[RFC PATCH v2 2/4] mm: restrict offload-only swap to proactive reclaim

[Matthias Goergens <matthias.goergens@gmail.com>]

Some swap backends need memory to accept writes, making them useful for
deliberate cold-page offload but unsuitable dependencies for reclaim under
acute memory pressure.

Add SWAP_FLAG_OFFLOAD_ONLY to reserve a swap area for explicitly admitted
reclaim. Cgroup v2 memory.reclaim, per-node reclaim and manual MGLRU
eviction establish admission; ordinary reclaim cannot initiate new,
non-zero backend writes to the area.

Track admission separately from scan_control.proactive and propagate it
through reclaim_state. The MGLRU debugfs interface is not stable ABI.

Filter swap allocation and reclaim capacity by eligibility, including
cached clusters and recovery of unused conventional swap-cache entries.
Preserve physical free-space reporting and workingset accounting.

Allocation filtering alone is insufficient: a partially swapped-in large
folio can retain its swap entry and reach writeout without allocating a
new slot. Refuse such ordinary-reclaim writes by redirtying and activating
the folio. Preserve architecture metadata before refusal, since sibling
faults can restore swap-indexed tags into the resident folio.

Bypass zswap and reject asynchronous page-cluster discard so later writes
cannot escape the admitting context. Exclude marked areas from hibernation
selection. Reads, swapoff and queued or in-flight I/O remain unaffected.

Refusing retained-entry writes can cause repeated reclaim or OOM; moving
these entries to conventional swap would require swap-entry migration.

Signed-off-by: Matthias Goergens <matthias.goergens@gmail.com>
---
 Documentation/mm/swap.rst                     | 105 +++++++++++++++
 drivers/gpu/drm/i915/gem/i915_gem_shrinker.c  |   2 +-
 .../gpu/drm/i915/gem/selftests/huge_pages.c   |   6 +-
 drivers/gpu/drm/msm/msm_gem_shrinker.c        |   2 +-
 drivers/gpu/drm/panthor/panthor_gem.c         |   2 +-
 drivers/gpu/drm/ttm/ttm_backup.c              |   2 +-
 drivers/gpu/drm/xe/tests/xe_bo.c              |   2 +-
 include/linux/swap.h                          |  29 ++++-
 include/linux/vm_event_item.h                 |   1 +
 mm/memcontrol.c                               |  20 ++-
 mm/page_io.c                                  |  24 +++-
 mm/swapfile.c                                 | 123 ++++++++++++++++--
 mm/vmscan.c                                   |  31 +++--
 mm/vmstat.c                                   |   1 +
 14 files changed, 312 insertions(+), 38 deletions(-)

