Merge tag 'vfs-6.14-rc1.netfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs

Pull vfs netfs updates from Christian Brauner:
 "This contains read performance improvements and support for monolithic
  single-blob objects that have to be read/written as such (e.g. AFS
  directory contents). The implementation of the two parts is interwoven
  as each makes the other possible.

   - Read performance improvements

     The read performance improvements are intended to speed up some
     loss of performance detected in cifs and to a lesser extend in afs.

     The problem is that we queue too many work items during the
     collection of read results: each individual subrequest is collected
     by its own work item, and then they have to interact with each
     other when a series of subrequests don't exactly align with the
     pattern of folios that are being read by the overall request.

     Whilst the processing of the pages covered by individual
     subrequests as they complete potentially allows folios to be woken
     in parallel and with minimum delay, it can shuffle wakeups for
     sequential reads out of order - and that is the most common I/O
     pattern.

     The final assessment and cleanup of an operation is then held up
     until the last I/O completes - and for a synchronous sequential
     operation, this means the bouncing around of work items just adds
     latency.

     Two changes have been made to make this work:

     (1) All collection is now done in a single "work item" that works
         progressively through the subrequests as they complete (and
         also dispatches retries as necessary).

     (2) For readahead and AIO, this work item be done on a workqueue
         and can run in parallel with the ultimate consumer of the data;
         for synchronous direct or unbuffered reads, the collection is
         run in the application thread and not offloaded.

     Functions such as smb2_readv_callback() then just tell netfslib
     that the subrequest has terminated; netfslib does a minimal bit of
     processing on the spot - stat counting and tracing mostly - and
     then queues/wakes up the worker. This simplifies the logic as the
     collector just walks sequentially through the subrequests as they
     complete and walks through the folios, if buffered, unlocking them
     as it goes. It also keeps to a minimum the amount of latency
     injected into the filesystem's low-level I/O handling

     The way netfs supports filesystems using the deprecated
     PG_private_2 flag is changed: folios are flagged and added to a
     write request as they complete and that takes care of scheduling
     the writes to the cache. The originating read request can then just
     unlock the pages whatever happens.

   - Single-blob object support

     Single-blob objects are files for which the content of the file
     must be read from or written to the server in a single operation
     because reading them in parts may yield inconsistent results. AFS
     directories are an example of this as there exists the possibility
     that the contents are generated on the fly and would differ between
     reads or might change due to third party interference.

     Such objects will be written to and retrieved from the cache if one
     is present, though we allow/may need to propose multiple
     subrequests to do so. The important part is that read from/write to
     the *server* is monolithic.

     Single blob reading is, for the moment, fully synchronous and does
     result collection in the application thread and, also for the
     moment, the API is supplied the buffer in the form of a folio_queue
     chain rather than using the pagecache.

   - Related afs changes

     This series makes a number of changes to the kafs filesystem,
     primarily in the area of directory handling:

      - AFS's FetchData RPC reply processing is made partially
        asynchronous which allows the netfs_io_request's outstanding
        operation counter to be removed as part of reducing the
        collection to a single work item.

      - Directory and symlink reading are plumbed through netfslib using
        the single-blob object API and are now cacheable with fscache.
        This also allows the afs_read struct to be eliminated and
        netfs_io_subrequest to be used directly instead.

      - Directory and symlink content are now stored in a folio_queue
        buffer rather than in the pagecache. This means we don't require
        the RCU read lock and xarray iteration to access it, and folios
        won't randomly disappear under us because the VM wants them
        back.

      - The vnode operation lock is changed from a mutex struct to a
        private lock implementation. The problem is that the lock now
        needs to be dropped in a separate thread and mutexes don't
        permit that.

      - When a new directory or symlink is created, we now initialise it
        locally and mark it valid rather than downloading it (we know
        what it's likely to look like).

      - We now use the in-directory hashtable to reduce the number of
        entries we need to scan when doing a lookup. The edit routines
        have to maintain the hash chains.

      - Cancellation (e.g. by signal) of an async call after the
        rxrpc_call has been set up is now offloaded to the worker thread
        as there will be a notification from rxrpc upon completion. This
        avoids a double cleanup.

   - A "rolling buffer" implementation is created to abstract out the
     two separate folio_queue chaining implementations I had (one for
     read and one for write).

   - Functions are provided to create/extend a buffer in a folio_queue
     chain and tear it down again.

     This is used to handle AFS directories, but could also be used to
     create bounce buffers for content crypto and transport crypto.

   - The was_async argument is dropped from netfs_read_subreq_terminated()

     Instead we wake the read collection work item by either queuing it
     or waking up the app thread.

   - We don't need to use BH-excluding locks when communicating between
     the issuing thread and the collection thread as neither of them now
     run in BH context.

   - Also included are a number of new tracepoints; a split of the
     netfslib write collection code to put retrying into its own file
     (it gets more complicated with content encryption).

   - There are also some minor fixes AFS included, including fixing the
     AFS directory format struct layout, reducing some directory
     over-invalidation and making afs_mkdir() translate EEXIST to
     ENOTEMPY (which is not available on all systems the servers
     support).

   - Finally, there's a patch to try and detect entry into the folio
     unlock function with no folio_queue structs in the buffer (which
     isn't allowed in the cases that can get there).

     This is a debugging patch, but should be minimal overhead"

* tag 'vfs-6.14-rc1.netfs' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (31 commits)
  netfs: Report on NULL folioq in netfs_writeback_unlock_folios()
  afs: Add a tracepoint for afs_read_receive()
  afs: Locally initialise the contents of a new symlink on creation
  afs: Use the contained hashtable to search a directory
  afs: Make afs_mkdir() locally initialise a new directory's content
  netfs: Change the read result collector to only use one work item
  afs: Make {Y,}FS.FetchData an asynchronous operation
  afs: Fix cleanup of immediately failed async calls
  afs: Eliminate afs_read
  afs: Use netfslib for symlinks, allowing them to be cached
  afs: Use netfslib for directories
  afs: Make afs_init_request() get a key if not given a file
  netfs: Add support for caching single monolithic objects such as AFS dirs
  netfs: Add functions to build/clean a buffer in a folio_queue
  afs: Add more tracepoints to do with tracking validity
  cachefiles: Add auxiliary data trace
  cachefiles: Add some subrequest tracepoints
  netfs: Remove some extraneous directory invalidations
  afs: Fix directory format encoding struct
  afs: Fix EEXIST error returned from afs_rmdir() to be ENOTEMPTY
  ...

17 files changed, 1863 insertions, 1108 deletions

diff --git a/fs/netfs/Makefile b/fs/netfs/Makefile
index d08b0bfb6756..b43188d64bd8 100644
--- a/fs/netfs/Makefile
+++ b/fs/netfs/Makefile
@@ -13,8 +13,11 @@ netfs-y := \
 	read_collect.o \
 	read_pgpriv2.o \
 	read_retry.o \
+	read_single.o \
+	rolling_buffer.o \
 	write_collect.o \
-	write_issue.o
+	write_issue.o \
+	write_retry.o
 
 netfs-$(CONFIG_NETFS_STATS) += stats.o
 
diff --git a/fs/netfs/buffered_read.c b/fs/netfs/buffered_read.c
index 4dc9b8286355..f761d44b3436 100644
--- a/fs/netfs/buffered_read.c
+++ b/fs/netfs/buffered_read.c
@@ -64,37 +64,6 @@ static int netfs_begin_cache_read(struct netfs_io_request *rreq, struct netfs_in
 }
 
 /*
- * Decant the list of folios to read into a rolling buffer.
- */
-static size_t netfs_load_buffer_from_ra(struct netfs_io_request *rreq,
-					struct folio_queue *folioq,
-					struct folio_batch *put_batch)
-{
-	unsigned int order, nr;
-	size_t size = 0;
-
-	nr = __readahead_batch(rreq->ractl, (struct page **)folioq->vec.folios,
-			       ARRAY_SIZE(folioq->vec.folios));
-	folioq->vec.nr = nr;
-	for (int i = 0; i < nr; i++) {
-		struct folio *folio = folioq_folio(folioq, i);
-
-		trace_netfs_folio(folio, netfs_folio_trace_read);
-		order = folio_order(folio);
-		folioq->orders[i] = order;
-		size += PAGE_SIZE << order;
-
-		if (!folio_batch_add(put_batch, folio))
-			folio_batch_release(put_batch);
-	}
-
-	for (int i = nr; i < folioq_nr_slots(folioq); i++)
-		folioq_clear(folioq, i);
-
-	return size;
-}
-
-/*
  * netfs_prepare_read_iterator - Prepare the subreq iterator for I/O
  * @subreq: The subrequest to be set up
  *
@@ -128,19 +97,12 @@ static ssize_t netfs_prepare_read_iterator(struct netfs_io_subrequest *subreq)
 
 		folio_batch_init(&put_batch);
 		while (rreq->submitted < subreq->start + rsize) {
-			struct folio_queue *tail = rreq->buffer_tail, *new;
-			size_t added;
-
-			new = kmalloc(sizeof(*new), GFP_NOFS);
-			if (!new)
-				return -ENOMEM;
-			netfs_stat(&netfs_n_folioq);
-			folioq_init(new);
-			new->prev = tail;
-			tail->next = new;
-			rreq->buffer_tail = new;
-			added = netfs_load_buffer_from_ra(rreq, new, &put_batch);
-			rreq->iter.count += added;
+			ssize_t added;
+
+			added = rolling_buffer_load_from_ra(&rreq->buffer, rreq->ractl,
+							    &put_batch);
+			if (added < 0)
+				return added;
 			rreq->submitted += added;
 		}
 		folio_batch_release(&put_batch);
@@ -148,7 +110,7 @@ static ssize_t netfs_prepare_read_iterator(struct netfs_io_subrequest *subreq)
 
 	subreq->len = rsize;
 	if (unlikely(rreq->io_streams[0].sreq_max_segs)) {
-		size_t limit = netfs_limit_iter(&rreq->iter, 0, rsize,
+		size_t limit = netfs_limit_iter(&rreq->buffer.iter, 0, rsize,
 						rreq->io_streams[0].sreq_max_segs);
 
 		if (limit < rsize) {
@@ -157,20 +119,10 @@ static ssize_t netfs_prepare_read_iterator(struct netfs_io_subrequest *subreq)
 		}
 	}
 
-	subreq->io_iter	= rreq->iter;
-
-	if (iov_iter_is_folioq(&subreq->io_iter)) {
-		if (subreq->io_iter.folioq_slot >= folioq_nr_slots(subreq->io_iter.folioq)) {
-			subreq->io_iter.folioq = subreq->io_iter.folioq->next;
-			subreq->io_iter.folioq_slot = 0;
-		}
-		subreq->curr_folioq = (struct folio_queue *)subreq->io_iter.folioq;
-		subreq->curr_folioq_slot = subreq->io_iter.folioq_slot;
-		subreq->curr_folio_order = subreq->curr_folioq->orders[subreq->curr_folioq_slot];
-	}
+	subreq->io_iter	= rreq->buffer.iter;
 
 	iov_iter_truncate(&subreq->io_iter, subreq->len);
-	iov_iter_advance(&rreq->iter, subreq->len);
+	rolling_buffer_advance(&rreq->buffer, subreq->len);
 	return subreq->len;
 }
 
@@ -179,25 +131,14 @@ static enum netfs_io_source netfs_cache_prepare_read(struct netfs_io_request *rr
 						     loff_t i_size)
 {
 	struct netfs_cache_resources *cres = &rreq->cache_resources;
+	enum netfs_io_source source;
 
 	if (!cres->ops)
 		return NETFS_DOWNLOAD_FROM_SERVER;
-	return cres->ops->prepare_read(subreq, i_size);
-}
-
-static void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error,
-					bool was_async)
-{
-	struct netfs_io_subrequest *subreq = priv;
-
-	if (transferred_or_error < 0) {
-		netfs_read_subreq_terminated(subreq, transferred_or_error, was_async);
-		return;
-	}
+	source = cres->ops->prepare_read(subreq, i_size);
+	trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
+	return source;
 
-	if (transferred_or_error > 0)
-		subreq->transferred += transferred_or_error;
-	netfs_read_subreq_terminated(subreq, 0, was_async);
 }
 
 /*
@@ -214,6 +155,47 @@ static void netfs_read_cache_to_pagecache(struct netfs_io_request *rreq,
 			netfs_cache_read_terminated, subreq);
 }
 
+static void netfs_issue_read(struct netfs_io_request *rreq,
+			     struct netfs_io_subrequest *subreq)
+{
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+
+	__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
+
+	/* We add to the end of the list whilst the collector may be walking
+	 * the list.  The collector only goes nextwards and uses the lock to
+	 * remove entries off of the front.
+	 */
+	spin_lock(&rreq->lock);
+	list_add_tail(&subreq->rreq_link, &stream->subrequests);
+	if (list_is_first(&subreq->rreq_link, &stream->subrequests)) {
+		stream->front = subreq;
+		if (!stream->active) {
+			stream->collected_to = stream->front->start;
+			/* Store list pointers before active flag */
+			smp_store_release(&stream->active, true);
+		}
+	}
+
+	spin_unlock(&rreq->lock);
+
+	switch (subreq->source) {
+	case NETFS_DOWNLOAD_FROM_SERVER:
+		rreq->netfs_ops->issue_read(subreq);
+		break;
+	case NETFS_READ_FROM_CACHE:
+		netfs_read_cache_to_pagecache(rreq, subreq);
+		break;
+	default:
+		__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
+		subreq->error = 0;
+		iov_iter_zero(subreq->len, &subreq->io_iter);
+		subreq->transferred = subreq->len;
+		netfs_read_subreq_terminated(subreq);
+		break;
+	}
+}
+
 /*
  * Perform a read to the pagecache from a series of sources of different types,
  * slicing up the region to be read according to available cache blocks and
@@ -226,11 +208,9 @@ static void netfs_read_to_pagecache(struct netfs_io_request *rreq)
 	ssize_t size = rreq->len;
 	int ret = 0;
 
-	atomic_inc(&rreq->nr_outstanding);
-
 	do {
 		struct netfs_io_subrequest *subreq;
-		enum netfs_io_source source = NETFS_DOWNLOAD_FROM_SERVER;
+		enum netfs_io_source source = NETFS_SOURCE_UNKNOWN;
 		ssize_t slice;
 
 		subreq = netfs_alloc_subrequest(rreq);
@@ -242,20 +222,14 @@ static void netfs_read_to_pagecache(struct netfs_io_request *rreq)
 		subreq->start	= start;
 		subreq->len	= size;
 
-		atomic_inc(&rreq->nr_outstanding);
-		spin_lock_bh(&rreq->lock);
-		list_add_tail(&subreq->rreq_link, &rreq->subrequests);
-		subreq->prev_donated = rreq->prev_donated;
-		rreq->prev_donated = 0;
-		trace_netfs_sreq(subreq, netfs_sreq_trace_added);
-		spin_unlock_bh(&rreq->lock);
-
 		source = netfs_cache_prepare_read(rreq, subreq, rreq->i_size);
 		subreq->source = source;
 		if (source == NETFS_DOWNLOAD_FROM_SERVER) {
 			unsigned long long zp = umin(ictx->zero_point, rreq->i_size);
 			size_t len = subreq->len;
 
+			if (unlikely(rreq->origin == NETFS_READ_SINGLE))
+				zp = rreq->i_size;
 			if (subreq->start >= zp) {
 				subreq->source = source = NETFS_FILL_WITH_ZEROES;
 				goto fill_with_zeroes;
@@ -275,17 +249,18 @@ static void netfs_read_to_pagecache(struct netfs_io_request *rreq)
 			netfs_stat(&netfs_n_rh_download);
 			if (rreq->netfs_ops->prepare_read) {
 				ret = rreq->netfs_ops->prepare_read(subreq);
-				if (ret < 0)
-					goto prep_failed;
+				if (ret < 0) {
+					subreq->error = ret;
+					/* Not queued - release both refs. */
+					netfs_put_subrequest(subreq, false,
+							     netfs_sreq_trace_put_cancel);
+					netfs_put_subrequest(subreq, false,
+							     netfs_sreq_trace_put_cancel);
+					break;
+				}
 				trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
 			}
-
-			slice = netfs_prepare_read_iterator(subreq);
-			if (slice < 0)
-				goto prep_iter_failed;
-
-			rreq->netfs_ops->issue_read(subreq);
-			goto done;
+			goto issue;
 		}
 
 	fill_with_zeroes:
@@ -293,94 +268,50 @@ static void netfs_read_to_pagecache(struct netfs_io_request *rreq)
 			subreq->source = NETFS_FILL_WITH_ZEROES;
 			trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
 			netfs_stat(&netfs_n_rh_zero);
-			slice = netfs_prepare_read_iterator(subreq);
-			if (slice < 0)
-				goto prep_iter_failed;
-			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
-			netfs_read_subreq_terminated(subreq, 0, false);
-			goto done;
+			goto issue;
 		}
 
 		if (source == NETFS_READ_FROM_CACHE) {
 			trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
-			slice = netfs_prepare_read_iterator(subreq);
-			if (slice < 0)
-				goto prep_iter_failed;
-			netfs_read_cache_to_pagecache(rreq, subreq);
-			goto done;
+			goto issue;
 		}
 
 		pr_err("Unexpected read source %u\n", source);
 		WARN_ON_ONCE(1);
 		break;
 
-	prep_iter_failed:
-		ret = slice;
-	prep_failed:
-		subreq->error = ret;
-		atomic_dec(&rreq->nr_outstanding);
-		netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
-		break;
-
-	done:
+	issue:
+		slice = netfs_prepare_read_iterator(subreq);
+		if (slice < 0) {
+			ret = slice;
+			subreq->error = ret;
+			trace_netfs_sreq(subreq, netfs_sreq_trace_cancel);
+			/* Not queued - release both refs. */
+			netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
+			netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
+			break;
+		}
 		size -= slice;
 		start += slice;
+		if (size <= 0) {
+			smp_wmb(); /* Write lists before ALL_QUEUED. */
+			set_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags);
+		}
+
+		netfs_issue_read(rreq, subreq);
 		cond_resched();
 	} while (size > 0);
 
-	if (atomic_dec_and_test(&rreq->nr_outstanding))
-		netfs_rreq_terminated(rreq, false);
+	if (unlikely(size > 0)) {
+		smp_wmb(); /* Write lists before ALL_QUEUED. */
+		set_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags);
+		netfs_wake_read_collector(rreq);
+	}
 
 	/* Defer error return as we may need to wait for outstanding I/O. */
 	cmpxchg(&rreq->error, 0, ret);
 }
 
-/*
- * Wait for the read operation to complete, successfully or otherwise.
- */
-static int netfs_wait_for_read(struct netfs_io_request *rreq)
-{
-	int ret;
-
-	trace_netfs_rreq(rreq, netfs_rreq_trace_wait_ip);
-	wait_on_bit(&rreq->flags, NETFS_RREQ_IN_PROGRESS, TASK_UNINTERRUPTIBLE);
-	ret = rreq->error;
-	if (ret == 0 && rreq->submitted < rreq->len) {
-		trace_netfs_failure(rreq, NULL, ret, netfs_fail_short_read);
-		ret = -EIO;
-	}
-
-	return ret;
-}
-
-/*
- * Set up the initial folioq of buffer folios in the rolling buffer and set the
- * iterator to refer to it.
- */
-static int netfs_prime_buffer(struct netfs_io_request *rreq)
-{
-	struct folio_queue *folioq;
-	struct folio_batch put_batch;
-	size_t added;
-
-	folioq = kmalloc(sizeof(*folioq), GFP_KERNEL);
-	if (!folioq)
-		return -ENOMEM;
-	netfs_stat(&netfs_n_folioq);
-	folioq_init(folioq);
-	rreq->buffer = folioq;
-	rreq->buffer_tail = folioq;
-	rreq->submitted = rreq->start;
-	iov_iter_folio_queue(&rreq->iter, ITER_DEST, folioq, 0, 0, 0);
-
-	folio_batch_init(&put_batch);
-	added = netfs_load_buffer_from_ra(rreq, folioq, &put_batch);
-	folio_batch_release(&put_batch);
-	rreq->iter.count += added;
-	rreq->submitted += added;
-	return 0;
-}
-
 /**
  * netfs_readahead - Helper to manage a read request
  * @ractl: The description of the readahead request
@@ -409,6 +340,8 @@ void netfs_readahead(struct readahead_control *ractl)
 	if (IS_ERR(rreq))
 		return;
 
