Merge branch 'pm-cpufreq'
[linux-2.6-microblaze.git] / mm / zswap.c
index fbb7829..182f6ad 100644 (file)
 #include <linux/rbtree.h>
 #include <linux/swap.h>
 #include <linux/crypto.h>
+#include <linux/scatterlist.h>
 #include <linux/mempool.h>
 #include <linux/zpool.h>
+#include <crypto/acompress.h>
 
 #include <linux/mm_types.h>
 #include <linux/page-flags.h>
@@ -81,7 +83,7 @@ static bool zswap_pool_reached_full;
 static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
 static int zswap_enabled_param_set(const char *,
                                   const struct kernel_param *);
-static struct kernel_param_ops zswap_enabled_param_ops = {
+static const struct kernel_param_ops zswap_enabled_param_ops = {
        .set =          zswap_enabled_param_set,
        .get =          param_get_bool,
 };
@@ -91,7 +93,7 @@ module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
 static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
 static int zswap_compressor_param_set(const char *,
                                      const struct kernel_param *);
-static struct kernel_param_ops zswap_compressor_param_ops = {
+static const struct kernel_param_ops zswap_compressor_param_ops = {
        .set =          zswap_compressor_param_set,
        .get =          param_get_charp,
        .free =         param_free_charp,
@@ -102,7 +104,7 @@ module_param_cb(compressor, &zswap_compressor_param_ops,
 /* Compressed storage zpool to use */
 static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
-static struct kernel_param_ops zswap_zpool_param_ops = {
+static const struct kernel_param_ops zswap_zpool_param_ops = {
        .set =          zswap_zpool_param_set,
        .get =          param_get_charp,
        .free =         param_free_charp,
@@ -127,9 +129,17 @@ module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
 * data structures
 **********************************/
 
+struct crypto_acomp_ctx {
+       struct crypto_acomp *acomp;
+       struct acomp_req *req;
+       struct crypto_wait wait;
+       u8 *dstmem;
+       struct mutex *mutex;
+};
+
 struct zswap_pool {
        struct zpool *zpool;
-       struct crypto_comp * __percpu *tfm;
+       struct crypto_acomp_ctx __percpu *acomp_ctx;
        struct kref kref;
        struct list_head list;
        struct work_struct release_work;
@@ -388,23 +398,43 @@ static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 * per-cpu code
 **********************************/
 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
+/*
+ * If users dynamically change the zpool type and compressor at runtime, i.e.
+ * zswap is running, zswap can have more than one zpool on one cpu, but they
+ * are sharing dtsmem. So we need this mutex to be per-cpu.
+ */
+static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
 
 static int zswap_dstmem_prepare(unsigned int cpu)
 {
+       struct mutex *mutex;
        u8 *dst;
 
        dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
        if (!dst)
                return -ENOMEM;
 
+       mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
+       if (!mutex) {
+               kfree(dst);
+               return -ENOMEM;
+       }
+
+       mutex_init(mutex);
        per_cpu(zswap_dstmem, cpu) = dst;
+       per_cpu(zswap_mutex, cpu) = mutex;
        return 0;
 }
 
 static int zswap_dstmem_dead(unsigned int cpu)
 {
+       struct mutex *mutex;
        u8 *dst;
 
+       mutex = per_cpu(zswap_mutex, cpu);
+       kfree(mutex);
+       per_cpu(zswap_mutex, cpu) = NULL;
+
        dst = per_cpu(zswap_dstmem, cpu);
        kfree(dst);
        per_cpu(zswap_dstmem, cpu) = NULL;
@@ -415,30 +445,54 @@ static int zswap_dstmem_dead(unsigned int cpu)
 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 {
        struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
-       struct crypto_comp *tfm;
-
-       if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
-               return 0;
+       struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
+       struct crypto_acomp *acomp;
+       struct acomp_req *req;
+
+       acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
+       if (IS_ERR(acomp)) {
+               pr_err("could not alloc crypto acomp %s : %ld\n",
+                               pool->tfm_name, PTR_ERR(acomp));
+               return PTR_ERR(acomp);
+       }
+       acomp_ctx->acomp = acomp;
 
