[PATCH 0/3] mm/zram: route block swap I/O through swap_ops

[PATCH 0/3] mm/zram: route block swap I/O through swap_ops

[Jianyue Wu]

This series builds on Christoph Hellwig's swap batching rework that
moves block swap onto struct swap_iocb and per-backend struct
swap_ops handlers [1].  Christoph's patches unify batching for
ordinary block devices and swap files.  zram still needs a custom
path because swap slots map to compressed pages, not disk sectors.

The first patch adds swap_register_block_ops() so a block driver can
install custom submit_read/submit_write handlers when swapon targets
its block device.  The default swap_bdev_ops path is unchanged for
devices that do not register.

The second patch registers zram_swap_ops at module init.  On write,
the swap core still batches folios into a swap_iocb.  zram maps each
folio to a slot index and stores it through zram_write_page() instead
of building one bio per page.  Read handling keeps slot_lock and
mark_slot_accessed() in one critical section.  Writeback-enabled zram
falls back to swap_bdev_submit_read() for ZRAM_WB slots.

The third patch moves slot_free_notify into swap_ops next to the
other zram swap callbacks, and documents the locking contract for
that hook.

Applied on top of Christoph Hellwig's "better block swap batching and
a different take on swap_ops" series [1].

[1] https://lore.kernel.org/linux-mm/?q=better+block+swap+batching

To: Andrew Morton <akpm@linux-foundation.org>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chris Li <chrisl@kernel.org>
Cc: Baoquan He <bhe@redhat.com>
Cc: Nhat Pham <nphamcs@gmail.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Kairui Song <kasong@tencent.com>
Cc: Kemeng Shi <shikemeng@huaweicloud.com>
Cc: Youngjun Park <youngjun.park@lge.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Sergey Senozhatsky <senozhatsky@chromium.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Jan Kara <jack@suse.cz>
Cc: linux-mm@kvack.org
Cc: linux-kernel@vger.kernel.org
Cc: linux-block@vger.kernel.org
Cc: linux-doc@vger.kernel.org

Signed-off-by: Jianyue Wu <wujianyue000@gmail.com>
---
Jianyue Wu (3):
      mm/page_io: let block drivers register custom swap I/O ops
      mm/zram: handle swap read/write via swap_ops
      mm/swap: route slot free notifications through swap_ops

 Documentation/filesystems/locking.rst |   5 -
 drivers/block/zram/zram_drv.c         | 215 +++++++++++++++++++++++++++-------
 include/linux/blkdev.h                |   2 -
 include/linux/swap.h                  |  47 ++++++++
 mm/page_io.c                          | 187 ++++++++++++++++++++++++++++-
 mm/swap.h                             |  18 +--
 mm/swapfile.c                         |  17 +--
 rust/kernel/block/mq/gen_disk.rs      |   1 -
 8 files changed, 414 insertions(+), 78 deletions(-)
---
base-commit: 842f51deada6449843f811bfa22e536a01ae5a0c
change-id: 20260614-zram-swap-ops-block-register-a1b2c3d4e5f6

Best regards,
-- 
Jianyue Wu <wujianyue000@gmail.com>

[PATCH 1/3] mm/page_io: let block drivers register custom swap I/O ops

[Jianyue Wu]

Add swap_register_block_ops() so a block driver can install custom
swap read/write handlers instead of always building bios.

When swapon targets a block device (S_ISBLK), setup_swap_extents()
checks whether that driver's block_device_operations were registered.
If yes, sis->ops points at the driver table. Otherwise sis->ops
stays on swap_bdev_ops.

Swap files are unchanged. They still use the filesystem path and
extent tree, because their page index is not a raw disk sector.

Register swap_ops in a single global slot keyed by the driver's
block_device_operations. lookup_swap_block_ops() matches sis->bdev
fops at swapon. -EBUSY if the slot is already taken. That is enough
while only zram needs custom swap I/O. Several block drivers would
need a per-fops lookup table instead.

swap_unregister_block_ops() must pass the same fops that
registered. Swap areas created before unregister keep the old ops
until swapoff. The driver module must remain loaded while they are
in use.

Signed-off-by: Jianyue Wu <wujianyue000@gmail.com>
---
 include/linux/swap.h |  35 +++++++++++++++++
 mm/page_io.c         | 106 +++++++++++++++++++++++++++++++++++++++++++++++++++
 mm/swap.h            |  18 +--------
 mm/swapfile.c        |   4 ++
 4 files changed, 147 insertions(+), 16 deletions(-)

diff --git a/include/linux/swap.h b/include/linux/swap.h
index 636d94108166..1d51df4179c1 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -19,6 +19,41 @@
 struct notifier_block;
 
 struct bio;
+struct block_device_operations;
+struct folio;
+struct swap_iocb;
+struct swap_info_struct;
+
+struct swap_io_ctx {
+	struct swap_iocb	*sio;
+	struct swap_info_struct	*sis;
+};
+
+/* Set when the swap backend requires GFP_NOFS allocations. */
+#define SWAP_OPS_F_NOFS		(1U << 0)
+
+/**
+ * struct swap_ops - per-swap-area I/O batching callbacks
+ * @can_merge: optional. Return true iff @folio can be appended to a ctx
+ *             that already holds @prev_folio of @prev_folio_size bytes.
+ *             When NULL, folios on the same swap area are batched until
+ *             the iocb is full or the plug is flushed.
+ * @submit_write: flush the accumulated write ctx to the backend.
+ * @submit_read: flush the accumulated read ctx to the backend.
+ */
+struct swap_ops {
+	unsigned int		flags;
+
+	bool			(*can_merge)(struct folio *folio,
+					     struct folio *prev_folio,
+					     size_t prev_folio_size, int rw);
+	void			(*submit_write)(struct swap_io_ctx *ctx);
+	void			(*submit_read)(struct swap_io_ctx *ctx);
+};
+
+int swap_register_block_ops(const struct block_device_operations *fops,
+			    const struct swap_ops *ops);
+void swap_unregister_block_ops(const struct block_device_operations *fops);
 
