staging: zcache: zero-filled pages awareness
authorWanpeng Li <liwanp@linux.vnet.ibm.com>
Tue, 2 Apr 2013 02:46:14 +0000 (10:46 +0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 2 Apr 2013 19:49:51 +0000 (12:49 -0700)
Compression of zero-filled pages can unneccessarily cause internal
fragmentation, and thus waste memory. This special case can be
optimized.

This patch captures zero-filled pages, and marks their corresponding
zcache backing page entry as zero-filled. Whenever such zero-filled
page is retrieved, we fill the page frame with zero.

Acked-by: Dan Magenheimer <dan.magenheimer@oracle.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/staging/zcache/zcache-main.c

index 274a48e94cfe2601aeb8b9b529a7ea4fbe83ad08..927d5fdaa3b104659a32143274c98d37035391c5 100644 (file)
@@ -61,6 +61,12 @@ static inline void frontswap_tmem_exclusive_gets(bool b)
 }
 #endif
 
+/*
+ * mark pampd to special value in order that later
+ * retrieve will identify zero-filled pages
+ */
+#define ZERO_FILLED 0x2
+
 /* enable (or fix code) when Seth's patches are accepted upstream */
 #define zcache_writeback_enabled 0
 
@@ -277,17 +283,23 @@ static void zcache_obj_free(struct tmem_obj *obj, struct tmem_pool *pool)
        kmem_cache_free(zcache_obj_cache, obj);
 }
 
-static bool page_is_zero_filled(void *ptr)
+/*
+ * Compressing zero-filled pages will waste memory and introduce
+ * serious fragmentation, skip it to avoid overhead.
+ */
+static bool page_is_zero_filled(struct page *p)
 {
        unsigned int pos;
-       unsigned long *page;
-
-       page = (unsigned long *)ptr;
+       char *page;
 
+       page = kmap_atomic(p);
        for (pos = 0; pos < PAGE_SIZE / sizeof(*page); pos++) {
-               if (page[pos])
+               if (page[pos]) {
+                       kunmap_atomic(page);
                        return false;
+               }
        }
+       kunmap_atomic(page);
 
        return true;
 }
@@ -356,8 +368,15 @@ static void *zcache_pampd_eph_create(char *data, size_t size, bool raw,
 {
        void *pampd = NULL, *cdata = data;
        unsigned clen = size;
+       bool zero_filled = false;
        struct page *page = (struct page *)(data), *newpage;
 
+       if (page_is_zero_filled(page)) {
+               clen = 0;
+               zero_filled = true;
+               goto got_pampd;
+       }
+
        if (!raw) {
                zcache_compress(page, &cdata, &clen);
                if (clen > zbud_max_buddy_size()) {
@@ -397,6 +416,8 @@ got_pampd:
        inc_zcache_eph_zpages();
        if (ramster_enabled && raw)
                ramster_count_foreign_pages(true, 1);
+       if (zero_filled)
+               pampd = (void *)ZERO_FILLED;
 out:
        return pampd;
 }
@@ -406,6 +427,7 @@ static void *zcache_pampd_pers_create(char *data, size_t size, bool raw,
 {
        void *pampd = NULL, *cdata = data;
        unsigned clen = size;
+       bool zero_filled = false;
        struct page *page = (struct page *)(data), *newpage;
        unsigned long zbud_mean_zsize;
        unsigned long curr_pers_zpages, total_zsize;
@@ -414,6 +436,13 @@ static void *zcache_pampd_pers_create(char *data, size_t size, bool raw,
                BUG_ON(!ramster_enabled);
                goto create_pampd;
        }
+
+       if (page_is_zero_filled(page)) {
+               clen = 0;
+               zero_filled = true;
+               goto got_pampd;
+       }
+
        curr_pers_zpages = zcache_pers_zpages;
 /* FIXME CONFIG_RAMSTER... subtract atomic remote_pers_pages here? */
        if (!raw)
@@ -471,6 +500,8 @@ got_pampd:
        inc_zcache_pers_zbytes(clen);
        if (ramster_enabled && raw)
                ramster_count_foreign_pages(false, 1);
+       if (zero_filled)
+               pampd = (void *)ZERO_FILLED;
 out:
        return pampd;
 }
@@ -532,7 +563,8 @@ out:
  */
 void zcache_pampd_create_finish(void *pampd, bool eph)
 {
-       zbud_create_finish((struct zbudref *)pampd, eph);
+       if (pampd != (void *)ZERO_FILLED)
+               zbud_create_finish((struct zbudref *)pampd, eph);
 }
 
