target/linux/generic/backport-5.15/020-v6.3-20-mm-support-POSIX_FADV_NOREUSE.patch

   1 From 686c3d4f71de9e0e7a27f03a5617a712385f90cd Mon Sep 17 00:00:00 2001
   2 From: Yu Zhao <yuzhao@google.com>
   3 Date: Fri, 30 Dec 2022 14:52:52 -0700
   4 Subject: [PATCH 20/29] mm: support POSIX_FADV_NOREUSE
   5
   6 This patch adds POSIX_FADV_NOREUSE to vma_has_recency() so that the LRU
   7 algorithm can ignore access to mapped files marked by this flag.
   8
   9 The advantages of POSIX_FADV_NOREUSE are:
  10 1. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not alter the
  11    default readahead behavior.
  12 2. Unlike MADV_SEQUENTIAL and MADV_RANDOM, it does not split VMAs and
  13    therefore does not take mmap_lock.
  14 3. Unlike MADV_COLD, setting it has a negligible cost, regardless of
  15    how many pages it affects.
  16
  17 Its limitations are:
  18 1. Like POSIX_FADV_RANDOM and POSIX_FADV_SEQUENTIAL, it currently does
  19    not support range. IOW, its scope is the entire file.
  20 2. It currently does not ignore access through file descriptors.
  21    Specifically, for the active/inactive LRU, given a file page shared
  22    by two users and one of them having set POSIX_FADV_NOREUSE on the
  23    file, this page will be activated upon the second user accessing
  24    it. This corner case can be covered by checking POSIX_FADV_NOREUSE
  25    before calling mark_page_accessed() on the read path. But it is
  26    considered not worth the effort.
  27
  28 There have been a few attempts to support POSIX_FADV_NOREUSE, e.g., [1].
  29 This time the goal is to fill a niche: a few desktop applications, e.g.,
  30 large file transferring and video encoding/decoding, want fast file
  31 streaming with mmap() rather than direct IO.  Among those applications, an
  32 SVT-AV1 regression was reported when running with MGLRU [2].  The
  33 following test can reproduce that regression.
  34
  35   kb=$(awk '/MemTotal/ { print $2 }' /proc/meminfo)
  36   kb=$((kb - 8*1024*1024))
  37
  38   modprobe brd rd_nr=1 rd_size=$kb
  39   dd if=/dev/zero of=/dev/ram0 bs=1M
  40
  41   mkfs.ext4 /dev/ram0
  42   mount /dev/ram0 /mnt/
  43   swapoff -a
  44
  45   fallocate -l 8G /mnt/swapfile
  46   mkswap /mnt/swapfile
  47   swapon /mnt/swapfile
  48
  49   wget http://ultravideo.cs.tut.fi/video/Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
  50   7z e -o/mnt/ Bosphorus_3840x2160_120fps_420_8bit_YUV_Y4M.7z
  51   SvtAv1EncApp --preset 12 -w 3840 -h 2160 \
  52                -i /mnt/Bosphorus_3840x2160.y4m
  53
  54 For MGLRU, the following change showed a [9-11]% increase in FPS,
  55 which makes it on par with the active/inactive LRU.
  56
  57   patch Source/App/EncApp/EbAppMain.c <<EOF
  58   31a32
  59   > #include <fcntl.h>
  60   35d35
  61   < #include <fcntl.h> /* _O_BINARY */
  62   117a118
  63   >             posix_fadvise(config->mmap.fd, 0, 0, POSIX_FADV_NOREUSE);
  64   EOF
  65
  66 [1] https://lore.kernel.org/r/1308923350-7932-1-git-send-email-andrea@betterlinux.com/
  67 [2] https://openbenchmarking.org/result/2209259-PTS-MGLRU8GB57
  68
  69 Link: https://lkml.kernel.org/r/20221230215252.2628425-2-yuzhao@google.com
  70 Signed-off-by: Yu Zhao <yuzhao@google.com>
  71 Cc: Alexander Viro <viro@zeniv.linux.org.uk>
  72 Cc: Andrea Righi <andrea.righi@canonical.com>
  73 Cc: Johannes Weiner <hannes@cmpxchg.org>
  74 Cc: Michael Larabel <Michael@MichaelLarabel.com>
  75 Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
  76 ---
  77  include/linux/fs.h        | 2 ++
  78  include/linux/mm_inline.h | 3 +++
  79  mm/fadvise.c              | 5 ++++-
  80  3 files changed, 9 insertions(+), 1 deletion(-)
  81
  82 --- a/include/linux/fs.h
  83 +++ b/include/linux/fs.h
  84 @@ -167,6 +167,8 @@ typedef int (dio_iodone_t)(struct kiocb
  85  /* File is stream-like */
  86  #define FMODE_STREAM           ((__force fmode_t)0x200000)
  87
  88 +#define        FMODE_NOREUSE           ((__force fmode_t)0x400000)
  89 +
  90  /* File was opened by fanotify and shouldn't generate fanotify events */
  91  #define FMODE_NONOTIFY         ((__force fmode_t)0x4000000)
  92
  93 --- a/include/linux/mm_inline.h
  94 +++ b/include/linux/mm_inline.h
  95 @@ -339,6 +339,9 @@ static inline bool vma_has_recency(struc
  96         if (vma->vm_flags & (VM_SEQ_READ | VM_RAND_READ))
  97                 return false;
  98
  99 +       if (vma->vm_file && (vma->vm_file->f_mode & FMODE_NOREUSE))
 100 +               return false;
 101 +
 102         return true;
 103  }
 104
 105 --- a/mm/fadvise.c
 106 +++ b/mm/fadvise.c
 107 @@ -80,7 +80,7 @@ int generic_fadvise(struct file *file, l
 108         case POSIX_FADV_NORMAL:
 109                 file->f_ra.ra_pages = bdi->ra_pages;
 110                 spin_lock(&file->f_lock);
 111 -               file->f_mode &= ~FMODE_RANDOM;
 112 +               file->f_mode &= ~(FMODE_RANDOM | FMODE_NOREUSE);
 113                 spin_unlock(&file->f_lock);
 114                 break;
 115         case POSIX_FADV_RANDOM:
 116 @@ -107,6 +107,9 @@ int generic_fadvise(struct file *file, l
 117                 force_page_cache_readahead(mapping, file, start_index, nrpages);
 118                 break;
 119         case POSIX_FADV_NOREUSE:
 120 +               spin_lock(&file->f_lock);
 121 +               file->f_mode |= FMODE_NOREUSE;
 122 +               spin_unlock(&file->f_lock);
 123                 break;
 124         case POSIX_FADV_DONTNEED:
 125                 if (!inode_write_congested(mapping->host))