-       tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
-       if (IS_ERR_OR_NULL(tfm)) {
-               pr_err("could not alloc crypto comp %s : %ld\n",
-                      pool->tfm_name, PTR_ERR(tfm));
+       req = acomp_request_alloc(acomp_ctx->acomp);
+       if (!req) {
+               pr_err("could not alloc crypto acomp_request %s\n",
+                      pool->tfm_name);
+               crypto_free_acomp(acomp_ctx->acomp);
                return -ENOMEM;
        }
-       *per_cpu_ptr(pool->tfm, cpu) = tfm;
+       acomp_ctx->req = req;
+
+       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(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
+                                  crypto_req_done, &acomp_ctx->wait);
+
+       acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
+       acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
+
        return 0;
 }
 
 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
 {
        struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
-       struct crypto_comp *tfm;
+       struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
+
+       if (!IS_ERR_OR_NULL(acomp_ctx)) {
+               if (!IS_ERR_OR_NULL(acomp_ctx->req))
+                       acomp_request_free(acomp_ctx->req);
+               if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
+                       crypto_free_acomp(acomp_ctx->acomp);
+       }
 
-       tfm = *per_cpu_ptr(pool->tfm, cpu);
-       if (!IS_ERR_OR_NULL(tfm))
-               crypto_free_comp(tfm);
-       *per_cpu_ptr(pool->tfm, cpu) = NULL;
        return 0;
 }
 
@@ -561,8 +615,9 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
        pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 
        strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
-       pool->tfm = alloc_percpu(struct crypto_comp *);
-       if (!pool->tfm) {
+
+       pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
+       if (!pool->acomp_ctx) {
                pr_err("percpu alloc failed\n");
                goto error;
        }
@@ -585,7 +640,8 @@ static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
        return pool;
 
 error:
-       free_percpu(pool->tfm);
+       if (pool->acomp_ctx)
+               free_percpu(pool->acomp_ctx);
        if (pool->zpool)
                zpool_destroy_pool(pool->zpool);
        kfree(pool);
@@ -596,14 +652,14 @@ static __init struct zswap_pool *__zswap_pool_create_fallback(void)
 {
        bool has_comp, has_zpool;
 
-       has_comp = crypto_has_comp(zswap_compressor, 0, 0);
+       has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
        if (!has_comp && strcmp(zswap_compressor,
                                CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
                pr_err("compressor %s not available, using default %s\n",
                       zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
                param_free_charp(&zswap_compressor);
                zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
-               has_comp = crypto_has_comp(zswap_compressor, 0, 0);
+               has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
        }
        if (!has_comp) {
                pr_err("default compressor %s not available\n",
@@ -639,7 +695,7 @@ static void zswap_pool_destroy(struct zswap_pool *pool)
        zswap_pool_debug("destroying", pool);
 
        cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
-       free_percpu(pool->tfm);
+       free_percpu(pool->acomp_ctx);
        zpool_destroy_pool(pool->zpool);
        kfree(pool);
 }
@@ -723,7 +779,7 @@ static int __zswap_param_set(const char *val, const struct kernel_param *kp,
                }
                type = s;
        } else if (!compressor) {
-               if (!crypto_has_comp(s, 0, 0)) {
+               if (!crypto_has_acomp(s, 0, 0)) {
                        pr_err("compressor %s not available\n", s);
                        return -ENOENT;
                }
@@ -774,7 +830,7 @@ static int __zswap_param_set(const char *val, const struct kernel_param *kp,
                 * failed, maybe both compressor and zpool params were bad.
                 * Allow changing this param, so pool creation will succeed
                 * when the other param is changed. We already verified this
-                * param is ok in the zpool_has_pool() or crypto_has_comp()
+                * param is ok in the zpool_has_pool() or crypto_has_acomp()
                 * checks above.
                 */
                ret = param_set_charp(s, kp);
@@ -876,8 +932,10 @@ static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
        pgoff_t offset;
        struct zswap_entry *entry;
        struct page *page;
-       struct crypto_comp *tfm;
-       u8 *src, *dst;
+       struct scatterlist input, output;
+       struct crypto_acomp_ctx *acomp_ctx;
+
+       u8 *src;
        unsigned int dlen;
        int ret;
        struct writeback_control wbc = {
@@ -916,14 +974,20 @@ static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 
        case ZSWAP_SWAPCACHE_NEW: /* page is locked */
                /* decompress */
+               acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
+
                dlen = PAGE_SIZE;
                src = (u8 *)zhdr + sizeof(struct zswap_header);
-               dst = kmap_atomic(page);
-               tfm = *get_cpu_ptr(entry->pool->tfm);
-               ret = crypto_comp_decompress(tfm, src, entry->length,
-                                            dst, &dlen);
-               put_cpu_ptr(entry->pool->tfm);
-               kunmap_atomic(dst);
+
+               mutex_lock(acomp_ctx->mutex);
+               sg_init_one(&input, src, entry->length);
+               sg_init_table(&output, 1);
+               sg_set_page(&output, page, PAGE_SIZE, 0);
+               acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
+               ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
+               dlen = acomp_ctx->req->dlen;
+               mutex_unlock(acomp_ctx->mutex);
+
                BUG_ON(ret);
                BUG_ON(dlen != PAGE_SIZE);
 