+	__set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &rreq->flags);
+
 	ret = netfs_begin_cache_read(rreq, ictx);
 	if (ret == -ENOMEM || ret == -EINTR || ret == -ERESTARTSYS)
 		goto cleanup_free;
@@ -420,7 +353,8 @@ void netfs_readahead(struct readahead_control *ractl)
 	netfs_rreq_expand(rreq, ractl);
 
 	rreq->ractl = ractl;
-	if (netfs_prime_buffer(rreq) < 0)
+	rreq->submitted = rreq->start;
+	if (rolling_buffer_init(&rreq->buffer, rreq->debug_id, ITER_DEST) < 0)
 		goto cleanup_free;
 	netfs_read_to_pagecache(rreq);
 
@@ -436,23 +370,18 @@ EXPORT_SYMBOL(netfs_readahead);
 /*
  * Create a rolling buffer with a single occupying folio.
  */
-static int netfs_create_singular_buffer(struct netfs_io_request *rreq, struct folio *folio)
+static int netfs_create_singular_buffer(struct netfs_io_request *rreq, struct folio *folio,
+					unsigned int rollbuf_flags)
 {
-	struct folio_queue *folioq;
+	ssize_t added;
 
-	folioq = kmalloc(sizeof(*folioq), GFP_KERNEL);
-	if (!folioq)
+	if (rolling_buffer_init(&rreq->buffer, rreq->debug_id, ITER_DEST) < 0)
 		return -ENOMEM;
 
-	netfs_stat(&netfs_n_folioq);
-	folioq_init(folioq);
-	folioq_append(folioq, folio);
-	BUG_ON(folioq_folio(folioq, 0) != folio);
-	BUG_ON(folioq_folio_order(folioq, 0) != folio_order(folio));
-	rreq->buffer = folioq;
-	rreq->buffer_tail = folioq;
-	rreq->submitted = rreq->start + rreq->len;
-	iov_iter_folio_queue(&rreq->iter, ITER_DEST, folioq, 0, 0, rreq->len);
+	added = rolling_buffer_append(&rreq->buffer, folio, rollbuf_flags);
+	if (added < 0)
+		return added;
+	rreq->submitted = rreq->start + added;
 	rreq->ractl = (struct readahead_control *)1UL;
 	return 0;
 }
@@ -520,7 +449,7 @@ static int netfs_read_gaps(struct file *file, struct folio *folio)
 	}
 	if (to < flen)
 		bvec_set_folio(&bvec[i++], folio, flen - to, to);
-	iov_iter_bvec(&rreq->iter, ITER_DEST, bvec, i, rreq->len);
+	iov_iter_bvec(&rreq->buffer.iter, ITER_DEST, bvec, i, rreq->len);
 	rreq->submitted = rreq->start + flen;
 
 	netfs_read_to_pagecache(rreq);
@@ -529,7 +458,7 @@ static int netfs_read_gaps(struct file *file, struct folio *folio)
 		folio_put(sink);
 
 	ret = netfs_wait_for_read(rreq);
-	if (ret == 0) {
+	if (ret >= 0) {
 		flush_dcache_folio(folio);
 		folio_mark_uptodate(folio);
 	}
@@ -588,7 +517,7 @@ int netfs_read_folio(struct file *file, struct folio *folio)
 	trace_netfs_read(rreq, rreq->start, rreq->len, netfs_read_trace_readpage);
 
 	/* Set up the output buffer */
-	ret = netfs_create_singular_buffer(rreq, folio);
+	ret = netfs_create_singular_buffer(rreq, folio, 0);
 	if (ret < 0)
 		goto discard;
 
@@ -745,7 +674,7 @@ retry:
 	trace_netfs_read(rreq, pos, len, netfs_read_trace_write_begin);
 
 	/* Set up the output buffer */
-	ret = netfs_create_singular_buffer(rreq, folio);
+	ret = netfs_create_singular_buffer(rreq, folio, 0);
 	if (ret < 0)
 		goto error_put;
 
@@ -810,15 +739,14 @@ int netfs_prefetch_for_write(struct file *file, struct folio *folio,
 	trace_netfs_read(rreq, start, flen, netfs_read_trace_prefetch_for_write);
 
 	/* Set up the output buffer */
-	ret = netfs_create_singular_buffer(rreq, folio);
+	ret = netfs_create_singular_buffer(rreq, folio, NETFS_ROLLBUF_PAGECACHE_MARK);
 	if (ret < 0)
 		goto error_put;
 
-	folioq_mark2(rreq->buffer, 0);
 	netfs_read_to_pagecache(rreq);
 	ret = netfs_wait_for_read(rreq);
 	netfs_put_request(rreq, false, netfs_rreq_trace_put_return);
-	return ret;
+	return ret < 0 ? ret : 0;
 
 error_put:
 	netfs_put_request(rreq, false, netfs_rreq_trace_put_discard);
diff --git a/fs/netfs/direct_read.c b/fs/netfs/direct_read.c
index b1a66a6e6bc2..0bf3c2f5a710 100644
--- a/fs/netfs/direct_read.c
+++ b/fs/netfs/direct_read.c
@@ -25,7 +25,7 @@ static void netfs_prepare_dio_read_iterator(struct netfs_io_subrequest *subreq)
 	subreq->len = rsize;
 
 	if (unlikely(rreq->io_streams[0].sreq_max_segs)) {
-		size_t limit = netfs_limit_iter(&rreq->iter, 0, rsize,
+		size_t limit = netfs_limit_iter(&rreq->buffer.iter, 0, rsize,
 						rreq->io_streams[0].sreq_max_segs);
 
 		if (limit < rsize) {
@@ -36,9 +36,9 @@ static void netfs_prepare_dio_read_iterator(struct netfs_io_subrequest *subreq)
 
 	trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
 
-	subreq->io_iter	= rreq->iter;
+	subreq->io_iter	= rreq->buffer.iter;
 	iov_iter_truncate(&subreq->io_iter, subreq->len);
-	iov_iter_advance(&rreq->iter, subreq->len);
+	iov_iter_advance(&rreq->buffer.iter, subreq->len);
 }
 
 /*
@@ -47,12 +47,11 @@ static void netfs_prepare_dio_read_iterator(struct netfs_io_subrequest *subreq)
  */
 static int netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
 {
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
 	unsigned long long start = rreq->start;
 	ssize_t size = rreq->len;
 	int ret = 0;
 
-	atomic_set(&rreq->nr_outstanding, 1);
-
 	do {
 		struct netfs_io_subrequest *subreq;
 		ssize_t slice;
@@ -67,19 +66,25 @@ static int netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
 		subreq->start	= start;
 		subreq->len	= size;
 
-		atomic_inc(&rreq->nr_outstanding);
-		spin_lock_bh(&rreq->lock);
-		list_add_tail(&subreq->rreq_link, &rreq->subrequests);
-		subreq->prev_donated = rreq->prev_donated;
-		rreq->prev_donated = 0;
+		__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
+
+		spin_lock(&rreq->lock);
+		list_add_tail(&subreq->rreq_link, &stream->subrequests);
+		if (list_is_first(&subreq->rreq_link, &stream->subrequests)) {
+			stream->front = subreq;
+			if (!stream->active) {
+				stream->collected_to = stream->front->start;
+				/* Store list pointers before active flag */
+				smp_store_release(&stream->active, true);
+			}
+		}
 		trace_netfs_sreq(subreq, netfs_sreq_trace_added);
-		spin_unlock_bh(&rreq->lock);
+		spin_unlock(&rreq->lock);
 
 		netfs_stat(&netfs_n_rh_download);
 		if (rreq->netfs_ops->prepare_read) {
 			ret = rreq->netfs_ops->prepare_read(subreq);
 			if (ret < 0) {
-				atomic_dec(&rreq->nr_outstanding);
 				netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
 				break;
 			}
@@ -87,20 +92,32 @@ static int netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
 
 		netfs_prepare_dio_read_iterator(subreq);
 		slice = subreq->len;
-		rreq->netfs_ops->issue_read(subreq);
-
 		size -= slice;
 		start += slice;
 		rreq->submitted += slice;
+		if (size <= 0) {
+			smp_wmb(); /* Write lists before ALL_QUEUED. */
+			set_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags);
+		}
+
+		rreq->netfs_ops->issue_read(subreq);
 
+		if (test_bit(NETFS_RREQ_PAUSE, &rreq->flags))
+			netfs_wait_for_pause(rreq);
+		if (test_bit(NETFS_RREQ_FAILED, &rreq->flags))
+			break;
 		if (test_bit(NETFS_RREQ_BLOCKED, &rreq->flags) &&
 		    test_bit(NETFS_RREQ_NONBLOCK, &rreq->flags))
 			break;
 		cond_resched();
 	} while (size > 0);
 
-	if (atomic_dec_and_test(&rreq->nr_outstanding))
-		netfs_rreq_terminated(rreq, false);
+	if (unlikely(size > 0)) {
+		smp_wmb(); /* Write lists before ALL_QUEUED. */
+		set_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags);
+		netfs_wake_read_collector(rreq);
+	}
+
 	return ret;
 }
 
@@ -133,21 +150,10 @@ static int netfs_unbuffered_read(struct netfs_io_request *rreq, bool sync)
 		goto out;
 	}
 
-	if (sync) {
-		trace_netfs_rreq(rreq, netfs_rreq_trace_wait_ip);
-		wait_on_bit(&rreq->flags, NETFS_RREQ_IN_PROGRESS,
-			    TASK_UNINTERRUPTIBLE);
-
-		ret = rreq->error;
-		if (ret == 0 && rreq->submitted < rreq->len &&
-		    rreq->origin != NETFS_DIO_READ) {
-			trace_netfs_failure(rreq, NULL, ret, netfs_fail_short_read);
-			ret = -EIO;
-		}
-	} else {
+	if (sync)
+		ret = netfs_wait_for_read(rreq);
+	else
 		ret = -EIOCBQUEUED;
-	}
-
 out:
 	_leave(" = %d", ret);
 	return ret;
@@ -199,15 +205,15 @@ ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *i
 	 * the request.
 	 */
 	if (user_backed_iter(iter)) {
-		ret = netfs_extract_user_iter(iter, rreq->len, &rreq->iter, 0);
+		ret = netfs_extract_user_iter(iter, rreq->len, &rreq->buffer.iter, 0);
 		if (ret < 0)
 			goto out;
-		rreq->direct_bv = (struct bio_vec *)rreq->iter.bvec;
+		rreq->direct_bv = (struct bio_vec *)rreq->buffer.iter.bvec;
 		rreq->direct_bv_count = ret;
 		rreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
-		rreq->len = iov_iter_count(&rreq->iter);
+		rreq->len = iov_iter_count(&rreq->buffer.iter);
 	} else {
-		rreq->iter = *iter;
+		rreq->buffer.iter = *iter;
 		rreq->len = orig_count;
 		rreq->direct_bv_unpin = false;
 		iov_iter_advance(iter, orig_count);
@@ -215,8 +221,10 @@ ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *i
 
 	// TODO: Set up bounce buffer if needed
 
-	if (!sync)
+	if (!sync) {
 		rreq->iocb = iocb;
+		__set_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &rreq->flags);
+	}
 
 	ret = netfs_unbuffered_read(rreq, sync);
 	if (ret < 0)
diff --git a/fs/netfs/direct_write.c b/fs/netfs/direct_write.c
index f9421f3e6d37..42ce53cc216e 100644
--- a/fs/netfs/direct_write.c
+++ b/fs/netfs/direct_write.c
@@ -68,12 +68,12 @@ ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *
 		 * request.
 		 */
 		if (user_backed_iter(iter)) {
-			n = netfs_extract_user_iter(iter, len, &wreq->iter, 0);
+			n = netfs_extract_user_iter(iter, len, &wreq->buffer.iter, 0);
 			if (n < 0) {
 				ret = n;
 				goto out;
 			}
-			wreq->direct_bv = (struct bio_vec *)wreq->iter.bvec;
+			wreq->direct_bv = (struct bio_vec *)wreq->buffer.iter.bvec;
 			wreq->direct_bv_count = n;
 			wreq->direct_bv_unpin = iov_iter_extract_will_pin(iter);
 		} else {
@@ -82,10 +82,8 @@ ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *
 			 * (eg. a bio_vec array) will persist till the end of
 			 * the op.
 			 */
-			wreq->iter = *iter;
+			wreq->buffer.iter = *iter;
 		}
-
-		wreq->io_iter = wreq->iter;
 	}
 
 	__set_bit(NETFS_RREQ_USE_IO_ITER, &wreq->flags);
@@ -97,7 +95,7 @@ ssize_t netfs_unbuffered_write_iter_locked(struct kiocb *iocb, struct iov_iter *
 	__set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags);
 	if (async)
 		wreq->iocb = iocb;
-	wreq->len = iov_iter_count(&wreq->io_iter);
+	wreq->len = iov_iter_count(&wreq->buffer.iter);
 	wreq->cleanup = netfs_cleanup_dio_write;
 	ret = netfs_unbuffered_write(wreq, is_sync_kiocb(iocb), wreq->len);
 	if (ret < 0) {
diff --git a/fs/netfs/internal.h b/fs/netfs/internal.h
index c562aec3b483..eb76f98c894b 100644
--- a/fs/netfs/internal.h
+++ b/fs/netfs/internal.h
@@ -23,6 +23,7 @@
 /*
  * buffered_read.c
  */
+void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error, bool was_async);
 int netfs_prefetch_for_write(struct file *file, struct folio *folio,
 			     size_t offset, size_t len);
 
@@ -58,11 +59,8 @@ static inline void netfs_proc_del_rreq(struct netfs_io_request *rreq) {}
 /*
  * misc.c
  */
-struct folio_queue *netfs_buffer_make_space(struct netfs_io_request *rreq);
-int netfs_buffer_append_folio(struct netfs_io_request *rreq, struct folio *folio,
-			      bool needs_put);
-struct folio_queue *netfs_delete_buffer_head(struct netfs_io_request *wreq);
-void netfs_clear_buffer(struct netfs_io_request *rreq);
+struct folio_queue *netfs_buffer_make_space(struct netfs_io_request *rreq,
+					    enum netfs_folioq_trace trace);
 void netfs_reset_iter(struct netfs_io_subrequest *subreq);
 
 /*
@@ -84,20 +82,27 @@ static inline void netfs_see_request(struct netfs_io_request *rreq,
 	trace_netfs_rreq_ref(rreq->debug_id, refcount_read(&rreq->ref), what);
 }
 
+static inline void netfs_see_subrequest(struct netfs_io_subrequest *subreq,
+					enum netfs_sreq_ref_trace what)
+{
+	trace_netfs_sreq_ref(subreq->rreq->debug_id, subreq->debug_index,
+			     refcount_read(&subreq->ref), what);
+}
+
 /*
  * read_collect.c
  */
-void netfs_read_termination_worker(struct work_struct *work);
-void netfs_rreq_terminated(struct netfs_io_request *rreq, bool was_async);
+void netfs_read_collection_worker(struct work_struct *work);
+void netfs_wake_read_collector(struct netfs_io_request *rreq);
+void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error, bool was_async);
+ssize_t netfs_wait_for_read(struct netfs_io_request *rreq);
+void netfs_wait_for_pause(struct netfs_io_request *rreq);
 
 /*
  * read_pgpriv2.c
  */
-void netfs_pgpriv2_mark_copy_to_cache(struct netfs_io_subrequest *subreq,
-				      struct netfs_io_request *rreq,
-				      struct folio_queue *folioq,
-				      int slot);
-void netfs_pgpriv2_write_to_the_cache(struct netfs_io_request *rreq);
+void netfs_pgpriv2_copy_to_cache(struct netfs_io_request *rreq, struct folio *folio);
+void netfs_pgpriv2_end_copy_to_cache(struct netfs_io_request *rreq);
 bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq);
 
 /*
@@ -113,6 +118,7 @@ void netfs_unlock_abandoned_read_pages(struct netfs_io_request *rreq);
 extern atomic_t netfs_n_rh_dio_read;
 extern atomic_t netfs_n_rh_readahead;
 extern atomic_t netfs_n_rh_read_folio;
+extern atomic_t netfs_n_rh_read_single;
 extern atomic_t netfs_n_rh_rreq;
 extern atomic_t netfs_n_rh_sreq;
 extern atomic_t netfs_n_rh_download;
@@ -181,9 +187,9 @@ void netfs_reissue_write(struct netfs_io_stream *stream,
 			 struct iov_iter *source);
 void netfs_issue_write(struct netfs_io_request *wreq,
 		       struct netfs_io_stream *stream);
-int netfs_advance_write(struct netfs_io_request *wreq,
-			struct netfs_io_stream *stream,
-			loff_t start, size_t len, bool to_eof);
+size_t netfs_advance_write(struct netfs_io_request *wreq,
+			   struct netfs_io_stream *stream,
+			   loff_t start, size_t len, bool to_eof);
 struct netfs_io_request *netfs_begin_writethrough(struct kiocb *iocb, size_t len);
 int netfs_advance_writethrough(struct netfs_io_request *wreq, struct writeback_control *wbc,
 			       struct folio *folio, size_t copied, bool to_page_end,
@@ -193,6 +199,11 @@ int netfs_end_writethrough(struct netfs_io_request *wreq, struct writeback_contr
 int netfs_unbuffered_write(struct netfs_io_request *wreq, bool may_wait, size_t len);
 
 /*
+ * write_retry.c
+ */
+void netfs_retry_writes(struct netfs_io_request *wreq);
+
+/*
  * Miscellaneous functions.
  */
 static inline bool netfs_is_cache_enabled(struct netfs_inode *ctx)
diff --git a/fs/netfs/main.c b/fs/netfs/main.c
index 6c7be1377ee0..4e3e62040831 100644
--- a/fs/netfs/main.c
+++ b/fs/netfs/main.c
@@ -37,9 +37,11 @@ static const char *netfs_origins[nr__netfs_io_origin] = {
 	[NETFS_READAHEAD]		= "RA",
 	[NETFS_READPAGE]		= "RP",
 	[NETFS_READ_GAPS]		= "RG",
+	[NETFS_READ_SINGLE]		= "R1",
 	[NETFS_READ_FOR_WRITE]		= "RW",
 	[NETFS_DIO_READ]		= "DR",
 	[NETFS_WRITEBACK]		= "WB",
+	[NETFS_WRITEBACK_SINGLE]	= "W1",
 	[NETFS_WRITETHROUGH]		= "WT",
 	[NETFS_UNBUFFERED_WRITE]	= "UW",
 	[NETFS_DIO_WRITE]		= "DW",
@@ -69,7 +71,7 @@ static int netfs_requests_seq_show(struct seq_file *m, void *v)
 		   refcount_read(&rreq->ref),
 		   rreq->flags,
 		   rreq->error,
-		   atomic_read(&rreq->nr_outstanding),
+		   0,
 		   rreq->start, rreq->submitted, rreq->len);
 	seq_putc(m, '\n');
 	return 0;
@@ -116,7 +118,7 @@ static int __init netfs_init(void)
 		goto error_reqpool;
 
 	netfs_subrequest_slab = kmem_cache_create("netfs_subrequest",
-						  sizeof(struct netfs_io_subrequest), 0,
+						  sizeof(struct netfs_io_subrequest) + 16, 0,
 						  SLAB_HWCACHE_ALIGN | SLAB_ACCOUNT,
 						  NULL);
 	if (!netfs_subrequest_slab)
diff --git a/fs/netfs/misc.c b/fs/netfs/misc.c
index 78fe5796b2b2..7099aa07737a 100644
--- a/fs/netfs/misc.c
+++ b/fs/netfs/misc.c
@@ -8,113 +8,101 @@
 #include <linux/swap.h>
 #include "internal.h"
 
