Merge tag 'erofs-for-5.19-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs

Pull erofs (and fscache) updates from Gao Xiang:
 "After working on it on the mailing list for more than half a year, we
  finally form 'erofs over fscache' feature into shape. Hopefully it
  could bring more possibility to the communities.

  The story mainly started from a new project what we called "RAFS v6" [1]
  for Nydus image service almost a year ago, which enhances EROFS to be
  a new form of one bootstrap (which includes metadata representing the
  whole fs tree) + several data-deduplicated content addressable blobs
  (actually treated as multiple devices). Each blob can represent one
  container image layer but not quite exactly since all new data can be
  fully existed in the previous blobs so no need to introduce another
  new blob.

  It is actually not a new idea (at least on my side it's much like a
  simpilied casync [2] for now) and has many benefits over per-file
  blobs or some other exist ways since typically each RAFS v6 image only
  has dozens of device blobs instead of thousands of per-file blobs.
  It's easy to be signed with user keys as a golden image, transfered
  untouchedly with minimal overhead over the network, kept in some type
  of storage conveniently, and run with (optional) runtime verification
  but without involving too many irrelevant features crossing the system
  beyond EROFS itself. At least it's our final goal and we're keeping
  working on it. There was also a good summary of this approach from the
  casync author [3].

  Regardless further optimizations, this work is almost done in the
  previous Linux release cycles. In this round, we'd like to introduce
  on-demand load for EROFS with the fscache/cachefiles infrastructure,
  considering the following advantages:

   - Introduce new file-based backend to EROFS. Although each image only
     contains dozens of blobs but in densely-deployed runC host for
     example, there could still be massive blobs on a machine, which is
     messy if each blob is treated as a device. In contrast, fscache and
     cachefiles are really great interfaces for us to make them work.

   - Introduce on-demand load to fscache and EROFS. Previously, fscache
     is mainly used to caching network-likewise filesystems, now it can
     support on-demand downloading for local fses too with the exact
     localfs on-disk format. It has many advantages which we're been
     described in the latest patchset cover letter [4]. In addition to
     that, most importantly, the cached data is still stored in the
     original local fs on-disk format so that it's still the one signed
     with private keys but only could be partially available. Users can
     fully trust it during running. Later, users can also back up
     cachefiles easily to another machine.

   - More reliable on-demand approach in principle. After data is all
     available locally, user daemon can be no longer online in some use
     cases, which helps daemon crash recovery (filesystems can still in
     service) and hot-upgrade (user daemon can be upgraded more
     frequently due to new features or protocols introduced.)

   - Other format can also be converted to EROFS filesystem format over
     the internet on the fly with the new on-demand load feature and
     mounted. That is entirely possible with on-demand load feature as
     long as such archive format metadata can be fetched in advance like
     stargz.

  In addition, although currently our target user is Nydus image service [5],
  but laterly, it can be used for other use cases like on-demand system
  booting, etc. As for the fscache on-demand load feature itself,
  strictly it can be used for other local fses too. Laterly we could
  promote most code to the iomap infrastructure and also enhance it in
  the read-write way if other local fses are interested.

  Thanks David Howells for taking so much time and patience on this
  these months, many thanks with great respect here again! Thanks Jeffle
  for working on this feature and Xin Yin from Bytedance for
  asynchronous I/O implementation as well as Zichen Tian, Jia Zhu, and
  Yan Song for testing, much appeciated. We're also exploring more
  possibly over fscache cache management over FSDAX for secure
  containers and working on more improvements and useful features for
  fscache, cachefiles, and on-demand load.

  In addition to "erofs over fscache", NFS export and idmapped mount are
  also completed in this cycle for container use cases as well.

  Summary:

   - Add erofs on-demand load support over fscache

   - Support NFS export for erofs

   - Support idmapped mounts for erofs

   - Don't prompt for risk any more when using big pcluster

   - Fix buffer copy overflow of ztailpacking feature

   - Several minor cleanups"

[1] https://lore.kernel.org/r/20210730194625.93856-1-hsiangkao@linux.alibaba.com
[2] https://github.com/systemd/casync
[3] http://0pointer.net/blog/casync-a-tool-for-distributing-file-system-images.html
[4] https://lore.kernel.org/r/20220509074028.74954-1-jefflexu@linux.alibaba.com
[5] https://github.com/dragonflyoss/image-service

* tag 'erofs-for-5.19-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/xiang/erofs: (29 commits)
  erofs: scan devices from device table
  erofs: change to use asynchronous io for fscache readpage/readahead
  erofs: add 'fsid' mount option
  erofs: implement fscache-based data readahead
  erofs: implement fscache-based data read for inline layout
  erofs: implement fscache-based data read for non-inline layout
  erofs: implement fscache-based metadata read
  erofs: register fscache context for extra data blobs
  erofs: register fscache context for primary data blob
  erofs: add erofs_fscache_read_folios() helper
  erofs: add anonymous inode caching metadata for data blobs
  erofs: add fscache context helper functions
  erofs: register fscache volume
  erofs: add fscache mode check helper
  erofs: make erofs_map_blocks() generally available
  cachefiles: document on-demand read mode
  cachefiles: add tracepoints for on-demand read mode
  cachefiles: enable on-demand read mode
  cachefiles: implement on-demand read
  cachefiles: notify the user daemon when withdrawing cookie
  ...

24 files changed, 1997 insertions, 164 deletions

diff --git a/Documentation/filesystems/caching/cachefiles.rst b/Documentation/filesystems/caching/cachefiles.rst
index 8bf396b76359b0..fc7abf712315aa 100644
--- a/Documentation/filesystems/caching/cachefiles.rst
+++ b/Documentation/filesystems/caching/cachefiles.rst
@@ -28,6 +28,7 @@ Cache on Already Mounted Filesystem
 
  (*) Debugging.
 
+ (*) On-demand Read.
 
 
 Overview
@@ -482,3 +483,180 @@ the control file.  For example::
 	echo $((1|4|8)) >/sys/module/cachefiles/parameters/debug
 
 will turn on all function entry debugging.
+
+
+On-demand Read
+==============
+
+When working in its original mode, CacheFiles serves as a local cache for a
+remote networking fs - while in on-demand read mode, CacheFiles can boost the
+scenario where on-demand read semantics are needed, e.g. container image
+distribution.
+
+The essential difference between these two modes is seen when a cache miss
+occurs: In the original mode, the netfs will fetch the data from the remote
+server and then write it to the cache file; in on-demand read mode, fetching
+the data and writing it into the cache is delegated to a user daemon.
+
+``CONFIG_CACHEFILES_ONDEMAND`` should be enabled to support on-demand read mode.
+
+
+Protocol Communication
+----------------------
+
+The on-demand read mode uses a simple protocol for communication between kernel
+and user daemon. The protocol can be modeled as::
+
+	kernel --[request]--> user daemon --[reply]--> kernel
+
+CacheFiles will send requests to the user daemon when needed.  The user daemon
+should poll the devnode ('/dev/cachefiles') to check if there's a pending
+request to be processed.  A POLLIN event will be returned when there's a pending
+request.
+
+The user daemon then reads the devnode to fetch a request to process.  It should
+be noted that each read only gets one request. When it has finished processing
+the request, the user daemon should write the reply to the devnode.
+
+Each request starts with a message header of the form::
+
+	struct cachefiles_msg {
+		__u32 msg_id;
+		__u32 opcode;
+		__u32 len;
+		__u32 object_id;
+		__u8  data[];
+	};
+
+where:
+
+	* ``msg_id`` is a unique ID identifying this request among all pending
+	  requests.
+
+	* ``opcode`` indicates the type of this request.
+
+	* ``object_id`` is a unique ID identifying the cache file operated on.
+
+	* ``data`` indicates the payload of this request.
+
+	* ``len`` indicates the whole length of this request, including the
+	  header and following type-specific payload.
+
+
+Turning on On-demand Mode
+-------------------------
+
+An optional parameter becomes available to the "bind" command::
+
+	bind [ondemand]
+
+When the "bind" command is given no argument, it defaults to the original mode.
+When it is given the "ondemand" argument, i.e. "bind ondemand", on-demand read
+mode will be enabled.
+
+
+The OPEN Request
+----------------
+
+When the netfs opens a cache file for the first time, a request with the
+CACHEFILES_OP_OPEN opcode, a.k.a an OPEN request will be sent to the user
+daemon.  The payload format is of the form::
+
+	struct cachefiles_open {
+		__u32 volume_key_size;
+		__u32 cookie_key_size;
+		__u32 fd;
+		__u32 flags;
+		__u8  data[];
+	};
+
+where:
+
+	* ``data`` contains the volume_key followed directly by the cookie_key.
+	  The volume key is a NUL-terminated string; the cookie key is binary
+	  data.
+
+	* ``volume_key_size`` indicates the size of the volume key in bytes.
+
+	* ``cookie_key_size`` indicates the size of the cookie key in bytes.
+
+	* ``fd`` indicates an anonymous fd referring to the cache file, through
+	  which the user daemon can perform write/llseek file operations on the
+	  cache file.
+
+
+The user daemon can use the given (volume_key, cookie_key) pair to distinguish
+the requested cache file.  With the given anonymous fd, the user daemon can
+fetch the data and write it to the cache file in the background, even when
+kernel has not triggered a cache miss yet.
+
+Be noted that each cache file has a unique object_id, while it may have multiple
+anonymous fds.  The user daemon may duplicate anonymous fds from the initial
+anonymous fd indicated by the @fd field through dup().  Thus each object_id can
+be mapped to multiple anonymous fds, while the usr daemon itself needs to
+maintain the mapping.
+
+When implementing a user daemon, please be careful of RLIMIT_NOFILE,
+``/proc/sys/fs/nr_open`` and ``/proc/sys/fs/file-max``.  Typically these needn't
+be huge since they're related to the number of open device blobs rather than
+open files of each individual filesystem.
+
+The user daemon should reply the OPEN request by issuing a "copen" (complete
+open) command on the devnode::
+
+	copen <msg_id>,<cache_size>
+
+where:
+
+	* ``msg_id`` must match the msg_id field of the OPEN request.
+
+	* When >= 0, ``cache_size`` indicates the size of the cache file;
+	  when < 0, ``cache_size`` indicates any error code encountered by the
+	  user daemon.
+
+
+The CLOSE Request
+-----------------
+
+When a cookie withdrawn, a CLOSE request (opcode CACHEFILES_OP_CLOSE) will be
+sent to the user daemon.  This tells the user daemon to close all anonymous fds
+associated with the given object_id.  The CLOSE request has no extra payload,
+and shouldn't be replied.
+
+
+The READ Request
+----------------
+
+When a cache miss is encountered in on-demand read mode, CacheFiles will send a
+READ request (opcode CACHEFILES_OP_READ) to the user daemon. This tells the user
+daemon to fetch the contents of the requested file range.  The payload is of the
+form::
+
+	struct cachefiles_read {
+		__u64 off;
+		__u64 len;
+	};
+
+where:
+
+	* ``off`` indicates the starting offset of the requested file range.
+
+	* ``len`` indicates the length of the requested file range.
+
+
+When it receives a READ request, the user daemon should fetch the requested data
+and write it to the cache file identified by object_id.
+
+When it has finished processing the READ request, the user daemon should reply
+by using the CACHEFILES_IOC_READ_COMPLETE ioctl on one of the anonymous fds
+associated with the object_id given in the READ request.  The ioctl is of the
+form::
+
+	ioctl(fd, CACHEFILES_IOC_READ_COMPLETE, msg_id);
+
+where:
+
+	* ``fd`` is one of the anonymous fds associated with the object_id
+	  given.
+
+	* ``msg_id`` must match the msg_id field of the READ request.
diff --git a/fs/cachefiles/Kconfig b/fs/cachefiles/Kconfig
index 719faeeda16881..8df715640a48f3 100644
--- a/fs/cachefiles/Kconfig
+++ b/fs/cachefiles/Kconfig
@@ -26,3 +26,15 @@ config CACHEFILES_ERROR_INJECTION
 	help
 	  This permits error injection to be enabled in cachefiles whilst a
 	  cache is in service.
+
+config CACHEFILES_ONDEMAND
+	bool "Support for on-demand read"
+	depends on CACHEFILES
+	default n
+	help
+	  This permits userspace to enable the cachefiles on-demand read mode.
+	  In this mode, when a cache miss occurs, responsibility for fetching
+	  the data lies with the cachefiles backend instead of with the netfs
+	  and is delegated to userspace.
+
+	  If unsure, say N.
diff --git a/fs/cachefiles/Makefile b/fs/cachefiles/Makefile
index 16d811f1a2faed..c37a7a9af10b60 100644
--- a/fs/cachefiles/Makefile
+++ b/fs/cachefiles/Makefile
@@ -16,5 +16,6 @@ cachefiles-y := \
 	xattr.o
 
 cachefiles-$(CONFIG_CACHEFILES_ERROR_INJECTION) += error_inject.o
+cachefiles-$(CONFIG_CACHEFILES_ONDEMAND) += ondemand.o
 
 obj-$(CONFIG_CACHEFILES) := cachefiles.o
diff --git a/fs/cachefiles/daemon.c b/fs/cachefiles/daemon.c
index 7ac04ee2c0a0a5..aa4efcabb5e37b 100644
--- a/fs/cachefiles/daemon.c
+++ b/fs/cachefiles/daemon.c
@@ -75,6 +75,9 @@ static const struct cachefiles_daemon_cmd cachefiles_daemon_cmds[] = {
 	{ "inuse",	cachefiles_daemon_inuse		},
 	{ "secctx",	cachefiles_daemon_secctx	},
 	{ "tag",	cachefiles_daemon_tag		},
+#ifdef CONFIG_CACHEFILES_ONDEMAND
+	{ "copen",	cachefiles_ondemand_copen	},
+#endif
 	{ "",		NULL				}
 };
 
@@ -108,6 +111,9 @@ static int cachefiles_daemon_open(struct inode *inode, struct file *file)
 	INIT_LIST_HEAD(&cache->volumes);
 	INIT_LIST_HEAD(&cache->object_list);
 	spin_lock_init(&cache->object_list_lock);
+	refcount_set(&cache->unbind_pincount, 1);
+	xa_init_flags(&cache->reqs, XA_FLAGS_ALLOC);
+	xa_init_flags(&cache->ondemand_ids, XA_FLAGS_ALLOC1);
 
