mm: shmem: fallback to page size splice if large folio has poisoned subpages

Commit Message

Baolin Wang Oct. 25, 2024, 3:26 a.m. UTC

The tmpfs has already supported the PMD-sized large folios, and splice()
can not read any subpages if the large folio has a poisoned subpage,
which is not good as we discussed in previous mail[1].

Thus adding a fallback to the PAGE_SIZE splice() still allows reading
normal subpages if the large folio has hwpoisoned subpages.

[1] https://lore.kernel.org/all/[e-mail]/
Signed-off-by: Baolin Wang <[e-mail]>
---
 mm/shmem.c | 39 +++++++++++++++++++++++++++++++--------
 1 file changed, 31 insertions(+), 8 deletions(-)

Comments

Matthew Wilcox (Oracle) Oct. 25, 2024, 3:31 a.m. UTC | #1

On Fri, Oct 25, 2024 at 11:26:39AM +0800, Baolin Wang wrote:
> The tmpfs has already supported the PMD-sized large folios, and splice()
> can not read any subpages if the large folio has a poisoned subpage,
> which is not good as we discussed in previous mail[1].

folios do not have subpages.  folios have pages.  do not use the term
"subpage" anywhere.  ever.

Baolin Wang Oct. 25, 2024, 3:41 a.m. UTC | #2

On 2024/10/25 11:31, Matthew Wilcox wrote:
> On Fri, Oct 25, 2024 at 11:26:39AM +0800, Baolin Wang wrote:
>> The tmpfs has already supported the PMD-sized large folios, and splice()
>> can not read any subpages if the large folio has a poisoned subpage,
>> which is not good as we discussed in previous mail[1].
> 
> folios do not have subpages.  folios have pages.  do not use the term
> "subpage" anywhere.  ever.

OK. This is my previous habit of naming it. Will change 'subpages' to 
'pages' for folios.

Patch

diff --git a/mm/shmem.c b/mm/shmem.c
index 1bef6e32a1fa..79010e636056 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -3291,11 +3291,16 @@  static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
 	len = min_t(size_t, len, npages * PAGE_SIZE);
 
 	do {
+		bool fallback_page_splice = false;
+		struct page *page = NULL;
+		pgoff_t index;
+		size_t size;
+
 		if (*ppos >= i_size_read(inode))
 			break;
 
-		error = shmem_get_folio(inode, *ppos / PAGE_SIZE, 0, &folio,
-					SGP_READ);
+		index = *ppos >> PAGE_SHIFT;
+		error = shmem_get_folio(inode, index, 0, &folio, SGP_READ);
 		if (error) {
 			if (error == -EINVAL)
 				error = 0;
@@ -3304,12 +3309,15 @@  static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
 		if (folio) {
 			folio_unlock(folio);
 
-			if (folio_test_hwpoison(folio) ||
-			    (folio_test_large(folio) &&
-			     folio_test_has_hwpoisoned(folio))) {
+			page = folio_file_page(folio, index);
+			if (PageHWPoison(page)) {
 				error = -EIO;
 				break;
 			}
+
+			if (folio_test_large(folio) &&
+			    folio_test_has_hwpoisoned(folio))
+				fallback_page_splice = true;
 		}
 
 		/*
@@ -3323,7 +3331,18 @@  static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
 		isize = i_size_read(inode);
 		if (unlikely(*ppos >= isize))
 			break;
-		part = min_t(loff_t, isize - *ppos, len);
+		/*
+		 * Fallback to PAGE_SIZE splice if the large folio has hwpoisoned
+		 * subpages.
+		 */
+		if (likely(!fallback_page_splice)) {
+			size = len;
+		} else {
+			size_t offset = *ppos & ~PAGE_MASK;
+
+			size = min_t(loff_t, PAGE_SIZE - offset, len);
+		}
+		part = min_t(loff_t, isize - *ppos, size);
 
 		if (folio) {
 			/*
@@ -3331,8 +3350,12 @@  static ssize_t shmem_file_splice_read(struct file *in, loff_t *ppos,
 			 * virtual addresses, take care about potential aliasing
 			 * before reading the page on the kernel side.
 			 */
-			if (mapping_writably_mapped(mapping))
-				flush_dcache_folio(folio);
+			if (mapping_writably_mapped(mapping)) {
+				if (likely(!fallback_page_splice))
+					flush_dcache_folio(folio);
+				else
+					flush_dcache_page(page);
+			}
 			folio_mark_accessed(folio);
 			/*
 			 * Ok, we have the page, and it's up-to-date, so we can