeable mapping is created), and the pages are file-backed. That can lead
to filesystem corruption. What happens is that when a file-backed page is
being written back, it is first mapped read-only in all of the CPU page
tables; the file system then assumes that nobody can write to the page, and
that the page content is therefore stable. Unfortunately, the GUP callers
generally do not monitor changes to the CPU pages tables; they instead
assume that the following pattern is safe (it's not):
get_user_pages()
Hardware can keep a reference to those pages for a very long time,
and write to it at any time. Because "hardware" here means "devices
that are not a CPU", this activity occurs without any interaction
with the kernel's file system code.
for each page
set_page_dirty
put_page()
In fact, the GUP documentation even recommends that pattern.
Anyway, the file system assumes that the page is stable (nothing is writing
to the page), and that is a problem: stable page content is necessary for
many filesystem actions during writeback, such as checksum, encryption,
RAID striping, etc. Furthermore, filesystem features like COW (copy on
write) or snapshot also rely on being able to use a new page for as memory
for that memory range inside the file.
Corruption during write back is clearly possible here. To solve that, one
idea is to identify pages that have active GUP, so that we can use a bounce
page to write stable data to the filesystem. The filesystem would work
on the bounce page, while any of the active GUP might write to the
original page. This would avoid the stable page violation problem, but note
that it is only part of the overall solution, because other problems
remain.
Other filesystem features that need to replace the page with a new one can
be inhibited for pages that are GUP-pinned. This will, however, alter and
limit some of those filesystem features. The only fix for that would be to
require GUP users monitor and respond to CPU page table updates. Subsystems
such as ODP and HMM do this, for example. This aspect of the problem is
still under discussion.
Direct IO
=========
Direct IO can cause corruption, if userspace does Direct-IO that writes to
a range of virtual addresses that are mmap'd to a file. The pages written
to are file-backed pages that can be under write back, while the Direct IO
is taking place. Here, Direct IO need races with a write back: it calls
GUP before page_mkclean() has replaced the CPU pte with a read-only entry.
The race window is pretty small, which is probably why years have gone by
before we noticed this problem: Direct IO is generally very quick, and
tends to finish up before the filesystem gets around to do anything with
the page contents. However, it's still a real problem. The solution is
to never let GUP return pages that are under write back, but instead,
force GUP to take a write fault on those pages. That way, GUP will
properly synchronize with the active write back. This does not change the
required GUP behavior, it just avoids that race.
What this patchset does
=======================
This patchset overloads page->_refcount, in order to track GUP-pinned
pages.
This patchset checks if the page is under write back, and if so, it backs
off and forces a page fault (via the GUP slow path). Before this patchset,
GUP might have returned the struct page because page_mkclean() had not yet
updated the CPU page table. After this patch, GUP no longer race with
page_mkclean() and thus any user of GUP properly synchronize on active
write back (this is not only useful to direct-IO, but also to other users
of GUP).
This patchset does not include any of the filesystem changes needed to
fix the issues. That is left as a separate patchset that will use the
new flag.
Changes from earlier versions
=============================
-- Fixed up kerneldoc issues in put_user_page*() functions, in response
to Mike Rapoport's review.
-- Use overloaded page->_refcount to track gup-pinned pages. This avoids the
need for an extra page flag, and also avoids the need for an extra counting
field.
[1] git@xxxxxxxxxx:johnhubbard/linux.git (branch: gup_dma_core)
[2] https://lwn.net/Articles/753027/ "The trouble with get_user_pages()"
Suggested-by: Jan Kara <jack@xxxxxxx>
Suggested-by: Jérôme Glisse <jglisse@xxxxxxxxxx>
Cc: Christian Benvenuti <benve@xxxxxxxxx>
Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Cc: Christopher Lameter <cl@xxxxxxxxx>
Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
Cc: Dave Chinner <david@xxxxxxxxxxxxx>
Cc: Dennis Dalessandro <dennis.dalessandro@xxxxxxxxx>
Cc: Doug Ledford <dledford@xxxxxxxxxx>
Cc: Jan Kara <jack@xxxxxxx>
Cc: Jason Gunthorpe <jgg@xxxxxxxx>
Cc: Jérôme Glisse <jglisse@xxxxxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Michal Hocko <mhocko@xxxxxxxxxx>
Cc: Mike Rapoport <rppt@xxxxxxxxxxxxx>
Cc: Mike Marciniszyn <mike.marciniszyn@xxxxxxxxx>
Cc: Ralph Campbell <rcampbell@xxxxxxxxxx>
Cc: Tom Talpey <tom@xxxxxxxxxx>
John Hubbard (6):
mm: introduce put_user_page*(), placeholder versions
infiniband/mm: convert put_page() to put_user_page*()
mm: page_cache_add_speculative(): refactoring
mm/gup: track gup-pinned pages
mm/gup: /proc/vmstat support for get/put user pages
mm/gup: Documentation/vm/get_user_pages.rst, MAINTAINERS
Documentation/vm/get_user_pages.rst | 197 ++++++++++++++++++++
Documentation/vm/index.rst | 1 +
MAINTAINERS | 10 +
drivers/infiniband/core/umem.c | 7 +-
drivers/infiniband/core/umem_odp.c | 2 +-
drivers/infiniband/hw/hfi1/user_pages.c | 11 +-
drivers/infiniband/hw/mthca/mthca_memfree.c | 6 +-
drivers/infiniband/hw/qib/qib_user_pages.c | 11 +-
drivers/infiniband/hw/qib/qib_user_sdma.c | 6 +-
drivers/infiniband/hw/usnic/usnic_uiom.c | 7 +-
include/linux/mm.h | 57 ++++++
include/linux/mmzone.h | 5 +
include/linux/pagemap.h | 36 ++--
mm/gup.c | 80 ++++++--
mm/swap.c | 104 +++++++++++
mm/vmstat.c | 5 +
16 files changed, 482 insertions(+), 63 deletions(-)
create mode 100644 Documentation/vm/get_user_pages.rst
--
2.20.1
