netfs: Speed up buffered reading

Improve the efficiency of buffered reads in a number of ways:

 (1) Overhaul the algorithm in general so that it's a lot more compact and
     split the read submission code between buffered and unbuffered
     versions.  The unbuffered version can be vastly simplified.

 (2) Read-result collection is handed off to a work queue rather than being
     done in the I/O thread.  Multiple subrequests can be processes
     simultaneously.

 (3) When a subrequest is collected, any folios it fully spans are
     collected and "spare" data on either side is donated to either the
     previous or the next subrequest in the sequence.

Notes:

 (*) Readahead expansion is massively slows down fio, presumably because it
     causes a load of extra allocations, both folio and xarray, up front
     before RPC requests can be transmitted.

 (*) RDMA with cifs does appear to work, both with SIW and RXE.

 (*) PG_private_2-based reading and copy-to-cache is split out into its own
     file and altered to use folio_queue.  Note that the copy to the cache
     now creates a new write transaction against the cache and adds the
     folios to be copied into it.  This allows it to use part of the
     writeback I/O code.

Signed-off-by: David Howells <dhowells@redhat.com>
cc: Jeff Layton <jlayton@kernel.org>
cc: netfs@lists.linux.dev
cc: linux-fsdevel@vger.kernel.org
Link: https://lore.kernel.org/r/20240814203850.2240469-20-dhowells@redhat.com/ # v2
Signed-off-by: Christian Brauner <brauner@kernel.org>

28 files changed, 2058 insertions, 470 deletions

diff --git a/fs/9p/vfs_addr.c b/fs/9p/vfs_addr.c
index 24fdc74caeba..819c75233235 100644
--- a/fs/9p/vfs_addr.c
+++ b/fs/9p/vfs_addr.c
@@ -68,17 +68,22 @@ static void v9fs_issue_read(struct netfs_io_subrequest *subreq)
 {
 	struct netfs_io_request *rreq = subreq->rreq;
 	struct p9_fid *fid = rreq->netfs_priv;
+	unsigned long long pos = subreq->start + subreq->transferred;
 	int total, err;
 
-	total = p9_client_read(fid, subreq->start + subreq->transferred,
-			       &subreq->io_iter, &err);
+	total = p9_client_read(fid, pos, &subreq->io_iter, &err);
 
 	/* if we just extended the file size, any portion not in
 	 * cache won't be on server and is zeroes */
 	if (subreq->rreq->origin != NETFS_DIO_READ)
 		__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
+	if (pos + total >= i_size_read(rreq->inode))
+		__set_bit(NETFS_SREQ_HIT_EOF, &subreq->flags);
 
-	netfs_subreq_terminated(subreq, err ?: total, false);
+	if (!err)
+		subreq->transferred += total;
+
+	netfs_read_subreq_terminated(subreq, err, false);
 }
 
 /**
diff --git a/fs/afs/file.c b/fs/afs/file.c
index 5a9d16848ad5..492d857a3fa0 100644
--- a/fs/afs/file.c
+++ b/fs/afs/file.c
@@ -16,6 +16,7 @@
 #include <linux/mm.h>
 #include <linux/swap.h>
 #include <linux/netfs.h>
+#include <trace/events/netfs.h>
 #include "internal.h"
 
 static int afs_file_mmap(struct file *file, struct vm_area_struct *vma);
@@ -242,9 +243,10 @@ static void afs_fetch_data_notify(struct afs_operation *op)
 
 	req->error = error;
 	if (subreq) {
-		if (subreq->rreq->origin != NETFS_DIO_READ)
-			__set_bit(NETFS_SREQ_CLEAR_TAIL, &subreq->flags);
-		netfs_subreq_terminated(subreq, error ?: req->actual_len, false);
+		subreq->rreq->i_size = req->file_size;
+		if (req->pos + req->actual_len >= req->file_size)
+			__set_bit(NETFS_SREQ_HIT_EOF, &subreq->flags);
+		netfs_read_subreq_terminated(subreq, error, false);
 		req->subreq = NULL;
 	} else if (req->done) {
 		req->done(req);
@@ -262,6 +264,12 @@ static void afs_fetch_data_success(struct afs_operation *op)
 	afs_fetch_data_notify(op);
 }
 
