f2fs: add block_age-based extent cache
authorJaegeuk Kim <jaegeuk@kernel.org>
Fri, 2 Dec 2022 01:37:15 +0000 (17:37 -0800)
committerJaegeuk Kim <jaegeuk@kernel.org>
Mon, 12 Dec 2022 22:53:56 +0000 (14:53 -0800)
This patch introduces a runtime hot/cold data separation method
for f2fs, in order to improve the accuracy for data temperature
classification, reduce the garbage collection overhead after
long-term data updates.

Enhanced hot/cold data separation can record data block update
frequency as "age" of the extent per inode, and take use of the age
info to indicate better temperature type for data block allocation:
 - It records total data blocks allocated since mount;
 - When file extent has been updated, it calculate the count of data
blocks allocated since last update as the age of the extent;
 - Before the data block allocated, it searches for the age info and
chooses the suitable segment for allocation.

Test and result:
 - Prepare: create about 30000 files
  * 3% for cold files (with cold file extension like .apk, from 3M to 10M)
  * 50% for warm files (with random file extension like .FcDxq, from 1K
to 4M)
  * 47% for hot files (with hot file extension like .db, from 1K to 256K)
 - create(5%)/random update(90%)/delete(5%) the files
  * total write amount is about 70G
  * fsync will be called for .db files, and buffered write will be used
for other files

The storage of test device is large enough(128G) so that it will not
switch to SSR mode during the test.

Benefit: dirty segment count increment reduce about 14%
 - before: Dirty +21110
 - after:  Dirty +18286

Signed-off-by: qixiaoyu1 <qixiaoyu1@xiaomi.com>
Signed-off-by: xiongping1 <xiongping1@xiaomi.com>
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
14 files changed:
Documentation/ABI/testing/sysfs-fs-f2fs
Documentation/filesystems/f2fs.rst
fs/f2fs/debug.c
fs/f2fs/extent_cache.c
fs/f2fs/f2fs.h
fs/f2fs/file.c
fs/f2fs/inode.c
fs/f2fs/node.c
fs/f2fs/node.h
fs/f2fs/segment.c
fs/f2fs/shrinker.c
fs/f2fs/super.c
fs/f2fs/sysfs.c
include/trace/events/f2fs.h

index 84a009a..9e37566 100644 (file)
@@ -655,3 +655,17 @@ Description:       When space utilization exceeds this, do background DISCARD aggressi
                Does DISCARD forcibly in a period of given min_discard_issue_time when the number
                of discards is not 0 and set discard granularity to 1.
                Default: 80
+
+What:          /sys/fs/f2fs/<disk>/hot_data_age_threshold
+Date:          November 2022
+Contact:       "Ping Xiong" <xiongping1@xiaomi.com>
+Description:   When DATA SEPARATION is on, it controls the age threshold to indicate
+               the data blocks as hot. By default it was initialized as 262144 blocks
+               (equals to 1GB).
+
+What:          /sys/fs/f2fs/<disk>/warm_data_age_threshold
+Date:          November 2022
+Contact:       "Ping Xiong" <xiongping1@xiaomi.com>
+Description:   When DATA SEPARATION is on, it controls the age threshold to indicate
+               the data blocks as warm. By default it was initialized as 2621440 blocks
+               (equals to 10GB).
index 67e1f3e..220f3e0 100644 (file)
@@ -347,6 +347,10 @@ memory=%s           Control memory mode. This supports "normal" and "low" modes.
                         Because of the nature of low memory devices, in this mode, f2fs
                         will try to save memory sometimes by sacrificing performance.
                         "normal" mode is the default mode and same as before.
+age_extent_cache        Enable an age extent cache based on rb-tree. It records
+                        data block update frequency of the extent per inode, in
+                        order to provide better temperature hints for data block
+                        allocation.
 ======================== ============================================================
 
 Debugfs Entries
index a9baa12..8f1ef74 100644 (file)
@@ -88,6 +88,9 @@ static void update_general_status(struct f2fs_sb_info *sbi)
        si->hit_largest = atomic64_read(&sbi->read_hit_largest);
        si->hit_total[EX_READ] += si->hit_largest;
 