 	/* set default caching limits
 	 * - limit at 1% free space and/or free files
@@ -126,6 +132,53 @@ static int cachefiles_daemon_open(struct inode *inode, struct file *file)
 	return 0;
 }
 
+static void cachefiles_flush_reqs(struct cachefiles_cache *cache)
+{
+	struct xarray *xa = &cache->reqs;
+	struct cachefiles_req *req;
+	unsigned long index;
+
+	/*
+	 * Make sure the following two operations won't be reordered.
+	 *   1) set CACHEFILES_DEAD bit
+	 *   2) flush requests in the xarray
+	 * Otherwise the request may be enqueued after xarray has been
+	 * flushed, leaving the orphan request never being completed.
+	 *
+	 * CPU 1			CPU 2
+	 * =====			=====
+	 * flush requests in the xarray
+	 *				test CACHEFILES_DEAD bit
+	 *				enqueue the request
+	 * set CACHEFILES_DEAD bit
+	 */
+	smp_mb();
+
+	xa_lock(xa);
+	xa_for_each(xa, index, req) {
+		req->error = -EIO;
+		complete(&req->done);
+	}
+	xa_unlock(xa);
+
+	xa_destroy(&cache->reqs);
+	xa_destroy(&cache->ondemand_ids);
+}
+
+void cachefiles_put_unbind_pincount(struct cachefiles_cache *cache)
+{
+	if (refcount_dec_and_test(&cache->unbind_pincount)) {
+		cachefiles_daemon_unbind(cache);
+		cachefiles_open = 0;
+		kfree(cache);
+	}
+}
+
+void cachefiles_get_unbind_pincount(struct cachefiles_cache *cache)
+{
+	refcount_inc(&cache->unbind_pincount);
+}
+
 /*
  * Release a cache.
  */
@@ -139,36 +192,27 @@ static int cachefiles_daemon_release(struct inode *inode, struct file *file)
 
 	set_bit(CACHEFILES_DEAD, &cache->flags);
 
-	cachefiles_daemon_unbind(cache);
+	if (cachefiles_in_ondemand_mode(cache))
+		cachefiles_flush_reqs(cache);
 
 	/* clean up the control file interface */
 	cache->cachefilesd = NULL;
 	file->private_data = NULL;
-	cachefiles_open = 0;
 
-	kfree(cache);
+	cachefiles_put_unbind_pincount(cache);
 
 	_leave("");
 	return 0;
 }
 
-/*
- * Read the cache state.
- */
-static ssize_t cachefiles_daemon_read(struct file *file, char __user *_buffer,
-				      size_t buflen, loff_t *pos)
+static ssize_t cachefiles_do_daemon_read(struct cachefiles_cache *cache,
+					 char __user *_buffer, size_t buflen)
 {
-	struct cachefiles_cache *cache = file->private_data;
 	unsigned long long b_released;
 	unsigned f_released;
 	char buffer[256];
 	int n;
 
-	//_enter(",,%zu,", buflen);
-
-	if (!test_bit(CACHEFILES_READY, &cache->flags))
-		return 0;
-
 	/* check how much space the cache has */
 	cachefiles_has_space(cache, 0, 0, cachefiles_has_space_check);
 
@@ -207,6 +251,25 @@ static ssize_t cachefiles_daemon_read(struct file *file, char __user *_buffer,
 }
 
 /*
+ * Read the cache state.
+ */
+static ssize_t cachefiles_daemon_read(struct file *file, char __user *_buffer,
+				      size_t buflen, loff_t *pos)
+{
+	struct cachefiles_cache *cache = file->private_data;
+
+	//_enter(",,%zu,", buflen);
+
+	if (!test_bit(CACHEFILES_READY, &cache->flags))
+		return 0;
+
+	if (cachefiles_in_ondemand_mode(cache))
+		return cachefiles_ondemand_daemon_read(cache, _buffer, buflen);
+	else
+		return cachefiles_do_daemon_read(cache, _buffer, buflen);
+}
+
+/*
  * Take a command from cachefilesd, parse it and act on it.
  */
 static ssize_t cachefiles_daemon_write(struct file *file,
@@ -297,8 +360,13 @@ static __poll_t cachefiles_daemon_poll(struct file *file,
 	poll_wait(file, &cache->daemon_pollwq, poll);
 	mask = 0;
 
-	if (test_bit(CACHEFILES_STATE_CHANGED, &cache->flags))
-		mask |= EPOLLIN;
+	if (cachefiles_in_ondemand_mode(cache)) {
+		if (!xa_empty(&cache->reqs))
+			mask |= EPOLLIN;
+	} else {
+		if (test_bit(CACHEFILES_STATE_CHANGED, &cache->flags))
+			mask |= EPOLLIN;
+	}
 
 	if (test_bit(CACHEFILES_CULLING, &cache->flags))
 		mask |= EPOLLOUT;
@@ -687,11 +755,6 @@ static int cachefiles_daemon_bind(struct cachefiles_cache *cache, char *args)
 	    cache->brun_percent  >= 100)
 		return -ERANGE;
 
-	if (*args) {
-		pr_err("'bind' command doesn't take an argument\n");
-		return -EINVAL;
-	}
-
 	if (!cache->rootdirname) {
 		pr_err("No cache directory specified\n");
 		return -EINVAL;
@@ -703,6 +766,18 @@ static int cachefiles_daemon_bind(struct cachefiles_cache *cache, char *args)
 		return -EBUSY;
 	}
 
+	if (IS_ENABLED(CONFIG_CACHEFILES_ONDEMAND)) {
+		if (!strcmp(args, "ondemand")) {
+			set_bit(CACHEFILES_ONDEMAND_MODE, &cache->flags);
+		} else if (*args) {
+			pr_err("Invalid argument to the 'bind' command\n");
+			return -EINVAL;
+		}
+	} else if (*args) {
+		pr_err("'bind' command doesn't take an argument\n");
+		return -EINVAL;
+	}
+
 	/* Make sure we have copies of the tag string */
 	if (!cache->tag) {
 		/*
diff --git a/fs/cachefiles/interface.c b/fs/cachefiles/interface.c
index ae93cee9d25ded..a69073a1d3f069 100644
--- a/fs/cachefiles/interface.c
+++ b/fs/cachefiles/interface.c
@@ -362,6 +362,8 @@ static void cachefiles_withdraw_cookie(struct fscache_cookie *cookie)
 		spin_unlock(&cache->object_list_lock);
 	}
 
+	cachefiles_ondemand_clean_object(object);
+
 	if (object->file) {
 		cachefiles_begin_secure(cache, &saved_cred);
 		cachefiles_clean_up_object(object, cache);
diff --git a/fs/cachefiles/internal.h b/fs/cachefiles/internal.h
index c793d33b0224ec..6cba2c6de2f963 100644
--- a/fs/cachefiles/internal.h
+++ b/fs/cachefiles/internal.h
@@ -15,6 +15,8 @@
 #include <linux/fscache-cache.h>
 #include <linux/cred.h>
 #include <linux/security.h>
+#include <linux/xarray.h>
+#include <linux/cachefiles.h>
 
 #define CACHEFILES_DIO_BLOCK_SIZE 4096
 
@@ -58,8 +60,13 @@ struct cachefiles_object {
 	enum cachefiles_content		content_info:8;	/* Info about content presence */
 	unsigned long			flags;
 #define CACHEFILES_OBJECT_USING_TMPFILE	0		/* Have an unlinked tmpfile */
+#ifdef CONFIG_CACHEFILES_ONDEMAND
+	int				ondemand_id;
+#endif
 };
 
+#define CACHEFILES_ONDEMAND_ID_CLOSED	-1
+
 /*
  * Cache files cache definition
  */
@@ -98,11 +105,31 @@ struct cachefiles_cache {
 #define CACHEFILES_DEAD			1	/* T if cache dead */
 #define CACHEFILES_CULLING		2	/* T if cull engaged */
 #define CACHEFILES_STATE_CHANGED	3	/* T if state changed (poll trigger) */
+#define CACHEFILES_ONDEMAND_MODE	4	/* T if in on-demand read mode */
 	char				*rootdirname;	/* name of cache root directory */
 	char				*secctx;	/* LSM security context */
 	char				*tag;		/* cache binding tag */
+	refcount_t			unbind_pincount;/* refcount to do daemon unbind */
+	struct xarray			reqs;		/* xarray of pending on-demand requests */
+	struct xarray			ondemand_ids;	/* xarray for ondemand_id allocation */
+	u32				ondemand_id_next;
+};
+
+static inline bool cachefiles_in_ondemand_mode(struct cachefiles_cache *cache)
+{
+	return IS_ENABLED(CONFIG_CACHEFILES_ONDEMAND) &&
+		test_bit(CACHEFILES_ONDEMAND_MODE, &cache->flags);
+}
+
+struct cachefiles_req {
+	struct cachefiles_object *object;
+	struct completion done;
+	int error;
+	struct cachefiles_msg msg;
 };
 
+#define CACHEFILES_REQ_NEW	XA_MARK_1
+
 #include <trace/events/cachefiles.h>
 
 static inline
@@ -145,6 +172,8 @@ extern int cachefiles_has_space(struct cachefiles_cache *cache,
  * daemon.c
  */
 extern const struct file_operations cachefiles_daemon_fops;
+extern void cachefiles_get_unbind_pincount(struct cachefiles_cache *cache);
+extern void cachefiles_put_unbind_pincount(struct cachefiles_cache *cache);
 
 /*
  * error_inject.c
@@ -201,6 +230,16 @@ extern void cachefiles_put_object(struct cachefiles_object *object,
  */
 extern bool cachefiles_begin_operation(struct netfs_cache_resources *cres,
 				       enum fscache_want_state want_state);
+extern int __cachefiles_prepare_write(struct cachefiles_object *object,
+				      struct file *file,
+				      loff_t *_start, size_t *_len,
+				      bool no_space_allocated_yet);
+extern int __cachefiles_write(struct cachefiles_object *object,
+			      struct file *file,
+			      loff_t start_pos,
+			      struct iov_iter *iter,
+			      netfs_io_terminated_t term_func,
+			      void *term_func_priv);
 
 /*
  * key.c
@@ -241,6 +280,45 @@ extern bool cachefiles_commit_tmpfile(struct cachefiles_cache *cache,
 				      struct cachefiles_object *object);
 
 /*
+ * ondemand.c
+ */
+#ifdef CONFIG_CACHEFILES_ONDEMAND
+extern ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
+					char __user *_buffer, size_t buflen);
+
+extern int cachefiles_ondemand_copen(struct cachefiles_cache *cache,
+				     char *args);
+
+extern int cachefiles_ondemand_init_object(struct cachefiles_object *object);
+extern void cachefiles_ondemand_clean_object(struct cachefiles_object *object);
+
+extern int cachefiles_ondemand_read(struct cachefiles_object *object,
+				    loff_t pos, size_t len);
+
+#else
+static inline ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
+					char __user *_buffer, size_t buflen)
+{
+	return -EOPNOTSUPP;
+}
+
+static inline int cachefiles_ondemand_init_object(struct cachefiles_object *object)
+{
+	return 0;
+}
+
+static inline void cachefiles_ondemand_clean_object(struct cachefiles_object *object)
+{
+}
+
+static inline int cachefiles_ondemand_read(struct cachefiles_object *object,
+					   loff_t pos, size_t len)
+{
+	return -EOPNOTSUPP;
+}
+#endif
+
+/*
  * security.c
  */
 extern int cachefiles_get_security_ID(struct cachefiles_cache *cache);
diff --git a/fs/cachefiles/io.c b/fs/cachefiles/io.c
index 9dc81e781f2b3b..000a28f46e59ef 100644
--- a/fs/cachefiles/io.c
+++ b/fs/cachefiles/io.c
@@ -277,36 +277,33 @@ static void cachefiles_write_complete(struct kiocb *iocb, long ret)
 /*
  * Initiate a write to the cache.
  */
-static int cachefiles_write(struct netfs_cache_resources *cres,
-			    loff_t start_pos,
-			    struct iov_iter *iter,
-			    netfs_io_terminated_t term_func,
-			    void *term_func_priv)
+int __cachefiles_write(struct cachefiles_object *object,
+		       struct file *file,
+		       loff_t start_pos,
+		       struct iov_iter *iter,
+		       netfs_io_terminated_t term_func,
+		       void *term_func_priv)
 {
-	struct cachefiles_object *object;
 	struct cachefiles_cache *cache;
 	struct cachefiles_kiocb *ki;
 	struct inode *inode;
-	struct file *file;
 	unsigned int old_nofs;
-	ssize_t ret = -ENOBUFS;
+	ssize_t ret;
 	size_t len = iov_iter_count(iter);
 
-	if (!fscache_wait_for_operation(cres, FSCACHE_WANT_WRITE))
-		goto presubmission_error;
 	fscache_count_write();
-	object = cachefiles_cres_object(cres);
 	cache = object->volume->cache;
-	file = cachefiles_cres_file(cres);
 
 	_enter("%pD,%li,%llx,%zx/%llx",
 	       file, file_inode(file)->i_ino, start_pos, len,
 	       i_size_read(file_inode(file)));
 
-	ret = -ENOMEM;
 	ki = kzalloc(sizeof(struct cachefiles_kiocb), GFP_KERNEL);
-	if (!ki)
-		goto presubmission_error;
+	if (!ki) {
+		if (term_func)
+			term_func(term_func_priv, -ENOMEM, false);
+		return -ENOMEM;
+	}
 
 	refcount_set(&ki->ki_refcnt, 2);
 	ki->iocb.ki_filp	= file;
@@ -314,7 +311,6 @@ static int cachefiles_write(struct netfs_cache_resources *cres,
 	ki->iocb.ki_flags	= IOCB_DIRECT | IOCB_WRITE;
 	ki->iocb.ki_ioprio	= get_current_ioprio();
 	ki->object		= object;
-	ki->inval_counter	= cres->inval_counter;
 	ki->start		= start_pos;
 	ki->len			= len;
 	ki->term_func		= term_func;
@@ -369,11 +365,24 @@ in_progress:
 	cachefiles_put_kiocb(ki);
 	_leave(" = %zd", ret);
 	return ret;
+}
 
