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>

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);
 }
 
 /**

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);
 }

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;
 

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;
 

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,
 };
 

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
 

fs/netfs/direct_read.c

@@ -16,6 +16,143 @@
 #include <linux/netfs.h>
 #include "internal.h"
 
+static void netfs_prepare_dio_read_iterator(struct netfs_io_subrequest *subreq)
+{
+	struct netfs_io_request *rreq = subreq->rreq;
+	size_t rsize;
+
+	rsize = umin(subreq->len, rreq->io_streams[0].sreq_max_len);
+	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 (limit < rsize) {
+			subreq->len = limit;
+			trace_netfs_sreq(subreq, netfs_sreq_trace_limited);
+		}
+	}
+
+	trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
+
+	subreq->io_iter	= rreq->iter;
+	iov_iter_truncate(&subreq->io_iter, subreq->len);
+	iov_iter_advance(&rreq->iter, subreq->len);
+}
+
+/*
+ * Perform a read to a buffer from the server, slicing up the region to be read
+ * according to the network rsize.
+ */
+static int netfs_dispatch_unbuffered_reads(struct netfs_io_request *rreq)
+{
+	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;
+
+		subreq = netfs_alloc_subrequest(rreq);
+		if (!subreq) {
+			ret = -ENOMEM;
+			break;
+		}
+
+		subreq->source	= NETFS_DOWNLOAD_FROM_SERVER;
+		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);
+
+		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;
+			}
+		}
+
+		netfs_prepare_dio_read_iterator(subreq);
+		slice = subreq->len;
+		rreq->netfs_ops->issue_read(subreq);
+
+		size -= slice;
+		start += slice;
+		rreq->submitted += slice;
+
+		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);
+	return ret;
+}
+
+/*
+ * Perform a read to an application buffer, bypassing the pagecache and the
+ * local disk cache.
+ */
+static int netfs_unbuffered_read(struct netfs_io_request *rreq, bool sync)
+{
+	int ret;
+
+	_enter("R=%x %llx-%llx",
+	       rreq->debug_id, rreq->start, rreq->start + rreq->len - 1);
+
+	if (rreq->len == 0) {
+		pr_err("Zero-sized read [R=%x]\n", rreq->debug_id);
+		return -EIO;
+	}
+
+	// TODO: Use bounce buffer if requested
+
+	inode_dio_begin(rreq->inode);
+
+	ret = netfs_dispatch_unbuffered_reads(rreq);
+
+	if (!rreq->submitted) {
+		netfs_put_request(rreq, false, netfs_rreq_trace_put_no_submit);
+		inode_dio_end(rreq->inode);
+		ret = 0;
+		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 {
+		ret = -EIOCBQUEUED;
+	}
+
+out:
+	_leave(" = %d", ret);
+	return ret;
+}
+
 /**
  * netfs_unbuffered_read_iter_locked - Perform an unbuffered or direct I/O read
  * @iocb: The I/O control descriptor describing the read
@@ -31,7 +168,7 @@ ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *i
 	struct netfs_io_request *rreq;
 	ssize_t ret;
 	size_t orig_count = iov_iter_count(iter);
-	bool async = !is_sync_kiocb(iocb);
+	bool sync = is_sync_kiocb(iocb);
 
 	_enter("");
 
@@ -78,13 +215,13 @@ ssize_t netfs_unbuffered_read_iter_locked(struct kiocb *iocb, struct iov_iter *i
 
 	// TODO: Set up bounce buffer if needed
 
-	if (async)
+	if (!sync)
 		rreq->iocb = iocb;
 
-	ret = netfs_begin_read(rreq, is_sync_kiocb(iocb));
+	ret = netfs_unbuffered_read(rreq, sync);
 	if (ret < 0)
 		goto out; /* May be -EIOCBQUEUED */
-	if (!async) {
+	if (sync) {
 		// TODO: Copy from bounce buffer
 		iocb->ki_pos += rreq->transferred;
 		ret = rreq->transferred;
@@ -94,8 +231,6 @@ out:
 	netfs_put_request(rreq, false, netfs_rreq_trace_put_return);
 	if (ret > 0)
 		orig_count -= ret;
-	if (ret != -EIOCBQUEUED)
-		iov_iter_revert(iter, orig_count - iov_iter_count(iter));
 	return ret;
 }
 EXPORT_SYMBOL(netfs_unbuffered_read_iter_locked);

fs/netfs/internal.h

@@ -23,16 +23,9 @@
 /*
  * buffered_read.c
  */
-void netfs_rreq_unlock_folios(struct netfs_io_request *rreq);
 int netfs_prefetch_for_write(struct file *file, struct folio *folio,
 			     size_t offset, size_t len);
 