+static void afs_fetch_data_aborted(struct afs_operation *op)
+{
+	afs_check_for_remote_deletion(op);
+	afs_fetch_data_notify(op);
+}
+
 static void afs_fetch_data_put(struct afs_operation *op)
 {
 	op->fetch.req->error = afs_op_error(op);
@@ -272,7 +280,7 @@ static const struct afs_operation_ops afs_fetch_data_operation = {
 	.issue_afs_rpc	= afs_fs_fetch_data,
 	.issue_yfs_rpc	= yfs_fs_fetch_data,
 	.success	= afs_fetch_data_success,
-	.aborted	= afs_check_for_remote_deletion,
+	.aborted	= afs_fetch_data_aborted,
 	.failed		= afs_fetch_data_notify,
 	.put		= afs_fetch_data_put,
 };
@@ -294,7 +302,7 @@ int afs_fetch_data(struct afs_vnode *vnode, struct afs_read *req)
 	op = afs_alloc_operation(req->key, vnode->volume);
 	if (IS_ERR(op)) {
 		if (req->subreq)
-			netfs_subreq_terminated(req->subreq, PTR_ERR(op), false);
+			netfs_read_subreq_terminated(req->subreq, PTR_ERR(op), false);
 		return PTR_ERR(op);
 	}
 
@@ -313,7 +321,7 @@ static void afs_read_worker(struct work_struct *work)
 
 	fsreq = afs_alloc_read(GFP_NOFS);
 	if (!fsreq)
-		return netfs_subreq_terminated(subreq, -ENOMEM, false);
+		return netfs_read_subreq_terminated(subreq, -ENOMEM, false);
 
 	fsreq->subreq	= subreq;
 	fsreq->pos	= subreq->start + subreq->transferred;
@@ -322,6 +330,7 @@ static void afs_read_worker(struct work_struct *work)
 	fsreq->vnode	= vnode;
 	fsreq->iter	= &subreq->io_iter;
 
+	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
 	afs_fetch_data(fsreq->vnode, fsreq);
 	afs_put_read(fsreq);
 }
diff --git a/fs/afs/fsclient.c b/fs/afs/fsclient.c
index 79cd30775b7a..098fa034a1cc 100644
--- a/fs/afs/fsclient.c
+++ b/fs/afs/fsclient.c
@@ -304,6 +304,7 @@ static int afs_deliver_fs_fetch_data(struct afs_call *call)
 	struct afs_vnode_param *vp = &op->file[0];
 	struct afs_read *req = op->fetch.req;
 	const __be32 *bp;
+	size_t count_before;
 	int ret;
 
 	_enter("{%u,%zu,%zu/%llu}",
@@ -345,10 +346,14 @@ static int afs_deliver_fs_fetch_data(struct afs_call *call)
 
 		/* extract the returned data */
 	case 2:
-		_debug("extract data %zu/%llu",
-		       iov_iter_count(call->iter), req->actual_len);
+		count_before = call->iov_len;
+		_debug("extract data %zu/%llu", count_before, req->actual_len);
 
 		ret = afs_extract_data(call, true);
+		if (req->subreq) {
+			req->subreq->transferred += count_before - call->iov_len;
+			netfs_read_subreq_progress(req->subreq, false);
+		}
 		if (ret < 0)
 			return ret;
 
diff --git a/fs/afs/yfsclient.c b/fs/afs/yfsclient.c
index f521e66d3bf6..024227aba4cd 100644
--- a/fs/afs/yfsclient.c
+++ b/fs/afs/yfsclient.c
@@ -355,6 +355,7 @@ static int yfs_deliver_fs_fetch_data64(struct afs_call *call)
 	struct afs_vnode_param *vp = &op->file[0];
 	struct afs_read *req = op->fetch.req;
 	const __be32 *bp;
+	size_t count_before;
 	int ret;
 
 	_enter("{%u,%zu, %zu/%llu}",
@@ -391,10 +392,14 @@ static int yfs_deliver_fs_fetch_data64(struct afs_call *call)
 
 		/* extract the returned data */
 	case 2:
-		_debug("extract data %zu/%llu",
-		       iov_iter_count(call->iter), req->actual_len);
+		count_before = call->iov_len;
+		_debug("extract data %zu/%llu", count_before, req->actual_len);
 
 		ret = afs_extract_data(call, true);
+		if (req->subreq) {
+			req->subreq->transferred += count_before - call->iov_len;
+			netfs_read_subreq_progress(req->subreq, false);
+		}
 		if (ret < 0)
 			return ret;
 
diff --git a/fs/ceph/addr.c b/fs/ceph/addr.c
index c4744a02db75..c500c1fd6b9f 100644
--- a/fs/ceph/addr.c
+++ b/fs/ceph/addr.c
@@ -13,6 +13,7 @@
 #include <linux/iversion.h>
 #include <linux/ktime.h>
 #include <linux/netfs.h>
+#include <trace/events/netfs.h>
 
