iomap: add a swapfile activation function

From: Darrick J. Wong <[email protected]>

Add a new iomap_swapfile_activate function so that filesystems can activate swap files without having to use the obsolete and slow bmap function. This enables XFS to support fallocate'd swap files and swap files on realtime devices.

Signed-off-by Darrick J. Wong <[email protected]>

---
v3: catch null iomap addr, fix too-short extent detection
v2: document the swap file layout requirements, combine adjacent
real/unwritten extents, align reported swap extents to physical page size boundaries, fix compiler errors when swap disabled

Quoted 7 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Add a new iomap_swapfile_activate function so that filesystems can activate swap files without having to use the obsolete and slow bmap function. This enables XFS to support fallocate'd swap files and swap files on realtime devices.

Shouldn't we also prevent the extents of an active swapfile from becoming shared? If I swapon(a) and reflink(a, b), swapout to a now has to break the reflink or corrupt b! In my old Btrfs swapfile series [1] I just forbid all reflink operations on active swapfiles.

One thing to note is that then this will need a matching ->swap_deactivate(), which currently isn't called if ->swap_activate() returned > 0.

1: https://github.com/osandov/linux/tree/btrfs-swap

Quoted 6 lines

Signed-off-by Darrick J. Wong <[email protected]>

---
v3: catch null iomap addr, fix too-short extent detection
v2: document the swap file layout requirements, combine adjacent
real/unwritten extents, align reported swap extents to physical page size boundaries, fix compiler errors when swap disabled

Quoted 9 lines · Omar Sandoval

Quoted 7 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Add a new iomap_swapfile_activate function so that filesystems can activate swap files without having to use the obsolete and slow bmap function. This enables XFS to support fallocate'd swap files and swap files on realtime devices.

Quoted 4 lines

Shouldn't we also prevent the extents of an active swapfile from becoming shared? If I swapon(a) and reflink(a, b), swapout to a now has to break the reflink or corrupt b! In my old Btrfs swapfile series [1] I just forbid all reflink operations on active swapfiles.

xfs already does this in its reflink handler: it takes the inode lock & bails out if IS_SWAPFILE(). swapon calls claim_swapfile to take the inode lock and sets S_SWAPFILE (if successful) so the two are effectively locked out from each other...

Quoted 3 lines

One thing to note is that then this will need a matching ->swap_deactivate(), which currently isn't called if ->swap_activate() returned > 0.

...so there shouldn't be any state to undo if the
iomap_swapfile_activate fails.

From: Darrick J. Wong <[email protected]>

Add a new iomap_swapfile_activate function so that filesystems can activate swap files without having to use the obsolete and slow bmap function. This enables XFS to support fallocate'd swap files and swap files on realtime devices.

Signed-off-by Darrick J. Wong <[email protected]>

---
v3: catch null iomap addr, fix too-short extent detection
v2: document the swap file layout requirements, combine adjacent
real/unwritten extents, align reported swap extents to physical page size boundaries, fix compiler errors when swap disabled

---
 fs/iomap.c            |  162 +++++++++++++++++++++++++++++++++++++++++++++++++
 fs/xfs/xfs_aops.c     |   12 ++++
 include/linux/iomap.h |   11 +++
 3 files changed, 185 insertions(+)

fs/iomap.c +26 -0

index afd163586aa0..ac7115492366 100644
@@ -27,6 +27,7 @@
 #include <linux/task_io_accounting_ops.h>
 #include <linux/dax.h>
 #include <linux/sched/signal.h>
+#include <linux/swap.h>
 
 #include "internal.h"
 