-/*
- * Make sure there's space in the rolling queue.
+/**
+ * netfs_alloc_folioq_buffer - Allocate buffer space into a folio queue
+ * @mapping: Address space to set on the folio (or NULL).
+ * @_buffer: Pointer to the folio queue to add to (may point to a NULL; updated).
+ * @_cur_size: Current size of the buffer (updated).
+ * @size: Target size of the buffer.
+ * @gfp: The allocation constraints.
  */
-struct folio_queue *netfs_buffer_make_space(struct netfs_io_request *rreq)
+int netfs_alloc_folioq_buffer(struct address_space *mapping,
+			      struct folio_queue **_buffer,
+			      size_t *_cur_size, ssize_t size, gfp_t gfp)
 {
-	struct folio_queue *tail = rreq->buffer_tail, *prev;
-	unsigned int prev_nr_slots = 0;
-
-	if (WARN_ON_ONCE(!rreq->buffer && tail) ||
-	    WARN_ON_ONCE(rreq->buffer && !tail))
-		return ERR_PTR(-EIO);
-
-	prev = tail;
-	if (prev) {
-		if (!folioq_full(tail))
-			return tail;
-		prev_nr_slots = folioq_nr_slots(tail);
-	}
-
-	tail = kmalloc(sizeof(*tail), GFP_NOFS);
-	if (!tail)
-		return ERR_PTR(-ENOMEM);
-	netfs_stat(&netfs_n_folioq);
-	folioq_init(tail);
-	tail->prev = prev;
-	if (prev)
-		/* [!] NOTE: After we set prev->next, the consumer is entirely
-		 * at liberty to delete prev.
-		 */
-		WRITE_ONCE(prev->next, tail);
-
-	rreq->buffer_tail = tail;
-	if (!rreq->buffer) {
-		rreq->buffer = tail;
-		iov_iter_folio_queue(&rreq->io_iter, ITER_SOURCE, tail, 0, 0, 0);
-	} else {
-		/* Make sure we don't leave the master iterator pointing to a
-		 * block that might get immediately consumed.
-		 */
-		if (rreq->io_iter.folioq == prev &&
-		    rreq->io_iter.folioq_slot == prev_nr_slots) {
-			rreq->io_iter.folioq = tail;
-			rreq->io_iter.folioq_slot = 0;
+	struct folio_queue *tail = *_buffer, *p;
+
+	size = round_up(size, PAGE_SIZE);
+	if (*_cur_size >= size)
+		return 0;
+
+	if (tail)
+		while (tail->next)
+			tail = tail->next;
+
+	do {
+		struct folio *folio;
+		int order = 0, slot;
+
+		if (!tail || folioq_full(tail)) {
+			p = netfs_folioq_alloc(0, GFP_NOFS, netfs_trace_folioq_alloc_buffer);
+			if (!p)
+				return -ENOMEM;
+			if (tail) {
+				tail->next = p;
+				p->prev = tail;
+			} else {
+				*_buffer = p;
+			}
+			tail = p;
 		}
-	}
-	rreq->buffer_tail_slot = 0;
-	return tail;
-}
 
-/*
- * Append a folio to the rolling queue.
- */
-int netfs_buffer_append_folio(struct netfs_io_request *rreq, struct folio *folio,
-			      bool needs_put)
-{
-	struct folio_queue *tail;
-	unsigned int slot, order = folio_order(folio);
+		if (size - *_cur_size > PAGE_SIZE)
+			order = umin(ilog2(size - *_cur_size) - PAGE_SHIFT,
+				     MAX_PAGECACHE_ORDER);
 
-	tail = netfs_buffer_make_space(rreq);
-	if (IS_ERR(tail))
-		return PTR_ERR(tail);
+		folio = folio_alloc(gfp, order);
+		if (!folio && order > 0)
+			folio = folio_alloc(gfp, 0);
+		if (!folio)
+			return -ENOMEM;
 
-	rreq->io_iter.count += PAGE_SIZE << order;
+		folio->mapping = mapping;
+		folio->index = *_cur_size / PAGE_SIZE;
+		trace_netfs_folio(folio, netfs_folio_trace_alloc_buffer);
+		slot = folioq_append_mark(tail, folio);
+		*_cur_size += folioq_folio_size(tail, slot);
+	} while (*_cur_size < size);
 
-	slot = folioq_append(tail, folio);
-	/* Store the counter after setting the slot. */
-	smp_store_release(&rreq->buffer_tail_slot, slot);
 	return 0;
 }
+EXPORT_SYMBOL(netfs_alloc_folioq_buffer);
 
-/*
- * Delete the head of a rolling queue.
+/**
+ * netfs_free_folioq_buffer - Free a folio queue.
+ * @fq: The start of the folio queue to free
+ *
+ * Free up a chain of folio_queues and, if marked, the marked folios they point
+ * to.
  */
-struct folio_queue *netfs_delete_buffer_head(struct netfs_io_request *wreq)
+void netfs_free_folioq_buffer(struct folio_queue *fq)
 {
-	struct folio_queue *head = wreq->buffer, *next = head->next;
-
-	if (next)
-		next->prev = NULL;
-	netfs_stat_d(&netfs_n_folioq);
-	kfree(head);
-	wreq->buffer = next;
-	return next;
-}
+	struct folio_queue *next;
+	struct folio_batch fbatch;
 
-/*
- * Clear out a rolling queue.
- */
-void netfs_clear_buffer(struct netfs_io_request *rreq)
-{
-	struct folio_queue *p;
+	folio_batch_init(&fbatch);
+
+	for (; fq; fq = next) {
+		for (int slot = 0; slot < folioq_count(fq); slot++) {
+			struct folio *folio = folioq_folio(fq, slot);
 
-	while ((p = rreq->buffer)) {
-		rreq->buffer = p->next;
-		for (int slot = 0; slot < folioq_count(p); slot++) {
-			struct folio *folio = folioq_folio(p, slot);
-			if (!folio)
+			if (!folio ||
+			    !folioq_is_marked(fq, slot))
 				continue;
-			if (folioq_is_marked(p, slot)) {
-				trace_netfs_folio(folio, netfs_folio_trace_put);
-				folio_put(folio);
-			}
+
+			trace_netfs_folio(folio, netfs_folio_trace_put);
+			if (folio_batch_add(&fbatch, folio))
+				folio_batch_release(&fbatch);
 		}
+
 		netfs_stat_d(&netfs_n_folioq);
-		kfree(p);
+		next = fq->next;
+		kfree(fq);
 	}
+
+	folio_batch_release(&fbatch);
 }
+EXPORT_SYMBOL(netfs_free_folioq_buffer);
 
 /*
  * Reset the subrequest iterator to refer just to the region remaining to be
diff --git a/fs/netfs/objects.c b/fs/netfs/objects.c
index 31e388ec6e48..dc6b41ef18b0 100644
--- a/fs/netfs/objects.c
+++ b/fs/netfs/objects.c
@@ -48,17 +48,20 @@ struct netfs_io_request *netfs_alloc_request(struct address_space *mapping,
 	spin_lock_init(&rreq->lock);
 	INIT_LIST_HEAD(&rreq->io_streams[0].subrequests);
 	INIT_LIST_HEAD(&rreq->io_streams[1].subrequests);
-	INIT_LIST_HEAD(&rreq->subrequests);
+	init_waitqueue_head(&rreq->waitq);
 	refcount_set(&rreq->ref, 1);
 
 	if (origin == NETFS_READAHEAD ||
 	    origin == NETFS_READPAGE ||
 	    origin == NETFS_READ_GAPS ||
+	    origin == NETFS_READ_SINGLE ||
 	    origin == NETFS_READ_FOR_WRITE ||
-	    origin == NETFS_DIO_READ)
-		INIT_WORK(&rreq->work, netfs_read_termination_worker);
-	else
+	    origin == NETFS_DIO_READ) {
+		INIT_WORK(&rreq->work, netfs_read_collection_worker);
+		rreq->io_streams[0].avail = true;
+	} else {
 		INIT_WORK(&rreq->work, netfs_write_collection_worker);
+	}
 
 	__set_bit(NETFS_RREQ_IN_PROGRESS, &rreq->flags);
 	if (file && file->f_flags & O_NONBLOCK)
@@ -92,14 +95,6 @@ void netfs_clear_subrequests(struct netfs_io_request *rreq, bool was_async)
 	struct netfs_io_stream *stream;
 	int s;
 
-	while (!list_empty(&rreq->subrequests)) {
-		subreq = list_first_entry(&rreq->subrequests,
-					  struct netfs_io_subrequest, rreq_link);
-		list_del(&subreq->rreq_link);
-		netfs_put_subrequest(subreq, was_async,
-				     netfs_sreq_trace_put_clear);
-	}
-
 	for (s = 0; s < ARRAY_SIZE(rreq->io_streams); s++) {
 		stream = &rreq->io_streams[s];
 		while (!list_empty(&stream->subrequests)) {
@@ -143,7 +138,7 @@ static void netfs_free_request(struct work_struct *work)
 		}
 		kvfree(rreq->direct_bv);
 	}
-	netfs_clear_buffer(rreq);
+	rolling_buffer_clear(&rreq->buffer);
 
 	if (atomic_dec_and_test(&ictx->io_count))
 		wake_up_var(&ictx->io_count);
diff --git a/fs/netfs/read_collect.c b/fs/netfs/read_collect.c
index e8624f5c7fcc..f65affa5a9e4 100644
--- a/fs/netfs/read_collect.c
+++ b/fs/netfs/read_collect.c
@@ -14,6 +14,14 @@
 #include <linux/task_io_accounting_ops.h>
 #include "internal.h"
 
+/* Notes made in the collector */
+#define HIT_PENDING	0x01	/* A front op was still pending */
+#define MADE_PROGRESS	0x04	/* Made progress cleaning up a stream or the folio set */
+#define BUFFERED	0x08	/* The pagecache needs cleaning up */
+#define NEED_RETRY	0x10	/* A front op requests retrying */
+#define COPY_TO_CACHE	0x40	/* Need to copy subrequest to cache */
+#define ABANDON_SREQ	0x80	/* Need to abandon untransferred part of subrequest */
+
 /*
  * Clear the unread part of an I/O request.
  */
@@ -31,14 +39,18 @@ static void netfs_clear_unread(struct netfs_io_subrequest *subreq)
  * cache the folio, we set the group to NETFS_FOLIO_COPY_TO_CACHE, mark it
  * dirty and let writeback handle it.
  */
-static void netfs_unlock_read_folio(struct netfs_io_subrequest *subreq,
-				    struct netfs_io_request *rreq,
+static void netfs_unlock_read_folio(struct netfs_io_request *rreq,
 				    struct folio_queue *folioq,
 				    int slot)
 {
 	struct netfs_folio *finfo;
 	struct folio *folio = folioq_folio(folioq, slot);
 
+	if (unlikely(folio_pos(folio) < rreq->abandon_to)) {
+		trace_netfs_folio(folio, netfs_folio_trace_abandon);
+		goto just_unlock;
+	}
+
 	flush_dcache_folio(folio);
 	folio_mark_uptodate(folio);
 
@@ -53,7 +65,7 @@ static void netfs_unlock_read_folio(struct netfs_io_subrequest *subreq,
 			kfree(finfo);
 		}
 
-		if (test_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags)) {
+		if (test_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags)) {
 			if (!WARN_ON_ONCE(folio_get_private(folio) != NULL)) {
 				trace_netfs_folio(folio, netfs_folio_trace_copy_to_cache);
 				folio_attach_private(folio, NETFS_FOLIO_COPY_TO_CACHE);
@@ -66,12 +78,11 @@ static void netfs_unlock_read_folio(struct netfs_io_subrequest *subreq,
 		folioq_clear(folioq, slot);
 	} else {
 		// TODO: Use of PG_private_2 is deprecated.
-		if (test_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags))
-			netfs_pgpriv2_mark_copy_to_cache(subreq, rreq, folioq, slot);
-		else
-			folioq_clear(folioq, slot);
+		if (test_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags))
+			netfs_pgpriv2_copy_to_cache(rreq, folio);
 	}
 
+just_unlock:
 	if (!test_bit(NETFS_RREQ_DONT_UNLOCK_FOLIOS, &rreq->flags)) {
 		if (folio->index == rreq->no_unlock_folio &&
 		    test_bit(NETFS_RREQ_NO_UNLOCK_FOLIO, &rreq->flags)) {
@@ -81,238 +92,249 @@ static void netfs_unlock_read_folio(struct netfs_io_subrequest *subreq,
 			folio_unlock(folio);
 		}
 	}
+
+	folioq_clear(folioq, slot);
 }
 
 /*
- * Unlock any folios that are now completely read.  Returns true if the
- * subrequest is removed from the list.
+ * Unlock any folios we've finished with.
  */
-static bool netfs_consume_read_data(struct netfs_io_subrequest *subreq, bool was_async)
+static void netfs_read_unlock_folios(struct netfs_io_request *rreq,
+				     unsigned int *notes)
 {
-	struct netfs_io_subrequest *prev, *next;
-	struct netfs_io_request *rreq = subreq->rreq;
-	struct folio_queue *folioq = subreq->curr_folioq;
-	size_t avail, prev_donated, next_donated, fsize, part, excess;
-	loff_t fpos, start;
-	loff_t fend;
-	int slot = subreq->curr_folioq_slot;
-
-	if (WARN(subreq->transferred > subreq->len,
-		 "Subreq overread: R%x[%x] %zu > %zu",
-		 rreq->debug_id, subreq->debug_index,
-		 subreq->transferred, subreq->len))
-		subreq->transferred = subreq->len;
-
-next_folio:
-	fsize = PAGE_SIZE << subreq->curr_folio_order;
-	fpos = round_down(subreq->start + subreq->consumed, fsize);
-	fend = fpos + fsize;
-
-	if (WARN_ON_ONCE(!folioq) ||
-	    WARN_ON_ONCE(!folioq_folio(folioq, slot)) ||
-	    WARN_ON_ONCE(folioq_folio(folioq, slot)->index != fpos / PAGE_SIZE)) {
-		pr_err("R=%08x[%x] s=%llx-%llx ctl=%zx/%zx/%zx sl=%u\n",
-		       rreq->debug_id, subreq->debug_index,
-		       subreq->start, subreq->start + subreq->transferred - 1,
-		       subreq->consumed, subreq->transferred, subreq->len,
-		       slot);
-		if (folioq) {
-			struct folio *folio = folioq_folio(folioq, slot);
-
-			pr_err("folioq: orders=%02x%02x%02x%02x\n",
-			       folioq->orders[0], folioq->orders[1],
-			       folioq->orders[2], folioq->orders[3]);
-			if (folio)
-				pr_err("folio: %llx-%llx ix=%llx o=%u qo=%u\n",
-				       fpos, fend - 1, folio_pos(folio), folio_order(folio),
-				       folioq_folio_order(folioq, slot));
-		}
-	}
+	struct folio_queue *folioq = rreq->buffer.tail;
+	unsigned long long collected_to = rreq->collected_to;
+	unsigned int slot = rreq->buffer.first_tail_slot;
 
-donation_changed:
-	/* Try to consume the current folio if we've hit or passed the end of
-	 * it.  There's a possibility that this subreq doesn't start at the
-	 * beginning of the folio, in which case we need to donate to/from the
-	 * preceding subreq.
-	 *
-	 * We also need to include any potential donation back from the
-	 * following subreq.
-	 */
-	prev_donated = READ_ONCE(subreq->prev_donated);
-	next_donated =  READ_ONCE(subreq->next_donated);
-	if (prev_donated || next_donated) {
-		spin_lock_bh(&rreq->lock);
-		prev_donated = subreq->prev_donated;
-		next_donated =  subreq->next_donated;
-		subreq->start -= prev_donated;
-		subreq->len += prev_donated;
-		subreq->transferred += prev_donated;
-		prev_donated = subreq->prev_donated = 0;
-		if (subreq->transferred == subreq->len) {
-			subreq->len += next_donated;
-			subreq->transferred += next_donated;
-			next_donated = subreq->next_donated = 0;
+	if (rreq->cleaned_to >= rreq->collected_to)
+		return;
+
+	// TODO: Begin decryption
+
+	if (slot >= folioq_nr_slots(folioq)) {
+		folioq = rolling_buffer_delete_spent(&rreq->buffer);
+		if (!folioq) {
+			rreq->front_folio_order = 0;
+			return;
 		}
-		trace_netfs_sreq(subreq, netfs_sreq_trace_add_donations);
-		spin_unlock_bh(&rreq->lock);
+		slot = 0;
 	}
 
-	avail = subreq->transferred;
-	if (avail == subreq->len)
-		avail += next_donated;
-	start = subreq->start;
-	if (subreq->consumed == 0) {
-		start -= prev_donated;
-		avail += prev_donated;
-	} else {
-		start += subreq->consumed;
-		avail -= subreq->consumed;
-	}
-	part = umin(avail, fsize);
-
-	trace_netfs_progress(subreq, start, avail, part);
-
-	if (start + avail >= fend) {
-		if (fpos == start) {
-			/* Flush, unlock and mark for caching any folio we've just read. */
-			subreq->consumed = fend - subreq->start;
-			netfs_unlock_read_folio(subreq, rreq, folioq, slot);
-			folioq_mark2(folioq, slot);
-			if (subreq->consumed >= subreq->len)
-				goto remove_subreq;
-		} else if (fpos < start) {
-			excess = fend - subreq->start;
-
-			spin_lock_bh(&rreq->lock);
-			/* If we complete first on a folio split with the
-			 * preceding subreq, donate to that subreq - otherwise
-			 * we get the responsibility.
-			 */
-			if (subreq->prev_donated != prev_donated) {
-				spin_unlock_bh(&rreq->lock);
-				goto donation_changed;
-			}
+	for (;;) {
+		struct folio *folio;
+		unsigned long long fpos, fend;
+		unsigned int order;
+		size_t fsize;
 
-			if (list_is_first(&subreq->rreq_link, &rreq->subrequests)) {
-				spin_unlock_bh(&rreq->lock);
-				pr_err("Can't donate prior to front\n");
-				goto bad;
-			}
+		if (*notes & COPY_TO_CACHE)
+			set_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags);
 
-			prev = list_prev_entry(subreq, rreq_link);
-			WRITE_ONCE(prev->next_donated, prev->next_donated + excess);
-			subreq->start += excess;
-			subreq->len -= excess;
-			subreq->transferred -= excess;
-			trace_netfs_donate(rreq, subreq, prev, excess,
-					   netfs_trace_donate_tail_to_prev);
-			trace_netfs_sreq(subreq, netfs_sreq_trace_donate_to_prev);
-
-			if (subreq->consumed >= subreq->len)
-				goto remove_subreq_locked;
-			spin_unlock_bh(&rreq->lock);
-		} else {
-			pr_err("fpos > start\n");
-			goto bad;
-		}
+		folio = folioq_folio(folioq, slot);
+		if (WARN_ONCE(!folio_test_locked(folio),
+			      "R=%08x: folio %lx is not locked\n",
+			      rreq->debug_id, folio->index))
+			trace_netfs_folio(folio, netfs_folio_trace_not_locked);
+
+		order = folioq_folio_order(folioq, slot);
+		rreq->front_folio_order = order;
+		fsize = PAGE_SIZE << order;
+		fpos = folio_pos(folio);
+		fend = umin(fpos + fsize, rreq->i_size);
+
+		trace_netfs_collect_folio(rreq, folio, fend, collected_to);
 
-		/* Advance the rolling buffer to the next folio. */
+		/* Unlock any folio we've transferred all of. */
+		if (collected_to < fend)
+			break;
+
+		netfs_unlock_read_folio(rreq, folioq, slot);
+		WRITE_ONCE(rreq->cleaned_to, fpos + fsize);
+		*notes |= MADE_PROGRESS;
+
+		clear_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags);
+
+		/* Clean up the head folioq.  If we clear an entire folioq, then
+		 * we can get rid of it provided it's not also the tail folioq
+		 * being filled by the issuer.
+		 */
+		folioq_clear(folioq, slot);
 		slot++;
 		if (slot >= folioq_nr_slots(folioq)) {
+			folioq = rolling_buffer_delete_spent(&rreq->buffer);
+			if (!folioq)
+				goto done;
 			slot = 0;
-			folioq = folioq->next;
-			subreq->curr_folioq = folioq;
+			trace_netfs_folioq(folioq, netfs_trace_folioq_read_progress);
 		}
-		subreq->curr_folioq_slot = slot;
-		if (folioq && folioq_folio(folioq, slot))
-			subreq->curr_folio_order = folioq->orders[slot];
-		if (!was_async)
-			cond_resched();
-		goto next_folio;
+
+		if (fpos + fsize >= collected_to)
+			break;
 	}
 
-	/* Deal with partial progress. */
-	if (subreq->transferred < subreq->len)
-		return false;
+	rreq->buffer.tail = folioq;
+done:
+	rreq->buffer.first_tail_slot = slot;
+}
 
-	/* Donate the remaining downloaded data to one of the neighbouring
-	 * subrequests.  Note that we may race with them doing the same thing.
+/*
+ * Collect and assess the results of various read subrequests.  We may need to
+ * retry some of the results.
+ *
+ * Note that we have a sequence of subrequests, which may be drawing on
+ * different sources and may or may not be the same size or starting position
+ * and may not even correspond in boundary alignment.
+ */
+static void netfs_collect_read_results(struct netfs_io_request *rreq)
+{
+	struct netfs_io_subrequest *front, *remove;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	unsigned int notes;
+
+	_enter("%llx-%llx", rreq->start, rreq->start + rreq->len);
+	trace_netfs_rreq(rreq, netfs_rreq_trace_collect);
+	trace_netfs_collect(rreq);
+
+reassess:
+	if (rreq->origin == NETFS_READAHEAD ||
+	    rreq->origin == NETFS_READPAGE ||
+	    rreq->origin == NETFS_READ_FOR_WRITE)
+		notes = BUFFERED;
+	else
+		notes = 0;
+
+	/* Remove completed subrequests from the front of the stream and
+	 * advance the completion point.  We stop when we hit something that's
+	 * in progress.  The issuer thread may be adding stuff to the tail
+	 * whilst we're doing this.
 	 */
-	spin_lock_bh(&rreq->lock);
+	front = READ_ONCE(stream->front);
+	while (front) {
+		size_t transferred;
 