 #include "super.h"
 #include "mds_client.h"
@@ -205,21 +206,6 @@ static void ceph_netfs_expand_readahead(struct netfs_io_request *rreq)
 	}
 }
 
-static bool ceph_netfs_clamp_length(struct netfs_io_subrequest *subreq)
-{
-	struct inode *inode = subreq->rreq->inode;
-	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
-	struct ceph_inode_info *ci = ceph_inode(inode);
-	u64 objno, objoff;
-	u32 xlen;
-
-	/* Truncate the extent at the end of the current block */
-	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
-				      &objno, &objoff, &xlen);
-	subreq->len = min(xlen, fsc->mount_options->rsize);
-	return true;
-}
-
 static void finish_netfs_read(struct ceph_osd_request *req)
 {
 	struct inode *inode = req->r_inode;
@@ -264,7 +250,12 @@ static void finish_netfs_read(struct ceph_osd_request *req)
 				     calc_pages_for(osd_data->alignment,
 					osd_data->length), false);
 	}
-	netfs_subreq_terminated(subreq, err, false);
+	if (err > 0) {
+		subreq->transferred = err;
+		err = 0;
+	}
+	trace_netfs_sreq(subreq, netfs_sreq_trace_io_progress);
+	netfs_read_subreq_terminated(subreq, err, false);
 	iput(req->r_inode);
 	ceph_dec_osd_stopping_blocker(fsc->mdsc);
 }
@@ -278,7 +269,6 @@ static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 	struct ceph_mds_request *req;
 	struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(inode->i_sb);
 	struct ceph_inode_info *ci = ceph_inode(inode);
-	struct iov_iter iter;
 	ssize_t err = 0;
 	size_t len;
 	int mode;
@@ -301,6 +291,7 @@ static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 	req->r_args.getattr.mask = cpu_to_le32(CEPH_STAT_CAP_INLINE_DATA);
 	req->r_num_caps = 2;
 
+	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
 	err = ceph_mdsc_do_request(mdsc, NULL, req);
 	if (err < 0)
 		goto out;
@@ -314,17 +305,36 @@ static bool ceph_netfs_issue_op_inline(struct netfs_io_subrequest *subreq)
 	}
 
 	len = min_t(size_t, iinfo->inline_len - subreq->start, subreq->len);
-	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
-	err = copy_to_iter(iinfo->inline_data + subreq->start, len, &iter);
-	if (err == 0)
+	err = copy_to_iter(iinfo->inline_data + subreq->start, len, &subreq->io_iter);
+	if (err == 0) {
 		err = -EFAULT;
+	} else {
+		subreq->transferred += err;
+		err = 0;
+	}
 
 	ceph_mdsc_put_request(req);
 out:
-	netfs_subreq_terminated(subreq, err, false);
+	netfs_read_subreq_terminated(subreq, err, false);
 	return true;
 }
 
+static int ceph_netfs_prepare_read(struct netfs_io_subrequest *subreq)
+{
+	struct netfs_io_request *rreq = subreq->rreq;
+	struct inode *inode = rreq->inode;
+	struct ceph_inode_info *ci = ceph_inode(inode);
+	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
+	u64 objno, objoff;
+	u32 xlen;
+
+	/* Truncate the extent at the end of the current block */
+	ceph_calc_file_object_mapping(&ci->i_layout, subreq->start, subreq->len,
+				      &objno, &objoff, &xlen);
+	rreq->io_streams[0].sreq_max_len = umin(xlen, fsc->mount_options->rsize);
+	return 0;
+}
+
 static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 {
 	struct netfs_io_request *rreq = subreq->rreq;
@@ -334,9 +344,8 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 	struct ceph_client *cl = fsc->client;
 	struct ceph_osd_request *req = NULL;
 	struct ceph_vino vino = ceph_vino(inode);
-	struct iov_iter iter;
-	int err = 0;
-	u64 len = subreq->len;
+	int err;
+	u64 len;
 	bool sparse = IS_ENCRYPTED(inode) || ceph_test_mount_opt(fsc, SPARSEREAD);
 	u64 off = subreq->start;
 	int extent_cnt;
@@ -349,6 +358,12 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 	if (ceph_has_inline_data(ci) && ceph_netfs_issue_op_inline(subreq))
 		return;
 