-presubmission_error:
-	if (term_func)
-		term_func(term_func_priv, ret, false);
-	return ret;
+static int cachefiles_write(struct netfs_cache_resources *cres,
+			    loff_t start_pos,
+			    struct iov_iter *iter,
+			    netfs_io_terminated_t term_func,
+			    void *term_func_priv)
+{
+	if (!fscache_wait_for_operation(cres, FSCACHE_WANT_WRITE)) {
+		if (term_func)
+			term_func(term_func_priv, -ENOBUFS, false);
+		return -ENOBUFS;
+	}
+
+	return __cachefiles_write(cachefiles_cres_object(cres),
+				  cachefiles_cres_file(cres),
+				  start_pos, iter,
+				  term_func, term_func_priv);
 }
 
 /*
@@ -394,6 +403,7 @@ static enum netfs_io_source cachefiles_prepare_read(struct netfs_io_subrequest *
 	enum netfs_io_source ret = NETFS_DOWNLOAD_FROM_SERVER;
 	loff_t off, to;
 	ino_t ino = file ? file_inode(file)->i_ino : 0;
+	int rc;
 
 	_enter("%zx @%llx/%llx", subreq->len, subreq->start, i_size);
 
@@ -406,7 +416,8 @@ static enum netfs_io_source cachefiles_prepare_read(struct netfs_io_subrequest *
 	if (test_bit(FSCACHE_COOKIE_NO_DATA_TO_READ, &cookie->flags)) {
 		__set_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
 		why = cachefiles_trace_read_no_data;
-		goto out_no_object;
+		if (!test_bit(NETFS_SREQ_ONDEMAND, &subreq->flags))
+			goto out_no_object;
 	}
 
 	/* The object and the file may be being created in the background. */
@@ -423,7 +434,7 @@ static enum netfs_io_source cachefiles_prepare_read(struct netfs_io_subrequest *
 	object = cachefiles_cres_object(cres);
 	cache = object->volume->cache;
 	cachefiles_begin_secure(cache, &saved_cred);
-
+retry:
 	off = cachefiles_inject_read_error();
 	if (off == 0)
 		off = vfs_llseek(file, subreq->start, SEEK_DATA);
@@ -474,6 +485,15 @@ static enum netfs_io_source cachefiles_prepare_read(struct netfs_io_subrequest *
 
 download_and_store:
 	__set_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
+	if (test_bit(NETFS_SREQ_ONDEMAND, &subreq->flags)) {
+		rc = cachefiles_ondemand_read(object, subreq->start,
+					      subreq->len);
+		if (!rc) {
+			__clear_bit(NETFS_SREQ_ONDEMAND, &subreq->flags);
+			goto retry;
+		}
+		ret = NETFS_INVALID_READ;
+	}
 out:
 	cachefiles_end_secure(cache, saved_cred);
 out_no_object:
@@ -484,13 +504,12 @@ out_no_object:
 /*
  * Prepare for a write to occur.
  */
-static int __cachefiles_prepare_write(struct netfs_cache_resources *cres,
-				      loff_t *_start, size_t *_len, loff_t i_size,
-				      bool no_space_allocated_yet)
+int __cachefiles_prepare_write(struct cachefiles_object *object,
+			       struct file *file,
+			       loff_t *_start, size_t *_len,
+			       bool no_space_allocated_yet)
 {
-	struct cachefiles_object *object = cachefiles_cres_object(cres);
 	struct cachefiles_cache *cache = object->volume->cache;
-	struct file *file = cachefiles_cres_file(cres);
 	loff_t start = *_start, pos;
 	size_t len = *_len, down;
 	int ret;
@@ -577,7 +596,8 @@ static int cachefiles_prepare_write(struct netfs_cache_resources *cres,
 	}
 
 	cachefiles_begin_secure(cache, &saved_cred);
-	ret = __cachefiles_prepare_write(cres, _start, _len, i_size,
+	ret = __cachefiles_prepare_write(object, cachefiles_cres_file(cres),
+					 _start, _len,
 					 no_space_allocated_yet);
 	cachefiles_end_secure(cache, saved_cred);
 	return ret;
diff --git a/fs/cachefiles/namei.c b/fs/cachefiles/namei.c
index ca9f3e4ec4b3fb..facf2ebe464b35 100644
--- a/fs/cachefiles/namei.c
+++ b/fs/cachefiles/namei.c
@@ -452,10 +452,9 @@ struct file *cachefiles_create_tmpfile(struct cachefiles_object *object)
 	struct dentry *fan = volume->fanout[(u8)object->cookie->key_hash];
 	struct file *file;
 	struct path path;
-	uint64_t ni_size = object->cookie->object_size;
+	uint64_t ni_size;
 	long ret;
 
-	ni_size = round_up(ni_size, CACHEFILES_DIO_BLOCK_SIZE);
 
 	cachefiles_begin_secure(cache, &saved_cred);
 
@@ -481,6 +480,15 @@ struct file *cachefiles_create_tmpfile(struct cachefiles_object *object)
 		goto out_dput;
 	}
 
+	ret = cachefiles_ondemand_init_object(object);
+	if (ret < 0) {
+		file = ERR_PTR(ret);
+		goto out_unuse;
+	}
+
+	ni_size = object->cookie->object_size;
+	ni_size = round_up(ni_size, CACHEFILES_DIO_BLOCK_SIZE);
+
 	if (ni_size > 0) {
 		trace_cachefiles_trunc(object, d_backing_inode(path.dentry), 0, ni_size,
 				       cachefiles_trunc_expand_tmpfile);
@@ -586,6 +594,10 @@ static bool cachefiles_open_file(struct cachefiles_object *object,
 	}
 	_debug("file -> %pd positive", dentry);
 
+	ret = cachefiles_ondemand_init_object(object);
+	if (ret < 0)
+		goto error_fput;
+
 	ret = cachefiles_check_auxdata(object, file);
 	if (ret < 0)
 		goto check_failed;
diff --git a/fs/cachefiles/ondemand.c b/fs/cachefiles/ondemand.c
new file mode 100644
index 00000000000000..a41ae6efc54545
--- /dev/null
+++ b/fs/cachefiles/ondemand.c
@@ -0,0 +1,503 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+#include <linux/fdtable.h>
+#include <linux/anon_inodes.h>
+#include <linux/uio.h>
+#include "internal.h"
+
+static int cachefiles_ondemand_fd_release(struct inode *inode,
+					  struct file *file)
+{
+	struct cachefiles_object *object = file->private_data;
+	struct cachefiles_cache *cache = object->volume->cache;
+	int object_id = object->ondemand_id;
+	struct cachefiles_req *req;
+	XA_STATE(xas, &cache->reqs, 0);
+
+	xa_lock(&cache->reqs);
+	object->ondemand_id = CACHEFILES_ONDEMAND_ID_CLOSED;
+
+	/*
+	 * Flush all pending READ requests since their completion depends on
+	 * anon_fd.
+	 */
+	xas_for_each(&xas, req, ULONG_MAX) {
+		if (req->msg.opcode == CACHEFILES_OP_READ) {
+			req->error = -EIO;
+			complete(&req->done);
+			xas_store(&xas, NULL);
+		}
+	}
+	xa_unlock(&cache->reqs);
+
+	xa_erase(&cache->ondemand_ids, object_id);
+	trace_cachefiles_ondemand_fd_release(object, object_id);
+	cachefiles_put_object(object, cachefiles_obj_put_ondemand_fd);
+	cachefiles_put_unbind_pincount(cache);
+	return 0;
+}
+
+static ssize_t cachefiles_ondemand_fd_write_iter(struct kiocb *kiocb,
+						 struct iov_iter *iter)
+{
+	struct cachefiles_object *object = kiocb->ki_filp->private_data;
+	struct cachefiles_cache *cache = object->volume->cache;
+	struct file *file = object->file;
+	size_t len = iter->count;
+	loff_t pos = kiocb->ki_pos;
+	const struct cred *saved_cred;
+	int ret;
+
+	if (!file)
+		return -ENOBUFS;
+
+	cachefiles_begin_secure(cache, &saved_cred);
+	ret = __cachefiles_prepare_write(object, file, &pos, &len, true);
+	cachefiles_end_secure(cache, saved_cred);
+	if (ret < 0)
+		return ret;
+
+	trace_cachefiles_ondemand_fd_write(object, file_inode(file), pos, len);
+	ret = __cachefiles_write(object, file, pos, iter, NULL, NULL);
+	if (!ret)
+		ret = len;
+
+	return ret;
+}
+
+static loff_t cachefiles_ondemand_fd_llseek(struct file *filp, loff_t pos,
+					    int whence)
+{
+	struct cachefiles_object *object = filp->private_data;
+	struct file *file = object->file;
+
+	if (!file)
+		return -ENOBUFS;
+
+	return vfs_llseek(file, pos, whence);
+}
+
+static long cachefiles_ondemand_fd_ioctl(struct file *filp, unsigned int ioctl,
+					 unsigned long arg)
+{
+	struct cachefiles_object *object = filp->private_data;
+	struct cachefiles_cache *cache = object->volume->cache;
+	struct cachefiles_req *req;
+	unsigned long id;
+
+	if (ioctl != CACHEFILES_IOC_READ_COMPLETE)
+		return -EINVAL;
+
+	if (!test_bit(CACHEFILES_ONDEMAND_MODE, &cache->flags))
+		return -EOPNOTSUPP;
+
+	id = arg;
+	req = xa_erase(&cache->reqs, id);
+	if (!req)
+		return -EINVAL;
+
+	trace_cachefiles_ondemand_cread(object, id);
+	complete(&req->done);
+	return 0;
+}
+
+static const struct file_operations cachefiles_ondemand_fd_fops = {
+	.owner		= THIS_MODULE,
+	.release	= cachefiles_ondemand_fd_release,
+	.write_iter	= cachefiles_ondemand_fd_write_iter,
+	.llseek		= cachefiles_ondemand_fd_llseek,
+	.unlocked_ioctl	= cachefiles_ondemand_fd_ioctl,
+};
+
+/*
+ * OPEN request Completion (copen)
+ * - command: "copen <id>,<cache_size>"
+ *   <cache_size> indicates the object size if >=0, error code if negative
+ */
+int cachefiles_ondemand_copen(struct cachefiles_cache *cache, char *args)
+{
+	struct cachefiles_req *req;
+	struct fscache_cookie *cookie;
+	char *pid, *psize;
+	unsigned long id;
+	long size;
+	int ret;
+
+	if (!test_bit(CACHEFILES_ONDEMAND_MODE, &cache->flags))
+		return -EOPNOTSUPP;
+
+	if (!*args) {
+		pr_err("Empty id specified\n");
+		return -EINVAL;
+	}
+
+	pid = args;
+	psize = strchr(args, ',');
+	if (!psize) {
+		pr_err("Cache size is not specified\n");
+		return -EINVAL;
+	}
+
+	*psize = 0;
+	psize++;
+
+	ret = kstrtoul(pid, 0, &id);
+	if (ret)
+		return ret;
+
+	req = xa_erase(&cache->reqs, id);
+	if (!req)
+		return -EINVAL;
+
+	/* fail OPEN request if copen format is invalid */
+	ret = kstrtol(psize, 0, &size);
+	if (ret) {
+		req->error = ret;
+		goto out;
+	}
+
+	/* fail OPEN request if daemon reports an error */
+	if (size < 0) {
+		if (!IS_ERR_VALUE(size))
+			size = -EINVAL;
+		req->error = size;
+		goto out;
+	}
+
+	cookie = req->object->cookie;
+	cookie->object_size = size;
+	if (size)
+		clear_bit(FSCACHE_COOKIE_NO_DATA_TO_READ, &cookie->flags);
+	else
+		set_bit(FSCACHE_COOKIE_NO_DATA_TO_READ, &cookie->flags);
+	trace_cachefiles_ondemand_copen(req->object, id, size);
+
+out:
+	complete(&req->done);
+	return ret;
+}
+
+static int cachefiles_ondemand_get_fd(struct cachefiles_req *req)
+{
+	struct cachefiles_object *object;
+	struct cachefiles_cache *cache;
+	struct cachefiles_open *load;
+	struct file *file;
+	u32 object_id;
+	int ret, fd;
+
+	object = cachefiles_grab_object(req->object,
+			cachefiles_obj_get_ondemand_fd);
+	cache = object->volume->cache;
+
+	ret = xa_alloc_cyclic(&cache->ondemand_ids, &object_id, NULL,
+			      XA_LIMIT(1, INT_MAX),
+			      &cache->ondemand_id_next, GFP_KERNEL);
+	if (ret < 0)
+		goto err;
+
+	fd = get_unused_fd_flags(O_WRONLY);
+	if (fd < 0) {
+		ret = fd;
+		goto err_free_id;
+	}
+
+	file = anon_inode_getfile("[cachefiles]", &cachefiles_ondemand_fd_fops,
+				  object, O_WRONLY);
+	if (IS_ERR(file)) {
+		ret = PTR_ERR(file);
+		goto err_put_fd;
+	}
+
+	file->f_mode |= FMODE_PWRITE | FMODE_LSEEK;
+	fd_install(fd, file);
+
+	load = (void *)req->msg.data;
+	load->fd = fd;
+	req->msg.object_id = object_id;
+	object->ondemand_id = object_id;
+
+	cachefiles_get_unbind_pincount(cache);
+	trace_cachefiles_ondemand_open(object, &req->msg, load);
+	return 0;
+
+err_put_fd:
+	put_unused_fd(fd);
+err_free_id:
+	xa_erase(&cache->ondemand_ids, object_id);
+err:
+	cachefiles_put_object(object, cachefiles_obj_put_ondemand_fd);
+	return ret;
+}
+
+ssize_t cachefiles_ondemand_daemon_read(struct cachefiles_cache *cache,
+					char __user *_buffer, size_t buflen)
+{
+	struct cachefiles_req *req;
+	struct cachefiles_msg *msg;
+	unsigned long id = 0;
+	size_t n;
+	int ret = 0;
+	XA_STATE(xas, &cache->reqs, 0);
+
+	/*
+	 * Search for a request that has not ever been processed, to prevent
+	 * requests from being processed repeatedly.
+	 */
+	xa_lock(&cache->reqs);
+	req = xas_find_marked(&xas, UINT_MAX, CACHEFILES_REQ_NEW);
+	if (!req) {
+		xa_unlock(&cache->reqs);
+		return 0;
+	}
+
+	msg = &req->msg;
+	n = msg->len;
+
+	if (n > buflen) {
+		xa_unlock(&cache->reqs);
+		return -EMSGSIZE;
+	}
+
+	xas_clear_mark(&xas, CACHEFILES_REQ_NEW);
+	xa_unlock(&cache->reqs);
+
+	id = xas.xa_index;
+	msg->msg_id = id;
+
+	if (msg->opcode == CACHEFILES_OP_OPEN) {
+		ret = cachefiles_ondemand_get_fd(req);
+		if (ret)
+			goto error;
+	}
+
+	if (copy_to_user(_buffer, msg, n) != 0) {
+		ret = -EFAULT;
+		goto err_put_fd;
+	}
+
+	/* CLOSE request has no reply */
+	if (msg->opcode == CACHEFILES_OP_CLOSE) {
+		xa_erase(&cache->reqs, id);
+		complete(&req->done);
+	}
+
+	return n;
+
+err_put_fd:
+	if (msg->opcode == CACHEFILES_OP_OPEN)
+		close_fd(((struct cachefiles_open *)msg->data)->fd);
+error:
+	xa_erase(&cache->reqs, id);
+	req->error = ret;
+	complete(&req->done);
+	return ret;
+}
+
+typedef int (*init_req_fn)(struct cachefiles_req *req, void *private);
+
+static int cachefiles_ondemand_send_req(struct cachefiles_object *object,
+					enum cachefiles_opcode opcode,
+					size_t data_len,
+					init_req_fn init_req,
+					void *private)
+{
+	struct cachefiles_cache *cache = object->volume->cache;
+	struct cachefiles_req *req;
+	XA_STATE(xas, &cache->reqs, 0);
+	int ret;
+
+	if (!test_bit(CACHEFILES_ONDEMAND_MODE, &cache->flags))
+		return 0;
+
+	if (test_bit(CACHEFILES_DEAD, &cache->flags))
+		return -EIO;
+
+	req = kzalloc(sizeof(*req) + data_len, GFP_KERNEL);
+	if (!req)
+		return -ENOMEM;
+
+	req->object = object;
+	init_completion(&req->done);
+	req->msg.opcode = opcode;
+	req->msg.len = sizeof(struct cachefiles_msg) + data_len;
+
+	ret = init_req(req, private);
+	if (ret)
+		goto out;
+
+	do {
+		/*
+		 * Stop enqueuing the request when daemon is dying. The
+		 * following two operations need to be atomic as a whole.
+		 *   1) check cache state, and
+		 *   2) enqueue request if cache is alive.
+		 * Otherwise the request may be enqueued after xarray has been
+		 * flushed, leaving the orphan request never being completed.
+		 *
+		 * CPU 1			CPU 2
+		 * =====			=====
+		 *				test CACHEFILES_DEAD bit
+		 * set CACHEFILES_DEAD bit
+		 * flush requests in the xarray
+		 *				enqueue the request
+		 */
+		xas_lock(&xas);
+
+		if (test_bit(CACHEFILES_DEAD, &cache->flags)) {
+			xas_unlock(&xas);
+			ret = -EIO;
+			goto out;
+		}
+
+		/* coupled with the barrier in cachefiles_flush_reqs() */
+		smp_mb();
+
+		if (opcode != CACHEFILES_OP_OPEN && object->ondemand_id <= 0) {
+			WARN_ON_ONCE(object->ondemand_id == 0);
+			xas_unlock(&xas);
+			ret = -EIO;
+			goto out;
+		}
+
+		xas.xa_index = 0;
+		xas_find_marked(&xas, UINT_MAX, XA_FREE_MARK);
+		if (xas.xa_node == XAS_RESTART)
+			xas_set_err(&xas, -EBUSY);
+		xas_store(&xas, req);
+		xas_clear_mark(&xas, XA_FREE_MARK);
+		xas_set_mark(&xas, CACHEFILES_REQ_NEW);
+		xas_unlock(&xas);
+	} while (xas_nomem(&xas, GFP_KERNEL));
+
+	ret = xas_error(&xas);
+	if (ret)
+		goto out;
+
+	wake_up_all(&cache->daemon_pollwq);
+	wait_for_completion(&req->done);
+	ret = req->error;
+out:
+	kfree(req);
+	return ret;
+}
+
+static int cachefiles_ondemand_init_open_req(struct cachefiles_req *req,
+					     void *private)
+{
+	struct cachefiles_object *object = req->object;
+	struct fscache_cookie *cookie = object->cookie;
+	struct fscache_volume *volume = object->volume->vcookie;
+	struct cachefiles_open *load = (void *)req->msg.data;
+	size_t volume_key_size, cookie_key_size;
+	void *volume_key, *cookie_key;
+
+	/*
+	 * Volume key is a NUL-terminated string. key[0] stores strlen() of the
+	 * string, followed by the content of the string (excluding '\0').
+	 */
+	volume_key_size = volume->key[0] + 1;
+	volume_key = volume->key + 1;
+
+	/* Cookie key is binary data, which is netfs specific. */
+	cookie_key_size = cookie->key_len;
+	cookie_key = fscache_get_key(cookie);
+
+	if (!(object->cookie->advice & FSCACHE_ADV_WANT_CACHE_SIZE)) {
+		pr_err("WANT_CACHE_SIZE is needed for on-demand mode\n");
+		return -EINVAL;
+	}
+
+	load->volume_key_size = volume_key_size;
+	load->cookie_key_size = cookie_key_size;
+	memcpy(load->data, volume_key, volume_key_size);
+	memcpy(load->data + volume_key_size, cookie_key, cookie_key_size);
+
+	return 0;
+}
+
+static int cachefiles_ondemand_init_close_req(struct cachefiles_req *req,
+					      void *private)
+{
+	struct cachefiles_object *object = req->object;
+	int object_id = object->ondemand_id;
+
+	/*
+	 * It's possible that object id is still 0 if the cookie looking up
+	 * phase failed before OPEN request has ever been sent. Also avoid
+	 * sending CLOSE request for CACHEFILES_ONDEMAND_ID_CLOSED, which means
+	 * anon_fd has already been closed.
+	 */
+	if (object_id <= 0)
+		return -ENOENT;
+
+	req->msg.object_id = object_id;
+	trace_cachefiles_ondemand_close(object, &req->msg);
+	return 0;
+}
+
+struct cachefiles_read_ctx {
+	loff_t off;
+	size_t len;
+};
+
+static int cachefiles_ondemand_init_read_req(struct cachefiles_req *req,
+					     void *private)
+{
+	struct cachefiles_object *object = req->object;
+	struct cachefiles_read *load = (void *)req->msg.data;
+	struct cachefiles_read_ctx *read_ctx = private;
+	int object_id = object->ondemand_id;
+
+	/* Stop enqueuing requests when daemon has closed anon_fd. */
+	if (object_id <= 0) {
+		WARN_ON_ONCE(object_id == 0);
+		pr_info_once("READ: anonymous fd closed prematurely.\n");
+		return -EIO;
+	}
+
+	req->msg.object_id = object_id;
+	load->off = read_ctx->off;
+	load->len = read_ctx->len;
+	trace_cachefiles_ondemand_read(object, &req->msg, load);
+	return 0;
+}
+
+int cachefiles_ondemand_init_object(struct cachefiles_object *object)
+{
+	struct fscache_cookie *cookie = object->cookie;
+	struct fscache_volume *volume = object->volume->vcookie;
+	size_t volume_key_size, cookie_key_size, data_len;
+
+	/*
+	 * CacheFiles will firstly check the cache file under the root cache
+	 * directory. If the coherency check failed, it will fallback to
+	 * creating a new tmpfile as the cache file. Reuse the previously
+	 * allocated object ID if any.
+	 */
+	if (object->ondemand_id > 0)
+		return 0;
+
+	volume_key_size = volume->key[0] + 1;
+	cookie_key_size = cookie->key_len;
+	data_len = sizeof(struct cachefiles_open) +
+		   volume_key_size + cookie_key_size;
+
+	return cachefiles_ondemand_send_req(object, CACHEFILES_OP_OPEN,
+			data_len, cachefiles_ondemand_init_open_req, NULL);
+}
+
+void cachefiles_ondemand_clean_object(struct cachefiles_object *object)
+{
+	cachefiles_ondemand_send_req(object, CACHEFILES_OP_CLOSE, 0,
+			cachefiles_ondemand_init_close_req, NULL);
+}
+
+int cachefiles_ondemand_read(struct cachefiles_object *object,
+			     loff_t pos, size_t len)
+{
+	struct cachefiles_read_ctx read_ctx = {pos, len};
+
+	return cachefiles_ondemand_send_req(object, CACHEFILES_OP_READ,
+			sizeof(struct cachefiles_read),
+			cachefiles_ondemand_init_read_req, &read_ctx);
+}
diff --git a/fs/erofs/Kconfig b/fs/erofs/Kconfig
index f57255ab88ed4a..85490370e0cad2 100644
--- a/fs/erofs/Kconfig
+++ b/fs/erofs/Kconfig
@@ -98,3 +98,13 @@ config EROFS_FS_ZIP_LZMA
 	  systems will be readable without selecting this option.
 