-	if (subreq->prev_donated != prev_donated ||
-	    subreq->next_donated != next_donated) {
-		spin_unlock_bh(&rreq->lock);
-		cond_resched();
-		goto donation_changed;
-	}
+		trace_netfs_collect_sreq(rreq, front);
+		_debug("sreq [%x] %llx %zx/%zx",
+		       front->debug_index, front->start, front->transferred, front->len);
 
-	/* Deal with the trickiest case: that this subreq is in the middle of a
-	 * folio, not touching either edge, but finishes first.  In such a
-	 * case, we donate to the previous subreq, if there is one and if it is
-	 * contiguous, so that the donation is only handled when that completes
-	 * - and remove this subreq from the list.
-	 *
-	 * If the previous subreq finished first, we will have acquired their
-	 * donation and should be able to unlock folios and/or donate nextwards.
-	 */
-	if (!subreq->consumed &&
-	    !prev_donated &&
-	    !list_is_first(&subreq->rreq_link, &rreq->subrequests) &&
-	    subreq->start == prev->start + prev->len) {
-		prev = list_prev_entry(subreq, rreq_link);
-		WRITE_ONCE(prev->next_donated, prev->next_donated + subreq->len);
-		subreq->start += subreq->len;
-		subreq->len = 0;
-		subreq->transferred = 0;
-		trace_netfs_donate(rreq, subreq, prev, subreq->len,
-				   netfs_trace_donate_to_prev);
-		trace_netfs_sreq(subreq, netfs_sreq_trace_donate_to_prev);
-		goto remove_subreq_locked;
+		if (stream->collected_to < front->start) {
+			trace_netfs_collect_gap(rreq, stream, front->start, 'F');
+			stream->collected_to = front->start;
+		}
+
+		if (test_bit(NETFS_SREQ_IN_PROGRESS, &front->flags))
+			notes |= HIT_PENDING;
+		smp_rmb(); /* Read counters after IN_PROGRESS flag. */
+		transferred = READ_ONCE(front->transferred);
+
+		/* If we can now collect the next folio, do so.  We don't want
+		 * to defer this as we have to decide whether we need to copy
+		 * to the cache or not, and that may differ between adjacent
+		 * subreqs.
+		 */
+		if (notes & BUFFERED) {
+			size_t fsize = PAGE_SIZE << rreq->front_folio_order;
+
+			/* Clear the tail of a short read. */
+			if (!(notes & HIT_PENDING) &&
+			    front->error == 0 &&
+			    transferred < front->len &&
+			    (test_bit(NETFS_SREQ_HIT_EOF, &front->flags) ||
+			     test_bit(NETFS_SREQ_CLEAR_TAIL, &front->flags))) {
+				netfs_clear_unread(front);
+				transferred = front->transferred = front->len;
+				trace_netfs_sreq(front, netfs_sreq_trace_clear);
+			}
+
+			stream->collected_to = front->start + transferred;
+			rreq->collected_to = stream->collected_to;
+
+			if (test_bit(NETFS_SREQ_COPY_TO_CACHE, &front->flags))
+				notes |= COPY_TO_CACHE;
+
+			if (test_bit(NETFS_SREQ_FAILED, &front->flags)) {
+				rreq->abandon_to = front->start + front->len;
+				front->transferred = front->len;
+				transferred = front->len;
+				trace_netfs_rreq(rreq, netfs_rreq_trace_set_abandon);
+			}
+			if (front->start + transferred >= rreq->cleaned_to + fsize ||
+			    test_bit(NETFS_SREQ_HIT_EOF, &front->flags))
+				netfs_read_unlock_folios(rreq, &notes);
+		} else {
+			stream->collected_to = front->start + transferred;
+			rreq->collected_to = stream->collected_to;
+		}
+
+		/* Stall if the front is still undergoing I/O. */
+		if (notes & HIT_PENDING)
+			break;
+
+		if (test_bit(NETFS_SREQ_FAILED, &front->flags)) {
+			if (!stream->failed) {
+				stream->error = front->error;
+				rreq->error = front->error;
+				set_bit(NETFS_RREQ_FAILED, &rreq->flags);
+				stream->failed = true;
+			}
+			notes |= MADE_PROGRESS | ABANDON_SREQ;
+		} else if (test_bit(NETFS_SREQ_NEED_RETRY, &front->flags)) {
+			stream->need_retry = true;
+			notes |= NEED_RETRY | MADE_PROGRESS;
+			break;
+		} else {
+			if (!stream->failed)
+				stream->transferred = stream->collected_to - rreq->start;
+			notes |= MADE_PROGRESS;
+		}
+
+		/* Remove if completely consumed. */
+		stream->source = front->source;
+		spin_lock(&rreq->lock);
+
+		remove = front;
+		trace_netfs_sreq(front, netfs_sreq_trace_discard);
+		list_del_init(&front->rreq_link);
+		front = list_first_entry_or_null(&stream->subrequests,
+						 struct netfs_io_subrequest, rreq_link);
+		stream->front = front;
+		spin_unlock(&rreq->lock);
+		netfs_put_subrequest(remove, false,
+				     notes & ABANDON_SREQ ?
+				     netfs_sreq_trace_put_abandon :
+				     netfs_sreq_trace_put_done);
 	}
 
-	/* If we can't donate down the chain, donate up the chain instead. */
-	excess = subreq->len - subreq->consumed + next_donated;
+	trace_netfs_collect_stream(rreq, stream);
+	trace_netfs_collect_state(rreq, rreq->collected_to, notes);
 
-	if (!subreq->consumed)
-		excess += prev_donated;
+	if (!(notes & BUFFERED))
+		rreq->cleaned_to = rreq->collected_to;
 
-	if (list_is_last(&subreq->rreq_link, &rreq->subrequests)) {
-		rreq->prev_donated = excess;
-		trace_netfs_donate(rreq, subreq, NULL, excess,
-				   netfs_trace_donate_to_deferred_next);
-	} else {
-		next = list_next_entry(subreq, rreq_link);
-		WRITE_ONCE(next->prev_donated, excess);
-		trace_netfs_donate(rreq, subreq, next, excess,
-				   netfs_trace_donate_to_next);
+	if (notes & NEED_RETRY)
+		goto need_retry;
+	if ((notes & MADE_PROGRESS) && test_bit(NETFS_RREQ_PAUSE, &rreq->flags)) {
+		trace_netfs_rreq(rreq, netfs_rreq_trace_unpause);
+		clear_bit_unlock(NETFS_RREQ_PAUSE, &rreq->flags);
+		smp_mb__after_atomic(); /* Set PAUSE before task state */
+		wake_up(&rreq->waitq);
+	}
+
+	if (notes & MADE_PROGRESS) {
+		//cond_resched();
+		goto reassess;
 	}
-	trace_netfs_sreq(subreq, netfs_sreq_trace_donate_to_next);
-	subreq->len = subreq->consumed;
-	subreq->transferred = subreq->consumed;
-	goto remove_subreq_locked;
-
-remove_subreq:
-	spin_lock_bh(&rreq->lock);
-remove_subreq_locked:
-	subreq->consumed = subreq->len;
-	list_del(&subreq->rreq_link);
-	spin_unlock_bh(&rreq->lock);
-	netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_consumed);
-	return true;
-
-bad:
-	/* Errr... prev and next both donated to us, but insufficient to finish
-	 * the folio.
+
+out:
+	_leave(" = %x", notes);
+	return;
+
+need_retry:
+	/* Okay...  We're going to have to retry parts of the stream.  Note
+	 * that any partially completed op will have had any wholly transferred
+	 * folios removed from it.
 	 */
-	printk("R=%08x[%x] s=%llx-%llx %zx/%zx/%zx\n",
-	       rreq->debug_id, subreq->debug_index,
-	       subreq->start, subreq->start + subreq->transferred - 1,
-	       subreq->consumed, subreq->transferred, subreq->len);
-	printk("folio: %llx-%llx\n", fpos, fend - 1);
-	printk("donated: prev=%zx next=%zx\n", prev_donated, next_donated);
-	printk("s=%llx av=%zx part=%zx\n", start, avail, part);
-	BUG();
+	_debug("retry");
+	netfs_retry_reads(rreq);
+	goto out;
 }
 
 /*
@@ -321,12 +343,13 @@ bad:
 static void netfs_rreq_assess_dio(struct netfs_io_request *rreq)
 {
 	struct netfs_io_subrequest *subreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
 	unsigned int i;
 
 	/* Collect unbuffered reads and direct reads, adding up the transfer
 	 * sizes until we find the first short or failed subrequest.
 	 */
-	list_for_each_entry(subreq, &rreq->subrequests, rreq_link) {
+	list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
 		rreq->transferred += subreq->transferred;
 
 		if (subreq->transferred < subreq->len ||
@@ -359,25 +382,67 @@ static void netfs_rreq_assess_dio(struct netfs_io_request *rreq)
 }
 
 /*
- * Assess the state of a read request and decide what to do next.
+ * Do processing after reading a monolithic single object.
+ */
+static void netfs_rreq_assess_single(struct netfs_io_request *rreq)
+{
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+
+	if (!rreq->error && stream->source == NETFS_DOWNLOAD_FROM_SERVER &&
+	    fscache_resources_valid(&rreq->cache_resources)) {
+		trace_netfs_rreq(rreq, netfs_rreq_trace_dirty);
+		netfs_single_mark_inode_dirty(rreq->inode);
+	}
+
+	if (rreq->iocb) {
+		rreq->iocb->ki_pos += rreq->transferred;
+		if (rreq->iocb->ki_complete)
+			rreq->iocb->ki_complete(
+				rreq->iocb, rreq->error ? rreq->error : rreq->transferred);
+	}
+	if (rreq->netfs_ops->done)
+		rreq->netfs_ops->done(rreq);
+}
+
+/*
+ * Perform the collection of subrequests and folios.
  *
  * Note that we're in normal kernel thread context at this point, possibly
  * running on a workqueue.
  */
-static void netfs_rreq_assess(struct netfs_io_request *rreq)
+static void netfs_read_collection(struct netfs_io_request *rreq)
 {
-	trace_netfs_rreq(rreq, netfs_rreq_trace_assess);
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
 
-	//netfs_rreq_is_still_valid(rreq);
+	netfs_collect_read_results(rreq);
 
-	if (test_and_clear_bit(NETFS_RREQ_NEED_RETRY, &rreq->flags)) {
-		netfs_retry_reads(rreq);
+	/* We're done when the app thread has finished posting subreqs and the
+	 * queue is empty.
+	 */
+	if (!test_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags))
+		return;
+	smp_rmb(); /* Read ALL_QUEUED before subreq lists. */
+
+	if (!list_empty(&stream->subrequests))
 		return;
-	}
 
-	if (rreq->origin == NETFS_DIO_READ ||
-	    rreq->origin == NETFS_READ_GAPS)
+	/* Okay, declare that all I/O is complete. */
+	rreq->transferred = stream->transferred;
+	trace_netfs_rreq(rreq, netfs_rreq_trace_complete);
+
+	//netfs_rreq_is_still_valid(rreq);
+
+	switch (rreq->origin) {
+	case NETFS_DIO_READ:
+	case NETFS_READ_GAPS:
 		netfs_rreq_assess_dio(rreq);
+		break;
+	case NETFS_READ_SINGLE:
+		netfs_rreq_assess_single(rreq);
+		break;
+	default:
+		break;
+	}
 	task_io_account_read(rreq->transferred);
 
 	trace_netfs_rreq(rreq, netfs_rreq_trace_wake_ip);
@@ -386,61 +451,66 @@ static void netfs_rreq_assess(struct netfs_io_request *rreq)
 	trace_netfs_rreq(rreq, netfs_rreq_trace_done);
 	netfs_clear_subrequests(rreq, false);
 	netfs_unlock_abandoned_read_pages(rreq);
-	if (unlikely(test_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags)))
-		netfs_pgpriv2_write_to_the_cache(rreq);
+	if (unlikely(rreq->copy_to_cache))
+		netfs_pgpriv2_end_copy_to_cache(rreq);
 }
 
-void netfs_read_termination_worker(struct work_struct *work)
+void netfs_read_collection_worker(struct work_struct *work)
 {
-	struct netfs_io_request *rreq =
-		container_of(work, struct netfs_io_request, work);
+	struct netfs_io_request *rreq = container_of(work, struct netfs_io_request, work);
+
 	netfs_see_request(rreq, netfs_rreq_trace_see_work);
-	netfs_rreq_assess(rreq);
-	netfs_put_request(rreq, false, netfs_rreq_trace_put_work_complete);
+	if (test_bit(NETFS_RREQ_IN_PROGRESS, &rreq->flags))
+		netfs_read_collection(rreq);
+	netfs_put_request(rreq, false, netfs_rreq_trace_put_work);
 }
 
 /*
- * Handle the completion of all outstanding I/O operations on a read request.
- * We inherit a ref from the caller.
+ * Wake the collection work item.
  */
-void netfs_rreq_terminated(struct netfs_io_request *rreq, bool was_async)
+void netfs_wake_read_collector(struct netfs_io_request *rreq)
 {
-	if (!was_async)
-		return netfs_rreq_assess(rreq);
-	if (!work_pending(&rreq->work)) {
-		netfs_get_request(rreq, netfs_rreq_trace_get_work);
-		if (!queue_work(system_unbound_wq, &rreq->work))
-			netfs_put_request(rreq, was_async, netfs_rreq_trace_put_work_nq);
+	if (test_bit(NETFS_RREQ_OFFLOAD_COLLECTION, &rreq->flags)) {
+		if (!work_pending(&rreq->work)) {
+			netfs_get_request(rreq, netfs_rreq_trace_get_work);
+			if (!queue_work(system_unbound_wq, &rreq->work))
+				netfs_put_request(rreq, true, netfs_rreq_trace_put_work_nq);
+		}
+	} else {
+		trace_netfs_rreq(rreq, netfs_rreq_trace_wake_queue);
+		wake_up(&rreq->waitq);
 	}
 }
 
 /**
  * netfs_read_subreq_progress - Note progress of a read operation.
  * @subreq: The read request that has terminated.
- * @was_async: True if we're in an asynchronous context.
  *
  * This tells the read side of netfs lib that a contributory I/O operation has
  * made some progress and that it may be possible to unlock some folios.
  *
  * Before calling, the filesystem should update subreq->transferred to track
  * the amount of data copied into the output buffer.
- *
- * If @was_async is true, the caller might be running in softirq or interrupt
- * context and we can't sleep.
  */
-void netfs_read_subreq_progress(struct netfs_io_subrequest *subreq,
-				bool was_async)
+void netfs_read_subreq_progress(struct netfs_io_subrequest *subreq)
 {
 	struct netfs_io_request *rreq = subreq->rreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	size_t fsize = PAGE_SIZE << rreq->front_folio_order;
 
 	trace_netfs_sreq(subreq, netfs_sreq_trace_progress);
 
-	if (subreq->transferred > subreq->consumed &&
+	/* If we are at the head of the queue, wake up the collector,
+	 * getting a ref to it if we were the ones to do so.
+	 */
+	if (subreq->start + subreq->transferred > rreq->cleaned_to + fsize &&
 	    (rreq->origin == NETFS_READAHEAD ||
 	     rreq->origin == NETFS_READPAGE ||
-	     rreq->origin == NETFS_READ_FOR_WRITE)) {
-		netfs_consume_read_data(subreq, was_async);
+	     rreq->origin == NETFS_READ_FOR_WRITE) &&
+	    list_is_first(&subreq->rreq_link, &stream->subrequests)
+	    ) {
 		__set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
+		netfs_wake_read_collector(rreq);
 	}
 }
 EXPORT_SYMBOL(netfs_read_subreq_progress);
@@ -448,27 +518,23 @@ EXPORT_SYMBOL(netfs_read_subreq_progress);
 /**
  * netfs_read_subreq_terminated - Note the termination of an I/O operation.
  * @subreq: The I/O request that has terminated.
- * @error: Error code indicating type of completion.
- * @was_async: The termination was asynchronous
  *
  * This tells the read helper that a contributory I/O operation has terminated,
  * one way or another, and that it should integrate the results.
  *
- * The caller indicates the outcome of the operation through @error, supplying
- * 0 to indicate a successful or retryable transfer (if NETFS_SREQ_NEED_RETRY
- * is set) or a negative error code.  The helper will look after reissuing I/O
- * operations as appropriate and writing downloaded data to the cache.
+ * The caller indicates the outcome of the operation through @subreq->error,
+ * supplying 0 to indicate a successful or retryable transfer (if
+ * NETFS_SREQ_NEED_RETRY is set) or a negative error code.  The helper will
+ * look after reissuing I/O operations as appropriate and writing downloaded
+ * data to the cache.
  *
  * Before calling, the filesystem should update subreq->transferred to track
  * the amount of data copied into the output buffer.
- *
- * If @was_async is true, the caller might be running in softirq or interrupt
- * context and we can't sleep.
  */
-void netfs_read_subreq_terminated(struct netfs_io_subrequest *subreq,
-				  int error, bool was_async)
+void netfs_read_subreq_terminated(struct netfs_io_subrequest *subreq)
 {
 	struct netfs_io_request *rreq = subreq->rreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
 
 	switch (subreq->source) {
 	case NETFS_READ_FROM_CACHE:
@@ -481,71 +547,156 @@ void netfs_read_subreq_terminated(struct netfs_io_subrequest *subreq,
 		break;
 	}
 
-	if (rreq->origin != NETFS_DIO_READ) {
-		/* Collect buffered reads.
-		 *
-		 * If the read completed validly short, then we can clear the
-		 * tail before going on to unlock the folios.
-		 */
-		if (error == 0 && subreq->transferred < subreq->len &&
-		    (test_bit(NETFS_SREQ_HIT_EOF, &subreq->flags) ||
-		     test_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags))) {
-			netfs_clear_unread(subreq);
-			subreq->transferred = subreq->len;
-			trace_netfs_sreq(subreq, netfs_sreq_trace_clear);
-		}
-		if (subreq->transferred > subreq->consumed &&
-		    (rreq->origin == NETFS_READAHEAD ||
-		     rreq->origin == NETFS_READPAGE ||
-		     rreq->origin == NETFS_READ_FOR_WRITE)) {
-			netfs_consume_read_data(subreq, was_async);
-			__set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
-		}
-		rreq->transferred += subreq->transferred;
-	}
-
 	/* Deal with retry requests, short reads and errors.  If we retry
 	 * but don't make progress, we abandon the attempt.
 	 */
-	if (!error && subreq->transferred < subreq->len) {
+	if (!subreq->error && subreq->transferred < subreq->len) {
 		if (test_bit(NETFS_SREQ_HIT_EOF, &subreq->flags)) {
 			trace_netfs_sreq(subreq, netfs_sreq_trace_hit_eof);
+		} else if (test_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags)) {
+			trace_netfs_sreq(subreq, netfs_sreq_trace_need_clear);
+		} else if (test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
+			trace_netfs_sreq(subreq, netfs_sreq_trace_need_retry);
+		} else if (test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags)) {
+			__set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
+			trace_netfs_sreq(subreq, netfs_sreq_trace_partial_read);
 		} else {
+			__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
+			subreq->error = -ENODATA;
 			trace_netfs_sreq(subreq, netfs_sreq_trace_short);
-			if (subreq->transferred > subreq->consumed) {
-				/* If we didn't read new data, abandon retry. */
-				if (subreq->retry_count &&
-				    test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags)) {
-					__set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
-					set_bit(NETFS_RREQ_NEED_RETRY, &rreq->flags);
-				}
-			} else if (test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags)) {
-				__set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
-				set_bit(NETFS_RREQ_NEED_RETRY, &rreq->flags);
-			} else {
-				__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
-				error = -ENODATA;
-			}
 		}
 	}
 