@@ -1089,3 +1090,164 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	return ret;
 }
 EXPORT_SYMBOL_GPL(iomap_dio_rw);
+
+/* Swapfile activation */
+
+#ifdef CONFIG_SWAP
+struct iomap_swapfile_info {
+	struct iomap iomap;		/* accumulated iomap */
+	struct swap_info_struct *sis;
+	uint64_t lowest_ppage;		/* lowest physical addr seen (pages) */
+	uint64_t highest_ppage;		/* highest physical addr seen (pages) */
+	unsigned long nr_pages;		/* number of pages collected */
+	int nr_extents;			/* extent count */
+};
+
+/*
+ * Collect physical extents for this swap file.  Physical extents reported to
+ * the swap code must be trimmed to align to a page boundary.  The logical
+ * offset within the file is irrelevant since the swapfile code maps logical
+ * page numbers of the swap device to the physical page-aligned extents.
+ */
+static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
+{
+	struct iomap *iomap = &isi->iomap;
+	unsigned long nr_pages;
+	uint64_t first_ppage;
+	uint64_t first_ppage_reported;

Quoted 65 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

Add a new iomap_swapfile_activate function so that filesystems can activate swap files without having to use the obsolete and slow bmap function. This enables XFS to support fallocate'd swap files and swap files on realtime devices.

Signed-off-by Darrick J. Wong <[email protected]>

---
v3: catch null iomap addr, fix too-short extent detection
v2: document the swap file layout requirements, combine adjacent
real/unwritten extents, align reported swap extents to physical page size boundaries, fix compiler errors when swap disabled

---
 fs/iomap.c            |  162 +++++++++++++++++++++++++++++++++++++++++++++++++
 fs/xfs/xfs_aops.c     |   12 ++++
 include/linux/iomap.h |   11 +++
 3 files changed, 185 insertions(+)

fs/iomap.c +31 -0

index afd163586aa0..ac7115492366 100644
@@ -27,6 +27,7 @@
 #include <linux/task_io_accounting_ops.h>
 #include <linux/dax.h>
 #include <linux/sched/signal.h>
+#include <linux/swap.h>
 
 #include "internal.h"
 

@@ -1089,3 +1090,164 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	return ret;
 }
 EXPORT_SYMBOL_GPL(iomap_dio_rw);
+
+/* Swapfile activation */
+
+#ifdef CONFIG_SWAP
+struct iomap_swapfile_info {
+	struct iomap iomap;		/* accumulated iomap */
+	struct swap_info_struct *sis;
+	uint64_t lowest_ppage;		/* lowest physical addr seen (pages) */
+	uint64_t highest_ppage;		/* highest physical addr seen (pages) */
+	unsigned long nr_pages;		/* number of pages collected */
+	int nr_extents;			/* extent count */
+};
+
+/*
+ * Collect physical extents for this swap file.  Physical extents reported to
+ * the swap code must be trimmed to align to a page boundary.  The logical
+ * offset within the file is irrelevant since the swapfile code maps logical
+ * page numbers of the swap device to the physical page-aligned extents.
+ */
+static int iomap_swapfile_add_extent(struct iomap_swapfile_info *isi)
+{
+	struct iomap *iomap = &isi->iomap;
+	unsigned long nr_pages;
+	uint64_t first_ppage;
+	uint64_t first_ppage_reported;
+	uint64_t last_ppage;
+	int error;
+
+	/*
+	 * Round the start up and the end down so that the physical

fs/iomap.c iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, · line 1127

Quoted 5 lines

fs/iomap.c +5 -0

+	 * extent aligns to a page boundary.
+	 */
+	first_ppage = ALIGN(iomap->addr, PAGE_SIZE) >> PAGE_SHIFT;
+	last_ppage = (ALIGN_DOWN(iomap->addr + iomap->length, PAGE_SIZE) >>
+			PAGE_SHIFT) - 1;

But this can still end up underflowing last_ppage to (unsigned long)-1 and the following test won't trigger?

fs/iomap.c iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter, · line 1131

Quoted 4 lines

fs/iomap.c +4 -0

+
+	/* Skip too-short physical extents. */
+	if (first_ppage > last_ppage)
+		return 0;

Honza

Signature · 2 lines

Jan Kara <[email protected]>
SUSE Labs, CR