+	// TODO: This rounding here is slightly dodgy.  It *should* work, for
+	// now, as the cache only deals in blocks that are a multiple of
+	// PAGE_SIZE and fscrypt blocks are at most PAGE_SIZE.  What needs to
+	// happen is for the fscrypt driving to be moved into netfslib and the
+	// data in the cache also to be stored encrypted.
+	len = subreq->len;
 	ceph_fscrypt_adjust_off_and_len(inode, &off, &len);
 
 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout, vino,
@@ -371,8 +386,6 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 	doutc(cl, "%llx.%llx pos=%llu orig_len=%zu len=%llu\n",
 	      ceph_vinop(inode), subreq->start, subreq->len, len);
 
-	iov_iter_xarray(&iter, ITER_DEST, &rreq->mapping->i_pages, subreq->start, len);
-
 	/*
 	 * FIXME: For now, use CEPH_OSD_DATA_TYPE_PAGES instead of _ITER for
 	 * encrypted inodes. We'd need infrastructure that handles an iov_iter
@@ -384,7 +397,7 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 		struct page **pages;
 		size_t page_off;
 
-		err = iov_iter_get_pages_alloc2(&iter, &pages, len, &page_off);
+		err = iov_iter_get_pages_alloc2(&subreq->io_iter, &pages, len, &page_off);
 		if (err < 0) {
 			doutc(cl, "%llx.%llx failed to allocate pages, %d\n",
 			      ceph_vinop(inode), err);
@@ -399,7 +412,7 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 		osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0, false,
 						 false);
 	} else {
-		osd_req_op_extent_osd_iter(req, 0, &iter);
+		osd_req_op_extent_osd_iter(req, 0, &subreq->io_iter);
 	}
 	if (!ceph_inc_osd_stopping_blocker(fsc->mdsc)) {
 		err = -EIO;
@@ -410,17 +423,19 @@ static void ceph_netfs_issue_read(struct netfs_io_subrequest *subreq)
 	req->r_inode = inode;
 	ihold(inode);
 
+	trace_netfs_sreq(subreq, netfs_sreq_trace_submit);
 	ceph_osdc_start_request(req->r_osdc, req);
 out:
 	ceph_osdc_put_request(req);
 	if (err)
-		netfs_subreq_terminated(subreq, err, false);
+		netfs_read_subreq_terminated(subreq, err, false);
 	doutc(cl, "%llx.%llx result %d\n", ceph_vinop(inode), err);
 }
 
 static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 {
 	struct inode *inode = rreq->inode;
+	struct ceph_fs_client *fsc = ceph_inode_to_fs_client(inode);
 	struct ceph_client *cl = ceph_inode_to_client(inode);
 	int got = 0, want = CEPH_CAP_FILE_CACHE;
 	struct ceph_netfs_request_data *priv;
@@ -472,6 +487,7 @@ static int ceph_init_request(struct netfs_io_request *rreq, struct file *file)
 
 	priv->caps = got;
 	rreq->netfs_priv = priv;
+	rreq->io_streams[0].sreq_max_len = fsc->mount_options->rsize;
 
 out:
 	if (ret < 0)
@@ -496,9 +512,9 @@ static void ceph_netfs_free_request(struct netfs_io_request *rreq)
 const struct netfs_request_ops ceph_netfs_ops = {
 	.init_request		= ceph_init_request,
 	.free_request		= ceph_netfs_free_request,
+	.prepare_read		= ceph_netfs_prepare_read,
 	.issue_read		= ceph_netfs_issue_read,
 	.expand_readahead	= ceph_netfs_expand_readahead,
-	.clamp_length		= ceph_netfs_clamp_length,
 	.check_write_begin	= ceph_netfs_check_write_begin,
 };
 
diff --git a/fs/netfs/Makefile b/fs/netfs/Makefile
index 8e6781e0b10b..d08b0bfb6756 100644
--- a/fs/netfs/Makefile
+++ b/fs/netfs/Makefile
@@ -5,12 +5,14 @@ netfs-y := \
 	buffered_write.o \
 	direct_read.o \
 	direct_write.o \
-	io.o \
 	iterator.o \
 	locking.o \
 	main.o \
 	misc.o \
 	objects.o \
+	read_collect.o \
+	read_pgpriv2.o \
+	read_retry.o \
 	write_collect.o \
 	write_issue.o
 
diff --git a/fs/netfs/buffered_read.c b/fs/netfs/buffered_read.c
index 27c750d39476..c40e226053cc 100644
--- a/fs/netfs/buffered_read.c
+++ b/fs/netfs/buffered_read.c
@@ -9,266 +9,388 @@
 #include <linux/task_io_accounting_ops.h>
 #include "internal.h"
 
