zram: partial IO refactoring
authorMinchan Kim <minchan@kernel.org>
Wed, 3 May 2017 21:55:41 +0000 (14:55 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 3 May 2017 22:52:11 +0000 (15:52 -0700)
For architecture(PAGE_SIZE > 4K), zram have supported partial IO.
However, the mixed code for handling normal/partial IO is too mess,
error-prone to modify IO handler functions with upcoming feature so this
patch aims for cleaning up zram's IO handling functions.

Link: http://lkml.kernel.org/r/1492052365-16169-3-git-send-email-minchan@kernel.org
Signed-off-by: Minchan Kim <minchan@kernel.org>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Johannes Thumshirn <jthumshirn@suse.de>
Cc: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/block/zram/zram_drv.c

index 8a38ff0c16a32e1f162fc2f1c16924c82e10b850..47e15fec3cd09ad6a5f902a314c61a3ecfc576dd 100644 (file)
@@ -45,6 +45,8 @@ static const char *default_compressor = "lzo";
 /* Module params (documentation at end) */
 static unsigned int num_devices = 1;
 
+static void zram_free_page(struct zram *zram, size_t index);
+
 static inline bool init_done(struct zram *zram)
 {
        return zram->disksize;
@@ -98,10 +100,17 @@ static void zram_set_obj_size(struct zram_meta *meta,
        meta->table[index].value = (flags << ZRAM_FLAG_SHIFT) | size;
 }
 
+#if PAGE_SIZE != 4096
 static inline bool is_partial_io(struct bio_vec *bvec)
 {
        return bvec->bv_len != PAGE_SIZE;
 }
+#else
+static inline bool is_partial_io(struct bio_vec *bvec)
+{
+       return false;
+}
+#endif
 
 static void zram_revalidate_disk(struct zram *zram)
 {
@@ -189,18 +198,6 @@ static bool page_same_filled(void *ptr, unsigned long *element)
        return true;
 }
 
-static void handle_same_page(struct bio_vec *bvec, unsigned long element)
-{
-       struct page *page = bvec->bv_page;
-       void *user_mem;
-
-       user_mem = kmap_atomic(page);
-       zram_fill_page(user_mem + bvec->bv_offset, bvec->bv_len, element);
-       kunmap_atomic(user_mem);
-
-       flush_dcache_page(page);
-}
-
 static ssize_t initstate_show(struct device *dev,
                struct device_attribute *attr, char *buf)
 {
@@ -416,6 +413,53 @@ static DEVICE_ATTR_RO(io_stat);
 static DEVICE_ATTR_RO(mm_stat);
 static DEVICE_ATTR_RO(debug_stat);
 
+static bool zram_same_page_read(struct zram *zram, u32 index,
+                               struct page *page,
+                               unsigned int offset, unsigned int len)
+{
+       struct zram_meta *meta = zram->meta;
+
+       bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
+       if (unlikely(!meta->table[index].handle) ||
+                       zram_test_flag(meta, index, ZRAM_SAME)) {
+               void *mem;
+
+               bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
+               mem = kmap_atomic(page);
+               zram_fill_page(mem + offset, len, meta->table[index].element);
+               kunmap_atomic(mem);
+               return true;
+       }
+       bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
+
+       return false;
+}
+
+static bool zram_same_page_write(struct zram *zram, u32 index,
+                                       struct page *page)
+{
+       unsigned long element;
+       void *mem = kmap_atomic(page);
+
+       if (page_same_filled(mem, &element)) {
+               struct zram_meta *meta = zram->meta;
+
+               kunmap_atomic(mem);
+               /* Free memory associated with this sector now. */
+               bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
+               zram_free_page(zram, index);
+               zram_set_flag(meta, index, ZRAM_SAME);
+               zram_set_element(meta, index, element);
+               bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
+
+               atomic64_inc(&zram->stats.same_pages);
+               return true;
+       }
+       kunmap_atomic(mem);
+
+       return false;
+}
+
 static void zram_meta_free(struct zram_meta *meta, u64 disksize)
 {
        size_t num_pages = disksize >> PAGE_SHIFT;
@@ -502,169 +546,103 @@ static void zram_free_page(struct zram *zram, size_t index)
        zram_set_obj_size(meta, index, 0);
 }
 
