[PATCH 0/2] mm: zswap: reduce request contention on loads

[PATCH 0/2] mm: zswap: reduce request contention on loads

[Usama Arif]

Stores and loads share a per-CPU acomp request and mutex. A low-priority
store can be preempted right after the compressor drops its stream
lock, while it still holds the zswap mutex, and a higher-priority load
on that CPU then waits for the store to run again. This follows the work
from Sergey Senozhatsky's zram series which splits it for the same
reason [1].

Patch 1 gives compression and decompression separate requests, waits
and mutexes, so loads no longer wait for stores, though they can still
wait for each other. Patch 2 decompresses with an on-stack request when
the algorithm is synchronous and needs no request context, which covers
all in-tree software compressors, so those loads take no zswap lock.
Asynchronous algorithms keep the per-CPU request and mutex. For software
compressors the series allocates the same number of requests as before;
each per-CPU context grows by 72 bytes, and the load path is about 270
bytes deeper on x86-64.

The series does not fix two related cases:
- Stores still serialize on the compression mutex, so a high-priority
  task that reclaims (direct reclaim, MADV_PAGEOUT) can still wait for
  a preempted store.
- On PREEMPT_RT the codec stream locks are preemptible, so a load can
  still wait for a preempted store inside the codec.

The numbers below are the slowest read per run, as a median (min-max)
of 5 runs. Each run is 12 seconds in a zstd VM with lazy preemption,
vm.page-cluster=0 and swap on /dev/ram0. With 1 vCPU, four nice +10
workers page memory out and read it back while a nice 0 task spins. A
nice -19 reader pages out its own buffer and measures how long each
read of it takes. With 8 vCPUs there are 16 workers, 8 spinning tasks
and 8 readers.

              Before series (ms)      With series (ms)
  1 vCPU      22.3 (21.6-22.6)        0.97 (0.72-1.4)
  8 vCPUs     314 (97-2542)           7.0 (5.0-98)

Reads over 10 ms fell from 26-35 per run to none with 1 vCPU, and from
3-18 per run to at most one with 8 vCPUs. The benchmark and test programs
were written with the help of an LLM.

[1] https://lore.kernel.org/all/20261005122036.718976-10-senozhatsky@chromium.org/
 
Usama Arif (2):
  mm: zswap: use separate compression and decompression requests
  mm: zswap: use stack requests for synchronous decompression

 mm/zswap.c | 136 ++++++++++++++++++++++++++++++++++-------------------
 1 file changed, 88 insertions(+), 48 deletions(-)

-- 
2.53.0-Meta

[PATCH 1/2] mm: zswap: use separate compression and decompression requests

[Usama Arif]

Stores and loads serialize on the same per-CPU acomp request and mutex.
A low-priority store can be preempted as soon as the compressor drops
its stream lock, while it still holds the mutex. A higher-priority load
on that CPU then waits until the store runs again, which can take a
long time when other tasks are runnable.

Give compression and decompression their own request, completion wait
and mutex. Since commit e2c3b6b21c77f ("mm: zswap: use SG list
decompression APIs from zsmalloc"), the per-CPU buffer is only used for
compression. The two requests can share the per-CPU transform: no
in-tree implementation modifies transform state while (de)compressing,
and shared codec state has its own locking. Loads can still wait for
each other on the decompression mutex, and stores still serialize on
the compression mutex.

This follows the proposal from Sergey Senozhatsky for the same split
for zram [1].

[1] https://lore.kernel.org/all/20261005122036.718976-10-senozhatsky@chromium.org/

Signed-off-by: Usama Arif <usama.arif@linux.dev>
---
 mm/zswap.c | 89 +++++++++++++++++++++++++++++++-----------------------
 1 file changed, 52 insertions(+), 37 deletions(-)

diff --git a/mm/zswap.c b/mm/zswap.c
index ae19e301fced7..54187b1ef751d 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -137,14 +137,20 @@ bool zswap_never_enabled(void)
 * data structures
 **********************************/
 
-struct crypto_acomp_ctx {
-	struct crypto_acomp *acomp;
+struct zswap_acomp_req {
 	struct acomp_req *req;
 	struct crypto_wait wait;
-	u8 *buffer;
 	struct mutex mutex;
 };
 
