[RFC v2 04/11] fs/buffer: reduce stack usage on bh_read_iter()

[Luis Chamberlain <mcgrof@kernel.org>]

Now that we can read asynchronously buffer heads from a folio in
chunks,  we can chop up bh_read_iter() with a smaller array size.
Use an array of 8 to avoid stack growth warnings on systems with
huge base page sizes.

Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 fs/buffer.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/fs/buffer.c b/fs/buffer.c
index f8e6a5454dbb..b4994c48e6ee 100644
--- a/fs/buffer.c
+++ b/fs/buffer.c
@@ -2410,7 +2410,10 @@ static void bh_read_batch_async(struct folio *folio,
          (__tmp);					\
          (__tmp) = bh_next(__tmp, __head))
 
+#define MAX_BUF_CHUNK 8
+
 struct bh_iter {
+	int chunk_number;
 	sector_t iblock;
 	get_block_t *get_block;
 	bool any_get_block_error;
@@ -2419,7 +2422,7 @@ struct bh_iter {
 };
 
 /*
- * Reads up to MAX_BUF_PER_PAGE buffer heads at a time on a folio on the given
+ * Reads up to MAX_BUF_CHUNK buffer heads at a time on a folio on the given
  * block range iblock to lblock and helps update the number of buffer-heads
  * which were not uptodate or unmapped for which we issued an async read for
  * on iter->bh_folio_reads for the full folio. Returns the last buffer-head we
@@ -2431,16 +2434,18 @@ static struct buffer_head *bh_read_iter(struct folio *folio,
 					struct inode *inode,
 					struct bh_iter *iter, sector_t lblock)
 {
-	struct buffer_head *arr[MAX_BUF_PER_PAGE];
+	struct buffer_head *arr[MAX_BUF_CHUNK];
 	struct buffer_head *bh = pivot, *last;
 	int nr = 0, i = 0;
 	size_t blocksize = head->b_size;
+	int chunk_idx = MAX_BUF_CHUNK * iter->chunk_number;
 	bool no_reads = false;
 	bool fully_mapped = false;
 
 	/* collect buffers not uptodate and not mapped yet */
 	for_each_bh_pivot(bh, last, head) {
-		BUG_ON(nr >= MAX_BUF_PER_PAGE);
+		if (nr >= MAX_BUF_CHUNK)
+			break;
 
 		if (buffer_uptodate(bh))
 			continue;
@@ -2457,7 +2462,8 @@ static struct buffer_head *bh_read_iter(struct folio *folio,
 					iter->any_get_block_error = true;
 			}
 			if (!buffer_mapped(bh)) {
-				folio_zero_range(folio, i * blocksize,
+				folio_zero_range(folio,
+						(i + chunk_idx) * blocksize,
 						blocksize);
 				if (!err)
 					set_buffer_uptodate(bh);
@@ -2476,8 +2482,7 @@ static struct buffer_head *bh_read_iter(struct folio *folio,
 	}
 
 	iter->bh_folio_reads += nr;
-
-	WARN_ON_ONCE(!bh_is_last(last, head));
+	iter->chunk_number++;
 
 	if (bh_is_last(last, head)) {
 		if (!iter->bh_folio_reads)
@@ -2507,6 +2512,7 @@ int block_read_full_folio(struct folio *folio, get_block_t *get_block)
 	struct buffer_head *bh, *head;
 	struct bh_iter iter = {
 		.get_block = get_block,
+		.chunk_number = 0,
 		.unmapped = 0,
 		.any_get_block_error = false,
 		.bh_folio_reads = 0,
-- 
2.43.0