+       /* block age extent_cache only */
+       si->allocated_data_blocks = atomic64_read(&sbi->allocated_data_blocks);
+
        /* validation check of the segment numbers */
        si->ndirty_node = get_pages(sbi, F2FS_DIRTY_NODES);
        si->ndirty_dent = get_pages(sbi, F2FS_DIRTY_DENTS);
@@ -516,6 +519,22 @@ static int stat_show(struct seq_file *s, void *v)
                seq_printf(s, "  - Inner Struct Count: tree: %d(%d), node: %d\n",
                                si->ext_tree[EX_READ], si->zombie_tree[EX_READ],
                                si->ext_node[EX_READ]);
+               seq_puts(s, "\nExtent Cache (Block Age):\n");
+               seq_printf(s, "  - Allocated Data Blocks: %llu\n",
+                               si->allocated_data_blocks);
+               seq_printf(s, "  - Hit Count: L1:%llu L2:%llu\n",
+                               si->hit_cached[EX_BLOCK_AGE],
+                               si->hit_rbtree[EX_BLOCK_AGE]);
+               seq_printf(s, "  - Hit Ratio: %llu%% (%llu / %llu)\n",
+                               !si->total_ext[EX_BLOCK_AGE] ? 0 :
+                               div64_u64(si->hit_total[EX_BLOCK_AGE] * 100,
+                               si->total_ext[EX_BLOCK_AGE]),
+                               si->hit_total[EX_BLOCK_AGE],
+                               si->total_ext[EX_BLOCK_AGE]);
+               seq_printf(s, "  - Inner Struct Count: tree: %d(%d), node: %d\n",
+                               si->ext_tree[EX_BLOCK_AGE],
+                               si->zombie_tree[EX_BLOCK_AGE],
+                               si->ext_node[EX_BLOCK_AGE]);
                seq_puts(s, "\nBalancing F2FS Async:\n");
                seq_printf(s, "  - DIO (R: %4d, W: %4d)\n",
                           si->nr_dio_read, si->nr_dio_write);
@@ -586,6 +605,8 @@ static int stat_show(struct seq_file *s, void *v)
                                si->cache_mem >> 10);
                seq_printf(s, "  - read extent cache: %llu KB\n",
                                si->ext_mem[EX_READ] >> 10);
+               seq_printf(s, "  - block age extent cache: %llu KB\n",
+                               si->ext_mem[EX_BLOCK_AGE] >> 10);
                seq_printf(s, "  - paged : %llu KB\n",
                                si->page_mem >> 10);
        }
index 305f969..1bd38a7 100644 (file)
@@ -6,6 +6,10 @@
  * Copyright (c) 2015 Samsung Electronics
  * Authors: Jaegeuk Kim <jaegeuk@kernel.org>
  *          Chao Yu <chao2.yu@samsung.com>
+ *
+ * block_age-based extent cache added by:
+ * Copyright (c) 2022 xiaomi Co., Ltd.
+ *             http://www.xiaomi.com/
  */
 
 #include <linux/fs.h>
@@ -18,6 +22,7 @@
 static void __set_extent_info(struct extent_info *ei,
                                unsigned int fofs, unsigned int len,
                                block_t blk, bool keep_clen,
+                               unsigned long age, unsigned long last_blocks,
                                enum extent_type type)
 {
        ei->fofs = fofs;
@@ -30,6 +35,9 @@ static void __set_extent_info(struct extent_info *ei,
 #ifdef CONFIG_F2FS_FS_COMPRESSION
                ei->c_len = 0;
 #endif
+       } else if (type == EX_BLOCK_AGE) {
+               ei->age = age;
+               ei->last_blocks = last_blocks;
        }
 }
 
@@ -47,10 +55,27 @@ static bool __may_read_extent_tree(struct inode *inode)
        return S_ISREG(inode->i_mode);
 }
 
+static bool __may_age_extent_tree(struct inode *inode)
+{
+       struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
+
+       if (!test_opt(sbi, AGE_EXTENT_CACHE))
+               return false;
+       /* don't cache block age info for cold file */
+       if (is_inode_flag_set(inode, FI_COMPRESSED_FILE))
+               return false;
+       if (file_is_cold(inode))
+               return false;
+
+       return S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode);
+}
+
 static bool __init_may_extent_tree(struct inode *inode, enum extent_type type)
 {
        if (type == EX_READ)
                return __may_read_extent_tree(inode);
+       else if (type == EX_BLOCK_AGE)
+               return __may_age_extent_tree(inode);
        return false;
 }
 
@@ -90,6 +115,11 @@ static bool __is_extent_mergeable(struct extent_info *back,
 #endif
                return (back->fofs + back->len == front->fofs &&
                                back->blk + back->len == front->blk);
+       } else if (type == EX_BLOCK_AGE) {
+               return (back->fofs + back->len == front->fofs &&
+                       abs(back->age - front->age) <= SAME_AGE_REGION &&
+                       abs(back->last_blocks - front->last_blocks) <=
+                                                       SAME_AGE_REGION);
        }
        return false;
 }
@@ -489,11 +519,22 @@ out:
                set_inode_flag(inode, FI_NO_EXTENT);
 }
 
+void f2fs_init_age_extent_tree(struct inode *inode)
+{
+       if (!__init_may_extent_tree(inode, EX_BLOCK_AGE))
+               return;
+       __grab_extent_tree(inode, EX_BLOCK_AGE);
+}
+
 void f2fs_init_extent_tree(struct inode *inode)
 {
        /* initialize read cache */
        if (__init_may_extent_tree(inode, EX_READ))
                __grab_extent_tree(inode, EX_READ);
+
+       /* initialize block age cache */
+       if (__init_may_extent_tree(inode, EX_BLOCK_AGE))
+               __grab_extent_tree(inode, EX_BLOCK_AGE);
 }
 
 static bool __lookup_extent_tree(struct inode *inode, pgoff_t pgofs,
@@ -544,6 +585,8 @@ out:
 
        if (type == EX_READ)
                trace_f2fs_lookup_read_extent_tree_end(inode, pgofs, ei);
+       else if (type == EX_BLOCK_AGE)
+               trace_f2fs_lookup_age_extent_tree_end(inode, pgofs, ei);
        return ret;
 }
 