-/*
- * io.c
- */
-void netfs_rreq_work(struct work_struct *work);
-int netfs_begin_read(struct netfs_io_request *rreq, bool sync);
-
 /*
  * main.c
  */
@@ -90,6 +83,28 @@ static inline void netfs_see_request(struct netfs_io_request *rreq,
 	trace_netfs_rreq_ref(rreq->debug_id, refcount_read(&rreq->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);
+
+/*
+ * 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);
+bool netfs_pgpriv2_unlock_copied_folios(struct netfs_io_request *wreq);
+
+/*
+ * read_retry.c
+ */
+void netfs_retry_reads(struct netfs_io_request *rreq);
+void netfs_unlock_abandoned_read_pages(struct netfs_io_request *rreq);
+
 /*
  * stats.c
  */
@@ -117,6 +132,7 @@ extern atomic_t netfs_n_wh_buffered_write;
 extern atomic_t netfs_n_wh_writethrough;
 extern atomic_t netfs_n_wh_dio_write;
 extern atomic_t netfs_n_wh_writepages;
+extern atomic_t netfs_n_wh_copy_to_cache;
 extern atomic_t netfs_n_wh_wstream_conflict;
 extern atomic_t netfs_n_wh_upload;
 extern atomic_t netfs_n_wh_upload_done;
@@ -162,6 +178,11 @@ struct netfs_io_request *netfs_create_write_req(struct address_space *mapping,
 void netfs_reissue_write(struct netfs_io_stream *stream,
 			 struct netfs_io_subrequest *subreq,
 			 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);
 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,

fs/netfs/iterator.c

@@ -188,9 +188,59 @@ static size_t netfs_limit_xarray(const struct iov_iter *iter, size_t start_offse
 	return min(span, max_size);
 }
 
+/*
+ * Select the span of a folio queue iterator we're going to use.  Limit it by
+ * both maximum size and maximum number of segments.  Returns the size of the
+ * span in bytes.
+ */
+static size_t netfs_limit_folioq(const struct iov_iter *iter, size_t start_offset,
+				 size_t max_size, size_t max_segs)
+{
+	const struct folio_queue *folioq = iter->folioq;
+	unsigned int nsegs = 0;
+	unsigned int slot = iter->folioq_slot;
+	size_t span = 0, n = iter->count;
+
+	if (WARN_ON(!iov_iter_is_folioq(iter)) ||
+	    WARN_ON(start_offset > n) ||
+	    n == 0)
+		return 0;
+	max_size = umin(max_size, n - start_offset);
+
+	if (slot >= folioq_nr_slots(folioq)) {
+		folioq = folioq->next;
+		slot = 0;
+	}
+
+	start_offset += iter->iov_offset;
+	do {
+		size_t flen = folioq_folio_size(folioq, slot);
+
+		if (start_offset < flen) {
+			span += flen - start_offset;
+			nsegs++;
+			start_offset = 0;
+		} else {
+			start_offset -= flen;
+		}
+		if (span >= max_size || nsegs >= max_segs)
+			break;
+
+		slot++;
+		if (slot >= folioq_nr_slots(folioq)) {
+			folioq = folioq->next;
+			slot = 0;
+		}
+	} while (folioq);
+
+	return umin(span, max_size);
+}
+
 size_t netfs_limit_iter(const struct iov_iter *iter, size_t start_offset,
 			size_t max_size, size_t max_segs)
 {
+	if (iov_iter_is_folioq(iter))
+		return netfs_limit_folioq(iter, start_offset, max_size, max_segs);
 	if (iov_iter_is_bvec(iter))
 		return netfs_limit_bvec(iter, start_offset, max_size, max_segs);
 	if (iov_iter_is_xarray(iter))