-static int zram_decompress_page(struct zram *zram, char *mem, u32 index)
+static int zram_decompress_page(struct zram *zram, struct page *page, u32 index)
 {
-       int ret = 0;
-       unsigned char *cmem;
-       struct zram_meta *meta = zram->meta;
+       int ret;
        unsigned long handle;
        unsigned int size;
+       void *src, *dst;
+       struct zram_meta *meta = zram->meta;
+
+       if (zram_same_page_read(zram, index, page, 0, PAGE_SIZE))
+               return 0;
 
        bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
        handle = meta->table[index].handle;
        size = zram_get_obj_size(meta, index);
 
-       if (!handle || zram_test_flag(meta, index, ZRAM_SAME)) {
-               bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
-               zram_fill_page(mem, PAGE_SIZE, meta->table[index].element);
-               return 0;
-       }
-
-       cmem = zs_map_object(meta->mem_pool, handle, ZS_MM_RO);
+       src = zs_map_object(meta->mem_pool, handle, ZS_MM_RO);
        if (size == PAGE_SIZE) {
-               memcpy(mem, cmem, PAGE_SIZE);
+               dst = kmap_atomic(page);
+               memcpy(dst, src, PAGE_SIZE);
+               kunmap_atomic(dst);
+               ret = 0;
        } else {
                struct zcomp_strm *zstrm = zcomp_stream_get(zram->comp);
 
-               ret = zcomp_decompress(zstrm, cmem, size, mem);
+               dst = kmap_atomic(page);
+               ret = zcomp_decompress(zstrm, src, size, dst);
+               kunmap_atomic(dst);
                zcomp_stream_put(zram->comp);
        }
        zs_unmap_object(meta->mem_pool, handle);
        bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
 
        /* Should NEVER happen. Return bio error if it does. */
-       if (unlikely(ret)) {
+       if (unlikely(ret))
                pr_err("Decompression failed! err=%d, page=%u\n", ret, index);
-               return ret;
-       }
 
-       return 0;
+       return ret;
 }
 
 static int zram_bvec_read(struct zram *zram, struct bio_vec *bvec,
-                         u32 index, int offset)
+                               u32 index, int offset)
 {
        int ret;
        struct page *page;
-       unsigned char *user_mem, *uncmem = NULL;
-       struct zram_meta *meta = zram->meta;
-       page = bvec->bv_page;
 
-       bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
-       if (unlikely(!meta->table[index].handle) ||
-                       zram_test_flag(meta, index, ZRAM_SAME)) {
-               bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
-               handle_same_page(bvec, meta->table[index].element);
-               return 0;
+       page = bvec->bv_page;
+       if (is_partial_io(bvec)) {
+               /* Use a temporary buffer to decompress the page */
+               page = alloc_page(GFP_NOIO|__GFP_HIGHMEM);
+               if (!page)
+                       return -ENOMEM;
        }
-       bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
 
-       if (is_partial_io(bvec))
-               /* Use  a temporary buffer to decompress the page */
-               uncmem = kmalloc(PAGE_SIZE, GFP_NOIO);
+       ret = zram_decompress_page(zram, page, index);
+       if (unlikely(ret))
+               goto out;
 
-       user_mem = kmap_atomic(page);
-       if (!is_partial_io(bvec))
-               uncmem = user_mem;
+       if (is_partial_io(bvec)) {
+               void *dst = kmap_atomic(bvec->bv_page);
+               void *src = kmap_atomic(page);
 
-       if (!uncmem) {
-               pr_err("Unable to allocate temp memory\n");
-               ret = -ENOMEM;
-               goto out_cleanup;
+               memcpy(dst + bvec->bv_offset, src + offset, bvec->bv_len);
+               kunmap_atomic(src);
+               kunmap_atomic(dst);
        }
-
-       ret = zram_decompress_page(zram, uncmem, index);
-       /* Should NEVER happen. Return bio error if it does. */
-       if (unlikely(ret))
-               goto out_cleanup;
-
+out:
        if (is_partial_io(bvec))
-               memcpy(user_mem + bvec->bv_offset, uncmem + offset,
-                               bvec->bv_len);
+               __free_page(page);
 
-       flush_dcache_page(page);
-       ret = 0;
-out_cleanup:
-       kunmap_atomic(user_mem);
-       if (is_partial_io(bvec))
-               kfree(uncmem);
        return ret;
 }
 