+/* Separate requests, so that decompression does not wait for compression. */
+struct crypto_acomp_ctx {
+	struct crypto_acomp *acomp;
+	struct zswap_acomp_req comp;
+	struct zswap_acomp_req decomp;
+	u8 *buffer;
+};
+
 /*
  * The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
  * The only case where lru_lock is not acquired while holding tree.lock is
@@ -270,14 +276,10 @@ static void acomp_ctx_free(struct crypto_acomp_ctx *acomp_ctx)
 	if (!acomp_ctx)
 		return;
 
-	/*
-	 * If there was an error in allocating @acomp_ctx->req, it
-	 * would be set to NULL.
-	 */
-	if (acomp_ctx->req)
-		acomp_request_free(acomp_ctx->req);
-
-	acomp_ctx->req = NULL;
+	acomp_request_free(acomp_ctx->comp.req);
+	acomp_ctx->comp.req = NULL;
+	acomp_request_free(acomp_ctx->decomp.req);
+	acomp_ctx->decomp.req = NULL;
 
 	/*
 	 * We have to handle both cases here: an error pointer return from
@@ -796,6 +798,28 @@ static void zswap_entry_free(struct zswap_entry *entry)
 /*********************************
 * compressed storage functions
 **********************************/
+static int zswap_acomp_req_init(struct zswap_acomp_req *areq,
+				struct crypto_acomp *acomp)
+{
+	/* acomp_request_alloc() returns NULL in case of an error. */
+	areq->req = acomp_request_alloc(acomp);
+	if (!areq->req)
+		return -ENOMEM;
+
+	crypto_init_wait(&areq->wait);
+
+	/*
+	 * if the backend of acomp is async zip, crypto_req_done() will wakeup
+	 * crypto_wait_req(); if the backend of acomp is scomp, the callback
+	 * won't be called, crypto_wait_req() will return without blocking.
+	 */
+	acomp_request_set_callback(areq->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
+				   crypto_req_done, &areq->wait);
+
+	mutex_init(&areq->mutex);
+	return 0;
+}
+
 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 {
 	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
@@ -827,25 +851,13 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 		goto fail;
 	}
 
-	/* acomp_request_alloc() returns NULL in case of an error. */
-	acomp_ctx->req = acomp_request_alloc(acomp_ctx->acomp);
-	if (!acomp_ctx->req) {
+	if (zswap_acomp_req_init(&acomp_ctx->comp, acomp_ctx->acomp) ||
+	    zswap_acomp_req_init(&acomp_ctx->decomp, acomp_ctx->acomp)) {
 		pr_err("could not alloc crypto acomp_request %s\n",
 		       pool->tfm_name);
 		goto fail;
 	}
 
-	crypto_init_wait(&acomp_ctx->wait);
-
-	/*
-	 * if the backend of acomp is async zip, crypto_req_done() will wakeup
-	 * crypto_wait_req(); if the backend of acomp is scomp, the callback
-	 * won't be called, crypto_wait_req() will return without blocking.
-	 */
-	acomp_request_set_callback(acomp_ctx->req, CRYPTO_TFM_REQ_MAY_BACKLOG,
-				   crypto_req_done, &acomp_ctx->wait);
-
-	mutex_init(&acomp_ctx->mutex);
 	return 0;
 
 fail:
@@ -866,14 +878,15 @@ static bool zswap_compress(struct folio *folio, long index,
 	bool mapped = false;
 
 	acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
-	mutex_lock(&acomp_ctx->mutex);
+	mutex_lock(&acomp_ctx->comp.mutex);
 
 	dst = acomp_ctx->buffer;
 	sg_init_table(&input, 1);
 	sg_set_folio(&input, folio, PAGE_SIZE, index * PAGE_SIZE);
 
 	sg_init_one(&output, dst, PAGE_SIZE);
-	acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
+	acomp_request_set_params(acomp_ctx->comp.req, &input, &output,
+				 PAGE_SIZE, dlen);
 