@@ -642,6 +685,10 @@ static void __update_extent_tree_range(struct inode *inode,
        if (type == EX_READ)
                trace_f2fs_update_read_extent_tree_range(inode, fofs, len,
                                                tei->blk, 0);
+       else if (type == EX_BLOCK_AGE)
+               trace_f2fs_update_age_extent_tree_range(inode, fofs, len,
+                                               tei->age, tei->last_blocks);
+
        write_lock(&et->lock);
 
        if (type == EX_READ) {
@@ -694,6 +741,7 @@ static void __update_extent_tree_range(struct inode *inode,
                                __set_extent_info(&ei,
                                        end, org_end - end,
                                        end - dei.fofs + dei.blk, false,
+                                       dei.age, dei.last_blocks,
                                        type);
                                en1 = __insert_extent_tree(sbi, et, &ei,
                                                        NULL, NULL, true);
@@ -702,6 +750,7 @@ static void __update_extent_tree_range(struct inode *inode,
                                __set_extent_info(&en->ei,
                                        end, en->ei.len - (end - dei.fofs),
                                        en->ei.blk + (end - dei.fofs), true,
+                                       dei.age, dei.last_blocks,
                                        type);
                                next_en = en;
                        }
@@ -732,11 +781,15 @@ static void __update_extent_tree_range(struct inode *inode,
                en = next_en;
        }
 
+       if (type == EX_BLOCK_AGE)
+               goto update_age_extent_cache;
+
        /* 3. update extent in read extent cache */
        BUG_ON(type != EX_READ);
 
        if (tei->blk) {
-               __set_extent_info(&ei, fofs, len, tei->blk, false, EX_READ);
+               __set_extent_info(&ei, fofs, len, tei->blk, false,
+                                 0, 0, EX_READ);
                if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en))
                        __insert_extent_tree(sbi, et, &ei,
                                        insert_p, insert_parent, leftmost);
@@ -758,7 +811,17 @@ static void __update_extent_tree_range(struct inode *inode,
                et->largest_updated = false;
                updated = true;
        }
+       goto out_read_extent_cache;
+update_age_extent_cache:
+       if (!tei->last_blocks)
+               goto out_read_extent_cache;
 
+       __set_extent_info(&ei, fofs, len, 0, false,
+                       tei->age, tei->last_blocks, EX_BLOCK_AGE);
+       if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en))
+               __insert_extent_tree(sbi, et, &ei,
+                                       insert_p, insert_parent, leftmost);
+out_read_extent_cache:
        write_unlock(&et->lock);
 
        if (updated)
@@ -796,7 +859,7 @@ void f2fs_update_read_extent_tree_range_compressed(struct inode *inode,
        if (en)
                goto unlock_out;
 
-       __set_extent_info(&ei, fofs, llen, blkaddr, true, EX_READ);
+       __set_extent_info(&ei, fofs, llen, blkaddr, true, 0, 0, EX_READ);
        ei.c_len = c_len;
 
        if (!__try_merge_extent_node(sbi, et, &ei, prev_en, next_en))
@@ -807,6 +870,72 @@ unlock_out:
 }
 #endif
 
+static unsigned long long __calculate_block_age(unsigned long long new,
+                                               unsigned long long old)
+{
+       unsigned long long diff;
+
+       diff = (new >= old) ? new - (new - old) : new + (old - new);
+
+       return div_u64(diff * LAST_AGE_WEIGHT, 100);
+}
+
+/* This returns a new age and allocated blocks in ei */
+static int __get_new_block_age(struct inode *inode, struct extent_info *ei)
+{
+       struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
+       loff_t f_size = i_size_read(inode);
+       unsigned long long cur_blocks =
+                               atomic64_read(&sbi->allocated_data_blocks);
+
+       /*
+        * When I/O is not aligned to a PAGE_SIZE, update will happen to the last
+        * file block even in seq write. So don't record age for newly last file
+        * block here.
+        */
+       if ((f_size >> PAGE_SHIFT) == ei->fofs && f_size & (PAGE_SIZE - 1) &&
+                       ei->blk == NEW_ADDR)
+               return -EINVAL;
+
+       if (__lookup_extent_tree(inode, ei->fofs, ei, EX_BLOCK_AGE)) {
+               unsigned long long cur_age;
+
+               if (cur_blocks >= ei->last_blocks)
+                       cur_age = cur_blocks - ei->last_blocks;
+               else
+                       /* allocated_data_blocks overflow */
+                       cur_age = ULLONG_MAX - ei->last_blocks + cur_blocks;
+
+               if (ei->age)
+                       ei->age = __calculate_block_age(cur_age, ei->age);
+               else
+                       ei->age = cur_age;
+               ei->last_blocks = cur_blocks;
+               WARN_ON(ei->age > cur_blocks);
+               return 0;
+       }
+
+       f2fs_bug_on(sbi, ei->blk == NULL_ADDR);
+
+       /* the data block was allocated for the first time */
+       if (ei->blk == NEW_ADDR)
+               goto out;
+
+       if (__is_valid_data_blkaddr(ei->blk) &&
+                       !f2fs_is_valid_blkaddr(sbi, ei->blk, DATA_GENERIC_ENHANCE)) {
+               f2fs_bug_on(sbi, 1);
+               return -EINVAL;
+       }
+out:
+       /*
+        * init block age with zero, this can happen when the block age extent
+        * was reclaimed due to memory constraint or system reboot
+        */
+       ei->age = 0;
+       ei->last_blocks = cur_blocks;
+       return 0;
+}
+
 static void __update_extent_cache(struct dnode_of_data *dn, enum extent_type type)
 {
        struct extent_info ei;
@@ -823,6 +952,10 @@ static void __update_extent_cache(struct dnode_of_data *dn, enum extent_type typ
                        ei.blk = NULL_ADDR;
                else
                        ei.blk = dn->data_blkaddr;
+       } else if (type == EX_BLOCK_AGE) {
+               ei.blk = dn->data_blkaddr;
+               if (__get_new_block_age(dn->inode, &ei))
+                       return;
        }
        __update_extent_tree_range(dn->inode, &ei, type);
 }
@@ -940,6 +1073,43 @@ unsigned int f2fs_shrink_read_extent_tree(struct f2fs_sb_info *sbi, int nr_shrin
        return __shrink_extent_tree(sbi, nr_shrink, EX_READ);
 }
 