-/*
- * [DEPRECATED] Unlock the folios in a read operation for when the filesystem
- * is using PG_private_2 and direct writing to the cache from here rather than
- * marking the page for writeback.
- *
- * Note that we don't touch folio->private in this code.
- */
-static void netfs_rreq_unlock_folios_pgpriv2(struct netfs_io_request *rreq,
-					     size_t *account)
+static void netfs_cache_expand_readahead(struct netfs_io_request *rreq,
+					 unsigned long long *_start,
+					 unsigned long long *_len,
+					 unsigned long long i_size)
 {
-	struct netfs_io_subrequest *subreq;
-	struct folio *folio;
-	pgoff_t start_page = rreq->start / PAGE_SIZE;
-	pgoff_t last_page = ((rreq->start + rreq->len) / PAGE_SIZE) - 1;
-	bool subreq_failed = false;
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
 
-	XA_STATE(xas, &rreq->mapping->i_pages, start_page);
+	if (cres->ops && cres->ops->expand_readahead)
+		cres->ops->expand_readahead(cres, _start, _len, i_size);
+}
 
-	/* Walk through the pagecache and the I/O request lists simultaneously.
-	 * We may have a mixture of cached and uncached sections and we only
-	 * really want to write out the uncached sections.  This is slightly
-	 * complicated by the possibility that we might have huge pages with a
-	 * mixture inside.
+static void netfs_rreq_expand(struct netfs_io_request *rreq,
+			      struct readahead_control *ractl)
+{
+	/* Give the cache a chance to change the request parameters.  The
+	 * resultant request must contain the original region.
 	 */
-	subreq = list_first_entry(&rreq->subrequests,
-				  struct netfs_io_subrequest, rreq_link);
-	subreq_failed = (subreq->error < 0);
+	netfs_cache_expand_readahead(rreq, &rreq->start, &rreq->len, rreq->i_size);
 
-	trace_netfs_rreq(rreq, netfs_rreq_trace_unlock_pgpriv2);
+	/* Give the netfs a chance to change the request parameters.  The
+	 * resultant request must contain the original region.
+	 */
+	if (rreq->netfs_ops->expand_readahead)
+		rreq->netfs_ops->expand_readahead(rreq);
 
-	rcu_read_lock();
-	xas_for_each(&xas, folio, last_page) {
-		loff_t pg_end;
-		bool pg_failed = false;
-		bool folio_started = false;
+	/* Expand the request if the cache wants it to start earlier.  Note
+	 * that the expansion may get further extended if the VM wishes to
+	 * insert THPs and the preferred start and/or end wind up in the middle
+	 * of THPs.
+	 *
+	 * If this is the case, however, the THP size should be an integer
+	 * multiple of the cache granule size, so we get a whole number of
+	 * granules to deal with.
+	 */
+	if (rreq->start  != readahead_pos(ractl) ||
+	    rreq->len != readahead_length(ractl)) {
+		readahead_expand(ractl, rreq->start, rreq->len);
+		rreq->start  = readahead_pos(ractl);
+		rreq->len = readahead_length(ractl);
 
-		if (xas_retry(&xas, folio))
-			continue;
+		trace_netfs_read(rreq, readahead_pos(ractl), readahead_length(ractl),
+				 netfs_read_trace_expanded);
+	}
+}
 
-		pg_end = folio_pos(folio) + folio_size(folio) - 1;
+/*
+ * Begin an operation, and fetch the stored zero point value from the cookie if
+ * available.
+ */
+static int netfs_begin_cache_read(struct netfs_io_request *rreq, struct netfs_inode *ctx)
+{
+	return fscache_begin_read_operation(&rreq->cache_resources, netfs_i_cookie(ctx));
+}
 
