[RFC] mm/madvise: enable files from read-only block for MADV_PAGEOUT

Commit Message

hailong Dec. 26, 2024, 1:22 p.m. UTC

From: Hailong Liu <[e-mail]>

Apps may load files from a read-only block after startup and then
switch to the background. In this case, the system daemon reclaim page
cache using process_madvise(vmas, MADV_PAGEOUT). However, currently
only inode_owner_or_capable() or file_permission(vma->vm_file, MAY_WRITE) == 0
meet the conditions of can_do_file_pageout().

In fact, for read-only block devices, we can directly discard these
pages and free up memory.

The test results are as follows:
Before
                   Pss  Private  Private  SwapPss      Rss     Heap     Heap     Heap
                 Total    Dirty    Clean    Dirty    Total     Size    Alloc     Free
                ------   ------   ------   ------   ------   ------   ------   ------
     .so mmap      199        0        0      128    27616
    .jar mmap      748        0       88        0    39444
    .apk mmap     6818        0     6188        0     8076
    .dex mmap      102        0       80       44     1120
    .oat mmap      148        0        0        0    11836
    .art mmap      341        0        8      652    30748
   Other mmap       42        0        0        4     2228
      Unknown        5        0        0     1012     1528
        TOTAL    17984        0     6372     9412   138096    31023    13628    17394

After
                   Pss  Private  Private  SwapPss      Rss     Heap     Heap     Heap
                 Total    Dirty    Clean    Dirty    Total     Size    Alloc     Free
                ------   ------   ------   ------   ------   ------   ------   ------
     .so mmap      206        0        0      132    27332
    .jar mmap      625        0        0        0    39288
    .apk mmap      613        0        0        0     1668
    .dex mmap       22        0        0       44     1040
    .oat mmap      151        0        0        0    11836
    .art mmap      340        0        0      636    30756
   Other mmap       44        0        0        4     2248
      Unknown        6        0        0     1004     1532
        TOTAL    11801        0        8     9624   131336    28939    13485    15453

From above we can see the *.apk mmap* is reclaimed.

Signed-off-by: Hailong Liu <[e-mail]>
---
 mm/madvise.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

--
2.30.0

Patch

diff --git a/mm/madvise.c b/mm/madvise.c
index 8e5bf11af1b2..503ee5e03b7e 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -337,12 +337,13 @@  static inline bool can_do_file_pageout(struct vm_area_struct *vma)
 		return false;
 	/*
 	 * paging out pagecache only for non-anonymous mappings that correspond
-	 * to the files the calling process could (if tried) open for writing;
-	 * otherwise we'd be including shared non-exclusive mappings, which
-	 * opens a side channel.
+	 * to the files the calling process could (if tried) open for writing or
+	 * file from read-only super block; otherwise we'd be including
+	 * shared non-exclusive mappings, which opens a side channel.
 	 */
 	return inode_owner_or_capable(&nop_mnt_idmap,
 				      file_inode(vma->vm_file)) ||
+	       sb_rdonly(file_inode(vma->vm_file)->i_sb) ||
 	       file_permission(vma->vm_file, MAY_WRITE) == 0;
 }