[PATCH v3 0/2] erofs: support large folios for fscache mode

[PATCH v3 0/2] erofs: support large folios for fscache mode

[Jingbo Xu]

v3:
- patch 1: when large folios supported, one folio or folio range can be
mapped into several slices, with each slice mapped to different cookies,
and thus each slice needs its own netfs_cache_resources.

In the implementation of v2, each .read_folio() or .readahead() calling
corresponds to only one request, and thus only one netfs_cache_resources
(embedded in the request).  In this case, fscache_begin_read_operation()
may be called multiple times on this cres, while cres->ops->end_operation()
is called only once when the whole request completes.  This can cause
the leakage of the corresponding cachefiles_object->n_accesses refcount,
which will cause the user daemon hangs there forever waiting for
cache->object_count decreasing to 0 when the user daemon exits.

Worsely, as we mentioned previously, when large folios supported, one
folio or folio range can be mapped to multiple chunks on different
cookies, in which case each mapped chunk needs its own cres. In the
implementation of v2, each .read_folio() or .readahead() calling
corresponds to only one request, and thus only one cres.  This will make
the only cres used by the first chunk gets overridden by the following
chunk.

To fix this, we introduce listed requests, where each .read_folio() or
.readahead() calling can correspond to a list of requests, with each
request corresponds to one cres.

v2: https://lore.kernel.org/all/20221128025011.36352-1-jefflexu@linux.alibaba.com/
v1: https://lore.kernel.org/all/20221126005756.7662-1-jefflexu@linux.alibaba.com/

v2:
- patch 2: keep the enabling for iomap and fscache mode in separate
  patches; don't enable the feature for the meta data routine for now
  (Gao Xiang)

Patch 1 is the main part of supporting large folios for fscache mode. It
relies on a pending patch[1] adding .prepare_ondemand_read() interface
in Cachefiles.

Patch 2 just turns the switch on and enables the feature for fscache
mode. It relies on a previous patch[2] which enables this feature for
iomap mode.

[1] https://lore.kernel.org/all/20221124034212.81892-1-jefflexu@linux.alibaba.com/
[2] https://lore.kernel.org/all/20221110074023.8059-1-jefflexu@linux.alibaba.com/



Jingbo Xu (2):
  erofs: support large folios for fscache mode
  erofs: enable large folios for fscache mode

 fs/erofs/fscache.c | 166 ++++++++++++++++++++++++---------------------
 fs/erofs/inode.c   |   3 +-
 2 files changed, 89 insertions(+), 80 deletions(-)

-- 
2.19.1.6.gb485710b

[PATCH v3 1/2] erofs: support large folios for fscache mode

[Jingbo Xu]

When large folios supported, one folio can be split into several slices,
each of which may be mapped to META/UNMAPPED/MAPPED, and the folio can
be unlocked as a whole only when all slices have completed.

Thus always allocate erofs_fscache_request for each .read_folio() or
.readahead(). In this case, only when all slices of the folio or folio
range have completed, the request will be marked as completed and the
folio or folio range will be unlocked then.

In this case, one folio (from .read_folio()) or one folio range (from
.readahead()) can be mapped into several slices, with each slice mapped
to different cookies, and thus each slice needs its own
netfs_cache_resources.  Here we introduce listed requests to support
this, where each .read_folio() or .readahead() calling can correspond to
a list of requests, with each request reading from the corresponding
cookie.

Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
---
 fs/erofs/fscache.c | 166 ++++++++++++++++++++++++---------------------
 1 file changed, 88 insertions(+), 78 deletions(-)

diff --git a/fs/erofs/fscache.c b/fs/erofs/fscache.c
index 3e794891cd91..86d5cd5f909f 100644
--- a/fs/erofs/fscache.c
+++ b/fs/erofs/fscache.c
@@ -13,6 +13,7 @@ static struct vfsmount *erofs_pseudo_mnt;
 
 struct erofs_fscache_request {
 	struct netfs_cache_resources cache_resources;
+	struct erofs_fscache_request *parent;
 	struct address_space	*mapping;	/* The mapping being accessed */
 	loff_t			start;		/* Start position */
 	size_t			len;		/* Length of the request */
@@ -22,7 +23,7 @@ struct erofs_fscache_request {
 };
 
 static struct erofs_fscache_request *erofs_fscache_req_alloc(struct address_space *mapping,
-					     loff_t start, size_t len)
+		loff_t start, size_t len, struct erofs_fscache_request *parent)
 {
 	struct erofs_fscache_request *req;
 
@@ -34,6 +35,10 @@ static struct erofs_fscache_request *erofs_fscache_req_alloc(struct address_spac
 	req->start   = start;
 	req->len     = len;
 	refcount_set(&req->ref, 1);
+	if (parent) {
+		req->parent = parent;
+		refcount_inc(&parent->ref);
+	}
 
 	return req;
 }
@@ -56,17 +61,22 @@ static void erofs_fscache_req_complete(struct erofs_fscache_request *req)
 		folio_unlock(folio);
 	}
 	rcu_read_unlock();
-
-	if (req->cache_resources.ops)
-		req->cache_resources.ops->end_operation(&req->cache_resources);
-
-	kfree(req);
 }
 
 static void erofs_fscache_req_put(struct erofs_fscache_request *req)
 {
-	if (refcount_dec_and_test(&req->ref))
-		erofs_fscache_req_complete(req);
+	struct erofs_fscache_request *parent = req->parent;
+
+	if (refcount_dec_and_test(&req->ref)) {
+		if (!parent)
+			erofs_fscache_req_complete(req);
+		if (req->cache_resources.ops)
+			req->cache_resources.ops->end_operation(&req->cache_resources);
+		kfree(req);
+
+		if (parent)
+			erofs_fscache_req_put(parent);
+	}
 }
 
 static void erofs_fscache_subreq_complete(void *priv,
@@ -74,8 +84,12 @@ static void erofs_fscache_subreq_complete(void *priv,
 {
 	struct erofs_fscache_request *req = priv;
 
-	if (IS_ERR_VALUE(transferred_or_error))
-		req->error = transferred_or_error;
+	if (IS_ERR_VALUE(transferred_or_error)) {
+		if (req->parent)
+			req->parent->error = transferred_or_error;
+		else
+			req->error = transferred_or_error;
+	}
 	erofs_fscache_req_put(req);
 }
 
@@ -152,7 +166,7 @@ static int erofs_fscache_meta_read_folio(struct file *data, struct folio *folio)
 	}
 
 	req = erofs_fscache_req_alloc(folio_mapping(folio),
-				folio_pos(folio), folio_size(folio));
+			folio_pos(folio), folio_size(folio), NULL);
 	if (IS_ERR(req)) {
 		folio_unlock(folio);
 		return PTR_ERR(req);
@@ -167,32 +181,20 @@ static int erofs_fscache_meta_read_folio(struct file *data, struct folio *folio)
 	return ret;
 }
 
-/*
- * Read into page cache in the range described by (@pos, @len).
- *
- * On return, if the output @unlock is true, the caller is responsible for page
- * unlocking; otherwise the callee will take this responsibility through request
- * completion.
- *
- * The return value is the number of bytes successfully handled, or negative
- * error code on failure. The only exception is that, the length of the range
- * instead of the error code is returned on failure after request is allocated,
- * so that .readahead() could advance rac accordingly.
- */
-static int erofs_fscache_data_read(struct address_space *mapping,
-				   loff_t pos, size_t len, bool *unlock)
+static int erofs_fscache_data_read_slice(struct erofs_fscache_request *req)
 {
+	struct address_space *mapping = req->mapping;
 	struct inode *inode = mapping->host;
 	struct super_block *sb = inode->i_sb;
-	struct erofs_fscache_request *req;
+	loff_t pos = req->start + req->submitted;
+	struct erofs_fscache_request *new;
 	struct erofs_map_blocks map;
 	struct erofs_map_dev mdev;
 	struct iov_iter iter;
+	loff_t pstart;
 	size_t count;
 	int ret;
 