fs/netfs/main.c

@@ -36,12 +36,14 @@ DEFINE_SPINLOCK(netfs_proc_lock);
 static const char *netfs_origins[nr__netfs_io_origin] = {
 	[NETFS_READAHEAD]		= "RA",
 	[NETFS_READPAGE]		= "RP",
+	[NETFS_READ_GAPS]		= "RG",
 	[NETFS_READ_FOR_WRITE]		= "RW",
 	[NETFS_DIO_READ]		= "DR",
 	[NETFS_WRITEBACK]		= "WB",
 	[NETFS_WRITETHROUGH]		= "WT",
 	[NETFS_UNBUFFERED_WRITE]	= "UW",
 	[NETFS_DIO_WRITE]		= "DW",
+	[NETFS_PGPRIV2_COPY_TO_CACHE]	= "2C",
 };
 
 /*
@@ -61,7 +63,7 @@ static int netfs_requests_seq_show(struct seq_file *m, void *v)
 
 	rreq = list_entry(v, struct netfs_io_request, proc_link);
 	seq_printf(m,
-		   "%08x %s %3d %2lx %4d %3d @%04llx %llx/%llx",
+		   "%08x %s %3d %2lx %4ld %3d @%04llx %llx/%llx",
 		   rreq->debug_id,
 		   netfs_origins[rreq->origin],
 		   refcount_read(&rreq->ref),

fs/netfs/objects.c

@@ -36,7 +36,6 @@ struct netfs_io_request *netfs_alloc_request(struct address_space *mapping,
 	memset(rreq, 0, kmem_cache_size(cache));
 	rreq->start	= start;
 	rreq->len	= len;
-	rreq->upper_len	= len;
 	rreq->origin	= origin;
 	rreq->netfs_ops	= ctx->ops;
 	rreq->mapping	= mapping;
@@ -44,6 +43,8 @@ struct netfs_io_request *netfs_alloc_request(struct address_space *mapping,
 	rreq->i_size	= i_size_read(inode);
 	rreq->debug_id	= atomic_inc_return(&debug_ids);
 	rreq->wsize	= INT_MAX;
+	rreq->io_streams[0].sreq_max_len = ULONG_MAX;
+	rreq->io_streams[0].sreq_max_segs = 0;
 	spin_lock_init(&rreq->lock);
 	INIT_LIST_HEAD(&rreq->io_streams[0].subrequests);
 	INIT_LIST_HEAD(&rreq->io_streams[1].subrequests);
@@ -52,9 +53,10 @@ struct netfs_io_request *netfs_alloc_request(struct address_space *mapping,
 
 	if (origin == NETFS_READAHEAD ||
 	    origin == NETFS_READPAGE ||
+	    origin == NETFS_READ_GAPS ||
 	    origin == NETFS_READ_FOR_WRITE ||
 	    origin == NETFS_DIO_READ)
-		INIT_WORK(&rreq->work, netfs_rreq_work);
+		INIT_WORK(&rreq->work, netfs_read_termination_worker);
 	else
 		INIT_WORK(&rreq->work, netfs_write_collection_worker);
 
@@ -163,7 +165,7 @@ void netfs_put_request(struct netfs_io_request *rreq, bool was_async,
 			if (was_async) {
 				rreq->work.func = netfs_free_request;
 				if (!queue_work(system_unbound_wq, &rreq->work))
-					BUG();
+					WARN_ON(1);
 			} else {
 				netfs_free_request(&rreq->work);
 			}