 /*
@@ -577,6 +609,14 @@ static int zcache_pampd_get_data(char *data, size_t *sizep, bool raw,
        BUG_ON(preemptible());
        BUG_ON(eph);    /* fix later if shared pools get implemented */
        BUG_ON(pampd_is_remote(pampd));
+
+       if (pampd == (void *)ZERO_FILLED) {
+               handle_zero_filled_page(data);
+               if (!raw)
+                       *sizep = PAGE_SIZE;
+               return 0;
+       }
+
        if (raw)
                ret = zbud_copy_from_zbud(data, (struct zbudref *)pampd,
                                                sizep, eph);
@@ -597,13 +637,22 @@ static int zcache_pampd_get_data_and_free(char *data, size_t *sizep, bool raw,
                                        void *pampd, struct tmem_pool *pool,
                                        struct tmem_oid *oid, uint32_t index)
 {
-       int ret;
-       bool eph = !is_persistent(pool);
+       int ret = 0;
+       bool eph = !is_persistent(pool), zero_filled = false;
        struct page *page = NULL;
        unsigned int zsize, zpages;
 
        BUG_ON(preemptible());
        BUG_ON(pampd_is_remote(pampd));
+
+       if (pampd == (void *)ZERO_FILLED) {
+               handle_zero_filled_page(data);
+               zero_filled = true;
+               if (!raw)
+                       *sizep = PAGE_SIZE;
+               goto zero_fill;
+       }
+
        if (raw)
                ret = zbud_copy_from_zbud(data, (struct zbudref *)pampd,
                                                sizep, eph);
@@ -615,6 +664,7 @@ static int zcache_pampd_get_data_and_free(char *data, size_t *sizep, bool raw,
        }
        page = zbud_free_and_delist((struct zbudref *)pampd, eph,
                                        &zsize, &zpages);
+zero_fill:
        if (eph) {
                if (page)
                        dec_zcache_eph_pageframes();
@@ -628,7 +678,7 @@ static int zcache_pampd_get_data_and_free(char *data, size_t *sizep, bool raw,
        }
        if (!is_local_client(pool->client))
                ramster_count_foreign_pages(eph, -1);
-       if (page)
+       if (page && !zero_filled)
                zcache_free_page(page);
        return ret;
 }
@@ -642,16 +692,22 @@ static void zcache_pampd_free(void *pampd, struct tmem_pool *pool,
 {
        struct page *page = NULL;
        unsigned int zsize, zpages;
+       bool zero_filled = false;
 
        BUG_ON(preemptible());
-       if (pampd_is_remote(pampd)) {
+
+       if (pampd == (void *)ZERO_FILLED)
+               zero_filled = true;
+
+       if (pampd_is_remote(pampd) && !zero_filled) {
                BUG_ON(!ramster_enabled);
                pampd = ramster_pampd_free(pampd, pool, oid, index, acct);
                if (pampd == NULL)
                        return;
        }
        if (is_ephemeral(pool)) {
-               page = zbud_free_and_delist((struct zbudref *)pampd,
+               if (!zero_filled)
+                       page = zbud_free_and_delist((struct zbudref *)pampd,
                                                true, &zsize, &zpages);
                if (page)
                        dec_zcache_eph_pageframes();
@@ -659,7 +715,8 @@ static void zcache_pampd_free(void *pampd, struct tmem_pool *pool,
                dec_zcache_eph_zbytes(zsize);
                /* FIXME CONFIG_RAMSTER... check acct parameter? */
        } else {
-               page = zbud_free_and_delist((struct zbudref *)pampd,
+               if (!zero_filled)
+                       page = zbud_free_and_delist((struct zbudref *)pampd,
                                                false, &zsize, &zpages);
                if (page)
                        dec_zcache_pers_pageframes();
@@ -668,7 +725,7 @@ static void zcache_pampd_free(void *pampd, struct tmem_pool *pool,
        }
        if (!is_local_client(pool->client))
                ramster_count_foreign_pages(is_ephemeral(pool), -1);
-       if (page)
+       if (page && !zero_filled)
                zcache_free_page(page);
 }