+/* block age extent cache operations */
+bool f2fs_lookup_age_extent_cache(struct inode *inode, pgoff_t pgofs,
+                               struct extent_info *ei)
+{
+       if (!__may_extent_tree(inode, EX_BLOCK_AGE))
+               return false;
+
+       return __lookup_extent_tree(inode, pgofs, ei, EX_BLOCK_AGE);
+}
+
+void f2fs_update_age_extent_cache(struct dnode_of_data *dn)
+{
+       return __update_extent_cache(dn, EX_BLOCK_AGE);
+}
+
+void f2fs_update_age_extent_cache_range(struct dnode_of_data *dn,
+                               pgoff_t fofs, unsigned int len)
+{
+       struct extent_info ei = {
+               .fofs = fofs,
+               .len = len,
+       };
+
+       if (!__may_extent_tree(dn->inode, EX_BLOCK_AGE))
+               return;
+
+       __update_extent_tree_range(dn->inode, &ei, EX_BLOCK_AGE);
+}
+
+unsigned int f2fs_shrink_age_extent_tree(struct f2fs_sb_info *sbi, int nr_shrink)
+{
+       if (!test_opt(sbi, AGE_EXTENT_CACHE))
+               return 0;
+
+       return __shrink_extent_tree(sbi, nr_shrink, EX_BLOCK_AGE);
+}
+
 static unsigned int __destroy_extent_node(struct inode *inode,
                                        enum extent_type type)
 {
@@ -960,6 +1130,7 @@ static unsigned int __destroy_extent_node(struct inode *inode,
 void f2fs_destroy_extent_node(struct inode *inode)
 {
        __destroy_extent_node(inode, EX_READ);
+       __destroy_extent_node(inode, EX_BLOCK_AGE);
 }
 
 static void __drop_extent_tree(struct inode *inode, enum extent_type type)
@@ -988,6 +1159,7 @@ static void __drop_extent_tree(struct inode *inode, enum extent_type type)
 void f2fs_drop_extent_tree(struct inode *inode)
 {
        __drop_extent_tree(inode, EX_READ);
+       __drop_extent_tree(inode, EX_BLOCK_AGE);
 }
 
 static void __destroy_extent_tree(struct inode *inode, enum extent_type type)
@@ -1028,6 +1200,7 @@ static void __destroy_extent_tree(struct inode *inode, enum extent_type type)
 void f2fs_destroy_extent_tree(struct inode *inode)
 {
        __destroy_extent_tree(inode, EX_READ);
+       __destroy_extent_tree(inode, EX_BLOCK_AGE);
 }
 
 static void __init_extent_tree_info(struct extent_tree_info *eti)
@@ -1045,6 +1218,12 @@ static void __init_extent_tree_info(struct extent_tree_info *eti)
 void f2fs_init_extent_cache_info(struct f2fs_sb_info *sbi)
 {
        __init_extent_tree_info(&sbi->extent_tree[EX_READ]);
+       __init_extent_tree_info(&sbi->extent_tree[EX_BLOCK_AGE]);
+
+       /* initialize for block age extents */
+       atomic64_set(&sbi->allocated_data_blocks, 0);
+       sbi->hot_data_age_threshold = DEF_HOT_DATA_AGE_THRESHOLD;
+       sbi->warm_data_age_threshold = DEF_WARM_DATA_AGE_THRESHOLD;
 }
 
 int __init f2fs_create_extent_cache(void)
index ec52e06..e8953c3 100644 (file)
@@ -107,6 +107,7 @@ extern const char *f2fs_fault_name[FAULT_MAX];
 #define F2FS_MOUNT_MERGE_CHECKPOINT    0x10000000
 #define        F2FS_MOUNT_GC_MERGE             0x20000000
 #define F2FS_MOUNT_COMPRESS_CACHE      0x40000000
+#define F2FS_MOUNT_AGE_EXTENT_CACHE    0x80000000
 
 #define F2FS_OPTION(sbi)       ((sbi)->mount_opt)
 #define clear_opt(sbi, option) (F2FS_OPTION(sbi).opt &= ~F2FS_MOUNT_##option)
@@ -607,9 +608,22 @@ enum {
 /* number of extent info in extent cache we try to shrink */
 #define READ_EXTENT_CACHE_SHRINK_NUMBER        128
 
