|
Mel Gorman |
fcc0d3 |
From d76e2382feea66e9b5a47de31d333f1a44932d55 Mon Sep 17 00:00:00 2001
|
|
Mel Gorman |
fcc0d3 |
From: Hugh Dickins <hughd@google.com>
|
|
Mel Gorman |
fcc0d3 |
Date: Thu, 2 Sep 2021 14:54:21 -0700
|
|
Mel Gorman |
fcc0d3 |
Subject: [PATCH] huge tmpfs: fix split_huge_page() after FALLOC_FL_KEEP_SIZE
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
References: git fixes (mm/shmem)
|
|
Mel Gorman |
fcc0d3 |
Patch-mainline: v5.15-rc1
|
|
Mel Gorman |
fcc0d3 |
Git-commit: d144bf6205342a4b5fed5d204ae18849a4de741b
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
A successful shmem_fallocate() guarantees that the extent has been
|
|
Mel Gorman |
fcc0d3 |
reserved, even beyond i_size when the FALLOC_FL_KEEP_SIZE flag was used.
|
|
Mel Gorman |
fcc0d3 |
But that guarantee is broken by shmem_unused_huge_shrink()'s attempts to
|
|
Mel Gorman |
fcc0d3 |
split huge pages and free their excess beyond i_size; and by other uses of
|
|
Mel Gorman |
fcc0d3 |
split_huge_page() near i_size.
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
It's sad to add a shmem inode field just for this, but I did not find a
|
|
Mel Gorman |
fcc0d3 |
better way to keep the guarantee. A flag to say KEEP_SIZE has been used
|
|
Mel Gorman |
fcc0d3 |
would be cheaper, but I'm averse to unclearable flags. The fallocend
|
|
Mel Gorman |
fcc0d3 |
field is not perfect either (many disjoint ranges might be fallocated),
|
|
Mel Gorman |
fcc0d3 |
but good enough; and gains another use later on.
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
Link: https://lkml.kernel.org/r/ca9a146-3a59-6cd3-7f28-e9a044bb1052@google.com
|
|
Mel Gorman |
fcc0d3 |
Fixes: 779750d20b93 ("shmem: split huge pages beyond i_size under memory pressure")
|
|
Mel Gorman |
fcc0d3 |
Signed-off-by: Hugh Dickins <hughd@google.com>
|
|
Mel Gorman |
fcc0d3 |
Reviewed-by: Yang Shi <shy828301@gmail.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: Matthew Wilcox <willy@infradead.org>
|
|
Mel Gorman |
fcc0d3 |
Cc: Miaohe Lin <linmiaohe@huawei.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: Michal Hocko <mhocko@suse.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: Mike Kravetz <mike.kravetz@oracle.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: Rik van Riel <riel@surriel.com>
|
|
Mel Gorman |
fcc0d3 |
Cc: Shakeel Butt <shakeelb@google.com>
|
|
Mel Gorman |
fcc0d3 |
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
|
|
Mel Gorman |
fcc0d3 |
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Mel Gorman |
fcc0d3 |
Signed-off-by: Mel Gorman <mgorman@suse.de>
|
|
Mel Gorman |
fcc0d3 |
---
|
|
Mel Gorman |
fcc0d3 |
include/linux/shmem_fs.h | 13 +++++++++++++
|
|
Mel Gorman |
fcc0d3 |
mm/huge_memory.c | 6 ++++--
|
|
Mel Gorman |
fcc0d3 |
mm/shmem.c | 15 ++++++++++++++-
|
|
Mel Gorman |
fcc0d3 |
3 files changed, 31 insertions(+), 3 deletions(-)
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
|
|
Mel Gorman |
fcc0d3 |
index 0a8499fb9c3c..bfc5899d18e0 100644
|
|
Mel Gorman |
fcc0d3 |
--- a/include/linux/shmem_fs.h
|
|
Mel Gorman |
fcc0d3 |
+++ b/include/linux/shmem_fs.h
|
|
Mel Gorman |
fcc0d3 |