fs/netfs/read_pgpriv2.c

@@ -0,0 +1,264 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* Read with PG_private_2 [DEPRECATED].
+ *
+ * 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/task_io_accounting_ops.h>
+#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)
+{
+	struct netfs_io_stream *cache  = &wreq->io_streams[1];
+	size_t fsize = folio_size(folio), flen = fsize;
+	loff_t fpos = folio_pos(folio), i_size;
+	bool to_eof = false;
+
+	_enter("");
+
+	/* 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.
+	 */
+	i_size = i_size_read(wreq->inode);
+
+	if (fpos >= i_size) {
+		/* mmap beyond eof. */
+		_debug("beyond eof");
+		folio_end_private_2(folio);
+		return 0;
+	}
+
+	if (fpos + fsize > wreq->i_size)
+		wreq->i_size = i_size;
+
+	if (flen > i_size - fpos) {
+		flen = i_size - fpos;
+		to_eof = true;
+	} else if (flen == i_size - fpos) {
+		to_eof = true;
+	}
+
+	_debug("folio %zx %zx", flen, fsize);
+
+	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;
+
+	cache->submit_max_len = fsize;
+	cache->submit_off = 0;
+	cache->submit_len = flen;
+
+	/* 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).
+	 */
+	do {
+		ssize_t part;
+
+		wreq->io_iter.iov_offset = cache->submit_off;
+
+		atomic64_set(&wreq->issued_to, fpos + cache->submit_off);
+		part = netfs_advance_write(wreq, cache, fpos + cache->submit_off,
+					   cache->submit_len, to_eof);
+		cache->submit_off += part;
+		cache->submit_max_len -= part;
+		if (part > cache->submit_len)
+			cache->submit_len = 0;
+		else
+			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);
+
+	if (flen < fsize)
+		netfs_issue_write(wreq, cache);
+
+	_leave(" = 0");
+	return 0;
+}
+
+/*
+ * [DEPRECATED] Go through the buffer and write any folios that are marked with
+ * the third mark to the cache.
+ */
+void netfs_pgpriv2_write_to_the_cache(struct netfs_io_request *rreq)
+{
+	struct netfs_io_request *wreq;
+	struct folio_queue *folioq;
+	struct folio *folio;
+	int error = 0;
+	int slot = 0;
+
+	_enter("");
+
+	if (!fscache_resources_valid(&rreq->cache_resources))
+		goto couldnt_start;
+
+	/* 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),
+				      NETFS_PGPRIV2_COPY_TO_CACHE);
+	if (IS_ERR(wreq)) {
+		kleave(" [create %ld]", PTR_ERR(wreq));
+		goto couldnt_start;
+	}
+
+	trace_netfs_write(wreq, netfs_write_trace_copy_to_cache);
+	netfs_stat(&netfs_n_wh_copy_to_cache);
+
+	for (;;) {
+		error = netfs_pgpriv2_copy_folio(wreq, folio);
+		if (error < 0)
+			break;
+
+		folioq_unmark3(folioq, slot);
+		if (!folioq->marks3) {
+			folioq = folioq->next;
+			if (!folioq)
+				break;
+		}
+
+		slot = __ffs(folioq->marks3);
+		folio = folioq_folio(folioq, slot);
+	}
+
+	netfs_issue_write(wreq, &wreq->io_streams[1]);
+	smp_wmb(); /* Write lists before ALL_QUEUED. */
+	set_bit(NETFS_RREQ_ALL_QUEUED, &wreq->flags);
+
+	netfs_put_request(wreq, false, netfs_rreq_trace_put_return);
+	_leave(" = %d", error);
+couldnt_start:
+	netfs_pgpriv2_cancel(rreq->buffer);
+}
+
+/*
+ * [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)
+{
+	struct folio_queue *folioq = wreq->buffer;
+	unsigned long long collected_to = wreq->collected_to;
+	unsigned int slot = wreq->buffer_head_slot;
+	bool made_progress = false;
+
+	if (slot >= folioq_nr_slots(folioq)) {
+		folioq = netfs_delete_buffer_head(wreq);
+		slot = 0;
+	}
+
+	for (;;) {
+		struct folio *folio;
+		unsigned long long fpos, fend;
+		size_t fsize, flen;
+
+		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))
+			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);
+
+		trace_netfs_collect_folio(wreq, folio, fend, collected_to);
+
+		/* Unlock any folio we've transferred all of. */
+		if (collected_to < fend)
+			break;
+
+		trace_netfs_folio(folio, netfs_folio_trace_end_copy);
+		folio_end_private_2(folio);
+		wreq->cleaned_to = fpos + fsize;
+		made_progress = true;
+
+		/* 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)) {
+			if (READ_ONCE(wreq->buffer_tail) == folioq)
+				break;
+			folioq = netfs_delete_buffer_head(wreq);
+			slot = 0;
+		}
+
+		if (fpos + fsize >= collected_to)
+			break;
+	}
+
+	wreq->buffer = folioq;
+	wreq->buffer_head_slot = slot;
+	return made_progress;
+}