diff --git a/Documentation/mm/swap.rst b/Documentation/mm/swap.rst
index 78819bd4d745..9acad2c5834f 100644
--- a/Documentation/mm/swap.rst
+++ b/Documentation/mm/swap.rst
@@ -3,3 +3,108 @@
 ====
 Swap
 ====
+
+Offload-only swap areas
+-----------------------
+
+Passing ``SWAP_FLAG_OFFLOAD_ONLY`` in the flags argument to ``swapon()``
+marks a swap area as eligible for explicit userspace proactive reclaim.
+The stable qualifying interfaces are cgroup v2
+``memory.reclaim`` and ``/sys/devices/system/node/nodeX/reclaim``.  The swap
+allocator excludes such an area from kswapd, direct reclaim, and other
+pressure-driven swap allocation.  Normal swap priority ordering still applies
+among the areas eligible for the current reclaim context.  Proactive reclaim
+can use both conventional and offload-only areas; the flag does not force it
+to choose an offload-only area or reserve conventional capacity exclusively
+for pressure reclaim.
+
+The MGLRU debugfs eviction interface currently establishes the same internal
+proactive-reclaim provenance and can therefore use an offload-only area.
+Debugfs is not a stable userspace ABI, however, so that behaviour is not part
+of this interface's permanent contract.
+
+This permits a system to combine a small conventional swap area, which is
+engineered for forward progress in emergency reclaim, with a larger or more
+complex area used for ordinary cold-page offload.  For example, the latter may
+be RAM-compressed or may use a filesystem with compression, checksums, or
+redundancy.  Making every such write path safe in direct reclaim can require
+backend-specific reserves, preallocation, non-blocking allocation, and
+recursion rules.  The flag restricts when writes may be initiated; backends
+still need to handle reads and complete previously admitted writes under
+memory pressure.
+
+For a RAM-compressed area such as zram, unused logical slots also do not imply
+that enough physical memory remains to store their future contents.  Static
+swap priority cannot express that distinction or provide late fallback after a
+selected area's write fails.
+
+The policy is attached to an activated swap area, not to its underlying
+physical storage.  A raw swap partition and a filesystem swapfile on the same
+device are separate areas and may use different policies.  The kernel does not
+infer this policy from the block driver, filesystem, or swap priority.
+Changing an active area's policy requires swapoff followed by reactivation.
+
+``/proc/swaps`` does not expose the policy.  Reported swap totals and free
+space include offload-only areas, so free swap space does not necessarily
+mean that pressure reclaim can allocate from it.
+
+This is a reclaim-provenance policy, not a measurement of current memory
+headroom.  Userspace should only request proactive offload while its own
+watermark or PSI policy considers memory pressure low.  Since
+``memory.reclaim`` can be delegated, that policy must also account for
+requests from delegated cgroups.
+
+DAMON reclaim and ``MADV_PAGEOUT`` do not currently establish the proactive
+reclaim context, so they cannot allocate slots from an offload-only area.
+Offload-only areas are also ineligible for hibernation image allocation.
+
+Offload-only areas bypass zswap stores.  Zswap writeback may run after the
+proactive context which selected the slot has ended, so admitting the folio to
+zswap would otherwise defer the backend write beyond that context.  The
+hierarchical cgroup v2 ``memory.zswap.writeback=0`` policy remains
+authoritative: when zswap is enabled, it also refuses direct proactive writes
+to an offload-only area.  Marking an area offload-only does not override a
+cgroup policy which disables all swapping attempts to devices.
+
+The flag controls both allocation of new swap slots and newly initiated
+non-zero backend writes.  A folio can retain its swap entry after swapin.  If
+ordinary reclaim later tries to rewrite such an offload-only entry, the VM
+redirties and activates the folio instead; proactive reclaim may retry the
+write.  Zero-filled folios may still update the in-memory swap zeromap without
+backend I/O.
+
+The flag does not prevent reads, swapoff, or writes which are already queued or
+in flight.  It therefore does not by itself provide a forward progress
+guarantee for an I/O path which allocates memory: earlier writes must still be
+able to complete, and reads must remain reclaim-safe.  Repeatedly refusing
+retained-entry writes can also reduce reclaim efficiency and lead to OOM while
+the dirty folios remain resident.
+
+Page-cluster discard is incompatible with an offload-only area because its
+work item can run after the context which freed the entries has ended.  Swapon
+therefore rejects a resolved page-cluster discard policy combined with
+``SWAP_FLAG_OFFLOAD_ONLY``.  Swapon-time discard is permitted because it
+completes synchronously during activation.  The existing discard precedence
+still applies: requesting both discard-once and discard-pages selects
+discard-once.  The bare ``SWAP_FLAG_DISCARD`` request enables page-cluster
+discard and is therefore rejected on a discard-capable offload-only area;
+add ``SWAP_FLAG_DISCARD_ONCE`` to ``SWAP_FLAG_DISCARD`` for synchronous
+activation-time discard.
+Discard requests which the swap area does not support remain ignored.
+
+Architecture-specific swap metadata preparation still runs before a retained
+write is refused, so that metadata remains coherent with the dirty resident
+folio.  This policy controls swap-backend I/O; it does not promise that core VM
+or architecture preparation performs no allocation.
+
+With ``CONFIG_VM_EVENT_COUNTERS``, ``/proc/vmstat`` reports
+``swpout_offload_refused`` in base pages.  The counter advances when ordinary
+reclaim refuses a newly initiated write through a retained offload-only
+entry.  Repeated refusals of the same folio are counted again.  It is not a
+count of skipped areas during new-slot allocation.
+
+An offload-only area should therefore be configured with a reclaim-safe swap
+area as fallback for new swap allocations.  This does not migrate retained
+offload-only entries or retry a failed backend write on another area.  If no
+eligible swap space remains, swap allocation fails and the existing reclaim
+and OOM policy applies.
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c
index e0d1f369a163..b58e61f15ab1 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c
@@ -21,7 +21,7 @@
 
 static bool swap_available(void)
 {
-	return get_nr_swap_pages() > 0;
+	return get_nr_swap_pages_eligible() > 0;
 }
 
 static bool can_release_pages(struct drm_i915_gem_object *obj)
diff --git a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
index 44718e728291..73b62b065510 100644
--- a/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
+++ b/drivers/gpu/drm/i915/gem/selftests/huge_pages.c
@@ -1895,8 +1895,8 @@ static int igt_shrink_thp(void *arg)
 	i915_gem_context_unlock_engines(ctx);
 	/*
 	 * Nuke everything *before* we unpin the pages so we can be reasonably
-	 * sure that when later checking get_nr_swap_pages() that some random
-	 * leftover object doesn't steal the remaining swap space.
+	 * sure that when later checking get_nr_swap_pages_eligible() that some
+	 * random leftover object doesn't steal the remaining swap space.
 	 */
 	i915_gem_shrink(NULL, i915, -1UL, NULL,
 			I915_SHRINK_BOUND |
@@ -1910,7 +1910,7 @@ static int igt_shrink_thp(void *arg)
 	 * Now that the pages are *unpinned* shrinking should invoke
 	 * shmem to truncate our pages, if we have available swap.
 	 */
-	should_swap = get_nr_swap_pages() > 0;
+	should_swap = get_nr_swap_pages_eligible() > 0;
 	i915_gem_shrink(NULL, i915, -1UL, NULL,
 			I915_SHRINK_BOUND |
 			I915_SHRINK_UNBOUND |
diff --git a/drivers/gpu/drm/msm/msm_gem_shrinker.c b/drivers/gpu/drm/msm/msm_gem_shrinker.c
index 9d2788f79ace..da6a66b747b0 100644
--- a/drivers/gpu/drm/msm/msm_gem_shrinker.c
+++ b/drivers/gpu/drm/msm/msm_gem_shrinker.c
@@ -21,7 +21,7 @@ module_param(enable_eviction, bool, 0600);
 