+/* number of age extent info in extent cache we try to shrink */
+#define AGE_EXTENT_CACHE_SHRINK_NUMBER 128
+#define LAST_AGE_WEIGHT                        30
+#define SAME_AGE_REGION                        1024
+
+/*
+ * Define data block with age less than 1GB as hot data
+ * define data block with age less than 10GB but more than 1GB as warm data
+ */
+#define DEF_HOT_DATA_AGE_THRESHOLD     262144
+#define DEF_WARM_DATA_AGE_THRESHOLD    2621440
+
 /* extent cache type */
 enum extent_type {
        EX_READ,
+       EX_BLOCK_AGE,
        NR_EXTENT_CACHES,
 };
 
@@ -637,6 +651,13 @@ struct extent_info {
                        unsigned int c_len;
 #endif
                };
+               /* block age extent_cache */
+               struct {
+                       /* block age of the extent */
+                       unsigned long long age;
+                       /* last total blocks allocated */
+                       unsigned long long last_blocks;
+               };
        };
 };
 
@@ -1653,6 +1674,11 @@ struct f2fs_sb_info {
 
        /* for extent tree cache */
        struct extent_tree_info extent_tree[NR_EXTENT_CACHES];
+       atomic64_t allocated_data_blocks;       /* for block age extent_cache */
+
+       /* The threshold used for hot and warm data seperation*/
+       unsigned int hot_data_age_threshold;
+       unsigned int warm_data_age_threshold;
 
        /* basic filesystem units */
        unsigned int log_sectors_per_block;     /* log2 sectors per block */
@@ -3857,6 +3883,8 @@ struct f2fs_stat_info {
        unsigned long long ext_mem[NR_EXTENT_CACHES];
        /* for read extent cache */
        unsigned long long hit_largest;
+       /* for block age extent cache */
+       unsigned long long allocated_data_blocks;
        int ndirty_node, ndirty_dent, ndirty_meta, ndirty_imeta;
        int ndirty_data, ndirty_qdata;
        unsigned int ndirty_dirs, ndirty_files, nquota_files, ndirty_all;
@@ -4168,6 +4196,16 @@ void f2fs_update_read_extent_cache_range(struct dnode_of_data *dn,
 unsigned int f2fs_shrink_read_extent_tree(struct f2fs_sb_info *sbi,
                        int nr_shrink);
 
+/* block age extent cache ops */
+void f2fs_init_age_extent_tree(struct inode *inode);
+bool f2fs_lookup_age_extent_cache(struct inode *inode, pgoff_t pgofs,
+                       struct extent_info *ei);
+void f2fs_update_age_extent_cache(struct dnode_of_data *dn);
+void f2fs_update_age_extent_cache_range(struct dnode_of_data *dn,
+                       pgoff_t fofs, unsigned int len);
+unsigned int f2fs_shrink_age_extent_tree(struct f2fs_sb_info *sbi,
+                       int nr_shrink);
+
 /*
  * sysfs.c
  */
index cbe7c24..56c23b5 100644 (file)
@@ -619,6 +619,7 @@ void f2fs_truncate_data_blocks_range(struct dnode_of_data *dn, int count)
                fofs = f2fs_start_bidx_of_node(ofs_of_node(dn->node_page),
                                                        dn->inode) + ofs;
                f2fs_update_read_extent_cache_range(dn, fofs, 0, len);
+               f2fs_update_age_extent_cache_range(dn, fofs, nr_free);
                dec_valid_block_count(sbi, dn->inode, nr_free);
        }
        dn->ofs_in_node = ofs;
index c845c16..ff6cf66 100644 (file)
@@ -480,6 +480,7 @@ static int do_read_inode(struct inode *inode)
 
        /* Need all the flag bits */
        f2fs_init_read_extent_tree(inode, node_page);
+       f2fs_init_age_extent_tree(inode);
 
        f2fs_put_page(node_page, 1);
 
index 07419c3..dde4c04 100644 (file)
@@ -60,7 +60,7 @@ bool f2fs_available_free_memory(struct f2fs_sb_info *sbi, int type)
        avail_ram = val.totalram - val.totalhigh;
 
        /*
-        * give 25%, 25%, 50%, 50%, 50% memory for each components respectively
+        * give 25%, 25%, 50%, 50%, 25%, 25% memory for each components respectively
         */
        if (type == FREE_NIDS) {
                mem_size = (nm_i->nid_cnt[FREE_NID] *
@@ -85,14 +85,16 @@ bool f2fs_available_free_memory(struct f2fs_sb_info *sbi, int type)
                                                sizeof(struct ino_entry);
                mem_size >>= PAGE_SHIFT;
                res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 1);
-       } else if (type == READ_EXTENT_CACHE) {
-               struct extent_tree_info *eti = &sbi->extent_tree[EX_READ];
+       } else if (type == READ_EXTENT_CACHE || type == AGE_EXTENT_CACHE) {
+               enum extent_type etype = type == READ_EXTENT_CACHE ?
+                                               EX_READ : EX_BLOCK_AGE;
+               struct extent_tree_info *eti = &sbi->extent_tree[etype];
 