 #define SWAP_FLAG_PREFER	0x8000	/* set if swap priority specified */
 #define SWAP_FLAG_PRIO_MASK	0x7fff
diff --git a/mm/page_io.c b/mm/page_io.c
index c020e8ebf966..3ab620860379 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -24,6 +24,8 @@
 #include <linux/uio.h>
 #include <linux/sched/task.h>
 #include <linux/delayacct.h>
+#include <linux/export.h>
+#include <linux/mutex.h>
 #include <linux/zswap.h>
 #include "swap.h"
 #include "swap_table.h"
@@ -325,6 +327,8 @@ static bool swap_can_merge(struct swap_io_ctx *ctx, struct folio *folio,
 
 	if (ctx->sis != sis)
 		return false;
+	if (!sis->ops->can_merge)
+		return true;
 	return sis->ops->can_merge(folio, prev_folio, prev_folio_size, rw);
 }
 
@@ -577,6 +581,18 @@ static void swap_bio_read_end_io(struct bio *bio)
 	swap_read_end(sio, failed);
 }
 
+/**
+ * swap_bdev_submit_write - default block-device write path for swap
+ * @ctx: in-progress submit_write context.
+ *
+ * Builds a bio for the accumulated ctx and submits it through the normal
+ * block layer. This is the submit_write implementation used by swap_bdev_ops
+ * for ordinary block swap areas. swap_ops providers that override submit_write
+ * (e.g. zram) but still fall back to the block layer for some I/Os should use
+ * their own bio construction, this function is not exported.
+ *
+ * Context: process context (may sleep if SWP_SYNCHRONOUS_IO is set).
+ */
 static void swap_bdev_submit_write(struct swap_io_ctx *ctx)
 {
 	struct swap_iocb *sio = ctx->sio;
@@ -640,6 +656,96 @@ const struct swap_ops swap_bdev_ops = {
 	.can_merge		= swap_bdev_can_merge,
 };
 
+static DEFINE_MUTEX(swap_block_ops_lock);
+static const struct block_device_operations *swap_block_fops;
+static const struct swap_ops *swap_block_ops;
+
+/**
+ * swap_register_block_ops - install swap callbacks for a block driver
+ * @fops: block_device_operations identifying the driver. Used as a
+ *        match key in setup_swap_extents(): a S_ISBLK swap area is
+ *        routed to @ops when its bdev's gendisk fops equals @fops.
+ * @ops:  swap_ops vtable selected for matching swap areas. Must populate
+ *        ->submit_read and ->submit_write. ->can_merge is optional.
+ *
+ * Lets a block driver (zram and similar) replace the default
+ * swap_bdev_ops with its own submit_read / submit_write implementation.
+ *
+ * Returns 0 on success, -EINVAL when @fops or @ops are bad (a required
+ * callback is missing), or -EBUSY when the single registration slot is
+ * already taken. That slot is enough while only zram needs custom swap I/O.
+ * Several block drivers would need a per-fops lookup table instead.
+ *
+ * Context: process context, may sleep.
+ */
+int swap_register_block_ops(const struct block_device_operations *fops,
+			    const struct swap_ops *ops)
+{
+	int ret;
+
+	if (WARN_ON_ONCE(!fops || !ops || !ops->submit_read ||
+			 !ops->submit_write))
+		return -EINVAL;
+
+	mutex_lock(&swap_block_ops_lock);
+	if (swap_block_fops || swap_block_ops) {
+		ret = -EBUSY;
+		goto out;
+	}
+	swap_block_fops = fops;
+	swap_block_ops = ops;
+	ret = 0;
+out:
+	mutex_unlock(&swap_block_ops_lock);
+	return ret;
+}
+EXPORT_SYMBOL_GPL(swap_register_block_ops);
+
+/**
+ * swap_unregister_block_ops - undo swap_register_block_ops()
+ * @fops: same block_device_operations passed to swap_register_block_ops().
+ *
+ * Clears the registered fops/ops slot so future swapon calls fall back
+ * to swap_bdev_ops. The @fops match acts as a soft owner check so a
+ * driver cannot accidentally tear down another driver's registration.
+ * A mismatch is treated as a bug and triggers WARN_ON_ONCE. Swap areas
+ * that already captured the registered ops keep their sis->ops pointer.
+ * The caller must ensure the module owning the ops outlives any such
+ * swap area. For block drivers this is guaranteed by the bdev open
+ * reference held across swapon.
+ * Calling unregister before a successful register is a no-op.
+ *
+ * Context: process context, may sleep.
+ */
+void swap_unregister_block_ops(const struct block_device_operations *fops)
+{
+	mutex_lock(&swap_block_ops_lock);
+	/* never registered or already unregistered. */
+	if (!swap_block_fops)
+		goto out;
+	if (WARN_ON_ONCE(swap_block_fops != fops))
+		goto out;
+	swap_block_fops = NULL;
+	swap_block_ops = NULL;
+out:
+	mutex_unlock(&swap_block_ops_lock);
+}
+EXPORT_SYMBOL_GPL(swap_unregister_block_ops);
+
+const struct swap_ops *lookup_swap_block_ops(struct swap_info_struct *sis)
+{
+	const struct swap_ops *ops = NULL;
+
+	if (!sis->bdev)
+		return NULL;
+
+	mutex_lock(&swap_block_ops_lock);
+	if (swap_block_fops && sis->bdev->bd_disk->fops == swap_block_fops)
+		ops = swap_block_ops;
+	mutex_unlock(&swap_block_ops_lock);
+	return ops;
+}
+
 static void swap_fs_submit(struct swap_io_ctx *ctx, int rw)
 {
 	struct swap_iocb *sio = ctx->sio;
diff --git a/mm/swap.h b/mm/swap.h
index edb512e619ee..4bdd38f7a5e8 100644
--- a/mm/swap.h
+++ b/mm/swap.h
@@ -4,6 +4,7 @@
 
 #include <linux/atomic.h> /* for atomic_long_t */
 #include <linux/mm.h> /* for PAGE_SHIFT */
+#include <linux/swap.h>
 
 struct mempolicy;
 struct swap_iocb;
@@ -79,22 +80,6 @@ enum swap_cluster_flags {
 	CLUSTER_FLAG_MAX,
 };
 