 static bool can_swap(void)
 {
-	return enable_eviction && get_nr_swap_pages() > 0;
+	return enable_eviction && get_nr_swap_pages_eligible() > 0;
 }
 
 static bool can_block(struct shrink_control *sc)
diff --git a/drivers/gpu/drm/panthor/panthor_gem.c b/drivers/gpu/drm/panthor/panthor_gem.c
index 72908be5e144..23488f417cf7 100644
--- a/drivers/gpu/drm/panthor/panthor_gem.c
+++ b/drivers/gpu/drm/panthor/panthor_gem.c
@@ -1371,7 +1371,7 @@ panthor_dummy_bo_create(struct panthor_device *ptdev)
 
 static bool can_swap(void)
 {
-	return get_nr_swap_pages() > 0;
+	return get_nr_swap_pages_eligible() > 0;
 }
 
 static bool can_block(struct shrink_control *sc)
diff --git a/drivers/gpu/drm/ttm/ttm_backup.c b/drivers/gpu/drm/ttm/ttm_backup.c
index 0c2d53a13b2a..bf9d41bdf652 100644
--- a/drivers/gpu/drm/ttm/ttm_backup.c
+++ b/drivers/gpu/drm/ttm/ttm_backup.c
@@ -206,7 +206,7 @@ u64 ttm_backup_bytes_avail(void)
 	 * number also depends on shmem actually swapping out backed-up
 	 * shmem objects without too much buffering.
 	 */
-	return (u64)get_nr_swap_pages() << PAGE_SHIFT;
+	return (u64)get_nr_swap_pages_eligible() << PAGE_SHIFT;
 }
 EXPORT_SYMBOL_GPL(ttm_backup_bytes_avail);
 
diff --git a/drivers/gpu/drm/xe/tests/xe_bo.c b/drivers/gpu/drm/xe/tests/xe_bo.c
index 6a17e13d58cf..14d6bb8e41c9 100644
--- a/drivers/gpu/drm/xe/tests/xe_bo.c
+++ b/drivers/gpu/drm/xe/tests/xe_bo.c
@@ -695,7 +695,7 @@ static int shrink_test_run_device(struct xe_device *xe)
 	}
 	to_alloc = ram * 2;
 
-	ram_and_swap = ram + get_nr_swap_pages() * PAGE_SIZE;
+	ram_and_swap = ram + get_nr_swap_pages_eligible() * PAGE_SIZE;
 	if (to_alloc > ram_and_swap)
 		purgeable = to_alloc - ram_and_swap;
 	purgeable += div64_u64(purgeable, 5);
diff --git a/include/linux/swap.h b/include/linux/swap.h
index 78974da6810e..75350c450c47 100644
--- a/include/linux/swap.h
+++ b/include/linux/swap.h
@@ -21,10 +21,11 @@
 #define SWAP_FLAG_DISCARD	0x10000 /* enable discard for swap */
 #define SWAP_FLAG_DISCARD_ONCE	0x20000 /* discard swap area at swapon-time */
 #define SWAP_FLAG_DISCARD_PAGES 0x40000 /* discard page-clusters after use */
+#define SWAP_FLAG_OFFLOAD_ONLY  0x80000 /* only use for proactive reclaim */
 
 #define SWAP_FLAGS_VALID	(SWAP_FLAG_PRIO_MASK | SWAP_FLAG_PREFER | \
 				 SWAP_FLAG_DISCARD | SWAP_FLAG_DISCARD_ONCE | \
-				 SWAP_FLAG_DISCARD_PAGES)
+				 SWAP_FLAG_DISCARD_PAGES | SWAP_FLAG_OFFLOAD_ONLY)
 /*
  * MAX_SWAPFILES defines the maximum number of swaptypes: things which can
  * be swapped to.  The swap type and the offset into that swap type are
@@ -140,12 +141,20 @@ union swap_header {
 struct reclaim_state {
 	/* pages reclaimed outside of LRU-based reclaim */
 	unsigned long reclaimed;
+	/* this reclaim context may use offload-only swap */
+	bool allow_offload_swap;
 #ifdef CONFIG_LRU_GEN
 	/* per-thread mm walk data */
 	struct lru_gen_mm_walk *mm_walk;
 #endif
 };
 
+static inline bool current_reclaim_allows_offload_swap(void)
+{
+	return current->reclaim_state &&
+	       current->reclaim_state->allow_offload_swap;
+}
+
 /*
  * mm_account_reclaimed_pages(): account reclaimed pages outside of LRU-based
  * reclaim
@@ -201,6 +210,7 @@ enum {
 	SWP_STABLE_WRITES = (1 << 11),	/* no overwrite PG_writeback pages */
 	SWP_SYNCHRONOUS_IO = (1 << 12),	/* synchronous IO is efficient */
 	SWP_HIBERNATION = (1 << 13),	/* pinned for hibernation */
+	SWP_OFFLOAD_ONLY = (1 << 14),	/* proactive-reclaim swap only */
 					/* add others here before... */
 };
 