-static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index,
-                          int offset)
+static int zram_compress(struct zram *zram, struct zcomp_strm **zstrm,
+                       struct page *page,
+                       unsigned long *out_handle, unsigned int *out_comp_len)
 {
-       int ret = 0;
-       unsigned int clen;
+       int ret;
+       unsigned int comp_len;
+       void *src;
+       unsigned long alloced_pages;
        unsigned long handle = 0;
-       struct page *page;
-       unsigned char *user_mem, *cmem, *src, *uncmem = NULL;
        struct zram_meta *meta = zram->meta;
-       struct zcomp_strm *zstrm = NULL;
-       unsigned long alloced_pages;
-       unsigned long element;
-
-       page = bvec->bv_page;
-       if (is_partial_io(bvec)) {
-               /*
-                * This is a partial IO. We need to read the full page
-                * before to write the changes.
-                */
-               uncmem = kmalloc(PAGE_SIZE, GFP_NOIO);
-               if (!uncmem) {
-                       ret = -ENOMEM;
-                       goto out;
-               }
-               ret = zram_decompress_page(zram, uncmem, index);
-               if (ret)
-                       goto out;
-       }
 
 compress_again:
-       user_mem = kmap_atomic(page);
-       if (is_partial_io(bvec)) {
-               memcpy(uncmem + offset, user_mem + bvec->bv_offset,
-                      bvec->bv_len);
-               kunmap_atomic(user_mem);
-               user_mem = NULL;
-       } else {
-               uncmem = user_mem;
-       }
-
-       if (page_same_filled(uncmem, &element)) {
-               if (user_mem)
-                       kunmap_atomic(user_mem);
-               /* Free memory associated with this sector now. */
-               bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
-               zram_free_page(zram, index);
-               zram_set_flag(meta, index, ZRAM_SAME);
-               zram_set_element(meta, index, element);
-               bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
-
-               atomic64_inc(&zram->stats.same_pages);
-               ret = 0;
-               goto out;
-       }
-
-       zstrm = zcomp_stream_get(zram->comp);
-       ret = zcomp_compress(zstrm, uncmem, &clen);
-       if (!is_partial_io(bvec)) {
-               kunmap_atomic(user_mem);
-               user_mem = NULL;
-               uncmem = NULL;
-       }
+       src = kmap_atomic(page);
+       ret = zcomp_compress(*zstrm, src, &comp_len);
+       kunmap_atomic(src);
 
        if (unlikely(ret)) {
                pr_err("Compression failed! err=%d\n", ret);
-               goto out;
+               if (handle)
+                       zs_free(meta->mem_pool, handle);
+               return ret;
        }
 
-       src = zstrm->buffer;
-       if (unlikely(clen > max_zpage_size)) {
-               clen = PAGE_SIZE;
-               if (is_partial_io(bvec))
-                       src = uncmem;
-       }
+       if (unlikely(comp_len > max_zpage_size))
+               comp_len = PAGE_SIZE;
 
        /*
         * handle allocation has 2 paths:
@@ -680,27 +658,21 @@ compress_again:
         * from the slow path and handle has already been allocated.
         */
        if (!handle)
-               handle = zs_malloc(meta->mem_pool, clen,
+               handle = zs_malloc(meta->mem_pool, comp_len,
                                __GFP_KSWAPD_RECLAIM |
                                __GFP_NOWARN |
                                __GFP_HIGHMEM |
                                __GFP_MOVABLE);
        if (!handle) {
                zcomp_stream_put(zram->comp);
-               zstrm = NULL;
-
                atomic64_inc(&zram->stats.writestall);
-
-               handle = zs_malloc(meta->mem_pool, clen,
+               handle = zs_malloc(meta->mem_pool, comp_len,
                                GFP_NOIO | __GFP_HIGHMEM |
                                __GFP_MOVABLE);
+               *zstrm = zcomp_stream_get(zram->comp);
                if (handle)
                        goto compress_again;
-
-               pr_err("Error allocating memory for compressed page: %u, size=%u\n",
-                       index, clen);
-               ret = -ENOMEM;
-               goto out;
+               return -ENOMEM;
        }
 
        alloced_pages = zs_get_total_pages(meta->mem_pool);
@@ -708,22 +680,45 @@ compress_again:
 
        if (zram->limit_pages && alloced_pages > zram->limit_pages) {
                zs_free(meta->mem_pool, handle);
-               ret = -ENOMEM;
-               goto out;
+               return -ENOMEM;
        }
 