-struct swap_io_ctx {
-	struct swap_iocb	*sio;
-	struct swap_info_struct	*sis;
-};
-
-#define SWAP_OPS_F_NOFS		(1U << 0)
-
-struct swap_ops {
-	unsigned int		flags;
-
-	bool (*can_merge)(struct folio *folio, struct folio *prev_folio,
-			size_t prev_folio_size, int rw);
-	void (*submit_write)(struct swap_io_ctx *ctx);
-	void (*submit_read)(struct swap_io_ctx *ctx);
-};
-
 #ifdef CONFIG_SWAP
 #include <linux/swapops.h> /* for swp_offset */
 #include <linux/blk_types.h> /* for bio_end_io_t */
@@ -472,6 +457,7 @@ static inline void __swap_cache_replace_folio(struct swap_cluster_info *ci,
 #endif /* CONFIG_SWAP */
 
 extern const struct swap_ops swap_bdev_ops;
+const struct swap_ops *lookup_swap_block_ops(struct swap_info_struct *sis);
 
 int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
 		struct list_head *folio_list);
diff --git a/mm/swapfile.c b/mm/swapfile.c
index 284eebc40a70..ebdc96092961 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -2849,6 +2849,10 @@ static int setup_swap_extents(struct swap_info_struct *sis,
 	sis->ops = &swap_bdev_ops;
 
 	if (S_ISBLK(inode->i_mode)) {
+		const struct swap_ops *block_ops = lookup_swap_block_ops(sis);
+
+		if (block_ops)
+			sis->ops = block_ops;
 		ret = add_swap_extent(sis, 0, sis->max, 0);
 		*span = sis->pages;
 		return ret;

-- 
2.43.0

[PATCH 2/3] mm/zram: handle swap read/write via swap_ops

[Jianyue Wu]

Register zram_swap_ops at module init.  The swap core still batches
folios into a swap_iocb; on flush, zram_swap_submit_write() maps each
folio page to its swap slot index and stores it via zram_write_page()
into the zspool, avoiding one bio per page.

For swap-in, zram_swap_submit_read() walks the same batch.  Without a
backing device, each slot is decompressed with read_from_zspool() while
slot_lock is held and mark_slot_accessed() runs in the same critical
section, so idle writeback cannot take the slot between read and mark.
When backing_dev is set, delegate the entire iocb to
swap_bdev_submit_read() because the batch may mix ZRAM_WB slots that
live on the backing block device.

Omit ->can_merge: zram batches through swap_iocb and compresses each
slot by index.  Block-sector merge rules do not apply.

Export swap_iocb_nr_folios(), swap_iocb_folio(), swap_read_end(),
swap_write_end(), and swap_bdev_submit_read() for the custom swap I/O
path.

Fail zram_init() if swap_register_block_ops() fails so the module
does not load without its swap path registered.

Signed-off-by: Jianyue Wu <wujianyue000@gmail.com>
---
 drivers/block/zram/zram_drv.c | 127 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/swap.h          |   5 ++
 mm/page_io.c                  |  81 ++++++++++++++++++++++++++-
 3 files changed, 210 insertions(+), 3 deletions(-)

diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 7917fc7a2a29..9b2bd0287402 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -34,6 +34,8 @@
 #include <linux/part_stat.h>
 #include <linux/kernel_read_file.h>
 #include <linux/rcupdate.h>
+#include <linux/swap.h>
+#include <linux/swapops.h>
 
 #include "zram_drv.h"
 
@@ -55,6 +57,9 @@ static unsigned int num_devices = 1;
 static size_t huge_class_size;
 
 static const struct block_device_operations zram_devops;
+#if IS_ENABLED(CONFIG_SWAP)
+static bool zram_swap_ops_registered;
+#endif
 
 static void slot_free(struct zram *zram, u32 index);
 #define slot_dep_map(zram, index) (&(zram)->table[(index)].dep_map)
@@ -2958,6 +2963,115 @@ static int zram_open(struct gendisk *disk, blk_mode_t mode)
 	return 0;
 }
 
+#if IS_ENABLED(CONFIG_SWAP)
+static void zram_swap_submit_read(struct swap_io_ctx *ctx)
+{
+	struct zram *zram = ctx->sis->bdev->bd_disk->private_data;
+	struct swap_iocb *sio = ctx->sio;
+	int nr = swap_iocb_nr_folios(sio);
+	bool failed = false;
+	int i, j;
+
+	/*
+	 * With a backing device configured, the batch may include ZRAM_WB
+	 * slots.  Fall back to the block read path for the whole iocb
+	 * instead of checking each slot.
+	 */
+#ifdef CONFIG_ZRAM_WRITEBACK
+	if (zram->backing_dev) {
+		swap_bdev_submit_read(ctx);
+		return;
+	}
+#endif
+
+	for (i = 0; i < nr; i++) {
+		struct folio *folio = swap_iocb_folio(sio, i);
+		u32 base = swp_offset(folio->swap);
+
+		for (j = 0; j < folio_nr_pages(folio); j++) {
+			u32 idx = base + j;
+			struct page *page = folio_page(folio, j);
+			int ret;
+
+			/*
+			 * read_from_zspool() and mark_slot_accessed() must run
+			 * under the same slot_lock.  zram_read_page() unlocks
+			 * before returning, which leaves a window where
+			 * writeback can pick an idle slot we just read.
+			 */
+			slot_lock(zram, idx);
+			ret = read_from_zspool(zram, page, idx);
+			if (!ret)
+				mark_slot_accessed(zram, idx);
+			slot_unlock(zram, idx);
+			if (ret) {
+				failed = true;
+				atomic64_inc(&zram->stats.failed_reads);
+				pr_alert_ratelimited("Read-error on swap-device %s at index %u: err=%d\n",
+						     zram->disk->disk_name, idx, ret);
+				goto out;
+			}
+			flush_dcache_page(page);
+		}
+	}
+out:
+	swap_read_end(sio, failed);
+}
+
+static void zram_swap_submit_write(struct swap_io_ctx *ctx)
+{
+	struct zram *zram = ctx->sis->bdev->bd_disk->private_data;
+	struct swap_iocb *sio = ctx->sio;
+	int nr = swap_iocb_nr_folios(sio);
+	bool failed = false;
+	int i, j, ret = 0;
+	u32 idx = 0;
+
+	for (i = 0; i < nr; i++) {
+		struct folio *folio = swap_iocb_folio(sio, i);
+		u32 base = swp_offset(folio->swap);
+
+		for (j = 0; j < folio_nr_pages(folio); j++) {
+			idx = base + j;
+			ret = zram_write_page(zram, folio_page(folio, j), idx);
+			if (ret) {
+				/*
+				 * Leave partial zram data in place, same as the bio
+				 * write path.  swap_write_end() re-dirties every
+				 * page in the batch so they stay in swapcache with
+				 * their swap entries.  Freeing zram slots here would
+				 * leave entries pointing at empty indices until
+				 * slot_free_notify runs.
+				 */
+				failed = true;
+				atomic64_inc(&zram->stats.failed_writes);
+				pr_alert_ratelimited("Write-error on swap-device %s at index %u: err=%d\n",
+						     zram->disk->disk_name, idx, ret);
+				goto out;
+			}
+			slot_lock(zram, idx);
+			mark_slot_accessed(zram, idx);
+			slot_unlock(zram, idx);
+		}
+	}
+out:
+	swap_write_end(sio, failed);
+}
+
+/*
+ * No ->can_merge: block rules exist to grow bios on contiguous sectors and
+ * matching blkcg.  zram already batches through swap_iocb, and
+ * submit_write() compresses each slot by index, not by sector layout.
+ * Reusing swap_bdev_can_merge() would only split batches without helping
+ * zspool I/O.
+ */
+static const struct swap_ops zram_swap_ops = {
+	.submit_read		= zram_swap_submit_read,
+	.submit_write		= zram_swap_submit_write,
+};
+
+#endif /* CONFIG_SWAP */
+
 static const struct block_device_operations zram_devops = {
 	.open = zram_open,
 	.submit_bio = zram_submit_bio,
@@ -3233,6 +3347,10 @@ static int zram_remove_cb(int id, void *ptr, void *data)
 