                mem_size = (atomic_read(&eti->total_ext_tree) *
                                sizeof(struct extent_tree) +
                                atomic_read(&eti->total_ext_node) *
                                sizeof(struct extent_node)) >> PAGE_SHIFT;
-               res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 1);
+               res = mem_size < ((avail_ram * nm_i->ram_thresh / 100) >> 2);
        } else if (type == DISCARD_CACHE) {
                mem_size = (atomic_read(&dcc->discard_cmd_cnt) *
                                sizeof(struct discard_cmd)) >> PAGE_SHIFT;
index 0aa4870..99454d4 100644 (file)
@@ -147,6 +147,7 @@ enum mem_type {
        DIRTY_DENTS,    /* indicates dirty dentry pages */
        INO_ENTRIES,    /* indicates inode entries */
        READ_EXTENT_CACHE,      /* indicates read extent cache */
+       AGE_EXTENT_CACHE,       /* indicates age extent cache */
        DISCARD_CACHE,  /* indicates memory of cached discard cmds */
        COMPRESS_PAGE,  /* indicates memory of cached compressed pages */
        BASE_CHECK,     /* check kernel status */
index 8722d1a..dee712f 100644 (file)
@@ -453,6 +453,11 @@ void f2fs_balance_fs_bg(struct f2fs_sb_info *sbi, bool from_bg)
                f2fs_shrink_read_extent_tree(sbi,
                                READ_EXTENT_CACHE_SHRINK_NUMBER);
 
+       /* try to shrink age extent cache when there is no enough memory */
+       if (!f2fs_available_free_memory(sbi, AGE_EXTENT_CACHE))
+               f2fs_shrink_age_extent_tree(sbi,
+                               AGE_EXTENT_CACHE_SHRINK_NUMBER);
+
        /* check the # of cached NAT entries */
        if (!f2fs_available_free_memory(sbi, NAT_ENTRIES))
                f2fs_try_to_free_nats(sbi, NAT_ENTRY_PER_BLOCK);
@@ -3151,10 +3156,28 @@ static int __get_segment_type_4(struct f2fs_io_info *fio)
        }
 }
 
+static int __get_age_segment_type(struct inode *inode, pgoff_t pgofs)
+{
+       struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
+       struct extent_info ei;
+
+       if (f2fs_lookup_age_extent_cache(inode, pgofs, &ei)) {
+               if (!ei.age)
+                       return NO_CHECK_TYPE;
+               if (ei.age <= sbi->hot_data_age_threshold)
+                       return CURSEG_HOT_DATA;
+               if (ei.age <= sbi->warm_data_age_threshold)
+                       return CURSEG_WARM_DATA;
+               return CURSEG_COLD_DATA;
+       }
+       return NO_CHECK_TYPE;
+}
+
 static int __get_segment_type_6(struct f2fs_io_info *fio)
 {
        if (fio->type == DATA) {
                struct inode *inode = fio->page->mapping->host;
+               int type;
 
                if (is_inode_flag_set(inode, FI_ALIGNED_WRITE))
                        return CURSEG_COLD_DATA_PINNED;
@@ -3169,6 +3192,11 @@ static int __get_segment_type_6(struct f2fs_io_info *fio)
                }
                if (file_is_cold(inode) || f2fs_need_compress_data(inode))
                        return CURSEG_COLD_DATA;
+
+               type = __get_age_segment_type(inode, fio->page->index);
+               if (type != NO_CHECK_TYPE)
+                       return type;
+
                if (file_is_hot(inode) ||
                                is_inode_flag_set(inode, FI_HOT_DATA) ||
                                f2fs_is_cow_file(inode))
@@ -3287,6 +3315,9 @@ void f2fs_allocate_data_block(struct f2fs_sb_info *sbi, struct page *page,
        locate_dirty_segment(sbi, GET_SEGNO(sbi, old_blkaddr));
        locate_dirty_segment(sbi, GET_SEGNO(sbi, *new_blkaddr));
 
+       if (IS_DATASEG(type))
+               atomic64_inc(&sbi->allocated_data_blocks);
+
        up_write(&sit_i->sentry_lock);
 
        if (page && IS_NODESEG(type)) {
@@ -3414,6 +3445,8 @@ void f2fs_outplace_write_data(struct dnode_of_data *dn,
        struct f2fs_summary sum;
 
        f2fs_bug_on(sbi, dn->data_blkaddr == NULL_ADDR);
+       if (fio->io_type == FS_DATA_IO || fio->io_type == FS_CP_DATA_IO)
+               f2fs_update_age_extent_cache(dn);
        set_summary(&sum, dn->nid, dn->ofs_in_node, fio->version);
        do_write_page(&sum, fio);
        f2fs_update_data_blkaddr(dn, fio->new_blkaddr);
index 33c490e..83d6fb9 100644 (file)
@@ -59,6 +59,9 @@ unsigned long f2fs_shrink_count(struct shrinker *shrink,
                /* count read extent cache entries */
                count += __count_extent_cache(sbi, EX_READ);
 
+               /* count block age extent cache entries */
+               count += __count_extent_cache(sbi, EX_BLOCK_AGE);
+
                /* count clean nat cache entries */
                count += __count_nat_entries(sbi);
 
@@ -102,8 +105,11 @@ unsigned long f2fs_shrink_scan(struct shrinker *shrink,
 
                sbi->shrinker_run_no = run_no;
 
+               /* shrink extent cache entries */
+               freed += f2fs_shrink_age_extent_tree(sbi, nr >> 2);
+
                /* shrink read extent cache entries */
-               freed += f2fs_shrink_read_extent_tree(sbi, nr >> 1);
+               freed += f2fs_shrink_read_extent_tree(sbi, nr >> 2);
 