-	*unlock = true;
-
 	map.m_la = pos;
 	ret = erofs_map_blocks(inode, &map, EROFS_GET_BLOCKS_RAW);
 	if (ret)
@@ -201,36 +203,37 @@ static int erofs_fscache_data_read(struct address_space *mapping,
 	if (map.m_flags & EROFS_MAP_META) {
 		struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
 		erofs_blk_t blknr;
-		size_t offset, size;
+		size_t offset;
 		void *src;
 
 		/* For tail packing layout, the offset may be non-zero. */
 		offset = erofs_blkoff(map.m_pa);
 		blknr = erofs_blknr(map.m_pa);
-		size = map.m_llen;
+		count = map.m_llen;
 
 		src = erofs_read_metabuf(&buf, sb, blknr, EROFS_KMAP);
 		if (IS_ERR(src))
 			return PTR_ERR(src);
 
-		iov_iter_xarray(&iter, READ, &mapping->i_pages, pos, PAGE_SIZE);
-		if (copy_to_iter(src + offset, size, &iter) != size) {
+		iov_iter_xarray(&iter, READ, &mapping->i_pages, pos, count);
+		if (copy_to_iter(src + offset, count, &iter) != count) {
 			erofs_put_metabuf(&buf);
 			return -EFAULT;
 		}
-		iov_iter_zero(PAGE_SIZE - size, &iter);
 		erofs_put_metabuf(&buf);
-		return PAGE_SIZE;
+		req->submitted += count;
+		return 0;
 	}
 
+	count = req->len - req->submitted;
 	if (!(map.m_flags & EROFS_MAP_MAPPED)) {
-		count = len;
 		iov_iter_xarray(&iter, READ, &mapping->i_pages, pos, count);
 		iov_iter_zero(count, &iter);
-		return count;
+		req->submitted += count;
+		return 0;
 	}
 
-	count = min_t(size_t, map.m_llen - (pos - map.m_la), len);
+	count = min_t(size_t, map.m_llen - (pos - map.m_la), count);
 	DBG_BUGON(!count || count % PAGE_SIZE);
 
 	mdev = (struct erofs_map_dev) {
@@ -241,68 +244,75 @@ static int erofs_fscache_data_read(struct address_space *mapping,
 	if (ret)
 		return ret;
 
-	req = erofs_fscache_req_alloc(mapping, pos, count);
-	if (IS_ERR(req))
-		return PTR_ERR(req);
+	pstart = mdev.m_pa + (pos - map.m_la);
+	if (!req->submitted)
+		return erofs_fscache_read_folios_async(mdev.m_fscache->cookie,
+				req, pstart, count);
+
+	/* allocate a child request if current request ever been submitted */
+	new = erofs_fscache_req_alloc(mapping,
+			req->start + req->submitted, count, req);
+	if (IS_ERR(new))
+		return PTR_ERR(new);
 
-	*unlock = false;
 	ret = erofs_fscache_read_folios_async(mdev.m_fscache->cookie,
-			req, mdev.m_pa + (pos - map.m_la), count);
-	if (ret)
-		req->error = ret;
+					      new, pstart, count);
+	req->submitted += count;
+	erofs_fscache_req_put(new);
+	return ret;
+}
 
-	erofs_fscache_req_put(req);
-	return count;
+/*
+ * Read into page cache in the range described by (req->start, req->len).
+ */
+static int erofs_fscache_data_read(struct erofs_fscache_request *req)
+{
+	int ret;
+
+	do {
+		ret = erofs_fscache_data_read_slice(req);
+		if (ret)
+			req->error = ret;
+	} while (!ret && req->submitted < req->len);
+
+	return ret;
 }
 
 static int erofs_fscache_read_folio(struct file *file, struct folio *folio)
 {
-	bool unlock;
+	struct erofs_fscache_request *req;
 	int ret;
 