 static void destroy_devices(void)
 {
+#if IS_ENABLED(CONFIG_SWAP)
+	if (zram_swap_ops_registered)
+		swap_unregister_block_ops(&zram_devops);
+#endif
 	class_unregister(&zram_control_class);
 	idr_for_each(&zram_index_idr, &zram_remove_cb, NULL);
 	zram_debugfs_destroy();
@@ -3269,6 +3387,15 @@ static int __init zram_init(void)
 		return -EBUSY;
 	}
 
+#if IS_ENABLED(CONFIG_SWAP)
+	ret = swap_register_block_ops(&zram_devops, &zram_swap_ops);
+	if (ret) {
+		pr_err("zram: failed to register swap ops (%d)\n", ret);
+		goto out_error;
+	}
+	zram_swap_ops_registered = true;
+#endif
+
 	while (num_devices != 0) {
 		mutex_lock(&zram_index_mutex);
 		ret = zram_add();
diff --git a/include/linux/swap.h b/include/linux/swap.h
index 1d51df4179c1..70bf6f3f04dc 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -54,6 +54,11 @@ struct swap_ops {
 int swap_register_block_ops(const struct block_device_operations *fops,
 			    const struct swap_ops *ops);
 void swap_unregister_block_ops(const struct block_device_operations *fops);
+int swap_iocb_nr_folios(struct swap_iocb *sio);
+struct folio *swap_iocb_folio(struct swap_iocb *sio, int idx);
+void swap_read_end(struct swap_iocb *sio, bool failed);
+void swap_write_end(struct swap_iocb *sio, bool failed);
+void swap_bdev_submit_read(struct swap_io_ctx *ctx);
 
 #define SWAP_FLAG_PREFER	0x8000	/* set if swap priority specified */
 #define SWAP_FLAG_PRIO_MASK	0x7fff
diff --git a/mm/page_io.c b/mm/page_io.c
index 3ab620860379..7c17e44823d1 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -486,7 +486,21 @@ void swap_read_folio(struct swap_io_ctx *ctx, struct folio *folio)
 	delayacct_swapin_end();
 }
 
-static void swap_write_end(struct swap_iocb *sio, bool failed)
+/**
+ * swap_write_end - finish a swap write iocb
+ * @sio:    swap_iocb whose pages were just written
+ * @failed: true if any of the underlying writes failed
+ *
+ * Ends writeback on every page captured by @sio. On failure each page
+ * is also re-dirtied and PG_reclaim is cleared, mirroring the bio
+ * write completion path. @sio is returned to the swap iocb mempool.
+ *
+ * swap_ops providers must call this exactly once per submit_write()
+ * ctx (typically at the end of their submit_write callback).
+ *
+ * Context: any context the submit_write() callback runs in.
+ */
+void swap_write_end(struct swap_iocb *sio, bool failed)
 {
 	int p;
 
@@ -501,6 +515,7 @@ static void swap_write_end(struct swap_iocb *sio, bool failed)
 	}
 	mempool_free(sio, sio_pool);
 }
+EXPORT_SYMBOL_GPL(swap_write_end);
 
 static void swap_fs_write_complete(struct kiocb *iocb, long ret)
 {
@@ -536,7 +551,26 @@ static void end_swap_bio_write(struct bio *bio)
 	swap_write_end(sio, failed);
 }
 
-static void swap_read_end(struct swap_iocb *sio, bool failed)
+/**
+ * swap_read_end - finish a swap read iocb
+ * @sio:    swap_iocb whose folios were just read in
+ * @failed: true if any of the underlying reads failed
+ *
+ * Unlocks every folio captured by @sio. On success each folio is also
+ * marked uptodate and swap-in counters (PSWPIN, mTHP, memcg) are bumped
+ * by folio_nr_pages(). On failure folios are left not-uptodate so the
+ * caller observes the failure and retries or surfaces an error. @sio is
+ * returned to the swap iocb mempool.
+ *
+ * swap_ops providers must call this exactly once per submit_read() ctx
+ * (typically at the end of their submit_read callback). If the provider
+ * defers to swap_bdev_ops.submit_read() for fallback, the bdev path
+ * will call swap_read_end() itself and the provider must not call it
+ * again for the same ctx.
+ *
+ * Context: any context the submit_read() callback runs in.
+ */
+void swap_read_end(struct swap_iocb *sio, bool failed)
 {
 	int p;
 
