[PATCH 1/2] libf2fs_io: Add user-space cache
From: Robin Hsu
Date: Tue Oct 29 2019 - 03:47:39 EST
Implemented cache options in F2FS configuration 'c':
* use c.cache_config.num_cache_entry to set the number of
cache entries (in block), minimum 1024, or 0 to disable cache.
* use c.cache_config.max_hash_collision to set maximum hash
collision (max 16).
* use c.cache_config.dbg_en to enable printing of debug messages.
Signed-off-by: Robin Hsu <robinhsu@xxxxxxxxxx>
---
include/f2fs_fs.h | 20 +++
lib/libf2fs_io.c | 317 ++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 337 insertions(+)
diff --git a/include/f2fs_fs.h b/include/f2fs_fs.h
index 84f3f3e..a386e61 100644
--- a/include/f2fs_fs.h
+++ b/include/f2fs_fs.h
@@ -3,6 +3,8 @@
*
* Copyright (c) 2012 Samsung Electronics Co., Ltd.
* http://www.samsung.com/
+ * Copyright (c) 2019 Google Inc.
+ * http://www.google.com/
*
* Dual licensed under the GPL or LGPL version 2 licenses.
*
@@ -329,6 +331,18 @@ struct device_info {
size_t zone_blocks;
};
+typedef off_t off64_t;
+
+typedef struct {
+ /* Value 0 means no cache, minimum 1024 */
+ off64_t num_cache_entry;
+
+ /* Value 0 means always overwrite (no collision allowed). maximum 16 */
+ unsigned max_hash_collision;
+
+ bool dbg_en;
+} dev_cache_config_t;
+
struct f2fs_configuration {
u_int32_t reserved_segments;
u_int32_t new_reserved_segments;
@@ -419,6 +433,9 @@ struct f2fs_configuration {
/* precomputed fs UUID checksum for seeding other checksums */
u_int32_t chksum_seed;
+
+ /* cache parameters */
+ dev_cache_config_t cache_config;
};
#ifdef CONFIG_64BIT
@@ -1185,6 +1202,9 @@ extern int f2fs_init_sparse_file(void);
extern int f2fs_finalize_device(void);
extern int f2fs_fsync_device(void);
+extern void dcache_init(void);
+extern void dcache_release(void);
+
extern int dev_read(void *, __u64, size_t);
#ifdef POSIX_FADV_WILLNEED
extern int dev_readahead(__u64, size_t);
diff --git a/lib/libf2fs_io.c b/lib/libf2fs_io.c
index 4d0ea0d..4888b6e 100644
--- a/lib/libf2fs_io.c
+++ b/lib/libf2fs_io.c
@@ -3,6 +3,8 @@
*
* Copyright (c) 2013 Samsung Electronics Co., Ltd.
* http://www.samsung.com/
+ * Copyright (c) 2019 Google Inc.
+ * http://www.google.com/
*
* Dual licensed under the GPL or LGPL version 2 licenses.
*/
@@ -64,6 +66,309 @@ static inline off64_t lseek64(int fd, __u64 offset, int set)
}
#endif
+/* ---------- dev_cache, Least Used First (LUF) policy ------------------- */
+/*
+ * Least used block will be the first victim to be replaced when max hash
+ * collision exceeds
+ */
+static bool *dcache_valid; /* is the cached block valid? */
+static off64_t *dcache_blk; /* which block it cached */
+static uint64_t *dcache_lastused; /* last used ticks for cache entries */
+static char *dcache_buf; /* cached block data */
+static uint64_t dcache_usetick; /* current use tick */
+
+static int64_t dcache_raccess;
+static int64_t dcache_rhit;
+static int64_t dcache_rmiss;
+static int64_t dcache_rreplace;
+
+static bool dcache_exit_registered = false;
+
+/*
+ * Shadow config:
+ *
+ * Active set of the configurations.