-		for (;;) {
-			loff_t sreq_end;
+/*
+ * 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)
+{
+	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 (!subreq) {
-				pg_failed = true;
-				break;
-			}
+	for (int i = nr; i < folioq_nr_slots(folioq); i++)
+		folioq_clear(folioq, i);
 
-			if (!folio_started &&
-			    test_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags) &&
-			    fscache_operation_valid(&rreq->cache_resources)) {
-				trace_netfs_folio(folio, netfs_folio_trace_copy_to_cache);
-				folio_start_private_2(folio);
-				folio_started = true;
-			}
+	return size;
+}
 
-			pg_failed |= subreq_failed;
-			sreq_end = subreq->start + subreq->len - 1;
-			if (pg_end < sreq_end)
-				break;
+/*
+ * netfs_prepare_read_iterator - Prepare the subreq iterator for I/O
+ * @subreq: The subrequest to be set up
+ *
+ * Prepare the I/O iterator representing the read buffer on a subrequest for
+ * the filesystem to use for I/O (it can be passed directly to a socket).  This
+ * is intended to be called from the ->issue_read() method once the filesystem
+ * has trimmed the request to the size it wants.
+ *
+ * Returns the limited size if successful and -ENOMEM if insufficient memory
+ * available.
+ *
+ * [!] NOTE: This must be run in the same thread as ->issue_read() was called
+ * in as we access the readahead_control struct.
+ */
+static ssize_t netfs_prepare_read_iterator(struct netfs_io_subrequest *subreq)
+{
+	struct netfs_io_request *rreq = subreq->rreq;
+	size_t rsize = subreq->len;
+
+	if (subreq->source == NETFS_DOWNLOAD_FROM_SERVER)
+		rsize = umin(rsize, rreq->io_streams[0].sreq_max_len);
+
+	if (rreq->ractl) {
+		/* If we don't have sufficient folios in the rolling buffer,
+		 * extract a folioq's worth from the readahead region at a time
+		 * into the buffer.  Note that this acquires a ref on each page
+		 * that we will need to release later - but we don't want to do
+		 * that until after we've started the I/O.
+		 */
+		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);
+			rreq->iter.count += added;
+			rreq->submitted += added;
+		}
+	}
 
-			*account += subreq->transferred;
-			if (!list_is_last(&subreq->rreq_link, &rreq->subrequests)) {
-				subreq = list_next_entry(subreq, rreq_link);
-				subreq_failed = (subreq->error < 0);
-			} else {
-				subreq = NULL;
-				subreq_failed = false;
-			}
+	subreq->len = rsize;
+	if (unlikely(rreq->io_streams[0].sreq_max_segs)) {
+		size_t limit = netfs_limit_iter(&rreq->iter, 0, rsize,
+						rreq->io_streams[0].sreq_max_segs);
 
-			if (pg_end == sreq_end)
-				break;
+		if (limit < rsize) {
+			subreq->len = limit;
+			trace_netfs_sreq(subreq, netfs_sreq_trace_limited);
 		}
+	}
 
-		if (!pg_failed) {
-			flush_dcache_folio(folio);
-			folio_mark_uptodate(folio);
-		}
+	subreq->io_iter	= rreq->iter;
 
-		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))
-				_debug("no unlock");
-			else
-				folio_unlock(folio);
+	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];
 	}
-	rcu_read_unlock();
+
+	iov_iter_truncate(&subreq->io_iter, subreq->len);
+	iov_iter_advance(&rreq->iter, subreq->len);
+	return subreq->len;
 }
 
-/*
- * Unlock the folios in a read operation.  We need to set PG_writeback on any
- * folios we're going to write back before we unlock them.
- *
- * Note that if the deprecated NETFS_RREQ_USE_PGPRIV2 is set then we use
- * PG_private_2 and do a direct write to the cache from here instead.
- */
-void netfs_rreq_unlock_folios(struct netfs_io_request *rreq)
+static enum netfs_io_source netfs_cache_prepare_read(struct netfs_io_request *rreq,
+						     struct netfs_io_subrequest *subreq,
+						     loff_t i_size)
 {
-	struct netfs_io_subrequest *subreq;
-	struct netfs_folio *finfo;
-	struct folio *folio;
-	pgoff_t start_page = rreq->start / PAGE_SIZE;
-	pgoff_t last_page = ((rreq->start + rreq->len) / PAGE_SIZE) - 1;
-	size_t account = 0;
-	bool subreq_failed = false;
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
 