@@ -557,6 +591,34 @@ static void swap_read_end(struct swap_iocb *sio, bool failed)
 
 	mempool_free(sio, sio_pool);
 }
+EXPORT_SYMBOL_GPL(swap_read_end);
+
+/**
+ * swap_iocb_nr_folios - number of folios in a swap I/O batch
+ * @sio: swap_iocb passed to a swap_ops submit callback.
+ *
+ * Returns how many folios the swap core has batched into @sio. Used
+ * together with swap_iocb_folio() so swap_ops providers can walk the
+ * batch without depending on the swap core's internal iocb layout.
+ */
+int swap_iocb_nr_folios(struct swap_iocb *sio)
+{
+	return sio->nr_bvecs;
+}
+EXPORT_SYMBOL_GPL(swap_iocb_nr_folios);
+
+/**
+ * swap_iocb_folio - folio at slot @idx in a swap I/O batch
+ * @sio: swap_iocb passed to a swap_ops submit callback.
+ * @idx: index in the range [0, swap_iocb_nr_folios(@sio)).
+ *
+ * Returns the folio at the given batch slot.
+ */
+struct folio *swap_iocb_folio(struct swap_iocb *sio, int idx)
+{
+	return page_folio(sio->bvecs[idx].bv_page);
+}
+EXPORT_SYMBOL_GPL(swap_iocb_folio);
 
 static void swap_fs_read_complete(struct kiocb *iocb, long ret)
 {
@@ -613,7 +675,19 @@ static void swap_bdev_submit_write(struct swap_io_ctx *ctx)
 	}
 }
 
-static void swap_bdev_submit_read(struct swap_io_ctx *ctx)
+/**
+ * swap_bdev_submit_read - fall back to the default block-device read path
+ * @ctx: in-progress submit_read context.
+ *
+ * Builds a bio for the accumulated ctx and submits it through the
+ * normal block layer. swap_ops providers can call this when they
+ * cannot serve a particular ctx themselves (for example zram folios
+ * stored on a backing device). The bio completion path takes care of
+ * calling swap_read_end() on @ctx. The caller must not call it again.
+ *
+ * Context: any context the submit_read() callback runs in.
+ */
+void swap_bdev_submit_read(struct swap_io_ctx *ctx)
 {
 	struct swap_iocb *sio = ctx->sio;
 	struct bio *bio = &sio->bio;
@@ -638,6 +712,7 @@ static void swap_bdev_submit_read(struct swap_io_ctx *ctx)
 		submit_bio(bio);
 	}
 }
+EXPORT_SYMBOL_GPL(swap_bdev_submit_read);
 
 static bool swap_bdev_can_merge(struct folio *folio, struct folio *prev_folio,
 		size_t prev_folio_size, int rw)

-- 
2.43.0

[PATCH 3/3] mm/swap: route slot free notifications through swap_ops

[Jianyue Wu]

Dispatch slot_free_notify through swap_ops instead of
block_device_operations. Zram keeps slot-free handling alongside its
other swap_ops methods.

Move slot_trylock into the CONFIG_SWAP block. With CONFIG_SWAP=n it
has no callers and the build fails on -Werror=unused-function.

Document the callback locking rules in include/linux/swap.h. Remove
the outdated locking.rst note for swap_slot_free_notify.

Signed-off-by: Jianyue Wu <wujianyue000@gmail.com>
---
 Documentation/filesystems/locking.rst |  5 --
 drivers/block/zram/zram_drv.c         | 88 ++++++++++++++++++-----------------
 include/linux/blkdev.h                |  2 -
 include/linux/swap.h                  |  7 +++
 mm/swapfile.c                         | 13 ++----
 rust/kernel/block/mq/gen_disk.rs      |  1 -
 6 files changed, 57 insertions(+), 59 deletions(-)

diff --git a/Documentation/filesystems/locking.rst b/Documentation/filesystems/locking.rst
index 70481bdc031d..964c841bf917 100644
--- a/Documentation/filesystems/locking.rst
+++ b/Documentation/filesystems/locking.rst
@@ -443,7 +443,6 @@ prototypes::
 				unsigned long *);
 	void (*unlock_native_capacity) (struct gendisk *);
 	int (*getgeo)(struct gendisk *, struct hd_geometry *);
-	void (*swap_slot_free_notify) (struct block_device *, unsigned long);
 
 locking rules:
 
@@ -457,12 +456,8 @@ compat_ioctl:		no
 direct_access:		no
 unlock_native_capacity:	no
 getgeo:			no
-swap_slot_free_notify:	no	(see below)
 ======================= ===================
 
-swap_slot_free_notify is called with swap_lock and sometimes the page lock
-held.
-
 
 file_operations
 ===============
diff --git a/drivers/block/zram/zram_drv.c b/drivers/block/zram/zram_drv.c
index 9b2bd0287402..b78246dc1746 100644
--- a/drivers/block/zram/zram_drv.c
+++ b/drivers/block/zram/zram_drv.c
@@ -72,31 +72,6 @@ static void slot_lock_init(struct zram *zram, u32 index)
 			 &__key, 0);
 }
 
-/*
- * entry locking rules:
- *
- * 1) Lock is exclusive
- *
- * 2) lock() function can sleep waiting for the lock
- *
- * 3) Lock owner can sleep
- *
- * 4) Use TRY lock variant when in atomic context
- *    - must check return value and handle locking failers
- */
-static __must_check bool slot_trylock(struct zram *zram, u32 index)
-{
-	unsigned long *lock = &zram->table[index].__lock;
-
-	if (!test_and_set_bit_lock(ZRAM_ENTRY_LOCK, lock)) {
-		mutex_acquire(slot_dep_map(zram, index), 0, 1, _RET_IP_);
-		lock_acquired(slot_dep_map(zram, index), _RET_IP_);
-		return true;
-	}
-
-	return false;
-}
-
 static void slot_lock(struct zram *zram, u32 index)
 {
 	unsigned long *lock = &zram->table[index].__lock;
@@ -2798,23 +2773,6 @@ static void zram_submit_bio(struct bio *bio)
 	}
 }
 
-static void zram_slot_free_notify(struct block_device *bdev,
-				unsigned long index)
-{
-	struct zram *zram;
-
-	zram = bdev->bd_disk->private_data;
-
-	atomic64_inc(&zram->stats.notify_free);
-	if (!slot_trylock(zram, index)) {
-		atomic64_inc(&zram->stats.miss_free);
-		return;
-	}
-
-	slot_free(zram, index);
-	slot_unlock(zram, index);
-}
-
 static void zram_comp_params_reset(struct zram *zram)
 {
 	u32 prio;
@@ -3058,6 +3016,50 @@ static void zram_swap_submit_write(struct swap_io_ctx *ctx)
 	swap_write_end(sio, failed);
 }
 