+ * Global configuration 'dcache_config' will be transferred here when
+ * when dcache_init() is called
+ */
+static dev_cache_config_t dcache_config = {0, 16, 1};
+static bool dcache_initialized = false;
+
+#define MIN_NUM_CACHE_ENTRY ((off64_t)1024)
+#define MAX_MAX_HASH_COLLISION 16
+
+static int dcache_relocate_offset0[] = {
+ 20, -20, 40, -40, 80, -80, 160, -160,
+ 320, -320, 640, -640, 1280, -1280, 2560, -2560,
+};
+static int dcache_relocate_offset[16];
+
+static void dcache_print_statistics(void)
+{
+ off64_t i;
+ off64_t useCnt;
+
+ /* Number of used cache entries */
+ useCnt = 0;
+ for (i = 0; i < dcache_config.num_cache_entry; i++)
+ if (dcache_valid[i])
+ ++useCnt;
+
+ /*
+ * c: number of cache entries
+ * u: used entries
+ * RA: number of read access blocks
+ * CH: cache hit
+ * CM: cache miss
+ * Repl: read cache replaced
+ */
+ printf ("\nc, u, RA, CH, CM, Repl=\n");
+ printf ("%lu %lu %ld %ld %ld %ld\n",
+ dcache_config.num_cache_entry,
+ useCnt,
+ dcache_raccess,
+ dcache_rhit,
+ dcache_rmiss,
+ dcache_rreplace);
+}
+
+void dcache_release(void)
+{
+ if (!dcache_initialized)
+ return;
+
+ dcache_initialized = false;
+
+ if (c.cache_config.dbg_en)
+ dcache_print_statistics();
+
+ if (dcache_blk != NULL)
+ free(dcache_blk);
+ if (dcache_lastused != NULL)
+ free(dcache_lastused);
+ if (dcache_buf != NULL)
+ free(dcache_buf);
+ if (dcache_valid != NULL)
+ free(dcache_valid);
+ dcache_config.num_cache_entry = 0;
+ dcache_blk = NULL;
+ dcache_lastused = NULL;
+ dcache_buf = NULL;
+ dcache_valid = NULL;
+}
+
+// return 0 for success, error code for failure.
+static int dcache_alloc_all(int n)
+{
+ if ((dcache_blk = (off64_t *) malloc(sizeof(off64_t) * n)) == NULL
+ || (dcache_lastused = (uint64_t *)
+ malloc(sizeof(uint64_t) * n)) == NULL
+ || (dcache_buf = (char *) malloc (F2FS_BLKSIZE * n)) == NULL
+ || (dcache_valid = (bool *) malloc(sizeof(bool) * n)) == NULL)
+ {
+ dcache_release();
+ return -1;
+ }
+ dcache_config.num_cache_entry = n;
+ return 0;
+}
+
+static void dcache_relocate_init(void)
+{
+ int i;
+ int n0 = (sizeof(dcache_relocate_offset0)
+ / sizeof(dcache_relocate_offset0[0]));
+ int n = (sizeof(dcache_relocate_offset)
+ / sizeof(dcache_relocate_offset[0]));
+
+ ASSERT(n == n0);
+ for (i = 0; i < n && i < dcache_config.max_hash_collision; i++) {
+ if (abs(dcache_relocate_offset0[i])
+ > dcache_config.num_cache_entry / 2) {
+ dcache_config.max_hash_collision = i;
+ break;
+ }
+ dcache_relocate_offset[i] =
+ dcache_config.num_cache_entry
+ + dcache_relocate_offset0[i];
+ }
+}
+
+void dcache_init(void)
+{
+ off64_t n;
+
+ if (c.cache_config.num_cache_entry == 0)
+ return;
+
+ /* release previous cache init, if any */
+ dcache_release();
+
+ dcache_blk = NULL;
+ dcache_lastused = NULL;
+ dcache_buf = NULL;
+ dcache_valid = NULL;
+
+ dcache_config = c.cache_config;
+
+ n = max(MIN_NUM_CACHE_ENTRY, dcache_config.num_cache_entry);
+
+ /* halve alloc size until alloc succeed, or min cache reached */
+ while (dcache_alloc_all(n) != 0 && n != MIN_NUM_CACHE_ENTRY)
+ n = max(MIN_NUM_CACHE_ENTRY, n/2);
+
+ /* must be the last: data dependent on num_cache_entry */
+ dcache_relocate_init();
+ dcache_initialized = true;
+
+ if (!dcache_exit_registered) {
+ dcache_exit_registered = true;
+ atexit(dcache_release); /* auto release */
+ }
+}
+
+static inline char *dcache_addr(off64_t entry)
+{
+ return dcache_buf + F2FS_BLKSIZE * entry;
+}
+
+/* relocate on (n+1)-th collision */
+static inline off64_t dcache_relocate(off64_t entry, int n)
+{
+ return (entry + dcache_relocate_offset[n]) %
+ dcache_config.num_cache_entry;
+}
+
+static off64_t dcache_find(off64_t blk)
+{
+ register off64_t n = dcache_config.num_cache_entry;
+ register unsigned m = dcache_config.max_hash_collision;
+ off64_t entry, least_used, target;
+ unsigned try;
+
+ target = least_used = entry = blk % n;
+
+ for (try = 0; try < m; try++) {
+ if (!dcache_valid[target] || dcache_blk[target] == blk)
+ return target; /* found target or empty cache slot */
+ if (dcache_lastused[target] < dcache_lastused[least_used])
+ least_used = target;
+ target = dcache_relocate(entry, try); /* next target */
+ }
+ return least_used; /* max search reached, return least used slot */
+}
+
+/* Physical read into cache */
+static int dcache_io_read(int fd, off64_t entry, off64_t offset, off64_t blk)
+{
+ if (lseek64(fd, (off64_t)offset, SEEK_SET) < 0) {
+ MSG(0, "\n lseek64 fail.\n");
+ return -1;
+ }
+ if (read(fd, dcache_buf + entry * F2FS_BLKSIZE, F2FS_BLKSIZE) < 0) {
+ MSG(0, "\n read() fail.\n");
+ return -1;
+ }
+ dcache_lastused[entry] = ++dcache_usetick;
+ dcache_valid[entry] = true;
+ dcache_blk[entry] = blk;
+ return 0;
+}
+
+/*
+ * - Note: Read/Write are not symmetric:
+ * For read, we need to do it block by block, due to the cache nature:
+ * some blocks may be cached, and others don't.