@@ -18,6 +18,7 @@ struct shmem_inode_info {
|
|
Mel Gorman |
fcc0d3 |
unsigned long flags;
|
|
Mel Gorman |
fcc0d3 |
unsigned long alloced; /* data pages alloced to file */
|
|
Mel Gorman |
fcc0d3 |
unsigned long swapped; /* subtotal assigned to swap */
|
|
Mel Gorman |
fcc0d3 |
+ pgoff_t fallocend; /* highest fallocate endindex */
|
|
Mel Gorman |
fcc0d3 |
struct list_head shrinklist; /* shrinkable hpage inodes */
|
|
Mel Gorman |
fcc0d3 |
struct list_head swaplist; /* chain of maybes on swap */
|
|
Mel Gorman |
fcc0d3 |
struct shared_policy policy; /* NUMA memory alloc policy */
|
|
Mel Gorman |
fcc0d3 |
@@ -119,6 +120,18 @@ static inline bool shmem_file(struct file *file)
|
|
Mel Gorman |
fcc0d3 |
return shmem_mapping(file->f_mapping);
|
|
Mel Gorman |
fcc0d3 |
}
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
+/*
|
|
Mel Gorman |
fcc0d3 |
+ * If fallocate(FALLOC_FL_KEEP_SIZE) has been used, there may be pages
|
|
Mel Gorman |
fcc0d3 |
+ * beyond i_size's notion of EOF, which fallocate has committed to reserving:
|
|
Mel Gorman |
fcc0d3 |
+ * which split_huge_page() must therefore not delete. This use of a single
|
|
Mel Gorman |
fcc0d3 |
+ * "fallocend" per inode errs on the side of not deleting a reservation when
|
|
Mel Gorman |
fcc0d3 |
+ * in doubt: there are plenty of cases when it preserves unreserved pages.
|
|
Mel Gorman |
fcc0d3 |
+ */
|
|
Mel Gorman |
fcc0d3 |
+static inline pgoff_t shmem_fallocend(struct inode *inode, pgoff_t eof)
|
|
Mel Gorman |
fcc0d3 |
+{
|
|
Mel Gorman |
fcc0d3 |
+ return max(eof, SHMEM_I(inode)->fallocend);
|
|
Mel Gorman |
fcc0d3 |
+}
|
|
Mel Gorman |
fcc0d3 |
+
|
|
Mel Gorman |
fcc0d3 |
extern bool shmem_charge(struct inode *inode, long pages);
|
|
Mel Gorman |
fcc0d3 |
extern void shmem_uncharge(struct inode *inode, long pages);
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
|
|
Mel Gorman |
fcc0d3 |
index 9f21e44c9030..5e9ef0fc261e 100644
|
|
Mel Gorman |
fcc0d3 |
--- a/mm/huge_memory.c
|
|
Mel Gorman |
fcc0d3 |
+++ b/mm/huge_memory.c
|
|
Mel Gorman |
fcc0d3 |
@@ -2428,11 +2428,11 @@ static void __split_huge_page(struct page *page, struct list_head *list,
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
for (i = nr - 1; i >= 1; i--) {
|
|
Mel Gorman |
fcc0d3 |
__split_huge_page_tail(head, i, lruvec, list);
|
|
Mel Gorman |
fcc0d3 |
- /* Some pages can be beyond i_size: drop them from page cache */
|
|
Mel Gorman |
fcc0d3 |
+ /* Some pages can be beyond EOF: drop them from page cache */
|
|
Mel Gorman |
fcc0d3 |
if (head[i].index >= end) {
|
|
Mel Gorman |
fcc0d3 |
ClearPageDirty(head + i);
|
|
Mel Gorman |
fcc0d3 |
__delete_from_page_cache(head + i, NULL);
|
|
Mel Gorman |
fcc0d3 |
- if (IS_ENABLED(CONFIG_SHMEM) && PageSwapBacked(head))
|
|
Mel Gorman |
fcc0d3 |
+ if (shmem_mapping(head->mapping))
|
|
Mel Gorman |
fcc0d3 |
shmem_uncharge(head->mapping->host, 1);
|
|
Mel Gorman |
fcc0d3 |
put_page(head + i);
|
|
Mel Gorman |
fcc0d3 |
} else if (!PageAnon(page)) {
|
|
Mel Gorman |
fcc0d3 |
@@ -2660,6 +2660,8 @@ int split_huge_page_to_list(struct page *page, struct list_head *list)
|
|
Mel Gorman |
fcc0d3 |
* head page lock is good enough to serialize the trimming.