-	subreq->error = error;
-	trace_netfs_sreq(subreq, netfs_sreq_trace_terminated);
-
-	if (unlikely(error < 0)) {
-		trace_netfs_failure(rreq, subreq, error, netfs_fail_read);
+	if (unlikely(subreq->error < 0)) {
+		trace_netfs_failure(rreq, subreq, subreq->error, netfs_fail_read);
 		if (subreq->source == NETFS_READ_FROM_CACHE) {
 			netfs_stat(&netfs_n_rh_read_failed);
+			__set_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
 		} else {
 			netfs_stat(&netfs_n_rh_download_failed);
-			set_bit(NETFS_RREQ_FAILED, &rreq->flags);
-			rreq->error = subreq->error;
+			__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
 		}
+		trace_netfs_rreq(rreq, netfs_rreq_trace_set_pause);
+		set_bit(NETFS_RREQ_PAUSE, &rreq->flags);
 	}
 
-	if (atomic_dec_and_test(&rreq->nr_outstanding))
-		netfs_rreq_terminated(rreq, was_async);
+	trace_netfs_sreq(subreq, netfs_sreq_trace_terminated);
+
+	clear_bit_unlock(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
+	smp_mb__after_atomic(); /* Clear IN_PROGRESS before task state */
+
+	/* If we are at the head of the queue, wake up the collector. */
+	if (list_is_first(&subreq->rreq_link, &stream->subrequests))
+		netfs_wake_read_collector(rreq);
 
-	netfs_put_subrequest(subreq, was_async, netfs_sreq_trace_put_terminated);
+	netfs_put_subrequest(subreq, true, netfs_sreq_trace_put_terminated);
 }
 EXPORT_SYMBOL(netfs_read_subreq_terminated);
+
+/*
+ * Handle termination of a read from the cache.
+ */
+void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error, bool was_async)
+{
+	struct netfs_io_subrequest *subreq = priv;
+
+	if (transferred_or_error > 0) {
+		subreq->error = 0;
+		if (transferred_or_error > 0) {
+			subreq->transferred += transferred_or_error;
+			__set_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
+		}
+	} else {
+		subreq->error = transferred_or_error;
+	}
+	netfs_read_subreq_terminated(subreq);
+}
+
+/*
+ * Wait for the read operation to complete, successfully or otherwise.
+ */
+ssize_t netfs_wait_for_read(struct netfs_io_request *rreq)
+{
+	struct netfs_io_subrequest *subreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	DEFINE_WAIT(myself);
+	ssize_t ret;
+
+	for (;;) {
+		trace_netfs_rreq(rreq, netfs_rreq_trace_wait_queue);
+		prepare_to_wait(&rreq->waitq, &myself, TASK_UNINTERRUPTIBLE);
+
+		subreq = list_first_entry_or_null(&stream->subrequests,
+						  struct netfs_io_subrequest, rreq_link);
+		if (subreq &&
+		    (!test_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags) ||
+		     test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags))) {
+			__set_current_state(TASK_RUNNING);
+			netfs_read_collection(rreq);
+			continue;
+		}
+
+		if (!test_bit(NETFS_RREQ_IN_PROGRESS, &rreq->flags))
+			break;
+
+		schedule();
+		trace_netfs_rreq(rreq, netfs_rreq_trace_woke_queue);
+	}
+
+	finish_wait(&rreq->waitq, &myself);
+
+	ret = rreq->error;
+	if (ret == 0) {
+		ret = rreq->transferred;
+		switch (rreq->origin) {
+		case NETFS_DIO_READ:
+		case NETFS_READ_SINGLE:
+			ret = rreq->transferred;
+			break;
+		default:
+			if (rreq->submitted < rreq->len) {
+				trace_netfs_failure(rreq, NULL, ret, netfs_fail_short_read);
+				ret = -EIO;
+			}
+			break;
+		}
+	}
+
+	return ret;
+}
+
+/*
+ * Wait for a paused read operation to unpause or complete in some manner.
+ */
+void netfs_wait_for_pause(struct netfs_io_request *rreq)
+{
+	struct netfs_io_subrequest *subreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	DEFINE_WAIT(myself);
+
+	trace_netfs_rreq(rreq, netfs_rreq_trace_wait_pause);
+
+	for (;;) {
+		trace_netfs_rreq(rreq, netfs_rreq_trace_wait_queue);
+		prepare_to_wait(&rreq->waitq, &myself, TASK_UNINTERRUPTIBLE);
+
+		subreq = list_first_entry_or_null(&stream->subrequests,
+						  struct netfs_io_subrequest, rreq_link);
+		if (subreq &&
+		    (!test_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags) ||
+		     test_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags))) {
+			__set_current_state(TASK_RUNNING);
+			netfs_read_collection(rreq);
+			continue;
+		}
+
+		if (!test_bit(NETFS_RREQ_IN_PROGRESS, &rreq->flags) ||
+		    !test_bit(NETFS_RREQ_PAUSE, &rreq->flags))
+			break;
+
+		schedule();
+		trace_netfs_rreq(rreq, netfs_rreq_trace_woke_queue);
+	}
+
+	finish_wait(&rreq->waitq, &myself);
+}
diff --git a/fs/netfs/read_pgpriv2.c b/fs/netfs/read_pgpriv2.c
index 54d5004fec18..cf7727060215 100644
--- a/fs/netfs/read_pgpriv2.c
+++ b/fs/netfs/read_pgpriv2.c
@@ -14,52 +14,11 @@
 #include "internal.h"
 
 /*
- * [DEPRECATED] Mark page as requiring copy-to-cache using PG_private_2.  The
- * third mark in the folio queue is used to indicate that this folio needs
- * writing.
- */
-void netfs_pgpriv2_mark_copy_to_cache(struct netfs_io_subrequest *subreq,
-				      struct netfs_io_request *rreq,
-				      struct folio_queue *folioq,
-				      int slot)
-{
-	struct folio *folio = folioq_folio(folioq, slot);
-
-	trace_netfs_folio(folio, netfs_folio_trace_copy_to_cache);
-	folio_start_private_2(folio);
-	folioq_mark3(folioq, slot);
-}
-
-/*
- * [DEPRECATED] Cancel PG_private_2 on all marked folios in the event of an
- * unrecoverable error.
- */
-static void netfs_pgpriv2_cancel(struct folio_queue *folioq)
-{
-	struct folio *folio;
-	int slot;
-
-	while (folioq) {
-		if (!folioq->marks3) {
-			folioq = folioq->next;
-			continue;
-		}
-
-		slot = __ffs(folioq->marks3);
-		folio = folioq_folio(folioq, slot);
-
-		trace_netfs_folio(folio, netfs_folio_trace_cancel_copy);
-		folio_end_private_2(folio);
-		folioq_unmark3(folioq, slot);
-	}
-}
-
-/*
  * [DEPRECATED] Copy a folio to the cache with PG_private_2 set.
  */
-static int netfs_pgpriv2_copy_folio(struct netfs_io_request *wreq, struct folio *folio)
+static void netfs_pgpriv2_copy_folio(struct netfs_io_request *creq, struct folio *folio)
 {
-	struct netfs_io_stream *cache  = &wreq->io_streams[1];
+	struct netfs_io_stream *cache = &creq->io_streams[1];
 	size_t fsize = folio_size(folio), flen = fsize;
 	loff_t fpos = folio_pos(folio), i_size;
 	bool to_eof = false;
@@ -70,17 +29,17 @@ static int netfs_pgpriv2_copy_folio(struct netfs_io_request *wreq, struct folio
 	 * of the page to beyond it, but cannot move i_size into or through the
 	 * page since we have it locked.
 	 */
-	i_size = i_size_read(wreq->inode);
+	i_size = i_size_read(creq->inode);
 
 	if (fpos >= i_size) {
 		/* mmap beyond eof. */
 		_debug("beyond eof");
 		folio_end_private_2(folio);
-		return 0;
+		return;
 	}
 
-	if (fpos + fsize > wreq->i_size)
-		wreq->i_size = i_size;
+	if (fpos + fsize > creq->i_size)
+		creq->i_size = i_size;
 
 	if (flen > i_size - fpos) {
 		flen = i_size - fpos;
@@ -94,8 +53,10 @@ static int netfs_pgpriv2_copy_folio(struct netfs_io_request *wreq, struct folio
 	trace_netfs_folio(folio, netfs_folio_trace_store_copy);
 
 	/* Attach the folio to the rolling buffer. */
-	if (netfs_buffer_append_folio(wreq, folio, false) < 0)
-		return -ENOMEM;
+	if (rolling_buffer_append(&creq->buffer, folio, 0) < 0) {
+		clear_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &creq->flags);
+		return;
+	}
 
 	cache->submit_extendable_to = fsize;
 	cache->submit_off = 0;
@@ -109,11 +70,11 @@ static int netfs_pgpriv2_copy_folio(struct netfs_io_request *wreq, struct folio
 	do {
 		ssize_t part;
 
-		wreq->io_iter.iov_offset = cache->submit_off;
+		creq->buffer.iter.iov_offset = cache->submit_off;
 
-		atomic64_set(&wreq->issued_to, fpos + cache->submit_off);
+		atomic64_set(&creq->issued_to, fpos + cache->submit_off);
 		cache->submit_extendable_to = fsize - cache->submit_off;
-		part = netfs_advance_write(wreq, cache, fpos + cache->submit_off,
+		part = netfs_advance_write(creq, cache, fpos + cache->submit_off,
 					   cache->submit_len, to_eof);
 		cache->submit_off += part;
 		if (part > cache->submit_len)
@@ -122,98 +83,95 @@ static int netfs_pgpriv2_copy_folio(struct netfs_io_request *wreq, struct folio
 			cache->submit_len -= part;
 	} while (cache->submit_len > 0);
 
-	wreq->io_iter.iov_offset = 0;
-	iov_iter_advance(&wreq->io_iter, fsize);
-	atomic64_set(&wreq->issued_to, fpos + fsize);
+	creq->buffer.iter.iov_offset = 0;
+	rolling_buffer_advance(&creq->buffer, fsize);
+	atomic64_set(&creq->issued_to, fpos + fsize);
 
 	if (flen < fsize)
-		netfs_issue_write(wreq, cache);
-
-	_leave(" = 0");
-	return 0;
+		netfs_issue_write(creq, cache);
 }
 
 /*
- * [DEPRECATED] Go through the buffer and write any folios that are marked with
- * the third mark to the cache.
+ * [DEPRECATED] Set up copying to the cache.
  */
-void netfs_pgpriv2_write_to_the_cache(struct netfs_io_request *rreq)
+static struct netfs_io_request *netfs_pgpriv2_begin_copy_to_cache(
+	struct netfs_io_request *rreq, struct folio *folio)
 {
-	struct netfs_io_request *wreq;
-	struct folio_queue *folioq;
-	struct folio *folio;
-	int error = 0;
-	int slot = 0;
-
-	_enter("");
+	struct netfs_io_request *creq;
 
 	if (!fscache_resources_valid(&rreq->cache_resources))
-		goto couldnt_start;
+		goto cancel;
 
-	/* Need the first folio to be able to set up the op. */
-	for (folioq = rreq->buffer; folioq; folioq = folioq->next) {
-		if (folioq->marks3) {
-			slot = __ffs(folioq->marks3);
-			break;
-		}
-	}
-	if (!folioq)
-		return;
-	folio = folioq_folio(folioq, slot);
-
-	wreq = netfs_create_write_req(rreq->mapping, NULL, folio_pos(folio),
+	creq = netfs_create_write_req(rreq->mapping, NULL, folio_pos(folio),
 				      NETFS_PGPRIV2_COPY_TO_CACHE);
-	if (IS_ERR(wreq)) {
-		kleave(" [create %ld]", PTR_ERR(wreq));
-		goto couldnt_start;
-	}
+	if (IS_ERR(creq))
+		goto cancel;
 
-	trace_netfs_write(wreq, netfs_write_trace_copy_to_cache);
+	if (!creq->io_streams[1].avail)
+		goto cancel_put;
+
+	trace_netfs_write(creq, netfs_write_trace_copy_to_cache);
 	netfs_stat(&netfs_n_wh_copy_to_cache);
-	if (!wreq->io_streams[1].avail) {
-		netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
-		goto couldnt_start;
-	}
+	rreq->copy_to_cache = creq;
+	return creq;
+
+cancel_put:
+	netfs_put_request(creq, false, netfs_rreq_trace_put_return);
+cancel:
+	rreq->copy_to_cache = ERR_PTR(-ENOBUFS);
+	clear_bit(NETFS_RREQ_FOLIO_COPY_TO_CACHE, &rreq->flags);
+	return ERR_PTR(-ENOBUFS);
+}
 
-	for (;;) {
-		error = netfs_pgpriv2_copy_folio(wreq, folio);
-		if (error < 0)
-			break;
+/*
+ * [DEPRECATED] Mark page as requiring copy-to-cache using PG_private_2 and add
+ * it to the copy write request.
+ */
+void netfs_pgpriv2_copy_to_cache(struct netfs_io_request *rreq, struct folio *folio)
+{
+	struct netfs_io_request *creq = rreq->copy_to_cache;
 
-		folioq_unmark3(folioq, slot);
-		if (!folioq->marks3) {
-			folioq = folioq->next;
-			if (!folioq)
-				break;
-		}
+	if (!creq)
+		creq = netfs_pgpriv2_begin_copy_to_cache(rreq, folio);
+	if (IS_ERR(creq))
+		return;
 
-		slot = __ffs(folioq->marks3);
-		folio = folioq_folio(folioq, slot);
-	}
+	trace_netfs_folio(folio, netfs_folio_trace_copy_to_cache);
+	folio_start_private_2(folio);
+	netfs_pgpriv2_copy_folio(creq, folio);
+}
 
-	netfs_issue_write(wreq, &wreq->io_streams[1]);
+/*
+ * [DEPRECATED] End writing to the cache, flushing out any outstanding writes.
+ */
+void netfs_pgpriv2_end_copy_to_cache(struct netfs_io_request *rreq)
+{
+	struct netfs_io_request *creq = rreq->copy_to_cache;
+
+	if (IS_ERR_OR_NULL(creq))
+		return;
+
+	netfs_issue_write(creq, &creq->io_streams[1]);
 	smp_wmb(); /* Write lists before ALL_QUEUED. */
-	set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
+	set_bit(NETFS_RREQ_ALL_QUEUED, &creq->flags);
 
-	netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
-	_leave(" = %d", error);
-couldnt_start:
-	netfs_pgpriv2_cancel(rreq->buffer);
+	netfs_put_request(creq, false, netfs_rreq_trace_put_return);
+	creq->copy_to_cache = NULL;
 }
 
 /*
  * [DEPRECATED] Remove the PG_private_2 mark from any folios we've finished
  * copying.
  */
-bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq)
+bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *creq)
 {
-	struct folio_queue *folioq = wreq->buffer;
-	unsigned long long collected_to = wreq->collected_to;
-	unsigned int slot = wreq->buffer_head_slot;
+	struct folio_queue *folioq = creq->buffer.tail;
+	unsigned long long collected_to = creq->collected_to;
+	unsigned int slot = creq->buffer.first_tail_slot;
 	bool made_progress = false;
 
 	if (slot >= folioq_nr_slots(folioq)) {
-		folioq = netfs_delete_buffer_head(wreq);
+		folioq = rolling_buffer_delete_spent(&creq->buffer);
 		slot = 0;
 	}
 
@@ -225,16 +183,16 @@ bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq)
 		folio = folioq_folio(folioq, slot);
 		if (WARN_ONCE(!folio_test_private_2(folio),
 			      "R=%08x: folio %lx is not marked private_2\n",
-			      wreq->debug_id, folio->index))
+			      creq->debug_id, folio->index))
 			trace_netfs_folio(folio, netfs_folio_trace_not_under_wback);
 
 		fpos = folio_pos(folio);
 		fsize = folio_size(folio);
 		flen = fsize;
 
-		fend = min_t(unsigned long long, fpos + flen, wreq->i_size);
+		fend = min_t(unsigned long long, fpos + flen, creq->i_size);
 
-		trace_netfs_collect_folio(wreq, folio, fend, collected_to);
+		trace_netfs_collect_folio(creq, folio, fend, collected_to);
 
 		/* Unlock any folio we've transferred all of. */
 		if (collected_to < fend)
@@ -242,7 +200,7 @@ bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq)
 
 		trace_netfs_folio(folio, netfs_folio_trace_end_copy);
 		folio_end_private_2(folio);
-		wreq->cleaned_to = fpos + fsize;
+		creq->cleaned_to = fpos + fsize;
 		made_progress = true;
 
 		/* Clean up the head folioq.  If we clear an entire folioq, then
@@ -252,9 +210,9 @@ bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq)
 		folioq_clear(folioq, slot);
 		slot++;
 		if (slot >= folioq_nr_slots(folioq)) {
-			if (READ_ONCE(wreq->buffer_tail) == folioq)
-				break;
-			folioq = netfs_delete_buffer_head(wreq);
+			folioq = rolling_buffer_delete_spent(&creq->buffer);
+			if (!folioq)
+				goto done;
 			slot = 0;
 		}
 
@@ -262,7 +220,8 @@ bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq)
 			break;
 	}
 
-	wreq->buffer = folioq;
-	wreq->buffer_head_slot = slot;
+	creq->buffer.tail = folioq;
+done:
+	creq->buffer.first_tail_slot = slot;
 	return made_progress;
 }
diff --git a/fs/netfs/read_retry.c b/fs/netfs/read_retry.c
index 16b676c68dcd..2290af0d51ac 100644
--- a/fs/netfs/read_retry.c
+++ b/fs/netfs/read_retry.c
@@ -12,15 +12,7 @@
 static void netfs_reissue_read(struct netfs_io_request *rreq,
 			       struct netfs_io_subrequest *subreq)
 {
-	struct iov_iter *io_iter = &subreq->io_iter;
-
-	if (iov_iter_is_folioq(io_iter)) {
-		subreq->curr_folioq = (struct folio_queue *)io_iter->folioq;
-		subreq->curr_folioq_slot = io_iter->folioq_slot;
-		subreq->curr_folio_order = subreq->curr_folioq->orders[subreq->curr_folioq_slot];
-	}
-
-	atomic_inc(&rreq->nr_outstanding);
+	__clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
 	__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
 	netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
 	subreq->rreq->netfs_ops->issue_read(subreq);
@@ -33,13 +25,12 @@ static void netfs_reissue_read(struct netfs_io_request *rreq,
 static void netfs_retry_read_subrequests(struct netfs_io_request *rreq)
 {
 	struct netfs_io_subrequest *subreq;
-	struct netfs_io_stream *stream0 = &rreq->io_streams[0];
-	LIST_HEAD(sublist);
-	LIST_HEAD(queue);
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	struct list_head *next;
 
 	_enter("R=%x", rreq->debug_id);
 
-	if (list_empty(&rreq->subrequests))
+	if (list_empty(&stream->subrequests))
 		return;
 
 	if (rreq->netfs_ops->retry_request)
@@ -50,9 +41,7 @@ static void netfs_retry_read_subrequests(struct netfs_io_request *rreq)
 	 */
 	if (!rreq->netfs_ops->prepare_read &&
 	    !rreq->cache_resources.ops) {
-		struct netfs_io_subrequest *subreq;
-
-		list_for_each_entry(subreq, &rreq->subrequests, rreq_link) {
+		list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
 			if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
 				break;
 			if (__test_and_clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
@@ -75,48 +64,44 @@ static void netfs_retry_read_subrequests(struct netfs_io_request *rreq)
 	 * populating with smaller subrequests.  In the event that the subreq
 	 * we just launched finishes before we insert the next subreq, it'll
 	 * fill in rreq->prev_donated instead.
-
+	 *
 	 * Note: Alternatively, we could split the tail subrequest right before
 	 * we reissue it and fix up the donations under lock.
 	 */
-	list_splice_init(&rreq->subrequests, &queue);
+	next = stream->subrequests.next;
 
 	do {
-		struct netfs_io_subrequest *from;
+		struct netfs_io_subrequest *from, *to, *tmp;
 		struct iov_iter source;
 		unsigned long long start, len;
-		size_t part, deferred_next_donated = 0;
+		size_t part;
 		bool boundary = false;
 
 		/* Go through the subreqs and find the next span of contiguous
 		 * buffer that we then rejig (cifs, for example, needs the
 		 * rsize renegotiating) and reissue.
 		 */
-		from = list_first_entry(&queue, struct netfs_io_subrequest, rreq_link);
-		list_move_tail(&from->rreq_link, &sublist);
+		from = list_entry(next, struct netfs_io_subrequest, rreq_link);
+		to = from;
 		start = from->start + from->transferred;
 		len   = from->len   - from->transferred;
 
-		_debug("from R=%08x[%x] s=%llx ctl=%zx/%zx/%zx",
+		_debug("from R=%08x[%x] s=%llx ctl=%zx/%zx",
 		       rreq->debug_id, from->debug_index,
-		       from->start, from->consumed, from->transferred, from->len);
+		       from->start, from->transferred, from->len);
 
 		if (test_bit(NETFS_SREQ_FAILED, &from->flags) ||
 		    !test_bit(NETFS_SREQ_NEED_RETRY, &from->flags))
 			goto abandon;
 