+/*
+ * entry locking rules:
+ *
+ * 1) Lock is exclusive
+ *
+ * 2) lock() function can sleep waiting for the lock
+ *
+ * 3) Lock owner can sleep
+ *
+ * 4) Use TRY lock variant when in atomic context
+ *    - must check return value and handle locking failers
+ */
+static __must_check bool slot_trylock(struct zram *zram, u32 index)
+{
+	unsigned long *lock = &zram->table[index].__lock;
+
+	if (!test_and_set_bit_lock(ZRAM_ENTRY_LOCK, lock)) {
+		mutex_acquire(slot_dep_map(zram, index), 0, 1, _RET_IP_);
+		lock_acquired(slot_dep_map(zram, index), _RET_IP_);
+		return true;
+	}
+
+	return false;
+}
+
+/*
+ * swap_range_free() holds the swap cluster lock. Use slot_trylock() so
+ * we never block on a slot that is already locked elsewhere.
+ */
+static void zram_swap_slot_free_notify(struct swap_info_struct *sis,
+				       unsigned long index)
+{
+	struct zram *zram = sis->bdev->bd_disk->private_data;
+
+	atomic64_inc(&zram->stats.notify_free);
+	if (!slot_trylock(zram, index)) {
+		atomic64_inc(&zram->stats.miss_free);
+		return;
+	}
+
+	slot_free(zram, index);
+	slot_unlock(zram, index);
+}
+
 /*
  * No ->can_merge: block rules exist to grow bios on contiguous sectors and
  * matching blkcg.  zram already batches through swap_iocb, and
@@ -3068,6 +3070,7 @@ static void zram_swap_submit_write(struct swap_io_ctx *ctx)
 static const struct swap_ops zram_swap_ops = {
 	.submit_read		= zram_swap_submit_read,
 	.submit_write		= zram_swap_submit_write,
+	.slot_free_notify	= zram_swap_slot_free_notify,
 };
 
 #endif /* CONFIG_SWAP */
@@ -3075,7 +3078,6 @@ static const struct swap_ops zram_swap_ops = {
 static const struct block_device_operations zram_devops = {
 	.open = zram_open,
 	.submit_bio = zram_submit_bio,
-	.swap_slot_free_notify = zram_slot_free_notify,
 	.owner = THIS_MODULE
 };
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 890128cdea1c..f861ceed39eb 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1669,8 +1669,6 @@ struct block_device_operations {
 	int (*getgeo)(struct gendisk *, struct hd_geometry *);
 	int (*set_read_only)(struct block_device *bdev, bool ro);
 	void (*free_disk)(struct gendisk *disk);
-	/* this callback is with swap_lock and sometimes page table lock held */
-	void (*swap_slot_free_notify) (struct block_device *, unsigned long);
 	int (*report_zones)(struct gendisk *, sector_t sector,
 			    unsigned int nr_zones,
 			    struct blk_report_zones_args *args);
diff --git a/include/linux/swap.h b/include/linux/swap.h
index 70bf6f3f04dc..09640eb5a45d 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -40,6 +40,11 @@ struct swap_io_ctx {
  *             the iocb is full or the plug is flushed.
  * @submit_write: flush the accumulated write ctx to the backend.
  * @submit_read: flush the accumulated read ctx to the backend.
+ * @slot_free_notify: optional callback invoked when a swap slot
+ *                    becomes free. swap_range_free() calls it with the
+ *                    swap cluster lock held. The folio lock may also be
+ *                    held on swap-cache teardown paths. Must not sleep
+ *                    or block.
  */
 struct swap_ops {
 	unsigned int		flags;
@@ -49,6 +54,8 @@ struct swap_ops {
 					     size_t prev_folio_size, int rw);
 	void			(*submit_write)(struct swap_io_ctx *ctx);
 	void			(*submit_read)(struct swap_io_ctx *ctx);
+	void			(*slot_free_notify)(struct swap_info_struct *sis,
+						    unsigned long offset);
 };
 
 int swap_register_block_ops(const struct block_device_operations *fops,
diff --git a/mm/swapfile.c b/mm/swapfile.c
index ebdc96092961..79a4166fb9bf 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -1311,21 +1311,18 @@ static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
 			    unsigned int nr_entries)
 {
 	unsigned long end = offset + nr_entries - 1;
-	void (*swap_slot_free_notify)(struct block_device *, unsigned long);
+	void (*slot_free_notify)(struct swap_info_struct *sis,
+				 unsigned long offset);
 	unsigned int i;
 
 	for (i = 0; i < nr_entries; i++)
 		zswap_invalidate(swp_entry(si->type, offset + i));
 
-	if (si->flags & SWP_BLKDEV)
-		swap_slot_free_notify =
-			si->bdev->bd_disk->fops->swap_slot_free_notify;
-	else
-		swap_slot_free_notify = NULL;
+	slot_free_notify = si->ops->slot_free_notify;
 	while (offset <= end) {
 		arch_swap_invalidate_page(si->type, offset);
-		if (swap_slot_free_notify)
-			swap_slot_free_notify(si->bdev, offset);
+		if (slot_free_notify)
+			slot_free_notify(si, offset);
 		offset++;
 	}
 
diff --git a/rust/kernel/block/mq/gen_disk.rs b/rust/kernel/block/mq/gen_disk.rs
index 912cb805caf5..25552d69f711 100644
--- a/rust/kernel/block/mq/gen_disk.rs
+++ b/rust/kernel/block/mq/gen_disk.rs
@@ -135,7 +135,6 @@ pub fn build<T: Operations>(
             unlock_native_capacity: None,
             getgeo: None,
             set_read_only: None,
-            swap_slot_free_notify: None,
             report_zones: None,
             devnode: None,
             alternative_gpt_sector: None,

-- 
2.43.0

Re: [PATCH 1/3] mm/page_io: let block drivers register custom swap I/O ops

[YoungJun Park]

On Sun, Jun 14, 2026 at 11:35:29PM +0800, Jianyue Wu wrote:

...

Hello Jianyue.

Currently, the patch commit log indicates only a single custom swap
registration is supported. Shouldn't we allow multiple block drivers to
register their custom ops simultaneously from the beginning?