@@ -389,6 +399,9 @@ static inline long get_nr_swap_pages(void)
 	return atomic_long_read(&nr_swap_pages);
 }
 
+long get_nr_swap_pages_eligible(void);
+bool folio_swap_full(struct folio *folio);
+
 extern void si_swapinfo(struct sysinfo *);
 extern int pin_hibernation_swap_type(dev_t device, sector_t offset);
 extern void unpin_hibernation_swap_type(int type);
@@ -443,10 +456,16 @@ static inline void put_swap_device(struct swap_info_struct *si)
 }
 
 #define get_nr_swap_pages()			0L
+#define get_nr_swap_pages_eligible()		0L
 #define total_swap_pages			0L
 #define total_swapcache_pages()			0UL
 #define vm_swap_full()				0
 
+static inline bool folio_swap_full(struct folio *folio)
+{
+	return false;
+}
+
 #define si_swapinfo(val) \
 	do { (val)->freeswap = (val)->totalswap = 0; } while (0)
 #define free_folio_and_swap_cache(folio) \
@@ -531,6 +550,7 @@ static inline void mem_cgroup_uncharge_swap(unsigned short id, unsigned int nr_p
 
 long mem_cgroup_get_folio_swap_margin(struct folio *folio);
 extern long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg);
+long mem_cgroup_get_nr_swap_pages_eligible(struct mem_cgroup *memcg);
 extern bool mem_cgroup_swap_full(struct folio *folio);
 #else
 static inline int mem_cgroup_try_charge_swap(struct folio *folio)
@@ -553,9 +573,14 @@ static inline long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
 	return get_nr_swap_pages();
 }
 
+static inline long mem_cgroup_get_nr_swap_pages_eligible(struct mem_cgroup *memcg)
+{
+	return get_nr_swap_pages_eligible();
+}
+
 static inline bool mem_cgroup_swap_full(struct folio *folio)
 {
-	return vm_swap_full();
+	return folio_swap_full(folio);
 }
 #endif
 
diff --git a/include/linux/vm_event_item.h b/include/linux/vm_event_item.h
index 2628ccda076a..fc8458b314a5 100644
--- a/include/linux/vm_event_item.h
+++ b/include/linux/vm_event_item.h
@@ -32,6 +32,7 @@
 	HIGHMEM_ZONE(xx) xx##_MOVABLE, DEVICE_ZONE(xx)
 
 enum vm_event_item { PGPGIN, PGPGOUT, PSWPIN, PSWPOUT,
+		SWPOUT_OFFLOAD_REFUSED,
 		FOR_ALL_ZONES(PGALLOC)
 		FOR_ALL_ZONES(ALLOCSTALL)
 		FOR_ALL_ZONES(PGSCAN_SKIP)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 1460cba53588..1412084d2f43 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -6002,16 +6002,28 @@ void __mem_cgroup_uncharge_swap(unsigned short id, unsigned int nr_pages)
 	rcu_read_unlock();
 }
 
-long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
+static long
+mem_cgroup_get_nr_swap_pages_with_limit(struct mem_cgroup *memcg,
+					long nr_swap_pages)
 {
-	long nr_swap_pages = get_nr_swap_pages();
-
 	if (!mem_cgroup_disabled() && !do_memsw_account())
 		nr_swap_pages = min(nr_swap_pages, page_counter_margin(&memcg->swap));
 
 	return nr_swap_pages;
 }
 
+long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
+{
+	return mem_cgroup_get_nr_swap_pages_with_limit(memcg,
+			get_nr_swap_pages());
+}
+
+long mem_cgroup_get_nr_swap_pages_eligible(struct mem_cgroup *memcg)
+{
+	return mem_cgroup_get_nr_swap_pages_with_limit(memcg,
+			get_nr_swap_pages_eligible());
+}
+
 /**
  * mem_cgroup_get_folio_swap_margin - get a folio's memcg swap margin
  * @folio: folio whose memcg margin is queried
@@ -6042,7 +6054,7 @@ bool mem_cgroup_swap_full(struct folio *folio)
 
 	VM_BUG_ON_FOLIO(!folio_test_locked(folio), folio);
 
-	if (vm_swap_full())
+	if (folio_swap_full(folio))
 		return true;
 	if (do_memsw_account() || !folio_memcg_charged(folio))
 		return ret;
diff --git a/mm/page_io.c b/mm/page_io.c
index 1da4ff484f09..1b281a1d6df0 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -203,6 +203,7 @@ static void swap_zeromap_folio_clear(struct folio *folio)
  */
 int swap_writeout(struct swap_io_ctx *ctx, struct folio *folio)
 {
+	struct swap_info_struct *sis = __swap_entry_to_info(folio->swap);
 	int ret = 0;
 
 	if (folio_free_swap(folio))
@@ -210,7 +211,9 @@ int swap_writeout(struct swap_io_ctx *ctx, struct folio *folio)
 
 	/*
 	 * Arch code may have to preserve more data than just the folio
-	 * contents, e.g. memory tags.
+	 * contents, e.g. memory tags.  Do this before refusing a retained
+	 * offload-only entry below: a later sibling swap-PTE fault can restore
+	 * swap-indexed metadata into this resident folio.
 	 */
 	ret = arch_prepare_to_swap(folio);
 	if (ret) {
@@ -228,6 +231,18 @@ int swap_writeout(struct swap_io_ctx *ctx, struct folio *folio)
 		goto out_unlock;
 	}
 