 	  If unsure, say N.
+
+config EROFS_FS_ONDEMAND
+	bool "EROFS fscache-based on-demand read support"
+	depends on CACHEFILES_ONDEMAND && (EROFS_FS=m && FSCACHE || EROFS_FS=y && FSCACHE=y)
+	default n
+	help
+	  This permits EROFS to use fscache-backed data blobs with on-demand
+	  read support.
+
+	  If unsure, say N.
diff --git a/fs/erofs/Makefile b/fs/erofs/Makefile
index 8a3317e38e5a8a..99bbc597a3e923 100644
--- a/fs/erofs/Makefile
+++ b/fs/erofs/Makefile
@@ -5,3 +5,4 @@ erofs-objs := super.o inode.o data.o namei.o dir.o utils.o pcpubuf.o sysfs.o
 erofs-$(CONFIG_EROFS_FS_XATTR) += xattr.o
 erofs-$(CONFIG_EROFS_FS_ZIP) += decompressor.o zmap.o zdata.o
 erofs-$(CONFIG_EROFS_FS_ZIP_LZMA) += decompressor_lzma.o
+erofs-$(CONFIG_EROFS_FS_ONDEMAND) += fscache.o
diff --git a/fs/erofs/data.c b/fs/erofs/data.c
index 780db1e5f4b723..bb9c1fd48c1936 100644
--- a/fs/erofs/data.c
+++ b/fs/erofs/data.c
@@ -6,6 +6,7 @@
  */
 #include "internal.h"
 #include <linux/prefetch.h>
+#include <linux/sched/mm.h>
 #include <linux/dax.h>
 #include <trace/events/erofs.h>
 
@@ -35,14 +36,20 @@ void *erofs_bread(struct erofs_buf *buf, struct inode *inode,
 	erofs_off_t offset = blknr_to_addr(blkaddr);
 	pgoff_t index = offset >> PAGE_SHIFT;
 	struct page *page = buf->page;
+	struct folio *folio;
+	unsigned int nofs_flag;
 
 	if (!page || page->index != index) {
 		erofs_put_metabuf(buf);
-		page = read_cache_page_gfp(mapping, index,
-				mapping_gfp_constraint(mapping, ~__GFP_FS));
-		if (IS_ERR(page))
-			return page;
+
+		nofs_flag = memalloc_nofs_save();
+		folio = read_cache_folio(mapping, index, NULL, NULL);
+		memalloc_nofs_restore(nofs_flag);
+		if (IS_ERR(folio))
+			return folio;
+
 		/* should already be PageUptodate, no need to lock page */
+		page = folio_file_page(folio, index);
 		buf->page = page;
 	}
 	if (buf->kmap_type == EROFS_NO_KMAP) {
@@ -63,6 +70,10 @@ void *erofs_bread(struct erofs_buf *buf, struct inode *inode,
 void *erofs_read_metabuf(struct erofs_buf *buf, struct super_block *sb,
 			 erofs_blk_t blkaddr, enum erofs_kmap_type type)
 {
+	if (erofs_is_fscache_mode(sb))
+		return erofs_bread(buf, EROFS_SB(sb)->s_fscache->inode,
+				   blkaddr, type);
+
 	return erofs_bread(buf, sb->s_bdev->bd_inode, blkaddr, type);
 }
 
@@ -110,8 +121,8 @@ static int erofs_map_blocks_flatmode(struct inode *inode,
 	return 0;
 }
 
-static int erofs_map_blocks(struct inode *inode,
-			    struct erofs_map_blocks *map, int flags)
+int erofs_map_blocks(struct inode *inode,
+		     struct erofs_map_blocks *map, int flags)
 {
 	struct super_block *sb = inode->i_sb;
 	struct erofs_inode *vi = EROFS_I(inode);
@@ -199,6 +210,7 @@ int erofs_map_dev(struct super_block *sb, struct erofs_map_dev *map)
 	map->m_bdev = sb->s_bdev;
 	map->m_daxdev = EROFS_SB(sb)->dax_dev;
 	map->m_dax_part_off = EROFS_SB(sb)->dax_part_off;
+	map->m_fscache = EROFS_SB(sb)->s_fscache;
 
 	if (map->m_deviceid) {
 		down_read(&devs->rwsem);
@@ -210,6 +222,7 @@ int erofs_map_dev(struct super_block *sb, struct erofs_map_dev *map)
 		map->m_bdev = dif->bdev;
 		map->m_daxdev = dif->dax_dev;
 		map->m_dax_part_off = dif->dax_part_off;
+		map->m_fscache = dif->fscache;
 		up_read(&devs->rwsem);
 	} else if (devs->extra_devices) {
 		down_read(&devs->rwsem);
@@ -227,6 +240,7 @@ int erofs_map_dev(struct super_block *sb, struct erofs_map_dev *map)
 				map->m_bdev = dif->bdev;
 				map->m_daxdev = dif->dax_dev;
 				map->m_dax_part_off = dif->dax_part_off;
+				map->m_fscache = dif->fscache;
 				break;
 			}
 		}
diff --git a/fs/erofs/decompressor.c b/fs/erofs/decompressor.c
index 3efa686c764417..6dca1900c7331d 100644
--- a/fs/erofs/decompressor.c
+++ b/fs/erofs/decompressor.c
@@ -46,8 +46,6 @@ int z_erofs_load_lz4_config(struct super_block *sb,
 			erofs_err(sb, "too large lz4 pclusterblks %u",
 				  sbi->lz4.max_pclusterblks);
 			return -EINVAL;
-		} else if (sbi->lz4.max_pclusterblks >= 2) {
-			erofs_info(sb, "EXPERIMENTAL big pcluster feature in use. Use at your own risk!");
 		}
 	} else {
 		distance = le16_to_cpu(dsb->u1.lz4_max_distance);
@@ -322,6 +320,7 @@ static int z_erofs_shifted_transform(struct z_erofs_decompress_req *rq,
 		PAGE_ALIGN(rq->pageofs_out + rq->outputsize) >> PAGE_SHIFT;
 	const unsigned int righthalf = min_t(unsigned int, rq->outputsize,
 					     PAGE_SIZE - rq->pageofs_out);
+	const unsigned int lefthalf = rq->outputsize - righthalf;
 	unsigned char *src, *dst;
 