-		deferred_next_donated = from->next_donated;
-		while ((subreq = list_first_entry_or_null(
-				&queue, struct netfs_io_subrequest, rreq_link))) {
-			if (subreq->start != start + len ||
-			    subreq->transferred > 0 ||
+		list_for_each_continue(next, &stream->subrequests) {
+			subreq = list_entry(next, struct netfs_io_subrequest, rreq_link);
+			if (subreq->start + subreq->transferred != start + len ||
+			    test_bit(NETFS_SREQ_BOUNDARY, &subreq->flags) ||
 			    !test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags))
 				break;
-			list_move_tail(&subreq->rreq_link, &sublist);
-			len += subreq->len;
-			deferred_next_donated = subreq->next_donated;
-			if (test_bit(NETFS_SREQ_BOUNDARY, &subreq->flags))
-				break;
+			to = subreq;
+			len += to->len;
 		}
 
 		_debug(" - range: %llx-%llx %llx", start, start + len - 1, len);
@@ -129,38 +114,31 @@ static void netfs_retry_read_subrequests(struct netfs_io_request *rreq)
 		source.count = len;
 
 		/* Work through the sublist. */
-		while ((subreq = list_first_entry_or_null(
-				&sublist, struct netfs_io_subrequest, rreq_link))) {
-			list_del(&subreq->rreq_link);
-
+		subreq = from;
+		list_for_each_entry_from(subreq, &stream->subrequests, rreq_link) {
+			if (!len)
+				break;
 			subreq->source	= NETFS_DOWNLOAD_FROM_SERVER;
 			subreq->start	= start - subreq->transferred;
 			subreq->len	= len   + subreq->transferred;
-			stream0->sreq_max_len = subreq->len;
-
 			__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
 			__clear_bit(NETFS_SREQ_MADE_PROGRESS, &subreq->flags);
 			subreq->retry_count++;
 
-			spin_lock_bh(&rreq->lock);
-			list_add_tail(&subreq->rreq_link, &rreq->subrequests);
-			subreq->prev_donated += rreq->prev_donated;
-			rreq->prev_donated = 0;
 			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
-			spin_unlock_bh(&rreq->lock);
-
-			BUG_ON(!len);
 
 			/* Renegotiate max_len (rsize) */
+			stream->sreq_max_len = subreq->len;
 			if (rreq->netfs_ops->prepare_read &&
 			    rreq->netfs_ops->prepare_read(subreq) < 0) {
 				trace_netfs_sreq(subreq, netfs_sreq_trace_reprep_failed);
 				__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
+				goto abandon;
 			}
 
-			part = umin(len, stream0->sreq_max_len);
-			if (unlikely(rreq->io_streams[0].sreq_max_segs))
-				part = netfs_limit_iter(&source, 0, part, stream0->sreq_max_segs);
+			part = umin(len, stream->sreq_max_len);
+			if (unlikely(stream->sreq_max_segs))
+				part = netfs_limit_iter(&source, 0, part, stream->sreq_max_segs);
 			subreq->len = subreq->transferred + part;
 			subreq->io_iter = source;
 			iov_iter_truncate(&subreq->io_iter, part);
@@ -170,57 +148,105 @@ static void netfs_retry_read_subrequests(struct netfs_io_request *rreq)
 			if (!len) {
 				if (boundary)
 					__set_bit(NETFS_SREQ_BOUNDARY, &subreq->flags);
-				subreq->next_donated = deferred_next_donated;
 			} else {
 				__clear_bit(NETFS_SREQ_BOUNDARY, &subreq->flags);
-				subreq->next_donated = 0;
 			}
 
+			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
 			netfs_reissue_read(rreq, subreq);
-			if (!len)
+			if (subreq == to)
 				break;
-
-			/* If we ran out of subrequests, allocate another. */
-			if (list_empty(&sublist)) {
-				subreq = netfs_alloc_subrequest(rreq);
-				if (!subreq)
-					goto abandon;
-				subreq->source = NETFS_DOWNLOAD_FROM_SERVER;
-				subreq->start = start;
-
-				/* We get two refs, but need just one. */
-				netfs_put_subrequest(subreq, false, netfs_sreq_trace_new);
-				trace_netfs_sreq(subreq, netfs_sreq_trace_split);
-				list_add_tail(&subreq->rreq_link, &sublist);
-			}
 		}
 
 		/* If we managed to use fewer subreqs, we can discard the
-		 * excess.
+		 * excess; if we used the same number, then we're done.
 		 */
-		while ((subreq = list_first_entry_or_null(
-				&sublist, struct netfs_io_subrequest, rreq_link))) {
-			trace_netfs_sreq(subreq, netfs_sreq_trace_discard);
-			list_del(&subreq->rreq_link);
-			netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_done);
+		if (!len) {
+			if (subreq == to)
+				continue;
+			list_for_each_entry_safe_from(subreq, tmp,
+						      &stream->subrequests, rreq_link) {
+				trace_netfs_sreq(subreq, netfs_sreq_trace_discard);
+				list_del(&subreq->rreq_link);
+				netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_done);
+				if (subreq == to)
+					break;
+			}
+			continue;
 		}
 
-	} while (!list_empty(&queue));
+		/* We ran out of subrequests, so we need to allocate some more
+		 * and insert them after.
+		 */
+		do {
+			subreq = netfs_alloc_subrequest(rreq);
+			if (!subreq) {
+				subreq = to;
+				goto abandon_after;
+			}
+			subreq->source		= NETFS_DOWNLOAD_FROM_SERVER;
+			subreq->start		= start;
+			subreq->len		= len;
+			subreq->debug_index	= atomic_inc_return(&rreq->subreq_counter);
+			subreq->stream_nr	= stream->stream_nr;
+			subreq->retry_count	= 1;
+
+			trace_netfs_sreq_ref(rreq->debug_id, subreq->debug_index,
+					     refcount_read(&subreq->ref),
+					     netfs_sreq_trace_new);
+			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
+
+			list_add(&subreq->rreq_link, &to->rreq_link);
+			to = list_next_entry(to, rreq_link);
+			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
+
+			stream->sreq_max_len	= umin(len, rreq->rsize);
+			stream->sreq_max_segs	= 0;
+			if (unlikely(stream->sreq_max_segs))
+				part = netfs_limit_iter(&source, 0, part, stream->sreq_max_segs);
+
+			netfs_stat(&netfs_n_rh_download);
+			if (rreq->netfs_ops->prepare_read(subreq) < 0) {
+				trace_netfs_sreq(subreq, netfs_sreq_trace_reprep_failed);
+				__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
+				goto abandon;
+			}
+
+			part = umin(len, stream->sreq_max_len);
+			subreq->len = subreq->transferred + part;
+			subreq->io_iter = source;
+			iov_iter_truncate(&subreq->io_iter, part);
+			iov_iter_advance(&source, part);
+
+			len -= part;
+			start += part;
+			if (!len && boundary) {
+				__set_bit(NETFS_SREQ_BOUNDARY, &to->flags);
+				boundary = false;
+			}
+
+			netfs_reissue_read(rreq, subreq);
+		} while (len);
+
+	} while (!list_is_head(next, &stream->subrequests));
 
 	return;
 
-	/* If we hit ENOMEM, fail all remaining subrequests */
+	/* If we hit an error, fail all remaining incomplete subrequests */
+abandon_after:
+	if (list_is_last(&subreq->rreq_link, &stream->subrequests))
+		return;
+	subreq = list_next_entry(subreq, rreq_link);
 abandon:
-	list_splice_init(&sublist, &queue);
-	list_for_each_entry(subreq, &queue, rreq_link) {
-		if (!subreq->error)
-			subreq->error = -ENOMEM;
-		__clear_bit(NETFS_SREQ_FAILED, &subreq->flags);
+	list_for_each_entry_from(subreq, &stream->subrequests, rreq_link) {
+		if (!subreq->error &&
+		    !test_bit(NETFS_SREQ_FAILED, &subreq->flags) &&
+		    !test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags))
+			continue;
+		subreq->error = -ENOMEM;
+		__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
 		__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
 	}
-	spin_lock_bh(&rreq->lock);
-	list_splice_tail_init(&queue, &rreq->subrequests);
-	spin_unlock_bh(&rreq->lock);
 }
 
 /*
@@ -228,14 +254,19 @@ abandon:
  */
 void netfs_retry_reads(struct netfs_io_request *rreq)
 {
-	trace_netfs_rreq(rreq, netfs_rreq_trace_resubmit);
+	struct netfs_io_subrequest *subreq;
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
 
-	atomic_inc(&rreq->nr_outstanding);
+	/* Wait for all outstanding I/O to quiesce before performing retries as
+	 * we may need to renegotiate the I/O sizes.
+	 */
+	list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
+		wait_on_bit(&subreq->flags, NETFS_SREQ_IN_PROGRESS,
+			    TASK_UNINTERRUPTIBLE);
+	}
 
+	trace_netfs_rreq(rreq, netfs_rreq_trace_resubmit);
 	netfs_retry_read_subrequests(rreq);
-
-	if (atomic_dec_and_test(&rreq->nr_outstanding))
-		netfs_rreq_terminated(rreq, false);
 }
 
 /*
@@ -246,7 +277,7 @@ void netfs_unlock_abandoned_read_pages(struct netfs_io_request *rreq)
 {
 	struct folio_queue *p;
 
-	for (p = rreq->buffer; p; p = p->next) {
+	for (p = rreq->buffer.tail; p; p = p->next) {
 		for (int slot = 0; slot < folioq_count(p); slot++) {
 			struct folio *folio = folioq_folio(p, slot);
 
diff --git a/fs/netfs/read_single.c b/fs/netfs/read_single.c
new file mode 100644
index 000000000000..fea0ecdecc53
--- /dev/null
+++ b/fs/netfs/read_single.c
@@ -0,0 +1,195 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/* Single, monolithic object support (e.g. AFS directory).
+ *
+ * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
+ * Written by David Howells (dhowells@redhat.com)
+ */
+
+#include <linux/export.h>
+#include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include <linux/uio.h>
+#include <linux/sched/mm.h>
+#include <linux/task_io_accounting_ops.h>
+#include <linux/netfs.h>
+#include "internal.h"
+
+/**
+ * netfs_single_mark_inode_dirty - Mark a single, monolithic object inode dirty
+ * @inode: The inode to mark
+ *
+ * Mark an inode that contains a single, monolithic object as dirty so that its
+ * writepages op will get called.  If set, the SINGLE_NO_UPLOAD flag indicates
+ * that the object will only be written to the cache and not uploaded (e.g. AFS
+ * directory contents).
+ */
+void netfs_single_mark_inode_dirty(struct inode *inode)
+{
+	struct netfs_inode *ictx = netfs_inode(inode);
+	bool cache_only = test_bit(NETFS_ICTX_SINGLE_NO_UPLOAD, &ictx->flags);
+	bool caching = fscache_cookie_enabled(netfs_i_cookie(netfs_inode(inode)));
+
+	if (cache_only && !caching)
+		return;
+
+	mark_inode_dirty(inode);
+
+	if (caching && !(inode->i_state & I_PINNING_NETFS_WB)) {
+		bool need_use = false;
+
+		spin_lock(&inode->i_lock);
+		if (!(inode->i_state & I_PINNING_NETFS_WB)) {
+			inode->i_state |= I_PINNING_NETFS_WB;
+			need_use = true;
+		}
+		spin_unlock(&inode->i_lock);
+
+		if (need_use)
+			fscache_use_cookie(netfs_i_cookie(ictx), true);
+	}
+
+}
+EXPORT_SYMBOL(netfs_single_mark_inode_dirty);
+
+static int netfs_single_begin_cache_read(struct netfs_io_request *rreq, struct netfs_inode *ctx)
+{
+	return fscache_begin_read_operation(&rreq->cache_resources, netfs_i_cookie(ctx));
+}
+
+static void netfs_single_cache_prepare_read(struct netfs_io_request *rreq,
+					    struct netfs_io_subrequest *subreq)
+{
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
+
+	if (!cres->ops) {
+		subreq->source = NETFS_DOWNLOAD_FROM_SERVER;
+		return;
+	}
+	subreq->source = cres->ops->prepare_read(subreq, rreq->i_size);
+	trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
+
+}
+
+static void netfs_single_read_cache(struct netfs_io_request *rreq,
+				    struct netfs_io_subrequest *subreq)
+{
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
+
+	_enter("R=%08x[%x]", rreq->debug_id, subreq->debug_index);
+	netfs_stat(&netfs_n_rh_read);
+	cres->ops->read(cres, subreq->start, &subreq->io_iter, NETFS_READ_HOLE_FAIL,
+			netfs_cache_read_terminated, subreq);
+}
+
+/*
+ * Perform a read to a buffer from the cache or the server.  Only a single
+ * subreq is permitted as the object must be fetched in a single transaction.
+ */
+static int netfs_single_dispatch_read(struct netfs_io_request *rreq)
+{
+	struct netfs_io_stream *stream = &rreq->io_streams[0];
+	struct netfs_io_subrequest *subreq;
+	int ret = 0;
+
+	subreq = netfs_alloc_subrequest(rreq);
+	if (!subreq)
+		return -ENOMEM;
+
+	subreq->source	= NETFS_SOURCE_UNKNOWN;
+	subreq->start	= 0;
+	subreq->len	= rreq->len;
+	subreq->io_iter	= rreq->buffer.iter;
+
+	__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
+
+	spin_lock(&rreq->lock);
+	list_add_tail(&subreq->rreq_link, &stream->subrequests);
+	trace_netfs_sreq(subreq, netfs_sreq_trace_added);
+	stream->front = subreq;
+	/* Store list pointers before active flag */
+	smp_store_release(&stream->active, true);
+	spin_unlock(&rreq->lock);
+
+	netfs_single_cache_prepare_read(rreq, subreq);
+	switch (subreq->source) {
+	case NETFS_DOWNLOAD_FROM_SERVER:
+		netfs_stat(&netfs_n_rh_download);
+		if (rreq->netfs_ops->prepare_read) {
+			ret = rreq->netfs_ops->prepare_read(subreq);
+			if (ret < 0)
+				goto cancel;
+		}
+
+		rreq->netfs_ops->issue_read(subreq);
+		rreq->submitted += subreq->len;
+		break;
+	case NETFS_READ_FROM_CACHE:
+		trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
+		netfs_single_read_cache(rreq, subreq);
+		rreq->submitted += subreq->len;
+		ret = 0;
+		break;
+	default:
+		pr_warn("Unexpected single-read source %u\n", subreq->source);
+		WARN_ON_ONCE(true);
+		ret = -EIO;
+		break;
+	}
+
+	smp_wmb(); /* Write lists before ALL_QUEUED. */
+	set_bit(NETFS_RREQ_ALL_QUEUED, &rreq->flags);
+	return ret;
+cancel:
+	netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
+	return ret;
+}
+
+/**
+ * netfs_read_single - Synchronously read a single blob of pages.
+ * @inode: The inode to read from.
+ * @file: The file we're using to read or NULL.
+ * @iter: The buffer we're reading into.
+ *
+ * Fulfil a read request for a single monolithic object by drawing data from
+ * the cache if possible, or the netfs if not.  The buffer may be larger than
+ * the file content; unused beyond the EOF will be zero-filled.  The content
+ * will be read with a single I/O request (though this may be retried).
+ *
+ * The calling netfs must initialise a netfs context contiguous to the vfs
+ * inode before calling this.
+ *
+ * This is usable whether or not caching is enabled.  If caching is enabled,
+ * the data will be stored as a single object into the cache.
+ */
+ssize_t netfs_read_single(struct inode *inode, struct file *file, struct iov_iter *iter)
+{
+	struct netfs_io_request *rreq;
+	struct netfs_inode *ictx = netfs_inode(inode);
+	ssize_t ret;
+
+	rreq = netfs_alloc_request(inode->i_mapping, file, 0, iov_iter_count(iter),
+				   NETFS_READ_SINGLE);
+	if (IS_ERR(rreq))
+		return PTR_ERR(rreq);
+
+	ret = netfs_single_begin_cache_read(rreq, ictx);
+	if (ret == -ENOMEM || ret == -EINTR || ret == -ERESTARTSYS)
+		goto cleanup_free;
+
+	netfs_stat(&netfs_n_rh_read_single);
+	trace_netfs_read(rreq, 0, rreq->len, netfs_read_trace_read_single);
+
+	rreq->buffer.iter = *iter;
+	netfs_single_dispatch_read(rreq);
+
+	ret = netfs_wait_for_read(rreq);
+	netfs_put_request(rreq, true, netfs_rreq_trace_put_return);
+	return ret;
+
+cleanup_free:
+	netfs_put_request(rreq, false, netfs_rreq_trace_put_failed);
+	return ret;
+}
+EXPORT_SYMBOL(netfs_read_single);
diff --git a/fs/netfs/rolling_buffer.c b/fs/netfs/rolling_buffer.c
new file mode 100644
index 000000000000..75d97af14b4a
--- /dev/null
+++ b/fs/netfs/rolling_buffer.c
@@ -0,0 +1,226 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/* Rolling buffer helpers
+ *
+ * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
+ * Written by David Howells (dhowells@redhat.com)
+ */
+
+#include <linux/bitops.h>
+#include <linux/pagemap.h>
+#include <linux/rolling_buffer.h>
+#include <linux/slab.h>
+#include "internal.h"
+
+static atomic_t debug_ids;
+
+/**
+ * netfs_folioq_alloc - Allocate a folio_queue struct
+ * @rreq_id: Associated debugging ID for tracing purposes
+ * @gfp: Allocation constraints
+ * @trace: Trace tag to indicate the purpose of the allocation
+ *
+ * Allocate, initialise and account the folio_queue struct and log a trace line
+ * to mark the allocation.
+ */
+struct folio_queue *netfs_folioq_alloc(unsigned int rreq_id, gfp_t gfp,
+				       unsigned int /*enum netfs_folioq_trace*/ trace)
+{
+	struct folio_queue *fq;
+
+	fq = kmalloc(sizeof(*fq), gfp);
+	if (fq) {
+		netfs_stat(&netfs_n_folioq);
+		folioq_init(fq, rreq_id);
+		fq->debug_id = atomic_inc_return(&debug_ids);
+		trace_netfs_folioq(fq, trace);
+	}
+	return fq;
+}
+EXPORT_SYMBOL(netfs_folioq_alloc);
+
+/**
+ * netfs_folioq_free - Free a folio_queue struct
+ * @folioq: The object to free
+ * @trace: Trace tag to indicate which free
+ *
+ * Free and unaccount the folio_queue struct.
+ */
+void netfs_folioq_free(struct folio_queue *folioq,
+		       unsigned int /*enum netfs_trace_folioq*/ trace)
+{
+	trace_netfs_folioq(folioq, trace);
+	netfs_stat_d(&netfs_n_folioq);
+	kfree(folioq);
+}
+EXPORT_SYMBOL(netfs_folioq_free);
+
+/*
+ * Initialise a rolling buffer.  We allocate an empty folio queue struct to so
+ * that the pointers can be independently driven by the producer and the
+ * consumer.
+ */
+int rolling_buffer_init(struct rolling_buffer *roll, unsigned int rreq_id,
+			unsigned int direction)
+{
+	struct folio_queue *fq;
+
+	fq = netfs_folioq_alloc(rreq_id, GFP_NOFS, netfs_trace_folioq_rollbuf_init);
+	if (!fq)
+		return -ENOMEM;
+
+	roll->head = fq;
+	roll->tail = fq;
+	iov_iter_folio_queue(&roll->iter, direction, fq, 0, 0, 0);
+	return 0;
+}
+
+/*
+ * Add another folio_queue to a rolling buffer if there's no space left.
+ */
+int rolling_buffer_make_space(struct rolling_buffer *roll)
+{
+	struct folio_queue *fq, *head = roll->head;
+
+	if (!folioq_full(head))
+		return 0;
+
+	fq = netfs_folioq_alloc(head->rreq_id, GFP_NOFS, netfs_trace_folioq_make_space);
+	if (!fq)
+		return -ENOMEM;
+	fq->prev = head;
+
+	roll->head = fq;
+	if (folioq_full(head)) {
+		/* Make sure we don't leave the master iterator pointing to a
+		 * block that might get immediately consumed.
+		 */
+		if (roll->iter.folioq == head &&
+		    roll->iter.folioq_slot == folioq_nr_slots(head)) {
+			roll->iter.folioq = fq;
+			roll->iter.folioq_slot = 0;
+		}
+	}
+
+	/* Make sure the initialisation is stored before the next pointer.
+	 *
+	 * [!] NOTE: After we set head->next, the consumer is at liberty to
+	 * immediately delete the old head.
+	 */
+	smp_store_release(&head->next, fq);
+	return 0;
+}
+
+/*
+ * Decant the list of folios to read into a rolling buffer.
+ */
+ssize_t rolling_buffer_load_from_ra(struct rolling_buffer *roll,
+				    struct readahead_control *ractl,
+				    struct folio_batch *put_batch)
+{
+	struct folio_queue *fq;
+	struct page **vec;
+	int nr, ix, to;
+	ssize_t size = 0;
+
+	if (rolling_buffer_make_space(roll) < 0)
+		return -ENOMEM;
+
+	fq = roll->head;
+	vec = (struct page **)fq->vec.folios;
+	nr = __readahead_batch(ractl, vec + folio_batch_count(&fq->vec),
+			       folio_batch_space(&fq->vec));
+	ix = fq->vec.nr;
+	to = ix + nr;
+	fq->vec.nr = to;
+	for (; ix < to; ix++) {
+		struct folio *folio = folioq_folio(fq, ix);
+		unsigned int order = folio_order(folio);
+
+		fq->orders[ix] = order;
+		size += PAGE_SIZE << order;
+		trace_netfs_folio(folio, netfs_folio_trace_read);
+		if (!folio_batch_add(put_batch, folio))
+			folio_batch_release(put_batch);
+	}
+	WRITE_ONCE(roll->iter.count, roll->iter.count + size);
+
+	/* Store the counter after setting the slot. */
+	smp_store_release(&roll->next_head_slot, to);
+
+	for (; ix < folioq_nr_slots(fq); ix++)
+		folioq_clear(fq, ix);
+
+	return size;
+}
+
+/*
+ * Append a folio to the rolling buffer.
+ */
+ssize_t rolling_buffer_append(struct rolling_buffer *roll, struct folio *folio,
+			      unsigned int flags)
+{
+	ssize_t size = folio_size(folio);
+	int slot;
+
+	if (rolling_buffer_make_space(roll) < 0)
+		return -ENOMEM;
+
+	slot = folioq_append(roll->head, folio);
+	if (flags & ROLLBUF_MARK_1)
+		folioq_mark(roll->head, slot);
+	if (flags & ROLLBUF_MARK_2)
+		folioq_mark2(roll->head, slot);
+
+	WRITE_ONCE(roll->iter.count, roll->iter.count + size);
+
+	/* Store the counter after setting the slot. */
+	smp_store_release(&roll->next_head_slot, slot);
+	return size;
+}
+
+/*
+ * Delete a spent buffer from a rolling queue and return the next in line.  We
+ * don't return the last buffer to keep the pointers independent, but return
+ * NULL instead.
+ */
+struct folio_queue *rolling_buffer_delete_spent(struct rolling_buffer *roll)
+{
+	struct folio_queue *spent = roll->tail, *next = READ_ONCE(spent->next);
+
+	if (!next)
+		return NULL;
+	next->prev = NULL;
+	netfs_folioq_free(spent, netfs_trace_folioq_delete);
+	roll->tail = next;
+	return next;
+}
+
+/*
+ * Clear out a rolling queue.  Folios that have mark 1 set are put.
+ */
+void rolling_buffer_clear(struct rolling_buffer *roll)
+{
+	struct folio_batch fbatch;
+	struct folio_queue *p;
+
+	folio_batch_init(&fbatch);
+
+	while ((p = roll->tail)) {
+		roll->tail = p->next;
+		for (int slot = 0; slot < folioq_count(p); slot++) {
+			struct folio *folio = folioq_folio(p, slot);
+
+			if (!folio)
+				continue;
+			if (folioq_is_marked(p, slot)) {
+				trace_netfs_folio(folio, netfs_folio_trace_put);
+				if (!folio_batch_add(&fbatch, folio))
+					folio_batch_release(&fbatch);
+			}
+		}
+
+		netfs_folioq_free(p, netfs_trace_folioq_clear);
+	}
+
+	folio_batch_release(&fbatch);
+}
diff --git a/fs/netfs/stats.c b/fs/netfs/stats.c
index 8e63516b40f6..f1af344266cc 100644
--- a/fs/netfs/stats.c
+++ b/fs/netfs/stats.c
@@ -12,6 +12,7 @@
 atomic_t netfs_n_rh_dio_read;
 atomic_t netfs_n_rh_readahead;
 atomic_t netfs_n_rh_read_folio;
+atomic_t netfs_n_rh_read_single;
 atomic_t netfs_n_rh_rreq;
 atomic_t netfs_n_rh_sreq;
 atomic_t netfs_n_rh_download;
@@ -46,10 +47,11 @@ atomic_t netfs_n_folioq;
 