>  int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
>  		struct list_head *folio_list);
> diff --git a/mm/swapfile.c b/mm/swapfile.c
> index 284eebc40a70..ebdc96092961 100644
> --- a/mm/swapfile.c
> +++ b/mm/swapfile.c
> @@ -2849,6 +2849,10 @@ static int setup_swap_extents(struct swap_info_struct *sis,
>  	sis->ops = &swap_bdev_ops;
>
>  	if (S_ISBLK(inode->i_mode)) {
> +		const struct swap_ops *block_ops = lookup_swap_block_ops(sis);

Also, just a personal thought on this part.

Instead of using `block_device_fops` as a lookup key, what if we handle
this similarly to how filesystems use the `a_ops->swap_activate` callback?

We could add a `swap_activate` callback directly into
struct block_device_operations (zram's zram_devops). This way, the
block device itself can set up and replace the swap `ops` directly without
needing a separate registration/lookup mechanism.

What are your thoughts on this approach?

Thanks,
Youngjun Park

Re: [PATCH 2/3] mm/zram: handle swap read/write via swap_ops

[YoungJun Park]

On Sun, Jun 14, 2026 at 11:35:30PM +0800, Jianyue Wu wrote:

Hello!

> +static void zram_swap_submit_read(struct swap_io_ctx *ctx)
> +{
> +	struct zram *zram = ctx->sis->bdev->bd_disk->private_data;

A passing thought. accessing `zram` here is too indirect. We might
need a `private_data` in the swap device struct someday?

(And If there is a real value like some swap-side only private data really needed.)

> +	struct swap_iocb *sio = ctx->sio;
> +	int nr = swap_iocb_nr_folios(sio);
> +	bool failed = false;
> +	int i, j;
> +			/*
> +			 * read_from_zspool() and mark_slot_accessed() must run
> +			 * under the same slot_lock.  zram_read_page() unlocks
> +			 * before returning, which leaves a window where
> +			 * writeback can pick an idle slot we just read.
> +			 */

Regarding the comment about the "window" where writeback can pick an
idle slot. I think this reasoning is a bit of a gray area. Writeback
could just as easily pick the slot right before entering this routine,
so the race condition seems fundamentally the same.

Isn't the actual justification here to separate the non-backend logic
and ensure mark_slot_accessed() is called under the lock, given that
zram_read_page() can call the backend device?

If the "window" mentioned in the comment is indeed a valid issue, then
zram_read_page() has the exact same problem and needs to be fixed as
well?

If not, IMHO I suggest revising or removing this comment to clarify
the true(?) intention. :)

> +			slot_lock(zram, idx);
> +			ret = read_from_zspool(zram, page, idx);
> +			if (!ret)
> +				mark_slot_accessed(zram, idx);
> +			slot_unlock(zram, idx);

Re: [PATCH 0/3] mm/zram: route block swap I/O through swap_ops

[Barry Song]

On Sun, Jun 14, 2026 at 11:35 PM Jianyue Wu <wujianyue000@gmail.com> wrote:
>
> This series builds on Christoph Hellwig's swap batching rework that
> moves block swap onto struct swap_iocb and per-backend struct
> swap_ops handlers [1].  Christoph's patches unify batching for
> ordinary block devices and swap files.  zram still needs a custom
> path because swap slots map to compressed pages, not disk sectors.
>
> The first patch adds swap_register_block_ops() so a block driver can
> install custom submit_read/submit_write handlers when swapon targets
> its block device.  The default swap_bdev_ops path is unchanged for
> devices that do not register.
>
> The second patch registers zram_swap_ops at module init.  On write,
> the swap core still batches folios into a swap_iocb.  zram maps each
> folio to a slot index and stores it through zram_write_page() instead
> of building one bio per page.  Read handling keeps slot_lock and
> mark_slot_accessed() in one critical section.  Writeback-enabled zram
> falls back to swap_bdev_submit_read() for ZRAM_WB slots.
>
> The third patch moves slot_free_notify into swap_ops next to the
> other zram swap callbacks, and documents the locking contract for
> that hook.
>
> Applied on top of Christoph Hellwig's "better block swap batching and
> a different take on swap_ops" series [1].

Nice. I think it's better to mark it as RFC at this stage.

By the way, besides the architectural refinements, have
you also observed any noticeable performance improvements?

>
> [1] https://lore.kernel.org/linux-mm/?q=better+block+swap+batching

Best Regards
Barry

Re: [PATCH 1/3] mm/page_io: let block drivers register custom swap I/O ops

[Jianyue Wu]

On 6/15/2026 9:50 AM, YoungJun Park wrote:
> On Sun, Jun 14, 2026 at 11:35:29PM +0800, Jianyue Wu wrote:
>
> ...
>
> Hello Jianyue.
>
> Currently, the patch commit log indicates only a single custom swap
> registration is supported. Shouldn't we allow multiple block drivers to
> register their custom ops simultaneously from the beginning?
>
>> int shmem_writeout(struct swap_io_ctx *ctx, struct folio *folio,
>> struct list_head *folio_list);
>> diff --git a/mm/swapfile.c b/mm/swapfile.c
>> index 284eebc40a70..ebdc96092961 100644
>> --- a/mm/swapfile.c
>> +++ b/mm/swapfile.c
>> @@ -2849,6 +2849,10 @@ static int setup_swap_extents(struct swap_info_struct *sis,
>> sis->ops = &swap_bdev_ops;
>>
>> if (S_ISBLK(inode->i_mode)) {
>> + const struct swap_ops *block_ops = lookup_swap_block_ops(sis);
>
> Also, just a personal thought on this part.
>
> Instead of using `block_device_fops` as a lookup key, what if we handle
> this similarly to how filesystems use the `a_ops->swap_activate` callback?
>
> We could add a `swap_activate` callback directly into
> struct block_device_operations (zram's zram_devops). This way, the
> block device itself can set up and replace the swap `ops` directly without
> needing a separate registration/lookup mechanism.
>
> What are your thoughts on this approach?
>
> Thanks,
> Youngjun Park
>

Hello Youngjun,

On multiple registrations:
Previously I was also a bit hesitate about this. Exactly, better to
support multiple block driver directly, I'll update it.

On swap_activate:
That's a very good idea, to use swap_activate callback, it is much
cleaner, I like this approach:) setup_swap_extents() would call it for
S_ISBLK swap targets, and the driver would install sis->ops at swapon
time. When the callback is NULL, the core can fall back to
swap_bdev_activate() and swap_bdev_ops. That removes the separate global
registration/lookup mechanism entirely, and multiple block drivers are
supported naturally because each device carries its own ops table.