@@ -1004,7 +1068,8 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 {
        struct zswap_tree *tree = zswap_trees[type];
        struct zswap_entry *entry, *dupentry;
-       struct crypto_comp *tfm;
+       struct scatterlist input, output;
+       struct crypto_acomp_ctx *acomp_ctx;
        int ret;
        unsigned int hlen, dlen = PAGE_SIZE;
        unsigned long handle, value;
@@ -1074,12 +1139,32 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
        }
 
        /* compress */
-       dst = get_cpu_var(zswap_dstmem);
-       tfm = *get_cpu_ptr(entry->pool->tfm);
-       src = kmap_atomic(page);
-       ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
-       kunmap_atomic(src);
-       put_cpu_ptr(entry->pool->tfm);
+       acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
+
+       mutex_lock(acomp_ctx->mutex);
+
+       dst = acomp_ctx->dstmem;
+       sg_init_table(&input, 1);
+       sg_set_page(&input, page, PAGE_SIZE, 0);
+
+       /* zswap_dstmem is of size (PAGE_SIZE * 2). Reflect same in sg_list */
+       sg_init_one(&output, dst, PAGE_SIZE * 2);
+       acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
+       /*
+        * it maybe looks a little bit silly that we send an asynchronous request,
+        * then wait for its completion synchronously. This makes the process look
+        * synchronous in fact.
+        * Theoretically, acomp supports users send multiple acomp requests in one
+        * acomp instance, then get those requests done simultaneously. but in this
+        * case, frontswap actually does store and load page by page, there is no
+        * existing method to send the second page before the first page is done
+        * in one thread doing frontswap.
+        * but in different threads running on different cpu, we have different
+        * acomp instance, so multiple threads can do (de)compression in parallel.
+        */
+       ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
+       dlen = acomp_ctx->req->dlen;
+
        if (ret) {
                ret = -EINVAL;
                goto put_dstmem;
@@ -1103,7 +1188,7 @@ static int zswap_frontswap_store(unsigned type, pgoff_t offset,
        memcpy(buf, &zhdr, hlen);
        memcpy(buf + hlen, dst, dlen);
        zpool_unmap_handle(entry->pool->zpool, handle);
-       put_cpu_var(zswap_dstmem);
+       mutex_unlock(acomp_ctx->mutex);
 
        /* populate entry */
        entry->offset = offset;
@@ -1131,7 +1216,7 @@ insert_entry:
        return 0;
 
 put_dstmem:
-       put_cpu_var(zswap_dstmem);
+       mutex_unlock(acomp_ctx->mutex);
        zswap_pool_put(entry->pool);
 freepage:
        zswap_entry_cache_free(entry);
@@ -1148,7 +1233,8 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
 {
        struct zswap_tree *tree = zswap_trees[type];
        struct zswap_entry *entry;
-       struct crypto_comp *tfm;
+       struct scatterlist input, output;
+       struct crypto_acomp_ctx *acomp_ctx;
        u8 *src, *dst;
        unsigned int dlen;
        int ret;
@@ -1175,11 +1261,16 @@ static int zswap_frontswap_load(unsigned type, pgoff_t offset,
        src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
        if (zpool_evictable(entry->pool->zpool))
                src += sizeof(struct zswap_header);
-       dst = kmap_atomic(page);
-       tfm = *get_cpu_ptr(entry->pool->tfm);
-       ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
-       put_cpu_ptr(entry->pool->tfm);
-       kunmap_atomic(dst);
+
+       acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
+       mutex_lock(acomp_ctx->mutex);
+       sg_init_one(&input, src, entry->length);
+       sg_init_table(&output, 1);
+       sg_set_page(&output, page, PAGE_SIZE, 0);
+       acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
+       ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
+       mutex_unlock(acomp_ctx->mutex);
+
        zpool_unmap_handle(entry->pool->zpool, entry->handle);
        BUG_ON(ret);