 int netfs_stats_show(struct seq_file *m, void *v)
 {
-	seq_printf(m, "Reads  : DR=%u RA=%u RF=%u WB=%u WBZ=%u\n",
+	seq_printf(m, "Reads  : DR=%u RA=%u RF=%u RS=%u WB=%u WBZ=%u\n",
 		   atomic_read(&netfs_n_rh_dio_read),
 		   atomic_read(&netfs_n_rh_readahead),
 		   atomic_read(&netfs_n_rh_read_folio),
+		   atomic_read(&netfs_n_rh_read_single),
 		   atomic_read(&netfs_n_rh_write_begin),
 		   atomic_read(&netfs_n_rh_write_zskip));
 	seq_printf(m, "Writes : BW=%u WT=%u DW=%u WP=%u 2C=%u\n",
diff --git a/fs/netfs/write_collect.c b/fs/netfs/write_collect.c
index ca3a11ed9b54..294f67795f79 100644
--- a/fs/netfs/write_collect.c
+++ b/fs/netfs/write_collect.c
@@ -17,10 +17,38 @@
 #define HIT_PENDING		0x01	/* A front op was still pending */
 #define NEED_REASSESS		0x02	/* Need to loop round and reassess */
 #define MADE_PROGRESS		0x04	/* Made progress cleaning up a stream or the folio set */
-#define BUFFERED		0x08	/* The pagecache needs cleaning up */
+#define NEED_UNLOCK		0x08	/* The pagecache needs unlocking */
 #define NEED_RETRY		0x10	/* A front op requests retrying */
 #define SAW_FAILURE		0x20	/* One stream or hit a permanent failure */
 
+static void netfs_dump_request(const struct netfs_io_request *rreq)
+{
+	pr_err("Request R=%08x r=%d fl=%lx or=%x e=%ld\n",
+	       rreq->debug_id, refcount_read(&rreq->ref), rreq->flags,
+	       rreq->origin, rreq->error);
+	pr_err("  st=%llx tsl=%zx/%llx/%llx\n",
+	       rreq->start, rreq->transferred, rreq->submitted, rreq->len);
+	pr_err("  cci=%llx/%llx/%llx\n",
+	       rreq->cleaned_to, rreq->collected_to, atomic64_read(&rreq->issued_to));
+	pr_err("  iw=%pSR\n", rreq->netfs_ops->issue_write);
+	for (int i = 0; i < NR_IO_STREAMS; i++) {
+		const struct netfs_io_subrequest *sreq;
+		const struct netfs_io_stream *s = &rreq->io_streams[i];
+
+		pr_err("  str[%x] s=%x e=%d acnf=%u,%u,%u,%u\n",
+		       s->stream_nr, s->source, s->error,
+		       s->avail, s->active, s->need_retry, s->failed);
+		pr_err("  str[%x] ct=%llx t=%zx\n",
+		       s->stream_nr, s->collected_to, s->transferred);
+		list_for_each_entry(sreq, &s->subrequests, rreq_link) {
+			pr_err("  sreq[%x:%x] sc=%u s=%llx t=%zx/%zx r=%d f=%lx\n",
+			       sreq->stream_nr, sreq->debug_index, sreq->source,
+			       sreq->start, sreq->transferred, sreq->len,
+			       refcount_read(&sreq->ref), sreq->flags);
+		}
+	}
+}
+
 /*
  * Successful completion of write of a folio to the server and/or cache.  Note
  * that we are not allowed to lock the folio here on pain of deadlocking with
@@ -83,9 +111,15 @@ end_wb:
 static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
 					  unsigned int *notes)
 {
-	struct folio_queue *folioq = wreq->buffer;
+	struct folio_queue *folioq = wreq->buffer.tail;
 	unsigned long long collected_to = wreq->collected_to;
-	unsigned int slot = wreq->buffer_head_slot;
+	unsigned int slot = wreq->buffer.first_tail_slot;
+
+	if (WARN_ON_ONCE(!folioq)) {
+		pr_err("[!] Writeback unlock found empty rolling buffer!\n");
+		netfs_dump_request(wreq);
+		return;
+	}
 
 	if (wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE) {
 		if (netfs_pgpriv2_unlock_copied_folios(wreq))
@@ -94,7 +128,9 @@ static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
 	}
 
 	if (slot >= folioq_nr_slots(folioq)) {
-		folioq = netfs_delete_buffer_head(wreq);
+		folioq = rolling_buffer_delete_spent(&wreq->buffer);
+		if (!folioq)
+			return;
 		slot = 0;
 	}
 
@@ -134,9 +170,9 @@ static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
 		folioq_clear(folioq, slot);
 		slot++;
 		if (slot >= folioq_nr_slots(folioq)) {
-			if (READ_ONCE(wreq->buffer_tail) == folioq)
-				break;
-			folioq = netfs_delete_buffer_head(wreq);
+			folioq = rolling_buffer_delete_spent(&wreq->buffer);
+			if (!folioq)
+				goto done;
 			slot = 0;
 		}
 
@@ -144,222 +180,9 @@ static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
 			break;
 	}
 
-	wreq->buffer = folioq;
-	wreq->buffer_head_slot = slot;
-}
-
-/*
- * Perform retries on the streams that need it.
- */
-static void netfs_retry_write_stream(struct netfs_io_request *wreq,
-				     struct netfs_io_stream *stream)
-{
-	struct list_head *next;
-
-	_enter("R=%x[%x:]", wreq->debug_id, stream->stream_nr);
-
-	if (list_empty(&stream->subrequests))
-		return;
-
-	if (stream->source == NETFS_UPLOAD_TO_SERVER &&
-	    wreq->netfs_ops->retry_request)
-		wreq->netfs_ops->retry_request(wreq, stream);
-
-	if (unlikely(stream->failed))
-		return;
-
-	/* If there's no renegotiation to do, just resend each failed subreq. */
-	if (!stream->prepare_write) {
-		struct netfs_io_subrequest *subreq;
-
-		list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
-			if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
-				break;
-			if (__test_and_clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
-				struct iov_iter source = subreq->io_iter;
-
-				iov_iter_revert(&source, subreq->len - source.count);
-				netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
-				netfs_reissue_write(stream, subreq, &source);
-			}
-		}
-		return;
-	}
-
-	next = stream->subrequests.next;
-
-	do {
-		struct netfs_io_subrequest *subreq = NULL, *from, *to, *tmp;
-		struct iov_iter source;
-		unsigned long long start, len;
-		size_t part;
-		bool boundary = false;
-
-		/* Go through the stream and find the next span of contiguous
-		 * data that we then rejig (cifs, for example, needs the wsize
-		 * renegotiating) and reissue.
-		 */
-		from = list_entry(next, struct netfs_io_subrequest, rreq_link);
-		to = from;
-		start = from->start + from->transferred;
-		len   = from->len   - from->transferred;
-
-		if (test_bit(NETFS_SREQ_FAILED, &from->flags) ||
-		    !test_bit(NETFS_SREQ_NEED_RETRY, &from->flags))
-			return;
-
-		list_for_each_continue(next, &stream->subrequests) {
-			subreq = list_entry(next, struct netfs_io_subrequest, rreq_link);
-			if (subreq->start + subreq->transferred != start + len ||
-			    test_bit(NETFS_SREQ_BOUNDARY, &subreq->flags) ||
-			    !test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags))
-				break;
-			to = subreq;
-			len += to->len;
-		}
-
-		/* Determine the set of buffers we're going to use.  Each
-		 * subreq gets a subset of a single overall contiguous buffer.
-		 */
-		netfs_reset_iter(from);
-		source = from->io_iter;
-		source.count = len;
-
-		/* Work through the sublist. */
-		subreq = from;
-		list_for_each_entry_from(subreq, &stream->subrequests, rreq_link) {
-			if (!len)
-				break;
-			/* Renegotiate max_len (wsize) */
-			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
-			__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
-			subreq->retry_count++;
-			stream->prepare_write(subreq);
-
-			part = min(len, stream->sreq_max_len);
-			subreq->len = part;
-			subreq->start = start;
-			subreq->transferred = 0;
-			len -= part;
-			start += part;
-			if (len && subreq == to &&
-			    __test_and_clear_bit(NETFS_SREQ_BOUNDARY, &to->flags))
-				boundary = true;
-
-			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
-			netfs_reissue_write(stream, subreq, &source);
-			if (subreq == to)
-				break;
-		}
-
-		/* If we managed to use fewer subreqs, we can discard the
-		 * excess; if we used the same number, then we're done.
-		 */
-		if (!len) {
-			if (subreq == to)
-				continue;
-			list_for_each_entry_safe_from(subreq, tmp,
-						      &stream->subrequests, rreq_link) {
-				trace_netfs_sreq(subreq, netfs_sreq_trace_discard);
-				list_del(&subreq->rreq_link);
-				netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_done);
-				if (subreq == to)
-					break;
-			}
-			continue;
-		}
-
-		/* We ran out of subrequests, so we need to allocate some more
-		 * and insert them after.
-		 */
-		do {
-			subreq = netfs_alloc_subrequest(wreq);
-			subreq->source		= to->source;
-			subreq->start		= start;
-			subreq->debug_index	= atomic_inc_return(&wreq->subreq_counter);
-			subreq->stream_nr	= to->stream_nr;
-			subreq->retry_count	= 1;
-
-			trace_netfs_sreq_ref(wreq->debug_id, subreq->debug_index,
-					     refcount_read(&subreq->ref),
-					     netfs_sreq_trace_new);
-			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
-
-			list_add(&subreq->rreq_link, &to->rreq_link);
-			to = list_next_entry(to, rreq_link);
-			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
-
-			stream->sreq_max_len	= len;
-			stream->sreq_max_segs	= INT_MAX;
-			switch (stream->source) {
-			case NETFS_UPLOAD_TO_SERVER:
-				netfs_stat(&netfs_n_wh_upload);
-				stream->sreq_max_len = umin(len, wreq->wsize);
-				break;
-			case NETFS_WRITE_TO_CACHE:
-				netfs_stat(&netfs_n_wh_write);
-				break;
-			default:
-				WARN_ON_ONCE(1);
-			}
-
-			stream->prepare_write(subreq);
-
-			part = umin(len, stream->sreq_max_len);
-			subreq->len = subreq->transferred + part;
-			len -= part;
-			start += part;
-			if (!len && boundary) {
-				__set_bit(NETFS_SREQ_BOUNDARY, &to->flags);
-				boundary = false;
-			}
-
-			netfs_reissue_write(stream, subreq, &source);
-			if (!len)
-				break;
-
-		} while (len);
-
-	} while (!list_is_head(next, &stream->subrequests));
-}
-
-/*
- * Perform retries on the streams that need it.  If we're doing content
- * encryption and the server copy changed due to a third-party write, we may
- * need to do an RMW cycle and also rewrite the data to the cache.
- */
-static void netfs_retry_writes(struct netfs_io_request *wreq)
-{
-	struct netfs_io_subrequest *subreq;
-	struct netfs_io_stream *stream;
-	int s;
-
-	/* Wait for all outstanding I/O to quiesce before performing retries as
-	 * we may need to renegotiate the I/O sizes.
-	 */
-	for (s = 0; s < NR_IO_STREAMS; s++) {
-		stream = &wreq->io_streams[s];
-		if (!stream->active)
-			continue;
-
-		list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
-			wait_on_bit(&subreq->flags, NETFS_SREQ_IN_PROGRESS,
-				    TASK_UNINTERRUPTIBLE);
-		}
-	}
-
-	// TODO: Enc: Fetch changed partial pages
-	// TODO: Enc: Reencrypt content if needed.
-	// TODO: Enc: Wind back transferred point.
-	// TODO: Enc: Mark cache pages for retry.
-
-	for (s = 0; s < NR_IO_STREAMS; s++) {
-		stream = &wreq->io_streams[s];
-		if (stream->need_retry) {
-			stream->need_retry = false;
-			netfs_retry_write_stream(wreq, stream);
-		}
-	}
+	wreq->buffer.tail = folioq;
+done:
+	wreq->buffer.first_tail_slot = slot;
 }
 
 /*
@@ -390,7 +213,7 @@ reassess_streams:
 	if (wreq->origin == NETFS_WRITEBACK ||
 	    wreq->origin == NETFS_WRITETHROUGH ||
 	    wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE)
-		notes = BUFFERED;
+		notes = NEED_UNLOCK;
 	else
 		notes = 0;
 
@@ -449,14 +272,14 @@ reassess_streams:
 
 		cancel:
 			/* Remove if completely consumed. */
-			spin_lock_bh(&wreq->lock);
+			spin_lock(&wreq->lock);
 
 			remove = front;
 			list_del_init(&front->rreq_link);
 			front = list_first_entry_or_null(&stream->subrequests,
 							 struct netfs_io_subrequest, rreq_link);
 			stream->front = front;
-			spin_unlock_bh(&wreq->lock);
+			spin_unlock(&wreq->lock);
 			netfs_put_subrequest(remove, false,
 					     notes & SAW_FAILURE ?
 					     netfs_sreq_trace_put_cancel :
@@ -487,7 +310,7 @@ reassess_streams:
 	trace_netfs_collect_state(wreq, wreq->collected_to, notes);
 
 	/* Unlock any folios that we have now finished with. */
-	if (notes & BUFFERED) {
+	if (notes & NEED_UNLOCK) {
 		if (wreq->cleaned_to < wreq->collected_to)
 			netfs_writeback_unlock_folios(wreq, &notes);
 	} else {
@@ -500,7 +323,9 @@ reassess_streams:
 		goto need_retry;
 	if ((notes & MADE_PROGRESS) && test_bit(NETFS_RREQ_PAUSE, &wreq->flags)) {
 		trace_netfs_rreq(wreq, netfs_rreq_trace_unpause);
-		clear_and_wake_up_bit(NETFS_RREQ_PAUSE, &wreq->flags);
+		clear_bit_unlock(NETFS_RREQ_PAUSE, &wreq->flags);
+		smp_mb__after_atomic(); /* Set PAUSE before task state */
+		wake_up(&wreq->waitq);
 	}
 
 	if (notes & NEED_REASSESS) {
diff --git a/fs/netfs/write_issue.c b/fs/netfs/write_issue.c
index ff0e82505a0b..69727411683e 100644
--- a/fs/netfs/write_issue.c
+++ b/fs/netfs/write_issue.c
@@ -94,9 +94,10 @@ struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
 {
 	struct netfs_io_request *wreq;
 	struct netfs_inode *ictx;
-	bool is_buffered = (origin == NETFS_WRITEBACK ||
-			    origin == NETFS_WRITETHROUGH ||
-			    origin == NETFS_PGPRIV2_COPY_TO_CACHE);
+	bool is_cacheable = (origin == NETFS_WRITEBACK ||
+			     origin == NETFS_WRITEBACK_SINGLE ||
+			     origin == NETFS_WRITETHROUGH ||
+			     origin == NETFS_PGPRIV2_COPY_TO_CACHE);
 
 	wreq = netfs_alloc_request(mapping, file, start, 0, origin);
 	if (IS_ERR(wreq))
@@ -105,8 +106,10 @@ struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
 	_enter("R=%x", wreq->debug_id);
 
 	ictx = netfs_inode(wreq->inode);
-	if (is_buffered && netfs_is_cache_enabled(ictx))
+	if (is_cacheable && netfs_is_cache_enabled(ictx))
 		fscache_begin_write_operation(&wreq->cache_resources, netfs_i_cookie(ictx));
+	if (rolling_buffer_init(&wreq->buffer, wreq->debug_id, ITER_SOURCE) < 0)
+		goto nomem;
 
 	wreq->cleaned_to = wreq->start;
 
@@ -129,6 +132,10 @@ struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
 	}
 
 	return wreq;
+nomem:
+	wreq->error = -ENOMEM;
+	netfs_put_request(wreq, false, netfs_rreq_trace_put_failed);
+	return ERR_PTR(-ENOMEM);
 }
 
