LINUX: Detect NULL page during write_begin 42/13242/2
authorAndrew Deason <adeason@sinenomine.net>
Thu, 5 Jul 2018 22:16:48 +0000 (17:16 -0500)
committerBenjamin Kaduk <kaduk@mit.edu>
Fri, 6 Jul 2018 03:26:41 +0000 (23:26 -0400)
commit89e80c354c404dedc0e5197f99710db0e5e08767
tree7feae1569860fd1a2bbae760cc9afd8de7e81523
parentb1ad473be01162fe9b3835544a835c4dcf0fcb35
LINUX: Detect NULL page during write_begin

In afs_linux_write_begin, we call grab_cache_page_write_begin to get a
page to use for writing data when servicing a write into AFS. Under
low-memory conditions, this can return NULL if Linux cannot find a
free page to use. Currently, we always try to reference the page
returned, and so this causes a BUG.

To avoid this, check if grab_cache_page_write_begin returns NULL, and
just return -ENOMEM, like other callers of grab_cache_page_write_begin
do.

Linux's fault injection framework is useful for testing code paths
like these. The following settings made it possible to
somewhat-reliably exercise the relevant code path on a test RHEL7
system:

    # grep ^ /sys/kernel/debug/fail_page_alloc/*
    /sys/kernel/debug/fail_page_alloc/ignore-gfp-highmem:Y
    /sys/kernel/debug/fail_page_alloc/ignore-gfp-wait:N
    /sys/kernel/debug/fail_page_alloc/interval:1
    /sys/kernel/debug/fail_page_alloc/min-order:0
    /sys/kernel/debug/fail_page_alloc/probability:100
    /sys/kernel/debug/fail_page_alloc/space:90
    /sys/kernel/debug/fail_page_alloc/task-filter:Y
    /sys/kernel/debug/fail_page_alloc/times:-1
    [...]

Change-Id: I00908658ae43aa3c8e12f2a0b956016d4441016c
Reviewed-on: https://gerrit.openafs.org/13242
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
src/afs/LINUX/osi_vnodeops.c