                /* shrink clean nat cache entries */
                if (freed < nr)
@@ -134,6 +140,8 @@ void f2fs_join_shrinker(struct f2fs_sb_info *sbi)
 void f2fs_leave_shrinker(struct f2fs_sb_info *sbi)
 {
        f2fs_shrink_read_extent_tree(sbi, __count_extent_cache(sbi, EX_READ));
+       f2fs_shrink_age_extent_tree(sbi,
+                               __count_extent_cache(sbi, EX_BLOCK_AGE));
 
        spin_lock(&f2fs_list_lock);
        list_del_init(&sbi->s_list);
index 412c2e7..180d8b8 100644 (file)
@@ -163,6 +163,7 @@ enum {
        Opt_nogc_merge,
        Opt_discard_unit,
        Opt_memory_mode,
+       Opt_age_extent_cache,
        Opt_err,
 };
 
@@ -241,6 +242,7 @@ static match_table_t f2fs_tokens = {
        {Opt_nogc_merge, "nogc_merge"},
        {Opt_discard_unit, "discard_unit=%s"},
        {Opt_memory_mode, "memory=%s"},
+       {Opt_age_extent_cache, "age_extent_cache"},
        {Opt_err, NULL},
 };
 
@@ -1257,6 +1259,9 @@ static int parse_options(struct super_block *sb, char *options, bool is_remount)
                        }
                        kfree(name);
                        break;
+               case Opt_age_extent_cache:
+                       set_opt(sbi, AGE_EXTENT_CACHE);
+                       break;
                default:
                        f2fs_err(sbi, "Unrecognized mount option \"%s\" or missing value",
                                 p);
@@ -1958,6 +1963,8 @@ static int f2fs_show_options(struct seq_file *seq, struct dentry *root)
                seq_puts(seq, ",extent_cache");
        else
                seq_puts(seq, ",noextent_cache");
+       if (test_opt(sbi, AGE_EXTENT_CACHE))
+               seq_puts(seq, ",age_extent_cache");
        if (test_opt(sbi, DATA_FLUSH))
                seq_puts(seq, ",data_flush");
 
@@ -2219,6 +2226,7 @@ static int f2fs_remount(struct super_block *sb, int *flags, char *data)
        bool need_restart_flush = false, need_stop_flush = false;
        bool need_restart_discard = false, need_stop_discard = false;
        bool no_read_extent_cache = !test_opt(sbi, READ_EXTENT_CACHE);
+       bool no_age_extent_cache = !test_opt(sbi, AGE_EXTENT_CACHE);
        bool enable_checkpoint = !test_opt(sbi, DISABLE_CHECKPOINT);
        bool no_io_align = !F2FS_IO_ALIGNED(sbi);
        bool no_atgc = !test_opt(sbi, ATGC);
@@ -2313,6 +2321,12 @@ static int f2fs_remount(struct super_block *sb, int *flags, char *data)
                f2fs_warn(sbi, "switch extent_cache option is not allowed");
                goto restore_opts;
        }
+       /* disallow enable/disable age extent_cache dynamically */
+       if (no_age_extent_cache == !!test_opt(sbi, AGE_EXTENT_CACHE)) {
+               err = -EINVAL;
+               f2fs_warn(sbi, "switch age_extent_cache option is not allowed");
+               goto restore_opts;
+       }
 
        if (no_io_align == !!F2FS_IO_ALIGNED(sbi)) {
                err = -EINVAL;
index a4745d5..2ab2151 100644 (file)
@@ -668,6 +668,24 @@ out:
                return count;
        }
 
+       if (!strcmp(a->attr.name, "hot_data_age_threshold")) {
+               if (t == 0 || t >= sbi->warm_data_age_threshold)
+                       return -EINVAL;
+               if (t == *ui)
+                       return count;
+               *ui = (unsigned int)t;
+               return count;
+       }
+
+       if (!strcmp(a->attr.name, "warm_data_age_threshold")) {
+               if (t == 0 || t <= sbi->hot_data_age_threshold)
+                       return -EINVAL;
+               if (t == *ui)
+                       return count;
+               *ui = (unsigned int)t;
+               return count;
+       }
+
        *ui = (unsigned int)t;
 
        return count;
@@ -923,6 +941,10 @@ F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, peak_atomic_write, peak_atomic_write);
 F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, committed_atomic_block, committed_atomic_block);
 F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, revoked_atomic_block, revoked_atomic_block);
 
+/* For block age extent cache */
+F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, hot_data_age_threshold, hot_data_age_threshold);
+F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, warm_data_age_threshold, warm_data_age_threshold);
+
 #define ATTR_LIST(name) (&f2fs_attr_##name.attr)
 static struct attribute *f2fs_attrs[] = {
        ATTR_LIST(gc_urgent_sleep_time),
@@ -1018,6 +1040,8 @@ static struct attribute *f2fs_attrs[] = {
        ATTR_LIST(peak_atomic_write),
        ATTR_LIST(committed_atomic_block),
        ATTR_LIST(revoked_atomic_block),
+       ATTR_LIST(hot_data_age_threshold),
+       ATTR_LIST(warm_data_age_threshold),
        NULL,
 };
 ATTRIBUTE_GROUPS(f2fs);
index 2bb3789..31d994e 100644 (file)
@@ -49,6 +49,7 @@ TRACE_DEFINE_ENUM(CP_TRIMMED);
 TRACE_DEFINE_ENUM(CP_PAUSE);
 TRACE_DEFINE_ENUM(CP_RESIZE);
 TRACE_DEFINE_ENUM(EX_READ);
+TRACE_DEFINE_ENUM(EX_BLOCK_AGE);
 