fs/netfs/read_retry.c

@@ -0,0 +1,256 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* Network filesystem read subrequest retrying.
+ *
+ * Copyright (C) 2024 Red Hat, Inc. All Rights Reserved.
+ * Written by David Howells (dhowells@redhat.com)
+ */
+
+#include <linux/fs.h>
+#include <linux/slab.h>
+#include "internal.h"
+
+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);
+	__set_bit(NETFS_SREQ_IN_PROGRESS, &subreq->flags);
+	netfs_get_subrequest(subreq, netfs_sreq_trace_get_resubmit);
+	subreq->rreq->netfs_ops->issue_read(subreq);
+}
+
+/*
+ * Go through the list of failed/short reads, retrying all retryable ones.  We
+ * need to switch failed cache reads to network downloads.
+ */
+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);
+
+	_enter("R=%x", rreq->debug_id);
+
+	if (list_empty(&rreq->subrequests))
+		return;
+
+	if (rreq->netfs_ops->retry_request)
+		rreq->netfs_ops->retry_request(rreq, NULL);
+
+	/* If there's no renegotiation to do, just resend each retryable subreq
+	 * up to the first permanently failed one.
+	 */
+	if (!rreq->netfs_ops->prepare_read &&
+	    !test_bit(NETFS_RREQ_COPY_TO_CACHE, &rreq->flags)) {
+		struct netfs_io_subrequest *subreq;
+
+		list_for_each_entry(subreq, &rreq->subrequests, rreq_link) {
+			if (test_bit(NETFS_SREQ_FAILED, &subreq->flags))
+				break;
+			if (__test_and_clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags)) {
+				netfs_reset_iter(subreq);
+				netfs_reissue_read(rreq, subreq);
+			}
+		}
+		return;
+	}
+
+	/* Okay, we need to renegotiate all the download requests and flip any
+	 * failed cache reads over to being download requests and negotiate
+	 * those also.  All fully successful subreqs have been removed from the
+	 * list and any spare data from those has been donated.
+	 *
+	 * What we do is decant the list and rebuild it one subreq at a time so
+	 * that we don't end up with donations jumping over a gap we're busy
+	 * 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);
+
+	do {
+		struct netfs_io_subrequest *from;
+		struct iov_iter source;
+		unsigned long long start, len;
+		size_t part, deferred_next_donated = 0;
+		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);
+		start = from->start + from->transferred;
+		len   = from->len   - from->transferred;
+
+		_debug("from R=%08x[%x] s=%llx ctl=%zx/%zx/%zx",
+		       rreq->debug_id, from->debug_index,
+		       from->start, from->consumed, 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 ||
+			    !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;
+		}
+
+		_debug(" - range: %llx-%llx %llx", start, start + len - 1, 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. */
+		while ((subreq = list_first_entry_or_null(
+				&sublist, struct netfs_io_subrequest, rreq_link))) {
+			list_del(&subreq->rreq_link);
+
+			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);
+			__set_bit(NETFS_SREQ_RETRYING, &subreq->flags);
+
+			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) */
+			if (rreq->netfs_ops->prepare_read(subreq) < 0) {
+				trace_netfs_sreq(subreq, netfs_sreq_trace_reprep_failed);
+				__set_bit(NETFS_SREQ_FAILED, &subreq->flags);
+			}
+
+			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);
+			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) {
+				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_reissue_read(rreq, subreq);
+			if (!len)
+				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.
+		 */
+		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);
+		}
+
+	} while (!list_empty(&queue));
+
+	return;
+
+	/* If we hit ENOMEM, fail all remaining subrequests */
+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);
+		__clear_bit(NETFS_SREQ_NEED_RETRY, &subreq->flags);
+		__clear_bit(NETFS_SREQ_RETRYING, &subreq->flags);
+	}
+	spin_lock_bh(&rreq->lock);
+	list_splice_tail_init(&queue, &rreq->subrequests);
+	spin_unlock_bh(&rreq->lock);
+}
+
+/*
+ * Retry reads.
+ */
+void netfs_retry_reads(struct netfs_io_request *rreq)
+{
+	trace_netfs_rreq(rreq, netfs_rreq_trace_resubmit);
+
+	atomic_inc(&rreq->nr_outstanding);
+
+	netfs_retry_read_subrequests(rreq);
+
+	if (atomic_dec_and_test(&rreq->nr_outstanding))
+		netfs_rreq_terminated(rreq, false);
+}
+
+/*
+ * Unlock any the pages that haven't been unlocked yet due to abandoned
+ * subrequests.
+ */
+void netfs_unlock_abandoned_read_pages(struct netfs_io_request *rreq)
+{
+	struct folio_queue *p;
+
+	for (p = rreq->buffer; p; p = p->next) {
+		for (int slot = 0; slot < folioq_count(p); slot++) {
+			struct folio *folio = folioq_folio(p, slot);
+
+			if (folio && !folioq_is_marked2(p, slot)) {
+				trace_netfs_folio(folio, netfs_folio_trace_abandon);
+				folio_unlock(folio);
+			}
+		}
+	}
+}