-	DBG_BUGON(folio_size(folio) != EROFS_BLKSIZ);
-
-	ret = erofs_fscache_data_read(folio_mapping(folio), folio_pos(folio),
-				      folio_size(folio), &unlock);
-	if (unlock) {
-		if (ret > 0)
-			folio_mark_uptodate(folio);
+	req = erofs_fscache_req_alloc(folio_mapping(folio),
+			folio_pos(folio), folio_size(folio), NULL);
+	if (IS_ERR(req)) {
 		folio_unlock(folio);
+		return PTR_ERR(req);
 	}
-	return ret < 0 ? ret : 0;
+
+	ret = erofs_fscache_data_read(req);
+	erofs_fscache_req_put(req);
+	return ret;
 }
 
 static void erofs_fscache_readahead(struct readahead_control *rac)
 {
-	struct folio *folio;
-	size_t len, done = 0;
-	loff_t start, pos;
-	bool unlock;
-	int ret, size;
+	struct erofs_fscache_request *req;
 
 	if (!readahead_count(rac))
 		return;
 
-	start = readahead_pos(rac);
-	len = readahead_length(rac);
+	req = erofs_fscache_req_alloc(rac->mapping,
+			readahead_pos(rac), readahead_length(rac), NULL);
+	if (IS_ERR(req))
+		return;
 
-	do {
-		pos = start + done;
-		ret = erofs_fscache_data_read(rac->mapping, pos,
-					      len - done, &unlock);
-		if (ret <= 0)
-			return;
+	/* The request completion will drop refs on the folios. */
+	while (readahead_folio(rac))
+		;
 
-		size = ret;
-		while (size) {
-			folio = readahead_folio(rac);
-			size -= folio_size(folio);
-			if (unlock) {
-				folio_mark_uptodate(folio);
-				folio_unlock(folio);
-			}
-		}
-	} while ((done += ret) < len);
+	erofs_fscache_data_read(req);
+	erofs_fscache_req_put(req);
 }
 
 static const struct address_space_operations erofs_fscache_meta_aops = {
-- 
2.19.1.6.gb485710b

[PATCH v3 2/2] erofs: enable large folios for fscache mode

[Jingbo Xu]

Enable large folios for fscache mode.  Enable this feature for
non-compressed format for now, until the compression part supports large
folios later.

One thing worth noting is that, the feature is not enabled for the meta
data routine since meta inodes don't need large folios for now, nor do
they support readahead yet.

Signed-off-by: Jingbo Xu <jefflexu@linux.alibaba.com>
Reviewed-by: Jia Zhu <zhujia.zj@bytedance.com>
Link: https://lore.kernel.org/r/20221128025011.36352-3-jefflexu@linux.alibaba.com
Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
---
 fs/erofs/inode.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/fs/erofs/inode.c b/fs/erofs/inode.c
index e457b8a59ee7..85932086d23f 100644
--- a/fs/erofs/inode.c
+++ b/fs/erofs/inode.c
@@ -295,8 +295,7 @@ static int erofs_fill_inode(struct inode *inode)
 		goto out_unlock;
 	}
 	inode->i_mapping->a_ops = &erofs_raw_access_aops;
-	if (!erofs_is_fscache_mode(inode->i_sb))
-		mapping_set_large_folios(inode->i_mapping);
+	mapping_set_large_folios(inode->i_mapping);
 #ifdef CONFIG_EROFS_FS_ONDEMAND
 	if (erofs_is_fscache_mode(inode->i_sb))
 		inode->i_mapping->a_ops = &erofs_fscache_access_aops;
-- 
2.19.1.6.gb485710b