 #define show_block_type(type)                                          \
        __print_symbolic(type,                                          \
@@ -155,6 +156,11 @@ TRACE_DEFINE_ENUM(EX_READ);
                { COMPRESS_ZSTD,        "ZSTD" },                       \
                { COMPRESS_LZORLE,      "LZO-RLE" })
 
+#define show_extent_type(type)                                         \
+       __print_symbolic(type,                                          \
+               { EX_READ,      "Read" },                               \
+               { EX_BLOCK_AGE, "Block Age" })
+
 struct f2fs_sb_info;
 struct f2fs_io_info;
 struct extent_info;
@@ -1544,7 +1550,7 @@ TRACE_EVENT(f2fs_lookup_extent_tree_start,
        TP_printk("dev = (%d,%d), ino = %lu, pgofs = %u, type = %s",
                show_dev_ino(__entry),
                __entry->pgofs,
-               __entry->type == EX_READ ? "Read" : "N/A")
+               show_extent_type(__entry->type))
 );
 
 TRACE_EVENT_CONDITION(f2fs_lookup_read_extent_tree_end,
@@ -1583,6 +1589,45 @@ TRACE_EVENT_CONDITION(f2fs_lookup_read_extent_tree_end,
                __entry->blk)
 );
 
+TRACE_EVENT_CONDITION(f2fs_lookup_age_extent_tree_end,
+
+       TP_PROTO(struct inode *inode, unsigned int pgofs,
+                                               struct extent_info *ei),
+
+       TP_ARGS(inode, pgofs, ei),
+
+       TP_CONDITION(ei),
+
+       TP_STRUCT__entry(
+               __field(dev_t,  dev)
+               __field(ino_t,  ino)
+               __field(unsigned int, pgofs)
+               __field(unsigned int, fofs)
+               __field(unsigned int, len)
+               __field(unsigned long long, age)
+               __field(unsigned long long, blocks)
+       ),
+
+       TP_fast_assign(
+               __entry->dev = inode->i_sb->s_dev;
+               __entry->ino = inode->i_ino;
+               __entry->pgofs = pgofs;
+               __entry->fofs = ei->fofs;
+               __entry->len = ei->len;
+               __entry->age = ei->age;
+               __entry->blocks = ei->last_blocks;
+       ),
+
+       TP_printk("dev = (%d,%d), ino = %lu, pgofs = %u, "
+               "age_ext_info(fofs: %u, len: %u, age: %llu, blocks: %llu)",
+               show_dev_ino(__entry),
+               __entry->pgofs,
+               __entry->fofs,
+               __entry->len,
+               __entry->age,
+               __entry->blocks)
+);
+
 TRACE_EVENT(f2fs_update_read_extent_tree_range,
 
        TP_PROTO(struct inode *inode, unsigned int pgofs, unsigned int len,
@@ -1618,6 +1663,41 @@ TRACE_EVENT(f2fs_update_read_extent_tree_range,
                __entry->c_len)
 );
 
+TRACE_EVENT(f2fs_update_age_extent_tree_range,
+
+       TP_PROTO(struct inode *inode, unsigned int pgofs, unsigned int len,
+                                       unsigned long long age,
+                                       unsigned long long last_blks),
+
+       TP_ARGS(inode, pgofs, len, age, last_blks),
+
+       TP_STRUCT__entry(
+               __field(dev_t,  dev)
+               __field(ino_t,  ino)
+               __field(unsigned int, pgofs)
+               __field(unsigned int, len)
+               __field(unsigned long long, age)
+               __field(unsigned long long, blocks)
+       ),
+
+       TP_fast_assign(
+               __entry->dev = inode->i_sb->s_dev;
+               __entry->ino = inode->i_ino;
+               __entry->pgofs = pgofs;
+               __entry->len = len;
+               __entry->age = age;
+               __entry->blocks = last_blks;
+       ),
+
+       TP_printk("dev = (%d,%d), ino = %lu, pgofs = %u, "
+                               "len = %u, age = %llu, blocks = %llu",
+               show_dev_ino(__entry),
+               __entry->pgofs,
+               __entry->len,
+               __entry->age,
+               __entry->blocks)
+);
+
 TRACE_EVENT(f2fs_shrink_extent_tree,
 
        TP_PROTO(struct f2fs_sb_info *sbi, unsigned int node_cnt,
@@ -1643,7 +1723,7 @@ TRACE_EVENT(f2fs_shrink_extent_tree,
                show_dev(__entry->dev),
                __entry->node_cnt,
                __entry->tree_cnt,
-               __entry->type == EX_READ ? "Read" : "N/A")
+               show_extent_type(__entry->type))
 );
 
 TRACE_EVENT(f2fs_destroy_extent_tree,
@@ -1670,7 +1750,7 @@ TRACE_EVENT(f2fs_destroy_extent_tree,
        TP_printk("dev = (%d,%d), ino = %lu, destroyed: node_cnt = %u, type = %s",
                show_dev_ino(__entry),
                __entry->node_cnt,
-               __entry->type == EX_READ ? "Read" : "N/A")
+               show_extent_type(__entry->type))
 );
 
 DECLARE_EVENT_CLASS(f2fs_sync_dirty_inodes,