fs/netfs/stats.c

@@ -32,6 +32,7 @@ atomic_t netfs_n_wh_buffered_write;
 atomic_t netfs_n_wh_writethrough;
 atomic_t netfs_n_wh_dio_write;
 atomic_t netfs_n_wh_writepages;
+atomic_t netfs_n_wh_copy_to_cache;
 atomic_t netfs_n_wh_wstream_conflict;
 atomic_t netfs_n_wh_upload;
 atomic_t netfs_n_wh_upload_done;
@@ -51,11 +52,12 @@ int netfs_stats_show(struct seq_file *m, void *v)
 		   atomic_read(&netfs_n_rh_read_folio),
 		   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\n",
+	seq_printf(m, "Writes : BW=%u WT=%u DW=%u WP=%u 2C=%u\n",
 		   atomic_read(&netfs_n_wh_buffered_write),
 		   atomic_read(&netfs_n_wh_writethrough),
 		   atomic_read(&netfs_n_wh_dio_write),
-		   atomic_read(&netfs_n_wh_writepages));
+		   atomic_read(&netfs_n_wh_writepages),
+		   atomic_read(&netfs_n_wh_copy_to_cache));
 	seq_printf(m, "ZeroOps: ZR=%u sh=%u sk=%u\n",
 		   atomic_read(&netfs_n_rh_zero),
 		   atomic_read(&netfs_n_rh_short_read),

fs/netfs/write_collect.c

@@ -87,6 +87,12 @@ static void netfs_writeback_unlock_folios(struct netfs_io_request *wreq,
 	unsigned long long collected_to = wreq->collected_to;
 	unsigned int slot = wreq->buffer_head_slot;
 
+	if (wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE) {
+		if (netfs_pgpriv2_unlock_copied_folios(wreq))
+			*notes |= MADE_PROGRESS;
+		return;
+	}
+
 	if (slot >= folioq_nr_slots(folioq)) {
 		folioq = netfs_delete_buffer_head(wreq);
 		slot = 0;
@@ -383,7 +389,8 @@ reassess_streams:
 	smp_rmb();
 	collected_to = ULLONG_MAX;
 	if (wreq->origin == NETFS_WRITEBACK ||
-	    wreq->origin == NETFS_WRITETHROUGH)
+	    wreq->origin == NETFS_WRITETHROUGH ||
+	    wreq->origin == NETFS_PGPRIV2_COPY_TO_CACHE)
 		notes = BUFFERED;
 	else
 		notes = 0;

fs/netfs/write_issue.c

@@ -95,7 +95,8 @@ 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_WRITETHROUGH ||
+			    origin == NETFS_PGPRIV2_COPY_TO_CACHE);
 
 	wreq = netfs_alloc_request(mapping, file, start, 0, origin);
 	if (IS_ERR(wreq))
@@ -161,10 +162,6 @@ static void netfs_prepare_write(struct netfs_io_request *wreq,
 
 	_enter("R=%x[%x]", wreq->debug_id, subreq->debug_index);
 
-	trace_netfs_sreq_ref(wreq->debug_id, subreq->debug_index,
-			     refcount_read(&subreq->ref),
-			     netfs_sreq_trace_new);
-
 	trace_netfs_sreq(subreq, netfs_sreq_trace_prepare);
 
 	stream->sreq_max_len	= UINT_MAX;
@@ -241,8 +238,8 @@ void netfs_reissue_write(struct netfs_io_stream *stream,
 	netfs_do_issue_write(stream, subreq);
 }
 
-static void netfs_issue_write(struct netfs_io_request *wreq,
-			      struct netfs_io_stream *stream)
+void netfs_issue_write(struct netfs_io_request *wreq,
+		       struct netfs_io_stream *stream)
 {
 	struct netfs_io_subrequest *subreq = stream->construct;
 