-       cmem = zs_map_object(meta->mem_pool, handle, ZS_MM_WO);
+       *out_handle = handle;
+       *out_comp_len = comp_len;
+       return 0;
+}
 
-       if ((clen == PAGE_SIZE) && !is_partial_io(bvec)) {
+static int __zram_bvec_write(struct zram *zram, struct bio_vec *bvec, u32 index)
+{
+       int ret;
+       unsigned long handle;
+       unsigned int comp_len;
+       void *src, *dst;
+       struct zcomp_strm *zstrm;
+       struct zram_meta *meta = zram->meta;
+       struct page *page = bvec->bv_page;
+
+       if (zram_same_page_write(zram, index, page))
+               return 0;
+
+       zstrm = zcomp_stream_get(zram->comp);
+       ret = zram_compress(zram, &zstrm, page, &handle, &comp_len);
+       if (ret) {
+               zcomp_stream_put(zram->comp);
+               return ret;
+       }
+
+
+       dst = zs_map_object(meta->mem_pool, handle, ZS_MM_WO);
+
+       src = zstrm->buffer;
+       if (comp_len == PAGE_SIZE)
                src = kmap_atomic(page);
-               memcpy(cmem, src, PAGE_SIZE);
+       memcpy(dst, src, comp_len);
+       if (comp_len == PAGE_SIZE)
                kunmap_atomic(src);
-       } else {
-               memcpy(cmem, src, clen);
-       }
 
        zcomp_stream_put(zram->comp);
-       zstrm = NULL;
        zs_unmap_object(meta->mem_pool, handle);
 
        /*
@@ -732,19 +727,54 @@ compress_again:
         */
        bit_spin_lock(ZRAM_ACCESS, &meta->table[index].value);
        zram_free_page(zram, index);
-
        meta->table[index].handle = handle;
-       zram_set_obj_size(meta, index, clen);
+       zram_set_obj_size(meta, index, comp_len);
        bit_spin_unlock(ZRAM_ACCESS, &meta->table[index].value);
 
        /* Update stats */
-       atomic64_add(clen, &zram->stats.compr_data_size);
+       atomic64_add(comp_len, &zram->stats.compr_data_size);
        atomic64_inc(&zram->stats.pages_stored);
+       return 0;
+}
+
+static int zram_bvec_write(struct zram *zram, struct bio_vec *bvec,
+                               u32 index, int offset)
+{
+       int ret;
+       struct page *page = NULL;
+       void *src;
+       struct bio_vec vec;
+
+       vec = *bvec;
+       if (is_partial_io(bvec)) {
+               void *dst;
+               /*
+                * This is a partial IO. We need to read the full page
+                * before to write the changes.
+                */
+               page = alloc_page(GFP_NOIO|__GFP_HIGHMEM);
+               if (!page)
+                       return -ENOMEM;
+
+               ret = zram_decompress_page(zram, page, index);
+               if (ret)
+                       goto out;
+
+               src = kmap_atomic(bvec->bv_page);
+               dst = kmap_atomic(page);
+               memcpy(dst + offset, src + bvec->bv_offset, bvec->bv_len);
+               kunmap_atomic(dst);
+               kunmap_atomic(src);
+
+               vec.bv_page = page;
+               vec.bv_len = PAGE_SIZE;
+               vec.bv_offset = 0;
+       }
+
+       ret = __zram_bvec_write(zram, &vec, index);
 out:
-       if (zstrm)
-               zcomp_stream_put(zram->comp);
        if (is_partial_io(bvec))
-               kfree(uncmem);
+               __free_page(page);
        return ret;
 }
 
@@ -800,6 +830,7 @@ static int zram_bvec_rw(struct zram *zram, struct bio_vec *bvec, u32 index,
        if (!is_write) {
                atomic64_inc(&zram->stats.num_reads);
                ret = zram_bvec_read(zram, bvec, index, offset);
+               flush_dcache_page(bvec->bv_page);
        } else {
                atomic64_inc(&zram->stats.num_writes);
                ret = zram_bvec_write(zram, bvec, index, offset);