+	/*
+	 * A folio can retain an existing swap entry after swapin.  Do not let
+	 * ordinary reclaim use an offload-only entry through that path.
+	 */
+	if ((READ_ONCE(sis->flags) & SWP_OFFLOAD_ONLY) &&
+	    !current_reclaim_allows_offload_swap()) {
+		count_vm_events(SWPOUT_OFFLOAD_REFUSED,
+				folio_nr_pages(folio));
+		folio_mark_dirty(folio);
+		return AOP_WRITEPAGE_ACTIVATE;
+	}
+
 	/*
 	 * Clear bits this folio occupies in the zeromap to prevent zero data
 	 * being read in from any previous zero writes that occupied the same
@@ -235,7 +250,12 @@ int swap_writeout(struct swap_io_ctx *ctx, struct folio *folio)
 	 */
 	swap_zeromap_folio_clear(folio);
 
-	if (zswap_store(folio)) {
+	/*
+	 * Zswap writeback can happen much later from pressure reclaim or its
+	 * shrinker workqueue.  Do not let it defer an offload-only backend write
+	 * beyond the proactive reclaim context which admitted the swap slot.
+	 */
+	if (!(READ_ONCE(sis->flags) & SWP_OFFLOAD_ONLY) && zswap_store(folio)) {
 		count_mthp_stat(folio_order(folio), MTHP_STAT_ZSWPOUT);
 		goto out_unlock;
 	}
diff --git a/mm/swapfile.c b/mm/swapfile.c
index 280dd906eb18..b72d598a9973 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -65,13 +65,17 @@ static void move_cluster(struct swap_info_struct *si,
  */
 static DEFINE_SPINLOCK(swap_lock);
 static unsigned int nr_swapfiles;
-atomic_long_t nr_swap_pages;
 /*
  * Some modules use swappable objects and may try to swap them out under
  * memory pressure (via the shrinker). Before doing so, they may wish to
  * check to see if any swap space is available.
+ *
+ * This remains the raw free-space counter for accounting users. Reclaim
+ * decisions subtract nr_swap_pages_offload_only below when necessary.
  */
+atomic_long_t nr_swap_pages;
 EXPORT_SYMBOL_GPL(nr_swap_pages);
+static atomic_long_t nr_swap_pages_offload_only;
 /* protected with swap_lock. reading in vm_swap_full() doesn't need lock */
 long total_swap_pages;
 #define DEF_SWAP_PRIO  -1
@@ -120,6 +124,7 @@ atomic_t nr_rotate_swap = ATOMIC_INIT(0);
 struct percpu_swap_cluster {
 	struct swap_info_struct *si[SWAP_NR_ORDERS];
 	unsigned long offset[SWAP_NR_ORDERS];
+	bool allow_offload_swap[SWAP_NR_ORDERS];
 	local_lock_t lock;
 };
 
@@ -163,6 +168,31 @@ static long swap_usage_in_pages(struct swap_info_struct *si)
 	return atomic_long_read(&si->inuse_pages) & SWAP_USAGE_COUNTER_MASK;
 }
 
+static bool swap_area_needs_reclaim(struct swap_info_struct *si)
+{
+	if (vm_swap_full())
+		return true;
+
+	/*
+	 * Free offload-only slots must not keep a full conventional area
+	 * pinned in swapcache.  Recover that area's unused cache entries
+	 * even when the raw pool is not full.  This is independent of the
+	 * current task: proactive reclaim can fill the conventional area too.
+	 */
+	return !(READ_ONCE(si->flags) & SWP_OFFLOAD_ONLY) &&
+		atomic_long_read(&nr_swap_pages_offload_only) > 0 &&
+		swap_usage_in_pages(si) == si->pages;
+}
+
+/* The caller must hold the lock on a folio in swapcache. */
+bool folio_swap_full(struct folio *folio)
+{
+	VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
+	VM_WARN_ON_FOLIO(!folio_test_swapcache(folio), folio);
+
+	return swap_area_needs_reclaim(__swap_entry_to_info(folio->swap));
+}
+
 /* Reclaim the swap entry anyway if possible */
 #define TTRS_ANYWAY		0x1
 /*
@@ -903,7 +933,7 @@ static bool cluster_scan_range(struct swap_info_struct *si,
 		if (swp_tb_is_null(swp_tb))
 			continue;
 		if (swp_tb_is_folio(swp_tb) && !__swp_tb_get_count(swp_tb)) {
-			if (!vm_swap_full())
+			if (!swap_area_needs_reclaim(si))
 				return false;
 			*need_reclaim = true;
 			continue;
@@ -1014,6 +1044,8 @@ static unsigned int alloc_swap_scan_cluster(struct swap_info_struct *si,
 	if (si->flags & SWP_SOLIDSTATE) {
 		this_cpu_write(percpu_swap_cluster.offset[order], next);
 		this_cpu_write(percpu_swap_cluster.si[order], si);
+		this_cpu_write(percpu_swap_cluster.allow_offload_swap[order],
+			       current_reclaim_allows_offload_swap());
 	} else {
 		si->global_cluster->next[order] = next;
 	}
@@ -1161,7 +1193,7 @@ static unsigned long cluster_alloc_swap_entry(struct swap_info_struct *si,
 	}
 