 	if (nrpages_out > 2) {
@@ -344,10 +343,10 @@ static int z_erofs_shifted_transform(struct z_erofs_decompress_req *rq,
 	if (nrpages_out == 2) {
 		DBG_BUGON(!rq->out[1]);
 		if (rq->out[1] == *rq->in) {
-			memmove(src, src + righthalf, rq->pageofs_out);
+			memmove(src, src + righthalf, lefthalf);
 		} else {
 			dst = kmap_atomic(rq->out[1]);
-			memcpy(dst, src + righthalf, rq->pageofs_out);
+			memcpy(dst, src + righthalf, lefthalf);
 			kunmap_atomic(dst);
 		}
 	}
diff --git a/fs/erofs/erofs_fs.h b/fs/erofs/erofs_fs.h
index 1238ca104f099a..2b48373f690b65 100644
--- a/fs/erofs/erofs_fs.h
+++ b/fs/erofs/erofs_fs.h
@@ -37,12 +37,9 @@
 #define EROFS_SB_EXTSLOT_SIZE	16
 
 struct erofs_deviceslot {
-	union {
-		u8 uuid[16];		/* used for device manager later */
-		u8 userdata[64];	/* digest(sha256), etc. */
-	} u;
-	__le32 blocks;			/* total fs blocks of this device */
-	__le32 mapped_blkaddr;		/* map starting at mapped_blkaddr */
+	u8 tag[64];		/* digest(sha256), etc. */
+	__le32 blocks;		/* total fs blocks of this device */
+	__le32 mapped_blkaddr;	/* map starting at mapped_blkaddr */
 	u8 reserved[56];
 };
 #define EROFS_DEVT_SLOT_SIZE	sizeof(struct erofs_deviceslot)
@@ -58,8 +55,8 @@ struct erofs_super_block {
 	__le16 root_nid;	/* nid of root directory */
 	__le64 inos;            /* total valid ino # (== f_files - f_favail) */
 
-	__le64 build_time;      /* inode v1 time derivation */
-	__le32 build_time_nsec;	/* inode v1 time derivation in nano scale */
+	__le64 build_time;      /* compact inode time derivation */
+	__le32 build_time_nsec;	/* compact inode time derivation in ns scale */
 	__le32 blocks;          /* used for statfs */
 	__le32 meta_blkaddr;	/* start block address of metadata area */
 	__le32 xattr_blkaddr;	/* start block address of shared xattr area */
@@ -79,15 +76,15 @@ struct erofs_super_block {
 
 /*
  * erofs inode datalayout (i_format in on-disk inode):
- * 0 - inode plain without inline data A:
+ * 0 - uncompressed flat inode without tail-packing inline data:
  * inode, [xattrs], ... | ... | no-holed data
- * 1 - inode VLE compression B (legacy):
- * inode, [xattrs], extents ... | ...
- * 2 - inode plain with inline data C:
- * inode, [xattrs], last_inline_data, ... | ... | no-holed data
- * 3 - inode compression D:
+ * 1 - compressed inode with non-compact indexes:
+ * inode, [xattrs], [map_header], extents ... | ...
+ * 2 - uncompressed flat inode with tail-packing inline data:
+ * inode, [xattrs], tailpacking data, ... | ... | no-holed data
+ * 3 - compressed inode with compact indexes:
  * inode, [xattrs], map_header, extents ... | ...
- * 4 - inode chunk-based E:
+ * 4 - chunk-based inode with (optional) multi-device support:
  * inode, [xattrs], chunk indexes ... | ...
  * 5~7 - reserved
  */
@@ -106,7 +103,7 @@ static inline bool erofs_inode_is_data_compressed(unsigned int datamode)
 		datamode == EROFS_INODE_FLAT_COMPRESSION_LEGACY;
 }
 
-/* bit definitions of inode i_advise */
+/* bit definitions of inode i_format */
 #define EROFS_I_VERSION_BITS            1
 #define EROFS_I_DATALAYOUT_BITS         3
 
@@ -140,8 +137,9 @@ struct erofs_inode_compact {
 	__le32 i_size;
 	__le32 i_reserved;
 	union {
-		/* file total compressed blocks for data mapping 1 */
+		/* total compressed blocks for compressed inodes */
 		__le32 compressed_blocks;
+		/* block address for uncompressed flat inodes */
 		__le32 raw_blkaddr;
 
 		/* for device files, used to indicate old/new device # */
@@ -156,9 +154,9 @@ struct erofs_inode_compact {
 	__le32 i_reserved2;
 };
 
-/* 32 bytes on-disk inode */
+/* 32-byte on-disk inode */
 #define EROFS_INODE_LAYOUT_COMPACT	0
-/* 64 bytes on-disk inode */
+/* 64-byte on-disk inode */
 #define EROFS_INODE_LAYOUT_EXTENDED	1
 
 /* 64-byte complete form of an ondisk inode */
@@ -171,8 +169,9 @@ struct erofs_inode_extended {
 	__le16 i_reserved;
 	__le64 i_size;
 	union {
-		/* file total compressed blocks for data mapping 1 */
+		/* total compressed blocks for compressed inodes */
 		__le32 compressed_blocks;
+		/* block address for uncompressed flat inodes */
 		__le32 raw_blkaddr;
 
 		/* for device files, used to indicate old/new device # */
@@ -365,17 +364,16 @@ enum {
 
 struct z_erofs_vle_decompressed_index {
 	__le16 di_advise;
-	/* where to decompress in the head cluster */
+	/* where to decompress in the head lcluster */
 	__le16 di_clusterofs;
 