+ * For write, since we always do a write-thru, we can join all writes into one,
+ * and write it once at the caller. This function updates the cache for write, but
+ * not the do a physical write. The caller is responsible for the physical write.
+ * - Note: We concentrate read/write together, due to the fact of similar structure to find
+ * the relevant cache entries
+ * - Return values:
+ * 0: success
+ * 1: cache not available (uninitialized)
+ * -1: error
+ */
+static int dcache_update_rw(int fd, void *buf, off64_t offset,
+ size_t byte_count, bool is_write)
+{
+ off64_t blk;
+ int32_t addr_in_blk;
+ off64_t start;
+
+ if (!dcache_initialized)
+ dcache_init(); /* auto initialize */
+
+ if (!dcache_initialized)
+ return 1; /* not available */
+
+ blk = offset / F2FS_BLKSIZE;
+ addr_in_blk = offset % F2FS_BLKSIZE;
+ start = blk * F2FS_BLKSIZE;
+
+ while (byte_count != 0) {
+ int32_t cur_size = min(byte_count,
+ (size_t)(F2FS_BLKSIZE - addr_in_blk));
+ off64_t entry = dcache_find(blk);
+
+ if (!is_write)
+ ++dcache_raccess;
+
+ if (dcache_valid[entry] && dcache_blk[entry] == blk) {
+ /* cache hit */
+ if (is_write) /* write: update cache */
+ memcpy(dcache_addr(entry) + addr_in_blk,
+ buf, cur_size);
+ else
+ ++dcache_rhit;
+ } else {
+ /* cache miss */
+ if (!is_write) {
+ int err;
+ ++dcache_rmiss;
+ if (dcache_valid[entry])
+ ++dcache_rreplace;
+ /* read: physical I/O read into cache */
+ err = dcache_io_read(fd, entry, start, blk);
+ if (err)
+ return err;
+ }
+ }
+
+ /* read: copy data from cache */
+ /* write: nothing to do, since we don't do physical write. */
+ if (!is_write)
+ memcpy(buf, dcache_addr(entry) + addr_in_blk,
+ cur_size);
+
+ /* next block */
+ ++blk;
+ buf += cur_size;
+ offset += F2FS_BLKSIZE;
+ byte_count -= cur_size;
+ addr_in_blk = 0;
+ }
+ return 0;
+}
+
+/*
+ * dcache_update_cache() just update cache, won't do physical I/O.
+ * Thus even no error, we need normal non-cache I/O for actual write
+ *
+ * return value: 1: cache not available
+ * 0: success, -1: I/O error
+ */
+inline int dcache_update_cache(int fd, void *buf, off64_t offset, size_t count)
+{
+ return dcache_update_rw(fd, buf, offset, count, true);
+}
+
+/* handles read into cache + read into buffer */
+inline int dcache_read(int fd, void *buf, off64_t offset, size_t count)
+{
+ return dcache_update_rw(fd, buf, offset, count, false);
+}
+
/*
* IO interfaces
*/
@@ -185,6 +490,7 @@ static int sparse_write_zeroed_blk(__u64 block, int count) { return 0; }
int dev_read(void *buf, __u64 offset, size_t len)
{
int fd;
+ int err;
if (c.sparse_mode)
return sparse_read_blk(offset / F2FS_BLKSIZE,
@@ -194,6 +500,11 @@ int dev_read(void *buf, __u64 offset, size_t len)
if (fd < 0)
return fd;
+ /* err = 1: cache not available, fall back to non-cache R/W */
+ /* err = 0: success, err=-1: I/O error */
+ err = dcache_read(fd, buf, (off64_t)offset, len);
+ if (err <= 0)
+ return err;
if (lseek64(fd, (off64_t)offset, SEEK_SET) < 0)
return -1;
if (read(fd, buf, len) < 0)
@@ -233,6 +544,12 @@ int dev_write(void *buf, __u64 offset, size_t len)
if (fd < 0)
return fd;
+ /*
+ * dcache_update_cache() just update cache, won't do I/O.
+ * Thus even no error, we need normal non-cache I/O for actual write
+ */
+ if (dcache_update_cache(fd, buf, (off64_t)offset, len) < 0)
+ return -1;
if (lseek64(fd, (off64_t)offset, SEEK_SET) < 0)
return -1;
if (write(fd, buf, len) < 0)
--
2.24.0.rc0.303.g954a862665-goog