 	/* Try reclaim full clusters if free and nonfull lists are drained */
-	if (vm_swap_full())
+	if (swap_area_needs_reclaim(si))
 		swap_reclaim_full_clusters(si, false);
 
 	if (order < PMD_ORDER) {
@@ -1315,9 +1347,11 @@ static void swap_range_alloc(struct swap_info_struct *si,
 			     unsigned int nr_entries)
 {
 	if (swap_usage_add(si, nr_entries)) {
-		if (vm_swap_full())
+		if (swap_area_needs_reclaim(si))
 			schedule_work(&si->reclaim_work);
 	}
+	if (si->flags & SWP_OFFLOAD_ONLY)
+		atomic_long_sub(nr_entries, &nr_swap_pages_offload_only);
 	atomic_long_sub(nr_entries, &nr_swap_pages);
 }
 
@@ -1346,6 +1380,8 @@ static void swap_range_free(struct swap_info_struct *si, unsigned long offset,
 	 * only after the above cleanups are done.
 	 */
 	smp_wmb();
+	if (si->flags & SWP_OFFLOAD_ONLY)
+		atomic_long_add(nr_entries, &nr_swap_pages_offload_only);
 	atomic_long_add(nr_entries, &nr_swap_pages);
 	swap_usage_sub(si, nr_entries);
 }
@@ -1366,6 +1402,31 @@ static bool get_swap_device_info(struct swap_info_struct *si)
 	return true;
 }
 
+static bool swap_area_eligible(struct swap_info_struct *si)
+{
+	if (!(READ_ONCE(si->flags) & SWP_OFFLOAD_ONLY))
+		return true;
+
+	return current_reclaim_allows_offload_swap();
+}
+
+long get_nr_swap_pages_eligible(void)
+{
+	long nr_pages;
+
+	if (current_reclaim_allows_offload_swap())
+		return get_nr_swap_pages();
+
+	/*
+	 * The reads are intentionally unpaired.  This is a capacity hint; the
+	 * allocator enforces eligibility.  Clamp a transient negative result.
+	 */
+	nr_pages = get_nr_swap_pages() -
+		   atomic_long_read(&nr_swap_pages_offload_only);
+	return max(nr_pages, 0L);
+}
+EXPORT_SYMBOL_GPL(get_nr_swap_pages_eligible);
+
 /*
  * Fast path try to get swap entries with specified order from current
  * CPU's swap entry pool (a cluster).
@@ -1385,6 +1446,20 @@ static bool swap_alloc_fast(struct folio *folio)
 	offset = this_cpu_read(percpu_swap_cluster.offset[order]);
 	if (!si || !offset || !get_swap_device_info(si))
 		return false;
+	if (!swap_area_eligible(si)) {
+		put_swap_device(si);
+		return false;
+	}
+	/*
+	 * Pressure reclaim may cache a lower-priority conventional area while
+	 * an offload-only area is ineligible.  Drop that cache on a context
+	 * change so proactive reclaim returns to the normal priority search.
+	 */
+	if (this_cpu_read(percpu_swap_cluster.allow_offload_swap[order]) !=
+	    current_reclaim_allows_offload_swap()) {
+		put_swap_device(si);
+		return false;
+	}
 
 	ci = swap_cluster_lock(si, offset);
 	if (cluster_is_usable(ci, order)) {
@@ -1407,6 +1482,9 @@ static void swap_alloc_slow(struct folio *folio)
 	spin_lock(&swap_avail_lock);
 start_over:
 	plist_for_each_entry_safe(si, next, &swap_avail_head, avail_list) {
+		if (!swap_area_eligible(si))
+			continue;
+
 		/* Rotate the device and switch to a new cluster */
 		plist_requeue(&si->avail_list, &swap_avail_head);
 		spin_unlock(&swap_avail_lock);
@@ -1738,8 +1816,8 @@ static int swap_dup_entries_cluster(struct swap_info_struct *si,
  *
  * Context: Caller needs to hold the folio lock.
  * Return: %0 on success, %-E2BIG if splitting the folio might allow swapout,
- * %-ENOSPC if no global swap space is available, or %-ENOMEM if splitting
- * would not help.
+ * %-ENOSPC if no global swap space is eligible for the caller, or %-ENOMEM
+ * if splitting would not help.
  */
 int folio_alloc_swap(struct folio *folio)
 {
@@ -1790,7 +1868,7 @@ int folio_alloc_swap(struct folio *folio)
 	return 0;
 
 failed:
-	if (get_nr_swap_pages() <= 0)
+	if (get_nr_swap_pages_eligible() <= 0)
 		return -ENOSPC;
 	if (mem_cgroup_get_folio_swap_margin(folio) <= 0)
 		return -ENOMEM;
@@ -2180,7 +2258,7 @@ swp_entry_t swap_alloc_hibernation_slot(int type)
 	struct swap_cluster_info *ci;
 	swp_entry_t entry = {0};
 
-	if (!si)
+	if (!si || (si->flags & SWP_OFFLOAD_ONLY))
 		goto fail;
 