 	union {
-		/* for the head cluster */
+		/* for the HEAD lclusters */
 		__le32 blkaddr;
 		/*
-		 * for the rest clusters
-		 * eg. for 4k page-sized cluster, maximum 4K*64k = 256M)
-		 * [0] - pointing to the head cluster
-		 * [1] - pointing to the tail cluster
+		 * for the NONHEAD lclusters
+		 * [0] - distance to its HEAD lcluster
+		 * [1] - distance to the next HEAD lcluster
 		 */
 		__le16 delta[2];
 	} di_u;
diff --git a/fs/erofs/fscache.c b/fs/erofs/fscache.c
new file mode 100644
index 00000000000000..7e4417167d0b41
--- /dev/null
+++ b/fs/erofs/fscache.c
@@ -0,0 +1,521 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (C) 2022, Alibaba Cloud
+ */
+#include <linux/fscache.h>
+#include "internal.h"
+
+static struct netfs_io_request *erofs_fscache_alloc_request(struct address_space *mapping,
+					     loff_t start, size_t len)
+{
+	struct netfs_io_request *rreq;
+
+	rreq = kzalloc(sizeof(struct netfs_io_request), GFP_KERNEL);
+	if (!rreq)
+		return ERR_PTR(-ENOMEM);
+
+	rreq->start	= start;
+	rreq->len	= len;
+	rreq->mapping	= mapping;
+	INIT_LIST_HEAD(&rreq->subrequests);
+	refcount_set(&rreq->ref, 1);
+	return rreq;
+}
+
+static void erofs_fscache_put_request(struct netfs_io_request *rreq)
+{
+	if (!refcount_dec_and_test(&rreq->ref))
+		return;
+	if (rreq->cache_resources.ops)
+		rreq->cache_resources.ops->end_operation(&rreq->cache_resources);
+	kfree(rreq);
+}
+
+static void erofs_fscache_put_subrequest(struct netfs_io_subrequest *subreq)
+{
+	if (!refcount_dec_and_test(&subreq->ref))
+		return;
+	erofs_fscache_put_request(subreq->rreq);
+	kfree(subreq);
+}
+
+static void erofs_fscache_clear_subrequests(struct netfs_io_request *rreq)
+{
+	struct netfs_io_subrequest *subreq;
+
+	while (!list_empty(&rreq->subrequests)) {
+		subreq = list_first_entry(&rreq->subrequests,
+				struct netfs_io_subrequest, rreq_link);
+		list_del(&subreq->rreq_link);
+		erofs_fscache_put_subrequest(subreq);
+	}
+}
+
+static void erofs_fscache_rreq_unlock_folios(struct netfs_io_request *rreq)
+{
+	struct netfs_io_subrequest *subreq;
+	struct folio *folio;
+	unsigned int iopos = 0;
+	pgoff_t start_page = rreq->start / PAGE_SIZE;
+	pgoff_t last_page = ((rreq->start + rreq->len) / PAGE_SIZE) - 1;
+	bool subreq_failed = false;
+
+	XA_STATE(xas, &rreq->mapping->i_pages, start_page);
+
+	subreq = list_first_entry(&rreq->subrequests,
+				  struct netfs_io_subrequest, rreq_link);
+	subreq_failed = (subreq->error < 0);
+
+	rcu_read_lock();
+	xas_for_each(&xas, folio, last_page) {
+		unsigned int pgpos =
+			(folio_index(folio) - start_page) * PAGE_SIZE;
+		unsigned int pgend = pgpos + folio_size(folio);
+		bool pg_failed = false;
+
+		for (;;) {
+			if (!subreq) {
+				pg_failed = true;
+				break;
+			}
+
+			pg_failed |= subreq_failed;
+			if (pgend < iopos + subreq->len)
+				break;
+
+			iopos += subreq->len;
+			if (!list_is_last(&subreq->rreq_link,
+					  &rreq->subrequests)) {
+				subreq = list_next_entry(subreq, rreq_link);
+				subreq_failed = (subreq->error < 0);
+			} else {
+				subreq = NULL;
+				subreq_failed = false;
+			}
+			if (pgend == iopos)
+				break;
+		}
+
+		if (!pg_failed)
+			folio_mark_uptodate(folio);
+
+		folio_unlock(folio);
+	}
+	rcu_read_unlock();
+}
+
+static void erofs_fscache_rreq_complete(struct netfs_io_request *rreq)
+{
+	erofs_fscache_rreq_unlock_folios(rreq);
+	erofs_fscache_clear_subrequests(rreq);
+	erofs_fscache_put_request(rreq);
+}
+
+static void erofc_fscache_subreq_complete(void *priv,
+		ssize_t transferred_or_error, bool was_async)
+{
+	struct netfs_io_subrequest *subreq = priv;
+	struct netfs_io_request *rreq = subreq->rreq;
+
+	if (IS_ERR_VALUE(transferred_or_error))
+		subreq->error = transferred_or_error;
+
+	if (atomic_dec_and_test(&rreq->nr_outstanding))
+		erofs_fscache_rreq_complete(rreq);
+
+	erofs_fscache_put_subrequest(subreq);
+}
+
+/*
+ * Read data from fscache and fill the read data into page cache described by
+ * @rreq, which shall be both aligned with PAGE_SIZE. @pstart describes
+ * the start physical address in the cache file.
+ */
+static int erofs_fscache_read_folios_async(struct fscache_cookie *cookie,
+				struct netfs_io_request *rreq, loff_t pstart)
+{
+	enum netfs_io_source source;
+	struct super_block *sb = rreq->mapping->host->i_sb;
+	struct netfs_io_subrequest *subreq;
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
+	struct iov_iter iter;
+	loff_t start = rreq->start;
+	size_t len = rreq->len;
+	size_t done = 0;
+	int ret;
+
+	atomic_set(&rreq->nr_outstanding, 1);
+
+	ret = fscache_begin_read_operation(cres, cookie);
+	if (ret)
+		goto out;
+
+	while (done < len) {
+		subreq = kzalloc(sizeof(struct netfs_io_subrequest),
+				 GFP_KERNEL);
+		if (subreq) {
+			INIT_LIST_HEAD(&subreq->rreq_link);
+			refcount_set(&subreq->ref, 2);
+			subreq->rreq = rreq;
+			refcount_inc(&rreq->ref);
+		} else {
+			ret = -ENOMEM;
+			goto out;
+		}
+
+		subreq->start = pstart + done;
+		subreq->len	=  len - done;
+		subreq->flags = 1 << NETFS_SREQ_ONDEMAND;
+
+		list_add_tail(&subreq->rreq_link, &rreq->subrequests);
+
+		source = cres->ops->prepare_read(subreq, LLONG_MAX);
+		if (WARN_ON(subreq->len == 0))
+			source = NETFS_INVALID_READ;
+		if (source != NETFS_READ_FROM_CACHE) {
+			erofs_err(sb, "failed to fscache prepare_read (source %d)",
+				  source);
+			ret = -EIO;
+			subreq->error = ret;
+			erofs_fscache_put_subrequest(subreq);
+			goto out;
+		}
+
+		atomic_inc(&rreq->nr_outstanding);
+
+		iov_iter_xarray(&iter, READ, &rreq->mapping->i_pages,
+				start + done, subreq->len);
+
+		ret = fscache_read(cres, subreq->start, &iter,
+				   NETFS_READ_HOLE_FAIL,
+				   erofc_fscache_subreq_complete, subreq);
+		if (ret == -EIOCBQUEUED)
+			ret = 0;
+		if (ret) {
+			erofs_err(sb, "failed to fscache_read (ret %d)", ret);
+			goto out;
+		}
+
+		done += subreq->len;
+	}
+out:
+	if (atomic_dec_and_test(&rreq->nr_outstanding))
+		erofs_fscache_rreq_complete(rreq);
+
+	return ret;
+}
+
+static int erofs_fscache_meta_readpage(struct file *data, struct page *page)
+{
+	int ret;
+	struct folio *folio = page_folio(page);
+	struct super_block *sb = folio_mapping(folio)->host->i_sb;
+	struct netfs_io_request *rreq;
+	struct erofs_map_dev mdev = {
+		.m_deviceid = 0,
+		.m_pa = folio_pos(folio),
+	};
+
+	ret = erofs_map_dev(sb, &mdev);
+	if (ret)
+		goto out;
+
+	rreq = erofs_fscache_alloc_request(folio_mapping(folio),
+				folio_pos(folio), folio_size(folio));
+	if (IS_ERR(rreq))
+		goto out;
+
+	return erofs_fscache_read_folios_async(mdev.m_fscache->cookie,
+				rreq, mdev.m_pa);
+out:
+	folio_unlock(folio);
+	return ret;
+}
+
+static int erofs_fscache_readpage_inline(struct folio *folio,
+					 struct erofs_map_blocks *map)
+{
+	struct super_block *sb = folio_mapping(folio)->host->i_sb;
+	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
+	erofs_blk_t blknr;
+	size_t offset, len;
+	void *src, *dst;
+
+	/* For tail packing layout, the offset may be non-zero. */
+	offset = erofs_blkoff(map->m_pa);
+	blknr = erofs_blknr(map->m_pa);
+	len = map->m_llen;
+
+	src = erofs_read_metabuf(&buf, sb, blknr, EROFS_KMAP);
+	if (IS_ERR(src))
+		return PTR_ERR(src);
+
+	dst = kmap_local_folio(folio, 0);
+	memcpy(dst, src + offset, len);
+	memset(dst + len, 0, PAGE_SIZE - len);
+	kunmap_local(dst);
+
+	erofs_put_metabuf(&buf);
+	return 0;
+}
+
+static int erofs_fscache_readpage(struct file *file, struct page *page)
+{
+	struct folio *folio = page_folio(page);
+	struct inode *inode = folio_mapping(folio)->host;
+	struct super_block *sb = inode->i_sb;
+	struct erofs_map_blocks map;
+	struct erofs_map_dev mdev;
+	struct netfs_io_request *rreq;
+	erofs_off_t pos;
+	loff_t pstart;
+	int ret;
+
+	DBG_BUGON(folio_size(folio) != EROFS_BLKSIZ);
+
+	pos = folio_pos(folio);
+	map.m_la = pos;
+
+	ret = erofs_map_blocks(inode, &map, EROFS_GET_BLOCKS_RAW);
+	if (ret)
+		goto out_unlock;
+
+	if (!(map.m_flags & EROFS_MAP_MAPPED)) {
+		folio_zero_range(folio, 0, folio_size(folio));
+		goto out_uptodate;
+	}
+
+	if (map.m_flags & EROFS_MAP_META) {
+		ret = erofs_fscache_readpage_inline(folio, &map);
+		goto out_uptodate;
+	}
+
+	mdev = (struct erofs_map_dev) {
+		.m_deviceid = map.m_deviceid,
+		.m_pa = map.m_pa,
+	};
+
+	ret = erofs_map_dev(sb, &mdev);
+	if (ret)
+		goto out_unlock;
+
+
+	rreq = erofs_fscache_alloc_request(folio_mapping(folio),
+				folio_pos(folio), folio_size(folio));
+	if (IS_ERR(rreq))
+		goto out_unlock;
+
+	pstart = mdev.m_pa + (pos - map.m_la);
+	return erofs_fscache_read_folios_async(mdev.m_fscache->cookie,
+				rreq, pstart);
+
+out_uptodate:
+	if (!ret)
+		folio_mark_uptodate(folio);
+out_unlock:
+	folio_unlock(folio);
+	return ret;
+}
+
+static void erofs_fscache_advance_folios(struct readahead_control *rac,
+					 size_t len, bool unlock)
+{
+	while (len) {
+		struct folio *folio = readahead_folio(rac);
+		len -= folio_size(folio);
+		if (unlock) {
+			folio_mark_uptodate(folio);
+			folio_unlock(folio);
+		}
+	}
+}
+
+static void erofs_fscache_readahead(struct readahead_control *rac)
+{
+	struct inode *inode = rac->mapping->host;
+	struct super_block *sb = inode->i_sb;
+	size_t len, count, done = 0;
+	erofs_off_t pos;
+	loff_t start, offset;
+	int ret;
+
+	if (!readahead_count(rac))
+		return;
+
+	start = readahead_pos(rac);
+	len = readahead_length(rac);
+
+	do {
+		struct erofs_map_blocks map;
+		struct erofs_map_dev mdev;
+		struct netfs_io_request *rreq;
+
+		pos = start + done;
+		map.m_la = pos;
+
+		ret = erofs_map_blocks(inode, &map, EROFS_GET_BLOCKS_RAW);
+		if (ret)
+			return;
+
+		offset = start + done;
+		count = min_t(size_t, map.m_llen - (pos - map.m_la),
+			      len - done);
+
+		if (!(map.m_flags & EROFS_MAP_MAPPED)) {
+			struct iov_iter iter;
+
+			iov_iter_xarray(&iter, READ, &rac->mapping->i_pages,
+					offset, count);
+			iov_iter_zero(count, &iter);
+
+			erofs_fscache_advance_folios(rac, count, true);
+			ret = count;
+			continue;
+		}
+
+		if (map.m_flags & EROFS_MAP_META) {
+			struct folio *folio = readahead_folio(rac);
+
+			ret = erofs_fscache_readpage_inline(folio, &map);
+			if (!ret) {
+				folio_mark_uptodate(folio);
+				ret = folio_size(folio);
+			}
+
+			folio_unlock(folio);
+			continue;
+		}
+
+		mdev = (struct erofs_map_dev) {
+			.m_deviceid = map.m_deviceid,
+			.m_pa = map.m_pa,
+		};
+		ret = erofs_map_dev(sb, &mdev);
+		if (ret)
+			return;
+
+		rreq = erofs_fscache_alloc_request(rac->mapping, offset, count);
+		if (IS_ERR(rreq))
+			return;
+		/*
+		 * Drop the ref of folios here. Unlock them in
+		 * rreq_unlock_folios() when rreq complete.
+		 */
+		erofs_fscache_advance_folios(rac, count, false);
+		ret = erofs_fscache_read_folios_async(mdev.m_fscache->cookie,
+					rreq, mdev.m_pa + (pos - map.m_la));
+		if (!ret)
+			ret = count;
+	} while (ret > 0 && ((done += ret) < len));
+}
+
+static const struct address_space_operations erofs_fscache_meta_aops = {
+	.readpage = erofs_fscache_meta_readpage,
+};
+
+const struct address_space_operations erofs_fscache_access_aops = {
+	.readpage = erofs_fscache_readpage,
+	.readahead = erofs_fscache_readahead,
+};
+
+int erofs_fscache_register_cookie(struct super_block *sb,
+				  struct erofs_fscache **fscache,
+				  char *name, bool need_inode)
+{
+	struct fscache_volume *volume = EROFS_SB(sb)->volume;
+	struct erofs_fscache *ctx;
+	struct fscache_cookie *cookie;
+	int ret;
+
+	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+	if (!ctx)
+		return -ENOMEM;
+
+	cookie = fscache_acquire_cookie(volume, FSCACHE_ADV_WANT_CACHE_SIZE,
+					name, strlen(name), NULL, 0, 0);
+	if (!cookie) {
+		erofs_err(sb, "failed to get cookie for %s", name);
+		ret = -EINVAL;
+		goto err;
+	}
+
+	fscache_use_cookie(cookie, false);
+	ctx->cookie = cookie;
+
+	if (need_inode) {
+		struct inode *const inode = new_inode(sb);
+
+		if (!inode) {
+			erofs_err(sb, "failed to get anon inode for %s", name);
+			ret = -ENOMEM;
+			goto err_cookie;
+		}
+
+		set_nlink(inode, 1);
+		inode->i_size = OFFSET_MAX;
+		inode->i_mapping->a_ops = &erofs_fscache_meta_aops;
+		mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS);
+
+		ctx->inode = inode;
+	}
+
+	*fscache = ctx;
+	return 0;
+
+err_cookie:
+	fscache_unuse_cookie(ctx->cookie, NULL, NULL);
+	fscache_relinquish_cookie(ctx->cookie, false);
+	ctx->cookie = NULL;
+err:
+	kfree(ctx);
+	return ret;
+}
+
+void erofs_fscache_unregister_cookie(struct erofs_fscache **fscache)
+{
+	struct erofs_fscache *ctx = *fscache;
+
+	if (!ctx)
+		return;
+
+	fscache_unuse_cookie(ctx->cookie, NULL, NULL);
+	fscache_relinquish_cookie(ctx->cookie, false);
+	ctx->cookie = NULL;
+
+	iput(ctx->inode);
+	ctx->inode = NULL;
+
+	kfree(ctx);
+	*fscache = NULL;
+}
+
+int erofs_fscache_register_fs(struct super_block *sb)
+{
+	struct erofs_sb_info *sbi = EROFS_SB(sb);
+	struct fscache_volume *volume;
+	char *name;
+	int ret = 0;
+
+	name = kasprintf(GFP_KERNEL, "erofs,%s", sbi->opt.fsid);
+	if (!name)
+		return -ENOMEM;
+
+	volume = fscache_acquire_volume(name, NULL, NULL, 0);
+	if (IS_ERR_OR_NULL(volume)) {
+		erofs_err(sb, "failed to register volume for %s", name);
+		ret = volume ? PTR_ERR(volume) : -EOPNOTSUPP;
+		volume = NULL;
+	}
+
+	sbi->volume = volume;
+	kfree(name);
+	return ret;
+}
+
+void erofs_fscache_unregister_fs(struct super_block *sb)
+{
+	struct erofs_sb_info *sbi = EROFS_SB(sb);
+
+	fscache_relinquish_volume(sbi->volume, NULL, false);
+	sbi->volume = NULL;
+}
diff --git a/fs/erofs/inode.c b/fs/erofs/inode.c
index e8b37ba5e9ad07..bcc8335b46b3a2 100644
--- a/fs/erofs/inode.c
+++ b/fs/erofs/inode.c
@@ -8,11 +8,6 @@
 
 #include <trace/events/erofs.h>
 
-/*
- * if inode is successfully read, return its inode page (or sometimes
- * the inode payload page if it's an extended inode) in order to fill
- * inline data if possible.
- */
 static void *erofs_read_inode(struct erofs_buf *buf,
 			      struct inode *inode, unsigned int *ofs)
 {
@@ -297,6 +292,10 @@ static int erofs_fill_inode(struct inode *inode, int isdir)
 		goto out_unlock;
 	}
 	inode->i_mapping->a_ops = &erofs_raw_access_aops;
+#ifdef CONFIG_EROFS_FS_ONDEMAND
+	if (erofs_is_fscache_mode(inode->i_sb))
+		inode->i_mapping->a_ops = &erofs_fscache_access_aops;
+#endif
 
 out_unlock:
 	erofs_put_metabuf(&buf);
@@ -370,7 +369,7 @@ int erofs_getattr(struct user_namespace *mnt_userns, const struct path *path,
 	stat->attributes_mask |= (STATX_ATTR_COMPRESSED |
 				  STATX_ATTR_IMMUTABLE);
 
-	generic_fillattr(&init_user_ns, inode, stat);
+	generic_fillattr(mnt_userns, inode, stat);
 	return 0;
 }
 
diff --git a/fs/erofs/internal.h b/fs/erofs/internal.h
index 5298c4ee277d72..cfee49d33b95a0 100644
--- a/fs/erofs/internal.h
+++ b/fs/erofs/internal.h
@@ -49,6 +49,7 @@ typedef u32 erofs_blk_t;
 
 struct erofs_device_info {
 	char *path;
+	struct erofs_fscache *fscache;
 	struct block_device *bdev;
 	struct dax_device *dax_dev;
 	u64 dax_part_off;
@@ -74,6 +75,7 @@ struct erofs_mount_opts {
 	unsigned int max_sync_decompress_pages;
 #endif
 	unsigned int mount_opt;
+	char *fsid;
 };
 
 struct erofs_dev_context {
@@ -96,6 +98,11 @@ struct erofs_sb_lz4_info {
 	u16 max_pclusterblks;
 };
 
+struct erofs_fscache {
+	struct fscache_cookie *cookie;
+	struct inode *inode;
+};
+
 struct erofs_sb_info {
 	struct erofs_mount_opts opt;	/* options */
 #ifdef CONFIG_EROFS_FS_ZIP
@@ -146,6 +153,10 @@ struct erofs_sb_info {
 	/* sysfs support */
 	struct kobject s_kobj;		/* /sys/fs/erofs/<devname> */
 	struct completion s_kobj_unregister;
+
+	/* fscache support */
+	struct fscache_volume *volume;
+	struct erofs_fscache *s_fscache;
 };
 
 #define EROFS_SB(sb) ((struct erofs_sb_info *)(sb)->s_fs_info)
@@ -161,6 +172,11 @@ struct erofs_sb_info {
 #define set_opt(opt, option)	((opt)->mount_opt |= EROFS_MOUNT_##option)
 #define test_opt(opt, option)	((opt)->mount_opt & EROFS_MOUNT_##option)
 
+static inline bool erofs_is_fscache_mode(struct super_block *sb)
+{
+	return IS_ENABLED(CONFIG_EROFS_FS_ONDEMAND) && !sb->s_bdev;
+}
+
 enum {
 	EROFS_ZIP_CACHE_DISABLED,
 	EROFS_ZIP_CACHE_READAHEAD,
@@ -381,31 +397,6 @@ extern const struct super_operations erofs_sops;
 extern const struct address_space_operations erofs_raw_access_aops;
 extern const struct address_space_operations z_erofs_aops;
 
-/*
- * Logical to physical block mapping
- *
- * Different with other file systems, it is used for 2 access modes:
- *
- * 1) RAW access mode:
- *
- * Users pass a valid (m_lblk, m_lofs -- usually 0) pair,
- * and get the valid m_pblk, m_pofs and the longest m_len(in bytes).
- *
- * Note that m_lblk in the RAW access mode refers to the number of
- * the compressed ondisk block rather than the uncompressed
- * in-memory block for the compressed file.
- *
- * m_pofs equals to m_lofs except for the inline data page.
- *
- * 2) Normal access mode:
- *
- * If the inode is not compressed, it has no difference with
- * the RAW access mode. However, if the inode is compressed,
- * users should pass a valid (m_lblk, m_lofs) pair, and get
- * the needed m_pblk, m_pofs, m_len to get the compressed data
- * and the updated m_lblk, m_lofs which indicates the start
- * of the corresponding uncompressed data in the file.
- */
 enum {
 	BH_Encoded = BH_PrivateStart,
 	BH_FullMapped,
@@ -467,6 +458,7 @@ static inline int z_erofs_map_blocks_iter(struct inode *inode,
 #endif	/* !CONFIG_EROFS_FS_ZIP */
 
 struct erofs_map_dev {
+	struct erofs_fscache *m_fscache;
 	struct block_device *m_bdev;
 	struct dax_device *m_daxdev;
 	u64 m_dax_part_off;
@@ -486,6 +478,8 @@ void *erofs_read_metabuf(struct erofs_buf *buf, struct super_block *sb,
 int erofs_map_dev(struct super_block *sb, struct erofs_map_dev *dev);
 int erofs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
 		 u64 start, u64 len);
+int erofs_map_blocks(struct inode *inode,
+		     struct erofs_map_blocks *map, int flags);
 