|
|
Mel Gorman |
fcc0d3 |
*/
|
|
Mel Gorman |
fcc0d3 |
end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE);
|
|
Mel Gorman |
fcc0d3 |
+ if (shmem_mapping(mapping))
|
|
Mel Gorman |
fcc0d3 |
+ end = shmem_fallocend(mapping->host, end);
|
|
Mel Gorman |
fcc0d3 |
}
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
/*
|
|
Mel Gorman |
fcc0d3 |
diff --git a/mm/shmem.c b/mm/shmem.c
|
|
Mel Gorman |
fcc0d3 |
index 9ef579f6cab3..e391325dfc21 100644
|
|
Mel Gorman |
fcc0d3 |
--- a/mm/shmem.c
|
|
Mel Gorman |
fcc0d3 |
+++ b/mm/shmem.c
|
|
Mel Gorman |
fcc0d3 |
@@ -902,6 +902,9 @@ static void shmem_undo_range(struct inode *inode, loff_t lstart, loff_t lend,
|
|
Mel Gorman |
fcc0d3 |
if (lend == -1)
|
|
Mel Gorman |
fcc0d3 |
end = -1; /* unsigned, so actually very big */
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
+ if (info->fallocend > start && info->fallocend <= end && !unfalloc)
|
|
Mel Gorman |
fcc0d3 |
+ info->fallocend = start;
|
|
Mel Gorman |
fcc0d3 |
+
|
|
Mel Gorman |
fcc0d3 |
pagevec_init(&pvec);
|
|
Mel Gorman |
fcc0d3 |
index = start;
|
|
Mel Gorman |
fcc0d3 |
while (index < end && find_lock_entries(mapping, index, end - 1,
|
|
Mel Gorman |
fcc0d3 |
@@ -2650,7 +2653,7 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
|
|
Mel Gorman |
fcc0d3 |
struct shmem_sb_info *sbinfo = SHMEM_SB(inode->i_sb);
|
|
Mel Gorman |
fcc0d3 |
struct shmem_inode_info *info = SHMEM_I(inode);
|
|
Mel Gorman |
fcc0d3 |
struct shmem_falloc shmem_falloc;
|
|
Mel Gorman |
fcc0d3 |
- pgoff_t start, index, end;
|
|
Mel Gorman |
fcc0d3 |
+ pgoff_t start, index, end, undo_fallocend;
|
|
Mel Gorman |
fcc0d3 |
int error;
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
|
|
Mel Gorman |
fcc0d3 |
@@ -2719,6 +2722,15 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
|
|
Mel Gorman |
fcc0d3 |
inode->i_private = &shmem_falloc;
|
|
Mel Gorman |
fcc0d3 |
spin_unlock(&inode->i_lock);
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
+ /*
|
|
Mel Gorman |
fcc0d3 |
+ * info->fallocend is only relevant when huge pages might be
|
|
Mel Gorman |
fcc0d3 |
+ * involved: to prevent split_huge_page() freeing fallocated
|
|
Mel Gorman |
fcc0d3 |
+ * pages when FALLOC_FL_KEEP_SIZE committed beyond i_size.
|
|
Mel Gorman |
fcc0d3 |
+ */
|
|
Mel Gorman |
fcc0d3 |
+ undo_fallocend = info->fallocend;
|
|
Mel Gorman |
fcc0d3 |
+ if (info->fallocend < end)
|
|
Mel Gorman |
fcc0d3 |
+ info->fallocend = end;
|
|
Mel Gorman |
fcc0d3 |
+
|
|
Mel Gorman |
fcc0d3 |
for (index = start; index < end; ) {
|
|
Mel Gorman |
fcc0d3 |
struct page *page;
|
|
Mel Gorman |
fcc0d3 |
|
|
Mel Gorman |
fcc0d3 |
@@ -2733,6 +2745,7 @@ static long shmem_fallocate(struct file *file, int mode, loff_t offset,
|
|
Mel Gorman |
fcc0d3 |
else
|
|
Mel Gorman |
fcc0d3 |
error = shmem_getpage(inode, index, &page, SGP_FALLOC);
|
|
Mel Gorman |
fcc0d3 |
if (error) {
|
|
Mel Gorman |
fcc0d3 |
+ info->fallocend = undo_fallocend;
|
|
Mel Gorman |
fcc0d3 |
/* Remove the !PageUptodate pages we added */
|
|
Mel Gorman |
fcc0d3 |
if (index > start) {
|
|
Mel Gorman |
fcc0d3 |
shmem_undo_range(inode,
|