Thanks,
Jianyue

Re: [PATCH 2/3] mm/zram: handle swap read/write via swap_ops

[Jianyue Wu]

On Mon, Jun 15, 2026 at 2:39 PM YoungJun Park <youngjun.park@lge.com> wrote:
>
> On Sun, Jun 14, 2026 at 11:35:30PM +0800, Jianyue Wu wrote:
>
> Hello!
>
> > +static void zram_swap_submit_read(struct swap_io_ctx *ctx)
> > +{
> > +     struct zram *zram = ctx->sis->bdev->bd_disk->private_data;
>
> A passing thought. accessing `zram` here is too indirect. We might
> need a `private_data` in the swap device struct someday?
>
> (And If there is a real value like some swap-side only private data really needed.)
>
> > +     struct swap_iocb *sio = ctx->sio;
> > +     int nr = swap_iocb_nr_folios(sio);
> > +     bool failed = false;
> > +     int i, j;
> > +                     /*
> > +                      * read_from_zspool() and mark_slot_accessed() must run
> > +                      * under the same slot_lock.  zram_read_page() unlocks
> > +                      * before returning, which leaves a window where
> > +                      * writeback can pick an idle slot we just read.
> > +                      */
>
> Regarding the comment about the "window" where writeback can pick an
> idle slot. I think this reasoning is a bit of a gray area. Writeback
> could just as easily pick the slot right before entering this routine,
> so the race condition seems fundamentally the same.
>
> Isn't the actual justification here to separate the non-backend logic
> and ensure mark_slot_accessed() is called under the lock, given that
> zram_read_page() can call the backend device?
>
> If the "window" mentioned in the comment is indeed a valid issue, then
> zram_read_page() has the exact same problem and needs to be fixed as
> well?
>
> If not, IMHO I suggest revising or removing this comment to clarify
> the true(?) intention. :)
>
> > +                     slot_lock(zram, idx);
> > +                     ret = read_from_zspool(zram, page, idx);
> > +                     if (!ret)
> > +                             mark_slot_accessed(zram, idx);
> > +                     slot_unlock(zram, idx);
>

Hello Youngjun,

Agree. Walking ctx->sis->bdev->bd_disk->private_data
from every swap_ops callback is too indirect. I will add an opaque
private_data field to struct swap_info_struct, set it from
->swap_activate() when the swap area is set up, and clear it on
swapoff. The zram callbacks will then use ctx->sis->private_data directly.

You are right. The writeback "window" reasoning was overstated.
Writeback could already have picked the slot before we enter the swap
read path, we have ZRAM_PP_SLOT to ensure it.
0. // condition 1 write pick the slot before the lock.
1. lock → read_from_zspool → unlock
2. // condition 2 write pick the slot inside the lock.
3. lock → mark_slot_accessed() → unlock // clear ZRAM_IDLE and ZRAM_PP_SLOT flag

I think simply removing this comment is good.

Thanks,
Jianyue

Re: [PATCH 0/3] mm/zram: route block swap I/O through swap_ops

[Jianyue Wu]

On Mon, Jun 15, 2026 at 5:14 PM Barry Song <baohua@kernel.org> wrote:
>
> On Sun, Jun 14, 2026 at 11:35 PM Jianyue Wu <wujianyue000@gmail.com> wrote:
> >
> > This series builds on Christoph Hellwig's swap batching rework that
> > moves block swap onto struct swap_iocb and per-backend struct
> > swap_ops handlers [1].  Christoph's patches unify batching for
> > ordinary block devices and swap files.  zram still needs a custom
> > path because swap slots map to compressed pages, not disk sectors.
> >
> > The first patch adds swap_register_block_ops() so a block driver can
> > install custom submit_read/submit_write handlers when swapon targets
> > its block device.  The default swap_bdev_ops path is unchanged for
> > devices that do not register.
> >
> > The second patch registers zram_swap_ops at module init.  On write,
> > the swap core still batches folios into a swap_iocb.  zram maps each
> > folio to a slot index and stores it through zram_write_page() instead
> > of building one bio per page.  Read handling keeps slot_lock and
> > mark_slot_accessed() in one critical section.  Writeback-enabled zram
> > falls back to swap_bdev_submit_read() for ZRAM_WB slots.
> >
> > The third patch moves slot_free_notify into swap_ops next to the
> > other zram swap callbacks, and documents the locking contract for
> > that hook.
> >
> > Applied on top of Christoph Hellwig's "better block swap batching and
> > a different take on swap_ops" series [1].
>
> Nice. I think it's better to mark it as RFC at this stage.
>
> By the way, besides the architectural refinements, have
> you also observed any noticeable performance improvements?
>
> >
> > [1] https://lore.kernel.org/linux-mm/?q=better+block+swap+batching
>
> Best Regards
> Barry

Hello Barry,

Thanks for the feedback:) I will mark the next revision as RFC.

I ran some local measurements on a zram swap workload.
Without a backing device (zspool-only swap read path), the swap_ops
path looks slightly better on average and median latency, while p99 is
roughly flat:
avg 1,750 ns vs 1,812 ns
p50 1,273 ns vs 1,504 ns
p99 6,318 ns vs 6,198 ns

With writeback/backing device enabled, the numbers are much noisier
(bd_reads per sample and cold-fault ratio varied a lot between runs),
so I would not read too much into them. Directionally, the swap_ops
path looked faster on avg/p50/p99 in the runs I captured, but I need
more controlled repeats before claiming a real win:
avg 39 µs vs 77 µs
p50 4.5 µs vs 90 µs
p99 116 µs vs 210 µs

bd_reads/sample 0.37 vs 0.75
cold-fault samples 62.5% vs 100%

So far I would describe the gain as modest for the common zspool case,
maybe because doesn't have merge benefit like bio side.
With the main motivation still being architectural fit (put zram
swap semantics under swap_ops) rather than a large performance jump.

Thanks,
Jianyue

Re: [PATCH 0/3] mm/zram: route block swap I/O through swap_ops

[Christoph Hellwig]

I fear this is going entirely in the wrong direction.

Yes, we have to keep zram around as a legacy interface for now,
but the right place to deal with compressed swap is in the core.

So please don't add more hacks for 'magic' block devices.