 /* inode.c */
 static inline unsigned long erofs_inode_hash(erofs_nid_t nid)
@@ -509,7 +503,7 @@ int erofs_getattr(struct user_namespace *mnt_userns, const struct path *path,
 /* namei.c */
 extern const struct inode_operations erofs_dir_iops;
 
-int erofs_namei(struct inode *dir, struct qstr *name,
+int erofs_namei(struct inode *dir, const struct qstr *name,
 		erofs_nid_t *nid, unsigned int *d_type);
 
 /* dir.c */
@@ -611,6 +605,36 @@ static inline int z_erofs_load_lzma_config(struct super_block *sb,
 }
 #endif	/* !CONFIG_EROFS_FS_ZIP */
 
+/* fscache.c */
+#ifdef CONFIG_EROFS_FS_ONDEMAND
+int erofs_fscache_register_fs(struct super_block *sb);
+void erofs_fscache_unregister_fs(struct super_block *sb);
+
+int erofs_fscache_register_cookie(struct super_block *sb,
+				  struct erofs_fscache **fscache,
+				  char *name, bool need_inode);
+void erofs_fscache_unregister_cookie(struct erofs_fscache **fscache);
+
+extern const struct address_space_operations erofs_fscache_access_aops;
+#else
+static inline int erofs_fscache_register_fs(struct super_block *sb)
+{
+	return 0;
+}
+static inline void erofs_fscache_unregister_fs(struct super_block *sb) {}
+
+static inline int erofs_fscache_register_cookie(struct super_block *sb,
+						struct erofs_fscache **fscache,
+						char *name, bool need_inode)
+{
+	return -EOPNOTSUPP;
+}
+
+static inline void erofs_fscache_unregister_cookie(struct erofs_fscache **fscache)
+{
+}
+#endif
+
 #define EFSCORRUPTED    EUCLEAN         /* Filesystem is corrupted */
 
 #endif	/* __EROFS_INTERNAL_H */
diff --git a/fs/erofs/namei.c b/fs/erofs/namei.c
index 554efa363317fb..fd75506799c403 100644
--- a/fs/erofs/namei.c
+++ b/fs/erofs/namei.c
@@ -165,9 +165,8 @@ out:		/* free if the candidate is valid */
 	return candidate;
 }
 
-int erofs_namei(struct inode *dir,
-		struct qstr *name,
-		erofs_nid_t *nid, unsigned int *d_type)
+int erofs_namei(struct inode *dir, const struct qstr *name, erofs_nid_t *nid,
+		unsigned int *d_type)
 {
 	int ndirents;
 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
diff --git a/fs/erofs/super.c b/fs/erofs/super.c
index 0c4b41130c2fd2..c6f5fa4ab24453 100644
--- a/fs/erofs/super.c
+++ b/fs/erofs/super.c
@@ -13,6 +13,7 @@
 #include <linux/fs_context.h>
 #include <linux/fs_parser.h>
 #include <linux/dax.h>
+#include <linux/exportfs.h>
 #include "xattr.h"
 
 #define CREATE_TRACE_POINTS
@@ -219,7 +220,52 @@ static int erofs_load_compr_cfgs(struct super_block *sb,
 }
 #endif
 
-static int erofs_init_devices(struct super_block *sb,
+static int erofs_init_device(struct erofs_buf *buf, struct super_block *sb,
+			     struct erofs_device_info *dif, erofs_off_t *pos)
+{
+	struct erofs_sb_info *sbi = EROFS_SB(sb);
+	struct erofs_deviceslot *dis;
+	struct block_device *bdev;
+	void *ptr;
+	int ret;
+
+	ptr = erofs_read_metabuf(buf, sb, erofs_blknr(*pos), EROFS_KMAP);
+	if (IS_ERR(ptr))
+		return PTR_ERR(ptr);
+	dis = ptr + erofs_blkoff(*pos);
+
+	if (!dif->path) {
+		if (!dis->tag[0]) {
+			erofs_err(sb, "empty device tag @ pos %llu", *pos);
+			return -EINVAL;
+		}
+		dif->path = kmemdup_nul(dis->tag, sizeof(dis->tag), GFP_KERNEL);
+		if (!dif->path)
+			return -ENOMEM;
+	}
+
+	if (erofs_is_fscache_mode(sb)) {
+		ret = erofs_fscache_register_cookie(sb, &dif->fscache,
+				dif->path, false);
+		if (ret)
+			return ret;
+	} else {
+		bdev = blkdev_get_by_path(dif->path, FMODE_READ | FMODE_EXCL,
+					  sb->s_type);
+		if (IS_ERR(bdev))
+			return PTR_ERR(bdev);
+		dif->bdev = bdev;
+		dif->dax_dev = fs_dax_get_by_bdev(bdev, &dif->dax_part_off);
+	}
+
+	dif->blocks = le32_to_cpu(dis->blocks);
+	dif->mapped_blkaddr = le32_to_cpu(dis->mapped_blkaddr);
+	sbi->total_blocks += dif->blocks;
+	*pos += EROFS_DEVT_SLOT_SIZE;
+	return 0;
+}
+
+static int erofs_scan_devices(struct super_block *sb,
 			      struct erofs_super_block *dsb)
 {
 	struct erofs_sb_info *sbi = EROFS_SB(sb);
@@ -227,8 +273,6 @@ static int erofs_init_devices(struct super_block *sb,
 	erofs_off_t pos;
 	struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
 	struct erofs_device_info *dif;
-	struct erofs_deviceslot *dis;
-	void *ptr;
 	int id, err = 0;
 
 	sbi->total_blocks = sbi->primarydevice_blocks;
@@ -237,7 +281,8 @@ static int erofs_init_devices(struct super_block *sb,
 	else
 		ondisk_extradevs = le16_to_cpu(dsb->extra_devices);
 
-	if (ondisk_extradevs != sbi->devs->extra_devices) {
+	if (sbi->devs->extra_devices &&
+	    ondisk_extradevs != sbi->devs->extra_devices) {
 		erofs_err(sb, "extra devices don't match (ondisk %u, given %u)",
 			  ondisk_extradevs, sbi->devs->extra_devices);
 		return -EINVAL;
@@ -248,30 +293,31 @@ static int erofs_init_devices(struct super_block *sb,
 	sbi->device_id_mask = roundup_pow_of_two(ondisk_extradevs + 1) - 1;
 	pos = le16_to_cpu(dsb->devt_slotoff) * EROFS_DEVT_SLOT_SIZE;
 	down_read(&sbi->devs->rwsem);
-	idr_for_each_entry(&sbi->devs->tree, dif, id) {
-		struct block_device *bdev;
-
-		ptr = erofs_read_metabuf(&buf, sb, erofs_blknr(pos),
-					 EROFS_KMAP);
-		if (IS_ERR(ptr)) {
-			err = PTR_ERR(ptr);
-			break;
+	if (sbi->devs->extra_devices) {
+		idr_for_each_entry(&sbi->devs->tree, dif, id) {
+			err = erofs_init_device(&buf, sb, dif, &pos);
+			if (err)
+				break;
 		}
-		dis = ptr + erofs_blkoff(pos);
-
-		bdev = blkdev_get_by_path(dif->path,
-					  FMODE_READ | FMODE_EXCL,
-					  sb->s_type);
-		if (IS_ERR(bdev)) {
-			err = PTR_ERR(bdev);
-			break;
+	} else {
+		for (id = 0; id < ondisk_extradevs; id++) {
+			dif = kzalloc(sizeof(*dif), GFP_KERNEL);
+			if (!dif) {
+				err = -ENOMEM;
+				break;
+			}
+
+			err = idr_alloc(&sbi->devs->tree, dif, 0, 0, GFP_KERNEL);
+			if (err < 0) {
+				kfree(dif);
+				break;
+			}
+			++sbi->devs->extra_devices;
+
+			err = erofs_init_device(&buf, sb, dif, &pos);
+			if (err)
+				break;
 		}
-		dif->bdev = bdev;
-		dif->dax_dev = fs_dax_get_by_bdev(bdev, &dif->dax_part_off);
-		dif->blocks = le32_to_cpu(dis->blocks);
-		dif->mapped_blkaddr = le32_to_cpu(dis->mapped_blkaddr);
-		sbi->total_blocks += dif->blocks;
-		pos += EROFS_DEVT_SLOT_SIZE;
 	}
 	up_read(&sbi->devs->rwsem);
 	erofs_put_metabuf(&buf);
@@ -358,10 +404,12 @@ static int erofs_read_superblock(struct super_block *sb)
 		goto out;
 
 	/* handle multiple devices */
-	ret = erofs_init_devices(sb, dsb);
+	ret = erofs_scan_devices(sb, dsb);
 
 	if (erofs_sb_has_ztailpacking(sbi))
 		erofs_info(sb, "EXPERIMENTAL compressed inline data feature in use. Use at your own risk!");
+	if (erofs_is_fscache_mode(sb))
+		erofs_info(sb, "EXPERIMENTAL fscache-based on-demand read feature in use. Use at your own risk!");
 out:
 	erofs_put_metabuf(&buf);
 	return ret;
@@ -390,6 +438,7 @@ enum {
 	Opt_dax,
 	Opt_dax_enum,
 	Opt_device,
+	Opt_fsid,
 	Opt_err
 };
 
@@ -414,6 +463,7 @@ static const struct fs_parameter_spec erofs_fs_parameters[] = {
 	fsparam_flag("dax",             Opt_dax),
 	fsparam_enum("dax",		Opt_dax_enum, erofs_dax_param_enums),
 	fsparam_string("device",	Opt_device),
+	fsparam_string("fsid",		Opt_fsid),
 	{}
 };
 
@@ -509,6 +559,16 @@ static int erofs_fc_parse_param(struct fs_context *fc,
 		}
 		++ctx->devs->extra_devices;
 		break;
+	case Opt_fsid:
+#ifdef CONFIG_EROFS_FS_ONDEMAND
+		kfree(ctx->opt.fsid);
+		ctx->opt.fsid = kstrdup(param->string, GFP_KERNEL);
+		if (!ctx->opt.fsid)
+			return -ENOMEM;
+#else
+		errorfc(fc, "fsid option not supported");
+#endif
+		break;
 	default:
 		return -ENOPARAM;
 	}
@@ -577,6 +637,44 @@ static int erofs_init_managed_cache(struct super_block *sb)
 static int erofs_init_managed_cache(struct super_block *sb) { return 0; }
 #endif
 
+static struct inode *erofs_nfs_get_inode(struct super_block *sb,
+					 u64 ino, u32 generation)
+{
+	return erofs_iget(sb, ino, false);
+}
+
+static struct dentry *erofs_fh_to_dentry(struct super_block *sb,
+		struct fid *fid, int fh_len, int fh_type)
+{
+	return generic_fh_to_dentry(sb, fid, fh_len, fh_type,
+				    erofs_nfs_get_inode);
+}
+
+static struct dentry *erofs_fh_to_parent(struct super_block *sb,
+		struct fid *fid, int fh_len, int fh_type)
+{
+	return generic_fh_to_parent(sb, fid, fh_len, fh_type,
+				    erofs_nfs_get_inode);
+}
+
+static struct dentry *erofs_get_parent(struct dentry *child)
+{
+	erofs_nid_t nid;
+	unsigned int d_type;
+	int err;
+
+	err = erofs_namei(d_inode(child), &dotdot_name, &nid, &d_type);
+	if (err)
+		return ERR_PTR(err);
+	return d_obtain_alias(erofs_iget(child->d_sb, nid, d_type == FT_DIR));
+}
+
+static const struct export_operations erofs_export_ops = {
+	.fh_to_dentry = erofs_fh_to_dentry,
+	.fh_to_parent = erofs_fh_to_parent,
+	.get_parent = erofs_get_parent,
+};
+
 static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
 {
 	struct inode *inode;
@@ -585,11 +683,9 @@ static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
 	int err;
 
 	sb->s_magic = EROFS_SUPER_MAGIC;
-
-	if (!sb_set_blocksize(sb, EROFS_BLKSIZ)) {
-		erofs_err(sb, "failed to set erofs blksize");
-		return -EINVAL;
-	}
+	sb->s_flags |= SB_RDONLY | SB_NOATIME;
+	sb->s_maxbytes = MAX_LFS_FILESIZE;
+	sb->s_op = &erofs_sops;
 
 	sbi = kzalloc(sizeof(*sbi), GFP_KERNEL);
 	if (!sbi)
@@ -597,10 +693,36 @@ static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
 
 	sb->s_fs_info = sbi;
 	sbi->opt = ctx->opt;
-	sbi->dax_dev = fs_dax_get_by_bdev(sb->s_bdev, &sbi->dax_part_off);
+	ctx->opt.fsid = NULL;
 	sbi->devs = ctx->devs;
 	ctx->devs = NULL;
 
+	if (erofs_is_fscache_mode(sb)) {
+		sb->s_blocksize = EROFS_BLKSIZ;
+		sb->s_blocksize_bits = LOG_BLOCK_SIZE;
+
+		err = erofs_fscache_register_fs(sb);
+		if (err)
+			return err;
+
+		err = erofs_fscache_register_cookie(sb, &sbi->s_fscache,
+						    sbi->opt.fsid, true);
+		if (err)
+			return err;
+
+		err = super_setup_bdi(sb);
+		if (err)
+			return err;
+	} else {
+		if (!sb_set_blocksize(sb, EROFS_BLKSIZ)) {
+			erofs_err(sb, "failed to set erofs blksize");
+			return -EINVAL;
+		}
+
+		sbi->dax_dev = fs_dax_get_by_bdev(sb->s_bdev,
+						  &sbi->dax_part_off);
+	}
+
 	err = erofs_read_superblock(sb);
 	if (err)
 		return err;
@@ -613,12 +735,10 @@ static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
 			clear_opt(&sbi->opt, DAX_ALWAYS);
 		}
 	}
-	sb->s_flags |= SB_RDONLY | SB_NOATIME;
-	sb->s_maxbytes = MAX_LFS_FILESIZE;
-	sb->s_time_gran = 1;
 