 	/*
@@ -2247,7 +2325,8 @@ static int __find_hibernation_swap_type(dev_t device, sector_t offset)
 	for (type = 0; type < nr_swapfiles; type++) {
 		struct swap_info_struct *sis = swap_info[type];
 
-		if (!(sis->flags & SWP_WRITEOK))
+		if (!(sis->flags & SWP_WRITEOK) ||
+		    (sis->flags & SWP_OFFLOAD_ONLY))
 			continue;
 
 		if (device == sis->bdev->bd_dev) {
@@ -2434,7 +2513,8 @@ int find_first_swap(dev_t *device)
 	for (type = 0; type < nr_swapfiles; type++) {
 		struct swap_info_struct *sis = swap_info[type];
 
-		if (!(sis->flags & SWP_WRITEOK))
+		if (!(sis->flags & SWP_WRITEOK) ||
+		    (sis->flags & SWP_OFFLOAD_ONLY))
 			continue;
 		*device = sis->bdev->bd_dev;
 		spin_unlock(&swap_lock);
@@ -2474,7 +2554,8 @@ unsigned int count_swap_pages(int type, int free)
 		struct swap_info_struct *sis = swap_info[type];
 
 		spin_lock(&sis->lock);
-		if (sis->flags & SWP_WRITEOK) {
+		if ((sis->flags & SWP_WRITEOK) &&
+		    !(sis->flags & SWP_OFFLOAD_ONLY)) {
 			n = sis->pages;
 			if (free)
 				n -= swap_usage_in_pages(sis);
@@ -3083,6 +3164,8 @@ static int setup_swap_extents(struct swap_info_struct *sis,
 
 static void _enable_swap_info(struct swap_info_struct *si)
 {
+	if (si->flags & SWP_OFFLOAD_ONLY)
+		atomic_long_add(si->pages, &nr_swap_pages_offload_only);
 	atomic_long_add(si->pages, &nr_swap_pages);
 	total_swap_pages += si->pages;
 
@@ -3231,6 +3314,8 @@ SYSCALL_DEFINE1(swapoff, const char __user *, specialfile)
 	spin_lock(&p->lock);
 	del_from_avail_list(p, true);
 	plist_del(&p->list, &swap_active_head);
+	if (p->flags & SWP_OFFLOAD_ONLY)
+		atomic_long_sub(p->pages, &nr_swap_pages_offload_only);
 	atomic_long_sub(p->pages, &nr_swap_pages);
 	total_swap_pages -= p->pages;
 	spin_unlock(&p->lock);
@@ -3727,7 +3812,6 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
 
 	if (swap_flags & ~SWAP_FLAGS_VALID)
 		return -EINVAL;
-
 	if (!capable(CAP_SYS_ADMIN))
 		return -EPERM;
 
@@ -3841,6 +3925,9 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
 	if (error)
 		goto bad_swap_unlock_inode;
 
+	if (swap_flags & SWAP_FLAG_OFFLOAD_ONLY)
+		si->flags |= SWP_OFFLOAD_ONLY;
+
 	if ((swap_flags & SWAP_FLAG_DISCARD) &&
 	    si->bdev && bdev_max_discard_sectors(si->bdev)) {
 		/*
@@ -3863,6 +3950,18 @@ SYSCALL_DEFINE2(swapon, const char __user *, specialfile, int, swap_flags)
 		else if (swap_flags & SWAP_FLAG_DISCARD_PAGES)
 			si->flags &= ~SWP_AREA_DISCARD;
 
+		/*
+		 * Cluster discard can run later from discard_work, after the
+		 * context which freed the entries has ended.  Swapon-time discard
+		 * is explicit and synchronous, but page discard cannot honour
+		 * offload provenance.
+		 */
+		if ((si->flags & SWP_OFFLOAD_ONLY) &&
+		    (si->flags & SWP_PAGE_DISCARD)) {
+			error = -EINVAL;
+			goto bad_swap_unlock_inode;
+		}
+
 		/* issue a swapon-time discard if it's still required */
 		if (si->flags & SWP_AREA_DISCARD) {
 			int err = discard_swap(si);
diff --git a/mm/vmscan.c b/mm/vmscan.c
index aaceed4759ee..0633feb5d88b 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -123,6 +123,9 @@ struct scan_control {
 	/* Proactive reclaim invoked by userspace */
 	unsigned int proactive:1;
 
+	/* This reclaim context may use offload-only swap */
+	unsigned int allow_offload_swap:1;
+
 	/*
 	 * Cgroup memory below memory.low is protected as long as we
 	 * don't threaten to OOM. If any cgroup is reclaimed at
@@ -291,14 +294,18 @@ static inline bool is_exec_file_folio(const struct folio *folio,
 }
 
 static void set_task_reclaim_state(struct task_struct *task,
-				   struct reclaim_state *rs)
+				   struct scan_control *sc)
 {
+	struct reclaim_state *rs = sc ? &sc->reclaim_state : NULL;
+
 	/* Check for an overwrite */
 	WARN_ON_ONCE(rs && task->reclaim_state);
 
 	/* Check for the nulling of an already-nulled member */
 	WARN_ON_ONCE(!rs && !task->reclaim_state);
 