 	/*
 	 * it maybe looks a little bit silly that we send an asynchronous request,
@@ -885,10 +898,12 @@ static bool zswap_compress(struct folio *folio, long index,
 	 * existing method to send the second page before the first page is done
 	 * in one thread doing zswap.
 	 * but in different threads running on different cpu, we have different
-	 * acomp instance, so multiple threads can do (de)compression in parallel.
+	 * acomp instance, and compression and decompression use separate
+	 * requests, so multiple threads can do (de)compression in parallel.
 	 */
-	comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
-	dlen = acomp_ctx->req->dlen;
+	comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->comp.req),
+				   &acomp_ctx->comp.wait);
+	dlen = acomp_ctx->comp.req->dlen;
 
 	/*
 	 * If a page cannot be compressed into a size smaller than PAGE_SIZE,
@@ -932,7 +947,7 @@ static bool zswap_compress(struct folio *folio, long index,
 	else if (alloc_ret)
 		zswap_reject_alloc_fail++;
 
-	mutex_unlock(&acomp_ctx->mutex);
+	mutex_unlock(&acomp_ctx->comp.mutex);
 	return comp_ret == 0 && alloc_ret == 0;
 }
 
@@ -948,7 +963,7 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
 		return false;
 
 	acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
-	mutex_lock(&acomp_ctx->mutex);
+	mutex_lock(&acomp_ctx->decomp.mutex);
 	zs_obj_read_sg_begin(pool->zs_pool, entry->handle, input, entry->length);
 
 	/* zswap entries of length PAGE_SIZE are not compressed. */
@@ -965,15 +980,15 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
 	} else {
 		sg_init_table(&output, 1);
 		sg_set_folio(&output, folio, PAGE_SIZE, 0);
-		acomp_request_set_params(acomp_ctx->req, input, &output,
+		acomp_request_set_params(acomp_ctx->decomp.req, input, &output,
 					 entry->length, PAGE_SIZE);
-		ret = crypto_acomp_decompress(acomp_ctx->req);
-		ret = crypto_wait_req(ret, &acomp_ctx->wait);
-		dlen = acomp_ctx->req->dlen;
+		ret = crypto_acomp_decompress(acomp_ctx->decomp.req);
+		ret = crypto_wait_req(ret, &acomp_ctx->decomp.wait);
+		dlen = acomp_ctx->decomp.req->dlen;
 	}
 
 	zs_obj_read_sg_end(pool->zs_pool, entry->handle);
-	mutex_unlock(&acomp_ctx->mutex);
+	mutex_unlock(&acomp_ctx->decomp.mutex);
 
 	if (!ret && dlen == PAGE_SIZE)
 		return true;
-- 
2.53.0-Meta

[PATCH 2/2] mm: zswap: use stack requests for synchronous decompression

[Usama Arif]

With separate requests for compression and decompression, loads still
serialize on the per-CPU decompression mutex. A low-priority load that
is preempted after the codec drops its stream lock keeps holding the mutex
and stalls every other load on that CPU, including higher-priority ones.

Synchronous algorithms whose requests need no extra context can use an
on-stack request, so decompress with one and take no zswap lock. All
in-tree software compressors qualify. Asynchronous algorithms, and
synchronous ones with request context, keep the per-CPU request and
mutex, which is still taken before the zsmalloc read lock.

Reading the per-CPU context without the mutex is safe. Since
commit ef3c0f6cb798e ("mm: zswap: tie per-CPU acomp_ctx lifetime to the
pool"), it is set up before its CPU comes online and is not torn down
until the pool is destroyed. The codecs keep their own stream locks, and
crypto_acomp_decompress() rejects on-stack requests only for
asynchronous transforms, which never take this path.

For software compressors this drops the heap request added by the
previous patch. The on-stack request and wait take 216 bytes, which
makes the load path about 270 bytes deeper on x86-64. Asynchronous
algorithms pay this too.

Signed-off-by: Usama Arif <usama.arif@linux.dev>
---
 mm/zswap.c | 65 +++++++++++++++++++++++++++++++++++++-----------------
 1 file changed, 45 insertions(+), 20 deletions(-)

diff --git a/mm/zswap.c b/mm/zswap.c
index 54187b1ef751d..7e7fb6e7ec24c 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -147,7 +147,7 @@ struct zswap_acomp_req {
 struct crypto_acomp_ctx {
 	struct crypto_acomp *acomp;
 	struct zswap_acomp_req comp;
-	struct zswap_acomp_req decomp;
+	struct zswap_acomp_req decomp;	/* unused by synchronous algorithms */
 	u8 *buffer;
 };
 