 /**
@@ -153,16 +160,15 @@ static void netfs_prepare_write(struct netfs_io_request *wreq,
 				loff_t start)
 {
 	struct netfs_io_subrequest *subreq;
-	struct iov_iter *wreq_iter = &wreq->io_iter;
+	struct iov_iter *wreq_iter = &wreq->buffer.iter;
 
 	/* Make sure we don't point the iterator at a used-up folio_queue
 	 * struct being used as a placeholder to prevent the queue from
 	 * collapsing.  In such a case, extend the queue.
 	 */
 	if (iov_iter_is_folioq(wreq_iter) &&
-	    wreq_iter->folioq_slot >= folioq_nr_slots(wreq_iter->folioq)) {
-		netfs_buffer_make_space(wreq);
-	}
+	    wreq_iter->folioq_slot >= folioq_nr_slots(wreq_iter->folioq))
+		rolling_buffer_make_space(&wreq->buffer);
 
 	subreq = netfs_alloc_subrequest(wreq);
 	subreq->source		= stream->source;
@@ -198,7 +204,7 @@ static void netfs_prepare_write(struct netfs_io_request *wreq,
 	 * the list.  The collector only goes nextwards and uses the lock to
 	 * remove entries off of the front.
 	 */
-	spin_lock_bh(&wreq->lock);
+	spin_lock(&wreq->lock);
 	list_add_tail(&subreq->rreq_link, &stream->subrequests);
 	if (list_is_first(&subreq->rreq_link, &stream->subrequests)) {
 		stream->front = subreq;
@@ -209,7 +215,7 @@ static void netfs_prepare_write(struct netfs_io_request *wreq,
 		}
 	}
 
-	spin_unlock_bh(&wreq->lock);
+	spin_unlock(&wreq->lock);
 
 	stream->construct = subreq;
 }
@@ -268,9 +274,9 @@ void netfs_issue_write(struct netfs_io_request *wreq,
  * we can avoid overrunning the credits obtained (cifs) and try to parallelise
  * content-crypto preparation with network writes.
  */
-int netfs_advance_write(struct netfs_io_request *wreq,
-			struct netfs_io_stream *stream,
-			loff_t start, size_t len, bool to_eof)
+size_t netfs_advance_write(struct netfs_io_request *wreq,
+			   struct netfs_io_stream *stream,
+			   loff_t start, size_t len, bool to_eof)
 {
 	struct netfs_io_subrequest *subreq = stream->construct;
 	size_t part;
@@ -327,6 +333,9 @@ static int netfs_write_folio(struct netfs_io_request *wreq,
 
 	_enter("");
 
+	if (rolling_buffer_make_space(&wreq->buffer) < 0)
+		return -ENOMEM;
+
 	/* netfs_perform_write() may shift i_size around the page or from out
 	 * of the page to beyond it, but cannot move i_size into or through the
 	 * page since we have it locked.
@@ -431,7 +440,7 @@ static int netfs_write_folio(struct netfs_io_request *wreq,
 	}
 
 	/* Attach the folio to the rolling buffer. */
-	netfs_buffer_append_folio(wreq, folio, false);
+	rolling_buffer_append(&wreq->buffer, folio, 0);
 
 	/* Move the submission point forward to allow for write-streaming data
 	 * not starting at the front of the page.  We don't do write-streaming
@@ -444,7 +453,8 @@ static int netfs_write_folio(struct netfs_io_request *wreq,
 		stream = &wreq->io_streams[s];
 		stream->submit_off = foff;
 		stream->submit_len = flen;
-		if ((stream->source == NETFS_WRITE_TO_CACHE && streamw) ||
+		if (!stream->avail ||
+		    (stream->source == NETFS_WRITE_TO_CACHE && streamw) ||
 		    (stream->source == NETFS_UPLOAD_TO_SERVER &&
 		     fgroup == NETFS_FOLIO_COPY_TO_CACHE)) {
 			stream->submit_off = UINT_MAX;
@@ -478,7 +488,7 @@ static int netfs_write_folio(struct netfs_io_request *wreq,
 
 		/* Advance the iterator(s). */
 		if (stream->submit_off > iter_off) {
-			iov_iter_advance(&wreq->io_iter, stream->submit_off - iter_off);
+			rolling_buffer_advance(&wreq->buffer, stream->submit_off - iter_off);
 			iter_off = stream->submit_off;
 		}
 
@@ -496,7 +506,7 @@ static int netfs_write_folio(struct netfs_io_request *wreq,
 	}
 
 	if (fsize > iter_off)
-		iov_iter_advance(&wreq->io_iter, fsize - iter_off);
+		rolling_buffer_advance(&wreq->buffer, fsize - iter_off);
 	atomic64_set(&wreq->issued_to, fpos + fsize);
 
 	if (!debug)
@@ -635,7 +645,7 @@ int netfs_advance_writethrough(struct netfs_io_request *wreq, struct writeback_c
 			       struct folio **writethrough_cache)
 {
 	_enter("R=%x ic=%zu ws=%u cp=%zu tp=%u",
-	       wreq->debug_id, wreq->iter.count, wreq->wsize, copied, to_page_end);
+	       wreq->debug_id, wreq->buffer.iter.count, wreq->wsize, copied, to_page_end);
 
 	if (!*writethrough_cache) {
 		if (folio_test_dirty(folio))
@@ -710,10 +720,10 @@ int netfs_unbuffered_write(struct netfs_io_request *wreq, bool may_wait, size_t
 		part = netfs_advance_write(wreq, upload, start, len, false);
 		start += part;
 		len -= part;
-		iov_iter_advance(&wreq->io_iter, part);
+		rolling_buffer_advance(&wreq->buffer, part);
 		if (test_bit(NETFS_RREQ_PAUSE, &wreq->flags)) {
 			trace_netfs_rreq(wreq, netfs_rreq_trace_wait_pause);
-			wait_on_bit(&wreq->flags, NETFS_RREQ_PAUSE, TASK_UNINTERRUPTIBLE);
+			wait_event(wreq->waitq, !test_bit(NETFS_RREQ_PAUSE, &wreq->flags));
 		}
 		if (test_bit(NETFS_RREQ_FAILED, &wreq->flags))
 			break;
@@ -723,3 +733,194 @@ int netfs_unbuffered_write(struct netfs_io_request *wreq, bool may_wait, size_t
 	_leave(" = %d", error);
 	return error;
 }
+
+/*
+ * Write some of a pending folio data back to the server and/or the cache.
+ */
+static int netfs_write_folio_single(struct netfs_io_request *wreq,
+				    struct folio *folio)
+{
+	struct netfs_io_stream *upload = &wreq->io_streams[0];
+	struct netfs_io_stream *cache  = &wreq->io_streams[1];
+	struct netfs_io_stream *stream;
+	size_t iter_off = 0;
+	size_t fsize = folio_size(folio), flen;
+	loff_t fpos = folio_pos(folio);
+	bool to_eof = false;
+	bool no_debug = false;
+
+	_enter("");
+
+	flen = folio_size(folio);
+	if (flen > wreq->i_size - fpos) {
+		flen = wreq->i_size - fpos;
+		folio_zero_segment(folio, flen, fsize);
+		to_eof = true;
+	} else if (flen == wreq->i_size - fpos) {
+		to_eof = true;
+	}
+
+	_debug("folio %zx/%zx", flen, fsize);
+
+	if (!upload->avail && !cache->avail) {
+		trace_netfs_folio(folio, netfs_folio_trace_cancel_store);
+		return 0;
+	}
+
+	if (!upload->construct)
+		trace_netfs_folio(folio, netfs_folio_trace_store);
+	else
+		trace_netfs_folio(folio, netfs_folio_trace_store_plus);
+
+	/* Attach the folio to the rolling buffer. */
+	folio_get(folio);
+	rolling_buffer_append(&wreq->buffer, folio, NETFS_ROLLBUF_PUT_MARK);
+
+	/* Move the submission point forward to allow for write-streaming data
+	 * not starting at the front of the page.  We don't do write-streaming
+	 * with the cache as the cache requires DIO alignment.
+	 *
+	 * Also skip uploading for data that's been read and just needs copying
+	 * to the cache.
+	 */
+	for (int s = 0; s < NR_IO_STREAMS; s++) {
+		stream = &wreq->io_streams[s];
+		stream->submit_off = 0;
+		stream->submit_len = flen;
+		if (!stream->avail) {
+			stream->submit_off = UINT_MAX;
+			stream->submit_len = 0;
+		}
+	}
+
+	/* Attach the folio to one or more subrequests.  For a big folio, we
+	 * could end up with thousands of subrequests if the wsize is small -
+	 * but we might need to wait during the creation of subrequests for
+	 * network resources (eg. SMB credits).
+	 */
+	for (;;) {
+		ssize_t part;
+		size_t lowest_off = ULONG_MAX;
+		int choose_s = -1;
+
+		/* Always add to the lowest-submitted stream first. */
+		for (int s = 0; s < NR_IO_STREAMS; s++) {
+			stream = &wreq->io_streams[s];
+			if (stream->submit_len > 0 &&
+			    stream->submit_off < lowest_off) {
+				lowest_off = stream->submit_off;
+				choose_s = s;
+			}
+		}
+
+		if (choose_s < 0)
+			break;
+		stream = &wreq->io_streams[choose_s];
+
+		/* Advance the iterator(s). */
+		if (stream->submit_off > iter_off) {
+			rolling_buffer_advance(&wreq->buffer, stream->submit_off - iter_off);
+			iter_off = stream->submit_off;
+		}
+
+		atomic64_set(&wreq->issued_to, fpos + stream->submit_off);
+		stream->submit_extendable_to = fsize - stream->submit_off;
+		part = netfs_advance_write(wreq, stream, fpos + stream->submit_off,
+					   stream->submit_len, to_eof);
+		stream->submit_off += part;
+		if (part > stream->submit_len)
+			stream->submit_len = 0;
+		else
+			stream->submit_len -= part;
+		if (part > 0)
+			no_debug = true;
+	}
+
+	wreq->buffer.iter.iov_offset = 0;
+	if (fsize > iter_off)
+		rolling_buffer_advance(&wreq->buffer, fsize - iter_off);
+	atomic64_set(&wreq->issued_to, fpos + fsize);
+
+	if (!no_debug)
+		kdebug("R=%x: No submit", wreq->debug_id);
+	_leave(" = 0");
+	return 0;
+}
+
+/**
+ * netfs_writeback_single - Write back a monolithic payload
+ * @mapping: The mapping to write from
+ * @wbc: Hints from the VM
+ * @iter: Data to write, must be ITER_FOLIOQ.
+ *
+ * Write a monolithic, non-pagecache object back to the server and/or
+ * the cache.
+ */
+int netfs_writeback_single(struct address_space *mapping,
+			   struct writeback_control *wbc,
+			   struct iov_iter *iter)
+{
+	struct netfs_io_request *wreq;
+	struct netfs_inode *ictx = netfs_inode(mapping->host);
+	struct folio_queue *fq;
+	size_t size = iov_iter_count(iter);
+	int ret;
+
+	if (WARN_ON_ONCE(!iov_iter_is_folioq(iter)))
+		return -EIO;
+
+	if (!mutex_trylock(&ictx->wb_lock)) {
+		if (wbc->sync_mode == WB_SYNC_NONE) {
+			netfs_stat(&netfs_n_wb_lock_skip);
+			return 0;
+		}
+		netfs_stat(&netfs_n_wb_lock_wait);
+		mutex_lock(&ictx->wb_lock);
+	}
+
+	wreq = netfs_create_write_req(mapping, NULL, 0, NETFS_WRITEBACK_SINGLE);
+	if (IS_ERR(wreq)) {
+		ret = PTR_ERR(wreq);
+		goto couldnt_start;
+	}
+
+	trace_netfs_write(wreq, netfs_write_trace_writeback);
+	netfs_stat(&netfs_n_wh_writepages);
+
+	if (__test_and_set_bit(NETFS_RREQ_UPLOAD_TO_SERVER, &wreq->flags))
+		wreq->netfs_ops->begin_writeback(wreq);
+
+	for (fq = (struct folio_queue *)iter->folioq; fq; fq = fq->next) {
+		for (int slot = 0; slot < folioq_count(fq); slot++) {
+			struct folio *folio = folioq_folio(fq, slot);
+			size_t part = umin(folioq_folio_size(fq, slot), size);
+
+			_debug("wbiter %lx %llx", folio->index, atomic64_read(&wreq->issued_to));
+
+			ret = netfs_write_folio_single(wreq, folio);
+			if (ret < 0)
+				goto stop;
+			size -= part;
+			if (size <= 0)
+				goto stop;
+		}
+	}
+
+stop:
+	for (int s = 0; s < NR_IO_STREAMS; s++)
+		netfs_issue_write(wreq, &wreq->io_streams[s]);
+	smp_wmb(); /* Write lists before ALL_QUEUED. */
+	set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
+
+	mutex_unlock(&ictx->wb_lock);
+
+	netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
+	_leave(" = %d", ret);
+	return ret;
+
+couldnt_start:
+	mutex_unlock(&ictx->wb_lock);
+	_leave(" = %d", ret);
+	return ret;
+}
+EXPORT_SYMBOL(netfs_writeback_single);
diff --git a/fs/netfs/write_retry.c b/fs/netfs/write_retry.c
new file mode 100644
index 000000000000..c841a851dd73
--- /dev/null
+++ b/fs/netfs/write_retry.c
@@ -0,0 +1,232 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* Network filesystem write retrying.
+ *
+ * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
+ * Written by David Howells (dhowells@redhat.com)
+ */
+
+#include <linux/fs.h>
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/slab.h>
+#include "internal.h"
+
+/*
+ * Perform retries on the streams that need it.
+ */
+static void netfs_retry_write_stream(struct netfs_io_request *wreq,
+				     struct netfs_io_stream *stream)
+{
+	struct list_head *next;
+
+	_enter("R=%x[%x:]", wreq->debug_id, stream->stream_nr);
+
+	if (list_empty(&stream->subrequests))
+		return;
+
+	if (stream->source == NETFS_UPLOAD_TO_SERVER &&
+	    wreq->netfs_ops->retry_request)
+		wreq->netfs_ops->retry_request(wreq, stream);
+
+	if (unlikely(stream->failed))
+		return;
+
+	/* If there's no renegotiation to do, just resend each failed subreq. */
+	if (!stream->prepare_write) {
+		struct netfs_io_subrequest *subreq;
+
+		list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
+			if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
+				break;
+			if (__test_and_clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
+				struct iov_iter source = subreq->io_iter;
+
+				iov_iter_revert(&source, subreq->len - source.count);
+				netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
+				netfs_reissue_write(stream, subreq, &source);
+			}
+		}
+		return;
+	}
+
+	next = stream->subrequests.next;
+
+	do {
+		struct netfs_io_subrequest *subreq = NULL, *from, *to, *tmp;
+		struct iov_iter source;
+		unsigned long long start, len;
+		size_t part;
+		bool boundary = false;
+
+		/* Go through the stream and find the next span of contiguous
+		 * data that we then rejig (cifs, for example, needs the wsize
+		 * renegotiating) and reissue.
+		 */
+		from = list_entry(next, struct netfs_io_subrequest, rreq_link);
+		to = from;
+		start = from->start + from->transferred;
+		len   = from->len   - from->transferred;
+
+		if (test_bit(NETFS_SREQ_FAILED, &from->flags) ||
+		    !test_bit(NETFS_SREQ_NEED_RETRY, &from->flags))
+			return;
+
+		list_for_each_continue(next, &stream->subrequests) {
+			subreq = list_entry(next, struct netfs_io_subrequest, rreq_link);
+			if (subreq->start + subreq->transferred != start + len ||
+			    test_bit(NETFS_SREQ_BOUNDARY, &subreq->flags) ||
+			    !test_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags))
+				break;
+			to = subreq;
+			len += to->len;
+		}
+
+		/* Determine the set of buffers we're going to use.  Each
+		 * subreq gets a subset of a single overall contiguous buffer.
+		 */
+		netfs_reset_iter(from);
+		source = from->io_iter;
+		source.count = len;
+
+		/* Work through the sublist. */
+		subreq = from;
+		list_for_each_entry_from(subreq, &stream->subrequests, rreq_link) {
+			if (!len)
+				break;
+
+			subreq->start	= start;
+			subreq->len	= len;
+			__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
+			subreq->retry_count++;
+			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
+
+			/* Renegotiate max_len (wsize) */
+			stream->sreq_max_len = len;
+			stream->prepare_write(subreq);
+
+			part = umin(len, stream->sreq_max_len);
+			if (unlikely(stream->sreq_max_segs))
+				part = netfs_limit_iter(&source, 0, part, stream->sreq_max_segs);
+			subreq->len = part;
+			subreq->transferred = 0;
+			len -= part;
+			start += part;
+			if (len && subreq == to &&
+			    __test_and_clear_bit(NETFS_SREQ_BOUNDARY, &to->flags))
+				boundary = true;
+
+			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
+			netfs_reissue_write(stream, subreq, &source);
+			if (subreq == to)
+				break;
+		}
+
+		/* If we managed to use fewer subreqs, we can discard the
+		 * excess; if we used the same number, then we're done.
+		 */
+		if (!len) {
+			if (subreq == to)
+				continue;
+			list_for_each_entry_safe_from(subreq, tmp,
+						      &stream->subrequests, rreq_link) {
+				trace_netfs_sreq(subreq, netfs_sreq_trace_discard);
+				list_del(&subreq->rreq_link);
+				netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_done);
+				if (subreq == to)
+					break;
+			}
+			continue;
+		}
+
+		/* We ran out of subrequests, so we need to allocate some more
+		 * and insert them after.
+		 */
+		do {
+			subreq = netfs_alloc_subrequest(wreq);
+			subreq->source		= to->source;
+			subreq->start		= start;
+			subreq->debug_index	= atomic_inc_return(&wreq->subreq_counter);
+			subreq->stream_nr	= to->stream_nr;
+			subreq->retry_count	= 1;
+
+			trace_netfs_sreq_ref(wreq->debug_id, subreq->debug_index,
+					     refcount_read(&subreq->ref),
+					     netfs_sreq_trace_new);
+			netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
+
+			list_add(&subreq->rreq_link, &to->rreq_link);
+			to = list_next_entry(to, rreq_link);
+			trace_netfs_sreq(subreq, netfs_sreq_trace_retry);
+
+			stream->sreq_max_len	= len;
+			stream->sreq_max_segs	= INT_MAX;
+			switch (stream->source) {
+			case NETFS_UPLOAD_TO_SERVER:
+				netfs_stat(&netfs_n_wh_upload);
+				stream->sreq_max_len = umin(len, wreq->wsize);
+				break;
+			case NETFS_WRITE_TO_CACHE:
+				netfs_stat(&netfs_n_wh_write);
+				break;
+			default:
+				WARN_ON_ONCE(1);
+			}
+
+			stream->prepare_write(subreq);
+
+			part = umin(len, stream->sreq_max_len);
+			subreq->len = subreq->transferred + part;
+			len -= part;
+			start += part;
+			if (!len && boundary) {
+				__set_bit(NETFS_SREQ_BOUNDARY, &to->flags);
+				boundary = false;
+			}
+
+			netfs_reissue_write(stream, subreq, &source);
+			if (!len)
+				break;
+
+		} while (len);
+
+	} while (!list_is_head(next, &stream->subrequests));
+}
+
+/*
+ * Perform retries on the streams that need it.  If we're doing content
+ * encryption and the server copy changed due to a third-party write, we may
+ * need to do an RMW cycle and also rewrite the data to the cache.
+ */
+void netfs_retry_writes(struct netfs_io_request *wreq)
+{
+	struct netfs_io_subrequest *subreq;
+	struct netfs_io_stream *stream;
+	int s;
+
+	/* Wait for all outstanding I/O to quiesce before performing retries as
+	 * we may need to renegotiate the I/O sizes.
+	 */
+	for (s = 0; s < NR_IO_STREAMS; s++) {
+		stream = &wreq->io_streams[s];
+		if (!stream->active)
+			continue;
+
+		list_for_each_entry(subreq, &stream->subrequests, rreq_link) {
+			wait_on_bit(&subreq->flags, NETFS_SREQ_IN_PROGRESS,
+				    TASK_UNINTERRUPTIBLE);
+		}
+	}
+
+	// TODO: Enc: Fetch changed partial pages
+	// TODO: Enc: Reencrypt content if needed.
+	// TODO: Enc: Wind back transferred point.
+	// TODO: Enc: Mark cache pages for retry.
+
+	for (s = 0; s < NR_IO_STREAMS; s++) {
+		stream = &wreq->io_streams[s];
+		if (stream->need_retry) {
+			stream->need_retry = false;
+			netfs_retry_write_stream(wreq, stream);
+		}
+	}
+}