+	if (rs)
+		rs->allow_offload_swap = sc->allow_offload_swap;
 	task->reclaim_state = rs;
 }
 
@@ -418,7 +425,7 @@ static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
 		 * And under GFP_NOIO, is there enough swapcached anon to make
 		 * scanning anon worthwhile?
 		 */
-		if (get_nr_swap_pages() > 0 &&
+		if (get_nr_swap_pages_eligible() > 0 &&
 		    !reclaimable_anon_is_low(memcg, nid, sc))
 			return true;
 	} else {
@@ -426,7 +433,7 @@ static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg,
 		 * Is the memcg below its swap limit, and under GFP_NOIO does
 		 * it have enough swapcached anon to make scanning worthwhile?
 		 */
-		if (mem_cgroup_get_nr_swap_pages(memcg) > 0 &&
+		if (mem_cgroup_get_nr_swap_pages_eligible(memcg) > 0 &&
 		    !reclaimable_anon_is_low(memcg, nid, sc))
 			return true;
 	}
@@ -2853,7 +2860,7 @@ static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc)
 		return 0;
 
 	if (!can_demote(pgdat->node_id, sc, memcg) &&
-	    mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH)
+	    mem_cgroup_get_nr_swap_pages_eligible(memcg) < MIN_LRU_BATCH)
 		return 0;
 
 	return swappiness;
@@ -5926,6 +5933,7 @@ static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
 		.reclaim_idx = MAX_NR_ZONES - 1,
 		.gfp_mask = GFP_KERNEL,
 		.proactive = true,
+		.allow_offload_swap = true,
 	};
 
 	buf = kvmalloc(len + 1, GFP_KERNEL);
@@ -5937,7 +5945,7 @@ static ssize_t lru_gen_seq_write(struct file *file, const char __user *src,
 		return -EFAULT;
 	}
 
-	set_task_reclaim_state(current, &sc.reclaim_state);
+	set_task_reclaim_state(current, &sc);
 	flags = memalloc_noreclaim_save();
 	blk_start_plug(&plug);
 	if (!set_mm_walk(NULL, true)) {
@@ -6941,7 +6949,7 @@ unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
 	if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask))
 		return 1;
 
-	set_task_reclaim_state(current, &sc.reclaim_state);
+	set_task_reclaim_state(current, &sc);
 	trace_mm_vmscan_direct_reclaim_begin(sc.gfp_mask, order, NULL);
 
 	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
@@ -6974,6 +6982,8 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
 		.may_unmap = 1,
 		.may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP),
 		.proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
+		.allow_offload_swap =
+			!!(reclaim_options & MEMCG_RECLAIM_PROACTIVE),
 	};
 	/*
 	 * Traverse the ZONELIST_FALLBACK zonelist of the current node to put
@@ -6982,7 +6992,7 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
 	 */
 	struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
 
-	set_task_reclaim_state(current, &sc.reclaim_state);
+	set_task_reclaim_state(current, &sc);
 	trace_mm_vmscan_memcg_reclaim_begin(sc.gfp_mask, 0, memcg);
 	noreclaim_flag = memalloc_noreclaim_save();
 
@@ -7272,7 +7282,7 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx)
 
 	trace_mm_vmscan_balance_pgdat_begin(pgdat->node_id, order,
 					    highest_zoneidx);
-	set_task_reclaim_state(current, &sc.reclaim_state);
+	set_task_reclaim_state(current, &sc);
 	psi_memstall_enter(&pflags);
 	__fs_reclaim_acquire(_THIS_IP_);
 
@@ -7769,7 +7779,7 @@ unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
 
 	fs_reclaim_acquire(sc.gfp_mask);
 	noreclaim_flag = memalloc_noreclaim_save();
-	set_task_reclaim_state(current, &sc.reclaim_state);
+	set_task_reclaim_state(current, &sc);
 
 	nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
 
@@ -7950,7 +7960,7 @@ static unsigned long __node_reclaim(struct pglist_data *pgdat,
 	 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
 	 */
 	noreclaim_flag = memalloc_noreclaim_save();
-	set_task_reclaim_state(p, &sc->reclaim_state);
+	set_task_reclaim_state(p, sc);
 
 	do {
 		shrink_node(pgdat, sc);
@@ -8136,6 +8146,7 @@ int user_proactive_reclaim(char *buf,
 				.may_unmap = 1,
 				.may_swap = 1,
 				.proactive = 1,
+				.allow_offload_swap = 1,
 			};
 
 			if (test_and_set_bit_lock(PGDAT_RECLAIM_LOCKED,
diff --git a/mm/vmstat.c b/mm/vmstat.c
index a3e809c57f29..b924c715886e 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c
@@ -1331,6 +1331,7 @@ const char * const vmstat_text[] = {
 	[I(PGPGOUT)]				= "pgpgout",
 	[I(PSWPIN)]				= "pswpin",
 	[I(PSWPOUT)]				= "pswpout",
+	[I(SWPOUT_OFFLOAD_REFUSED)]		= "swpout_offload_refused",
 
 #define OFF (NR_VM_ZONE_STAT_ITEMS + NR_VM_NUMA_EVENT_ITEMS + \
 	     NR_VM_NODE_STAT_ITEMS + NR_VM_STAT_ITEMS)
-- 
2.55.0