-	XA_STATE(xas, &rreq->mapping->i_pages, start_page);
+	if (!cres->ops)
+		return NETFS_DOWNLOAD_FROM_SERVER;
+	return cres->ops->prepare_read(subreq, i_size);
+}
 
-	if (test_bit(NETFS_RREQ_FAILED, &rreq->flags)) {
-		__clear_bit(NETFS_RREQ_COPY_TO_CACHE, &rreq->flags);
-		list_for_each_entry(subreq, &rreq->subrequests, rreq_link) {
-			__clear_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
-		}
-	}
+static void netfs_cache_read_terminated(void *priv, ssize_t transferred_or_error,
+					bool was_async)
+{
+	struct netfs_io_subrequest *subreq = priv;
 
-	/* Handle deprecated PG_private_2 case. */
-	if (test_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags)) {
-		netfs_rreq_unlock_folios_pgpriv2(rreq, &account);
-		goto out;
+	if (transferred_or_error < 0) {
+		netfs_read_subreq_terminated(subreq, transferred_or_error, was_async);
+		return;
 	}
 
-	/* Walk through the pagecache and the I/O request lists simultaneously.
-	 * We may have a mixture of cached and uncached sections and we only
-	 * really want to write out the uncached sections.  This is slightly
-	 * complicated by the possibility that we might have huge pages with a
-	 * mixture inside.
-	 */
-	subreq = list_first_entry(&rreq->subrequests,
-				  struct netfs_io_subrequest, rreq_link);
-	subreq_failed = (subreq->error < 0);
-
-	trace_netfs_rreq(rreq, netfs_rreq_trace_unlock);
+	if (transferred_or_error > 0)
+		subreq->transferred += transferred_or_error;
+	netfs_read_subreq_terminated(subreq, 0, was_async);
+}
 
-	rcu_read_lock();
-	xas_for_each(&xas, folio, last_page) {
-		loff_t pg_end;
-		bool pg_failed = false;
-		bool wback_to_cache = false;
+/*
+ * Issue a read against the cache.
+ * - Eats the caller's ref on subreq.
+ */
+static void netfs_read_cache_to_pagecache(struct netfs_io_request *rreq,
+					  struct netfs_io_subrequest *subreq)
+{
+	struct netfs_cache_resources *cres = &rreq->cache_resources;
 
-		if (xas_retry(&xas, folio))
-			continue;
+	netfs_stat(&netfs_n_rh_read);
+	cres->ops->read(cres, subreq->start, &subreq->io_iter, NETFS_READ_HOLE_IGNORE,
+			netfs_cache_read_terminated, subreq);
+}
 
-		pg_end = folio_pos(folio) + folio_size(folio) - 1;
+/*
+ * 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
+ * network rsize.
+ */
+static void netfs_read_to_pagecache(struct netfs_io_request *rreq)
+{
+	struct netfs_inode *ictx = netfs_inode(rreq->inode);
+	unsigned long long start = rreq->start;
+	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;
+		ssize_t slice;
+
+		subreq = netfs_alloc_subrequest(rreq);
+		if (!subreq) {
+			ret = -ENOMEM;
+			break;
+		}
 
-		for (;;) {
-			loff_t sreq_end;
+		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 (subreq->start >= zp) {
+				subreq->source = source = NETFS_FILL_WITH_ZEROES;
+				goto fill_with_zeroes;
+			}
 
-			if (!subreq) {
-				pg_failed = true;
+			if (len > zp - subreq->start)
+				len = zp - subreq->start;
+			if (len == 0) {
+				pr_err("ZERO-LEN READ: R=%08x[%x] l=%zx/%zx s=%llx z=%llx i=%llx",
+				       rreq->debug_id, subreq->debug_index,
+				       subreq->len, size,
+				       subreq->start, ictx->zero_point, rreq->i_size);
 				break;
 			}
+			subreq->len = len;
+
+			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;
+				}
+				trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
+			}
 
-			wback_to_cache |= test_bit(NETFS_SREQ_COPY_TO_CACHE, &subreq->flags);
-			pg_failed |= subreq_failed;
-			sreq_end = subreq->start + subreq->len - 1;
-			if (pg_end < sreq_end)
+			slice = netfs_prepare_read_iterator(subreq);
+			if (slice < 0) {
+				atomic_dec(&rreq->nr_outstanding);
+				netfs_put_subrequest(subreq, false, netfs_sreq_trace_put_cancel);
+				ret = slice;
 				break;
-
-