-	sb->s_op = &erofs_sops;
+	sb->s_time_gran = 1;
 	sb->s_xattr = erofs_xattr_handlers;
+	sb->s_export_op = &erofs_export_ops;
 
 	if (test_opt(&sbi->opt, POSIX_ACL))
 		sb->s_flags |= SB_POSIXACL;
@@ -661,6 +781,11 @@ static int erofs_fc_fill_super(struct super_block *sb, struct fs_context *fc)
 
 static int erofs_fc_get_tree(struct fs_context *fc)
 {
+	struct erofs_fs_context *ctx = fc->fs_private;
+
+	if (IS_ENABLED(CONFIG_EROFS_FS_ONDEMAND) && ctx->opt.fsid)
+		return get_tree_nodev(fc, erofs_fc_fill_super);
+
 	return get_tree_bdev(fc, erofs_fc_fill_super);
 }
 
@@ -690,6 +815,7 @@ static int erofs_release_device_info(int id, void *ptr, void *data)
 	fs_put_dax(dif->dax_dev);
 	if (dif->bdev)
 		blkdev_put(dif->bdev, FMODE_READ | FMODE_EXCL);
+	erofs_fscache_unregister_cookie(&dif->fscache);
 	kfree(dif->path);
 	kfree(dif);
 	return 0;
@@ -709,6 +835,7 @@ static void erofs_fc_free(struct fs_context *fc)
 	struct erofs_fs_context *ctx = fc->fs_private;
 
 	erofs_free_dev_context(ctx->devs);
+	kfree(ctx->opt.fsid);
 	kfree(ctx);
 }
 
@@ -749,7 +876,10 @@ static void erofs_kill_sb(struct super_block *sb)
 
 	WARN_ON(sb->s_magic != EROFS_SUPER_MAGIC);
 
-	kill_block_super(sb);
+	if (erofs_is_fscache_mode(sb))
+		generic_shutdown_super(sb);
+	else
+		kill_block_super(sb);
 
 	sbi = EROFS_SB(sb);
 	if (!sbi)
@@ -757,6 +887,9 @@ static void erofs_kill_sb(struct super_block *sb)
 
 	erofs_free_dev_context(sbi->devs);
 	fs_put_dax(sbi->dax_dev);
+	erofs_fscache_unregister_cookie(&sbi->s_fscache);
+	erofs_fscache_unregister_fs(sb);
+	kfree(sbi->opt.fsid);
 	kfree(sbi);
 	sb->s_fs_info = NULL;
 }
@@ -774,6 +907,7 @@ static void erofs_put_super(struct super_block *sb)
 	iput(sbi->managed_cache);
 	sbi->managed_cache = NULL;
 #endif
+	erofs_fscache_unregister_cookie(&sbi->s_fscache);
 }
 
 static struct file_system_type erofs_fs_type = {
@@ -781,7 +915,7 @@ static struct file_system_type erofs_fs_type = {
 	.name           = "erofs",
 	.init_fs_context = erofs_init_fs_context,
 	.kill_sb        = erofs_kill_sb,
-	.fs_flags       = FS_REQUIRES_DEV,
+	.fs_flags       = FS_REQUIRES_DEV | FS_ALLOW_IDMAP,
 };
 MODULE_ALIAS_FS("erofs");
 
@@ -857,7 +991,10 @@ static int erofs_statfs(struct dentry *dentry, struct kstatfs *buf)
 {
 	struct super_block *sb = dentry->d_sb;
 	struct erofs_sb_info *sbi = EROFS_SB(sb);
-	u64 id = huge_encode_dev(sb->s_bdev->bd_dev);
+	u64 id = 0;
+
+	if (!erofs_is_fscache_mode(sb))
+		id = huge_encode_dev(sb->s_bdev->bd_dev);
 
 	buf->f_type = sb->s_magic;
 	buf->f_bsize = EROFS_BLKSIZ;
@@ -902,6 +1039,10 @@ static int erofs_show_options(struct seq_file *seq, struct dentry *root)
 		seq_puts(seq, ",dax=always");
 	if (test_opt(opt, DAX_NEVER))
 		seq_puts(seq, ",dax=never");
+#ifdef CONFIG_EROFS_FS_ONDEMAND
+	if (opt->fsid)
+		seq_printf(seq, ",fsid=%s", opt->fsid);
+#endif
 	return 0;
 }
 
diff --git a/fs/erofs/sysfs.c b/fs/erofs/sysfs.c
index f3babf1e660832..c1383e508bbecf 100644
--- a/fs/erofs/sysfs.c
+++ b/fs/erofs/sysfs.c
@@ -205,8 +205,8 @@ int erofs_register_sysfs(struct super_block *sb)
 
 	sbi->s_kobj.kset = &erofs_root;
 	init_completion(&sbi->s_kobj_unregister);
-	err = kobject_init_and_add(&sbi->s_kobj, &erofs_sb_ktype, NULL,
-				   "%s", sb->s_id);
+	err = kobject_init_and_add(&sbi->s_kobj, &erofs_sb_ktype, NULL, "%s",
+			erofs_is_fscache_mode(sb) ? sbi->opt.fsid : sb->s_id);
 	if (err)
 		goto put_sb_kobj;
 	return 0;
diff --git a/include/linux/fscache.h b/include/linux/fscache.h
index e25539072463ba..72585c9729a2c4 100644
--- a/include/linux/fscache.h
+++ b/include/linux/fscache.h
@@ -39,6 +39,7 @@ struct fscache_cookie;
 #define FSCACHE_ADV_SINGLE_CHUNK	0x01 /* The object is a single chunk of data */
 #define FSCACHE_ADV_WRITE_CACHE		0x00 /* Do cache if written to locally */
 #define FSCACHE_ADV_WRITE_NOCACHE	0x02 /* Don't cache if written to locally */
+#define FSCACHE_ADV_WANT_CACHE_SIZE	0x04 /* Retrieve cache size at runtime */
 
 #define FSCACHE_INVAL_DIO_WRITE		0x01 /* Invalidate due to DIO write */
 
diff --git a/include/linux/netfs.h b/include/linux/netfs.h
index 0c33b715cbfd26..80b7728277b1b1 100644
--- a/include/linux/netfs.h
+++ b/include/linux/netfs.h
@@ -159,6 +159,7 @@ struct netfs_io_subrequest {
 #define NETFS_SREQ_SHORT_IO		2	/* Set if the I/O was short */
 #define NETFS_SREQ_SEEK_DATA_READ	3	/* Set if ->read() should SEEK_DATA first */
 #define NETFS_SREQ_NO_PROGRESS		4	/* Set if we didn't manage to read any data */
+#define NETFS_SREQ_ONDEMAND		5	/* Set if it's from on-demand read mode */
 };
 
 enum netfs_io_origin {
diff --git a/include/trace/events/cachefiles.h b/include/trace/events/cachefiles.h
index 311c14a20e7036..d8d4d73fe7b6ad 100644
--- a/include/trace/events/cachefiles.h
+++ b/include/trace/events/cachefiles.h
@@ -31,6 +31,8 @@ enum cachefiles_obj_ref_trace {
 	cachefiles_obj_see_lookup_failed,
 	cachefiles_obj_see_withdraw_cookie,
 	cachefiles_obj_see_withdrawal,
+	cachefiles_obj_get_ondemand_fd,
+	cachefiles_obj_put_ondemand_fd,
 };
 
 enum fscache_why_object_killed {
@@ -671,6 +673,180 @@ TRACE_EVENT(cachefiles_io_error,
 		      __entry->error)
 	    );
 
+TRACE_EVENT(cachefiles_ondemand_open,
+	    TP_PROTO(struct cachefiles_object *obj, struct cachefiles_msg *msg,
+		     struct cachefiles_open *load),
+
+	    TP_ARGS(obj, msg, load),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj		)
+		    __field(unsigned int,	msg_id		)
+		    __field(unsigned int,	object_id	)
+		    __field(unsigned int,	fd		)
+		    __field(unsigned int,	flags		)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->msg_id	= msg->msg_id;
+		    __entry->object_id	= msg->object_id;
+		    __entry->fd		= load->fd;
+		    __entry->flags	= load->flags;
+			   ),
+
+	    TP_printk("o=%08x mid=%x oid=%x fd=%d f=%x",
+		      __entry->obj,
+		      __entry->msg_id,
+		      __entry->object_id,
+		      __entry->fd,
+		      __entry->flags)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_copen,
+	    TP_PROTO(struct cachefiles_object *obj, unsigned int msg_id,
+		     long len),
+
+	    TP_ARGS(obj, msg_id, len),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj	)
+		    __field(unsigned int,	msg_id	)
+		    __field(long,		len	)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->msg_id	= msg_id;
+		    __entry->len	= len;
+			   ),
+
+	    TP_printk("o=%08x mid=%x l=%lx",
+		      __entry->obj,
+		      __entry->msg_id,
+		      __entry->len)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_close,
+	    TP_PROTO(struct cachefiles_object *obj, struct cachefiles_msg *msg),
+
+	    TP_ARGS(obj, msg),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj		)
+		    __field(unsigned int,	msg_id		)
+		    __field(unsigned int,	object_id	)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->msg_id	= msg->msg_id;
+		    __entry->object_id	= msg->object_id;
+			   ),
+
+	    TP_printk("o=%08x mid=%x oid=%x",
+		      __entry->obj,
+		      __entry->msg_id,
+		      __entry->object_id)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_read,
+	    TP_PROTO(struct cachefiles_object *obj, struct cachefiles_msg *msg,
+		     struct cachefiles_read *load),
+
+	    TP_ARGS(obj, msg, load),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj		)
+		    __field(unsigned int,	msg_id		)
+		    __field(unsigned int,	object_id	)
+		    __field(loff_t,		start		)
+		    __field(size_t,		len		)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->msg_id	= msg->msg_id;
+		    __entry->object_id	= msg->object_id;
+		    __entry->start	= load->off;
+		    __entry->len	= load->len;
+			   ),
+
+	    TP_printk("o=%08x mid=%x oid=%x s=%llx l=%zx",
+		      __entry->obj,
+		      __entry->msg_id,
+		      __entry->object_id,
+		      __entry->start,
+		      __entry->len)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_cread,
+	    TP_PROTO(struct cachefiles_object *obj, unsigned int msg_id),
+
+	    TP_ARGS(obj, msg_id),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj	)
+		    __field(unsigned int,	msg_id	)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->msg_id	= msg_id;
+			   ),
+
+	    TP_printk("o=%08x mid=%x",
+		      __entry->obj,
+		      __entry->msg_id)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_fd_write,
+	    TP_PROTO(struct cachefiles_object *obj, struct inode *backer,
+		     loff_t start, size_t len),
+
+	    TP_ARGS(obj, backer, start, len),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj	)
+		    __field(unsigned int,	backer	)
+		    __field(loff_t,		start	)
+		    __field(size_t,		len	)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->backer	= backer->i_ino;
+		    __entry->start	= start;
+		    __entry->len	= len;
+			   ),
+
+	    TP_printk("o=%08x iB=%x s=%llx l=%zx",
+		      __entry->obj,
+		      __entry->backer,
+		      __entry->start,
+		      __entry->len)
+	    );
+
+TRACE_EVENT(cachefiles_ondemand_fd_release,
+	    TP_PROTO(struct cachefiles_object *obj, int object_id),
+
+	    TP_ARGS(obj, object_id),
+
+	    TP_STRUCT__entry(
+		    __field(unsigned int,	obj		)
+		    __field(unsigned int,	object_id	)
+			     ),
+
+	    TP_fast_assign(
+		    __entry->obj	= obj ? obj->debug_id : 0;
+		    __entry->object_id	= object_id;
+			   ),
+
+	    TP_printk("o=%08x oid=%x",
+		      __entry->obj,
+		      __entry->object_id)
+	    );
+
 #endif /* _TRACE_CACHEFILES_H */
 
 /* This part must be outside protection */
diff --git a/include/uapi/linux/cachefiles.h b/include/uapi/linux/cachefiles.h
new file mode 100644
index 00000000000000..78caa73e534341
--- /dev/null
+++ b/include/uapi/linux/cachefiles.h
@@ -0,0 +1,68 @@
+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
+#ifndef _LINUX_CACHEFILES_H
+#define _LINUX_CACHEFILES_H
+
+#include <linux/types.h>
+#include <linux/ioctl.h>
+
+/*
+ * Fscache ensures that the maximum length of cookie key is 255. The volume key
+ * is controlled by netfs, and generally no bigger than 255.
+ */
+#define CACHEFILES_MSG_MAX_SIZE	1024
+
+enum cachefiles_opcode {
+	CACHEFILES_OP_OPEN,
+	CACHEFILES_OP_CLOSE,
+	CACHEFILES_OP_READ,
+};
+
+/*
+ * Message Header
+ *
+ * @msg_id	a unique ID identifying this message
+ * @opcode	message type, CACHEFILE_OP_*
+ * @len		message length, including message header and following data
+ * @object_id	a unique ID identifying a cache file
+ * @data	message type specific payload
+ */
+struct cachefiles_msg {
+	__u32 msg_id;
+	__u32 opcode;
+	__u32 len;
+	__u32 object_id;
+	__u8  data[];
+};
+
+/*
+ * @data contains the volume_key followed directly by the cookie_key. volume_key
+ * is a NUL-terminated string; @volume_key_size indicates the size of the volume
+ * key in bytes. cookie_key is binary data, which is netfs specific;
+ * @cookie_key_size indicates the size of the cookie key in bytes.
+ *
+ * @fd identifies an anon_fd referring to the cache file.
+ */
+struct cachefiles_open {
+	__u32 volume_key_size;
+	__u32 cookie_key_size;
+	__u32 fd;
+	__u32 flags;
+	__u8  data[];
+};
+
+/*
+ * @off		indicates the starting offset of the requested file range
+ * @len		indicates the length of the requested file range
+ */
+struct cachefiles_read {
+	__u64 off;
+	__u64 len;
+};
+
+/*
+ * Reply for READ request
+ * @arg for this ioctl is the @id field of READ request.
+ */
+#define CACHEFILES_IOC_READ_COMPLETE	_IOW(0x98, 1, int)
+
+#endif