@@ -851,15 +851,20 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 		goto fail;
 	}
 
-	if (zswap_acomp_req_init(&acomp_ctx->comp, acomp_ctx->acomp) ||
-	    zswap_acomp_req_init(&acomp_ctx->decomp, acomp_ctx->acomp)) {
-		pr_err("could not alloc crypto acomp_request %s\n",
-		       pool->tfm_name);
-		goto fail;
+	if (zswap_acomp_req_init(&acomp_ctx->comp, acomp_ctx->acomp))
+		goto req_fail;
+
+	/* Synchronous algorithms decompress with an on-stack request. */
+	if (acomp_is_async(acomp_ctx->acomp) ||
+	    crypto_acomp_reqsize(acomp_ctx->acomp) > MAX_SYNC_COMP_REQSIZE) {
+		if (zswap_acomp_req_init(&acomp_ctx->decomp, acomp_ctx->acomp))
+			goto req_fail;
 	}
 
 	return 0;
 
+req_fail:
+	pr_err("could not alloc crypto acomp_request %s\n", pool->tfm_name);
 fail:
 	acomp_ctx_free(acomp_ctx);
 	return ret;
@@ -951,19 +956,14 @@ static bool zswap_compress(struct folio *folio, long index,
 	return comp_ret == 0 && alloc_ret == 0;
 }
 
-static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
+static bool __zswap_decompress(struct zswap_entry *entry,
+			       struct zswap_pool *pool, struct acomp_req *req,
+			       struct crypto_wait *wait, struct folio *folio)
 {
-	struct zswap_pool *pool = zswap_entry_pool(entry);
 	struct scatterlist input[2]; /* zsmalloc returns an SG list 1-2 entries */
 	struct scatterlist output;
-	struct crypto_acomp_ctx *acomp_ctx;
 	int ret = 0, dlen;
 
-	if (WARN_ON_ONCE(!pool))
-		return false;
-
-	acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
-	mutex_lock(&acomp_ctx->decomp.mutex);
 	zs_obj_read_sg_begin(pool->zs_pool, entry->handle, input, entry->length);
 
 	/* zswap entries of length PAGE_SIZE are not compressed. */
@@ -980,15 +980,14 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
 	} else {
 		sg_init_table(&output, 1);
 		sg_set_folio(&output, folio, PAGE_SIZE, 0);
-		acomp_request_set_params(acomp_ctx->decomp.req, input, &output,
-					 entry->length, PAGE_SIZE);
-		ret = crypto_acomp_decompress(acomp_ctx->decomp.req);
-		ret = crypto_wait_req(ret, &acomp_ctx->decomp.wait);
-		dlen = acomp_ctx->decomp.req->dlen;
+		acomp_request_set_params(req, input, &output, entry->length,
+					 PAGE_SIZE);
+		ret = crypto_acomp_decompress(req);
+		ret = crypto_wait_req(ret, wait);
+		dlen = req->dlen;
 	}
 
 	zs_obj_read_sg_end(pool->zs_pool, entry->handle);
-	mutex_unlock(&acomp_ctx->decomp.mutex);
 
 	if (!ret && dlen == PAGE_SIZE)
 		return true;
@@ -1002,6 +1001,32 @@ static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
 	return false;
 }
 
+static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
+{
+	struct zswap_pool *pool = zswap_entry_pool(entry);
+	struct crypto_acomp_ctx *acomp_ctx;
+	bool ret;
+
+	if (WARN_ON_ONCE(!pool))
+		return false;
+
+	acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
+	if (!acomp_ctx->decomp.req) {
+		ACOMP_REQUEST_ON_STACK(req, acomp_ctx->acomp);
+		DECLARE_CRYPTO_WAIT(wait);
+
+		acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
+					   crypto_req_done, &wait);
+		return __zswap_decompress(entry, pool, req, &wait, folio);
+	}
+
+	mutex_lock(&acomp_ctx->decomp.mutex);
+	ret = __zswap_decompress(entry, pool, acomp_ctx->decomp.req,
+				 &acomp_ctx->decomp.wait, folio);
+	mutex_unlock(&acomp_ctx->decomp.mutex);
+	return ret;
+}
+
 /*********************************
 * writeback code
 **********************************/
-- 
2.53.0-Meta