@@ -259,9 +256,9 @@ static 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.
  */
-static int netfs_advance_write(struct netfs_io_request *wreq,
-			       struct netfs_io_stream *stream,
-			       loff_t start, size_t len, bool to_eof)
+int 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;

fs/nfs/fscache.c

@@ -267,6 +267,7 @@ static int nfs_netfs_init_request(struct netfs_io_request *rreq, struct file *fi
 	rreq->debug_id = atomic_inc_return(&nfs_netfs_debug_id);
 	/* [DEPRECATED] Use PG_private_2 to mark folio being written to the cache. */
 	__set_bit(NETFS_RREQ_USE_PGPRIV2, &rreq->flags);
+	rreq->io_streams[0].sreq_max_len = NFS_SB(rreq->inode->i_sb)->rsize;
 
 	return 0;
 }
@@ -288,14 +289,6 @@ static struct nfs_netfs_io_data *nfs_netfs_alloc(struct netfs_io_subrequest *sre
 	return netfs;
 }
 
-static bool nfs_netfs_clamp_length(struct netfs_io_subrequest *sreq)
-{
-	size_t	rsize = NFS_SB(sreq->rreq->inode->i_sb)->rsize;
-
-	sreq->len = min(sreq->len, rsize);
-	return true;
-}
-
 static void nfs_netfs_issue_read(struct netfs_io_subrequest *sreq)
 {
 	struct nfs_netfs_io_data	*netfs;
@@ -304,17 +297,18 @@ static void nfs_netfs_issue_read(struct netfs_io_subrequest *sreq)
 	struct nfs_open_context *ctx = sreq->rreq->netfs_priv;
 	struct page *page;
 	unsigned long idx;
+	pgoff_t start, last;
 	int err;
-	pgoff_t start = (sreq->start + sreq->transferred) >> PAGE_SHIFT;
-	pgoff_t last = ((sreq->start + sreq->len -
-			 sreq->transferred - 1) >> PAGE_SHIFT);
+
+	start = (sreq->start + sreq->transferred) >> PAGE_SHIFT;
+	last = ((sreq->start + sreq->len - sreq->transferred - 1) >> PAGE_SHIFT);
 
 	nfs_pageio_init_read(&pgio, inode, false,
 			     &nfs_async_read_completion_ops);
 
 	netfs = nfs_netfs_alloc(sreq);
 	if (!netfs)
-		return netfs_subreq_terminated(sreq, -ENOMEM, false);
+		return netfs_read_subreq_terminated(sreq, -ENOMEM, false);
 
 	pgio.pg_netfs = netfs; /* used in completion */
 
@@ -380,5 +374,4 @@ const struct netfs_request_ops nfs_netfs_ops = {
 	.init_request		= nfs_netfs_init_request,
 	.free_request		= nfs_netfs_free_request,
 	.issue_read		= nfs_netfs_issue_read,
-	.clamp_length		= nfs_netfs_clamp_length
 };

fs/nfs/fscache.h

@@ -60,8 +60,6 @@ static inline void nfs_netfs_get(struct nfs_netfs_io_data *netfs)
 
 static inline void nfs_netfs_put(struct nfs_netfs_io_data *netfs)
 {
-	ssize_t final_len;
-
 	/* Only the last RPC completion should call netfs_subreq_terminated() */
 	if (!refcount_dec_and_test(&netfs->refcount))
 		return;
@@ -74,8 +72,9 @@ static inline void nfs_netfs_put(struct nfs_netfs_io_data *netfs)
 	 * Correct the final length here to be no larger than the netfs subrequest
 	 * length, and thus avoid netfs's "Subreq overread" warning message.
 	 */
-	final_len = min_t(s64, netfs->sreq->len, atomic64_read(&netfs->transferred));
-	netfs_subreq_terminated(netfs->sreq, netfs->error ?: final_len, false);
+	netfs->sreq->transferred = min_t(s64, netfs->sreq->len,
+					 atomic64_read(&netfs->transferred));
+	netfs_read_subreq_terminated(netfs->sreq, netfs->error, false);
 	kfree(netfs);
 }
 static inline void nfs_netfs_inode_init(struct nfs_inode *nfsi)