f2fs: support passing down write hints given by users to block layer
authorHyunchul Lee <cheol.lee@lge.com>
Wed, 31 Jan 2018 02:36:57 +0000 (11:36 +0900)
committerJaegeuk Kim <jaegeuk@kernel.org>
Mon, 12 Mar 2018 23:05:35 +0000 (08:05 +0900)
Add the 'whint_mode' mount option that controls which write
hints are passed down to block layer. There are "off" and
"user-based" mode. The default mode is "off".

1) whint_mode=off. F2FS only passes down WRITE_LIFE_NOT_SET.

2) whint_mode=user-based. F2FS tries to pass down hints given
by users.

User                  F2FS                     Block
----                  ----                     -----
                      META                     WRITE_LIFE_NOT_SET
                      HOT_NODE                 "
                      WARM_NODE                "
                      COLD_NODE                "
ioctl(COLD)           COLD_DATA                WRITE_LIFE_EXTREME
extension list        "                        "

-- buffered io
WRITE_LIFE_EXTREME    COLD_DATA                WRITE_LIFE_EXTREME
WRITE_LIFE_SHORT      HOT_DATA                 WRITE_LIFE_SHORT
WRITE_LIFE_NOT_SET    WARM_DATA                WRITE_LIFE_NOT_SET
WRITE_LIFE_NONE       "                        "
WRITE_LIFE_MEDIUM     "                        "
WRITE_LIFE_LONG       "                        "

-- direct io
WRITE_LIFE_EXTREME    COLD_DATA                WRITE_LIFE_EXTREME
WRITE_LIFE_SHORT      HOT_DATA                 WRITE_LIFE_SHORT
WRITE_LIFE_NOT_SET    WARM_DATA                WRITE_LIFE_NOT_SET
WRITE_LIFE_NONE       "                        WRITE_LIFE_NONE
WRITE_LIFE_MEDIUM     "                        WRITE_LIFE_MEDIUM
WRITE_LIFE_LONG       "                        WRITE_LIFE_LONG

Many thanks to Chao Yu and Jaegeuk Kim for comments to
implement this patch.

Signed-off-by: Hyunchul Lee <cheol.lee@lge.com>
Reviewed-by: Chao Yu <yuchao0@huawei.com>
[Jaegeuk Kim: avoid build warning]
[Chao Yu: fix to restore whint_mode in ->remount_fs]
Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org>
fs/f2fs/data.c
fs/f2fs/f2fs.h
fs/f2fs/segment.c
fs/f2fs/super.c

index 7578ed1..0dae856 100644 (file)
@@ -175,15 +175,22 @@ static bool __same_bdev(struct f2fs_sb_info *sbi,
  */
 static struct bio *__bio_alloc(struct f2fs_sb_info *sbi, block_t blk_addr,
                                struct writeback_control *wbc,
-                               int npages, bool is_read)
+                               int npages, bool is_read,
+                               enum page_type type, enum temp_type temp)
 {
        struct bio *bio;
 
        bio = f2fs_bio_alloc(sbi, npages, true);
 
        f2fs_target_device(sbi, blk_addr, bio);
-       bio->bi_end_io = is_read ? f2fs_read_end_io : f2fs_write_end_io;
-       bio->bi_private = is_read ? NULL : sbi;
+       if (is_read) {
+               bio->bi_end_io = f2fs_read_end_io;
+               bio->bi_private = NULL;
+       } else {
+               bio->bi_end_io = f2fs_write_end_io;
+               bio->bi_private = sbi;
+               bio->bi_write_hint = io_type_to_rw_hint(sbi, type, temp);
+       }
        if (wbc)
                wbc_init_bio(wbc, bio);
 
@@ -382,7 +389,7 @@ int f2fs_submit_page_bio(struct f2fs_io_info *fio)
 
        /* Allocate a new bio */
        bio = __bio_alloc(fio->sbi, fio->new_blkaddr, fio->io_wbc,
-                               1, is_read_io(fio->op));
+                               1, is_read_io(fio->op), fio->type, fio->temp);
 
        if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) {
                bio_put(bio);
@@ -445,7 +452,8 @@ alloc_new:
                        goto out_fail;
                }
                io->bio = __bio_alloc(sbi, fio->new_blkaddr, fio->io_wbc,
-                                               BIO_MAX_PAGES, false);
+                                               BIO_MAX_PAGES, false,
+                                               fio->type, fio->temp);
                io->fio = *fio;
        }
 
@@ -2287,10 +2295,13 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 {
        struct address_space *mapping = iocb->ki_filp->f_mapping;
        struct inode *inode = mapping->host;
+       struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
        size_t count = iov_iter_count(iter);
        loff_t offset = iocb->ki_pos;
        int rw = iov_iter_rw(iter);
        int err;
+       enum rw_hint hint = iocb->ki_hint;
+       int whint_mode = sbi->whint_mode;
 
        err = check_direct_IO(inode, iter, offset);
        if (err)
@@ -2301,11 +2312,16 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
 
        trace_f2fs_direct_IO_enter(inode, offset, count, rw);
 
+       if (rw == WRITE && whint_mode == WHINT_MODE_OFF)
+               iocb->ki_hint = WRITE_LIFE_NOT_SET;
+
        down_read(&F2FS_I(inode)->dio_rwsem[rw]);
        err = blockdev_direct_IO(iocb, inode, iter, get_data_block_dio);
        up_read(&F2FS_I(inode)->dio_rwsem[rw]);
 
        if (rw == WRITE) {
+               if (whint_mode == WHINT_MODE_OFF)
+                       iocb->ki_hint = hint;
                if (err > 0) {
                        f2fs_update_iostat(F2FS_I_SB(inode), APP_DIRECT_IO,
                                                                        err);
index 3f4dafe..6c4236d 100644 (file)
@@ -1037,6 +1037,11 @@ enum {
        MAX_TIME,
 };
 
+enum {
+       WHINT_MODE_OFF,         /* not pass down write hints */
+       WHINT_MODE_USER,        /* try to pass down hints given by users */
+};
+
 struct f2fs_sb_info {
        struct super_block *sb;                 /* pointer to VFS super block */
        struct proc_dir_entry *s_proc;          /* proc entry */
@@ -1220,6 +1225,8 @@ struct f2fs_sb_info {
        char *s_qf_names[MAXQUOTAS];
        int s_jquota_fmt;                       /* Format of quota to use */
 #endif
+       /* For which write hints are passed down to block layer */
+       int whint_mode;
 };
 
 #ifdef CONFIG_F2FS_FAULT_INJECTION
@@ -2768,6 +2775,8 @@ void destroy_segment_manager(struct f2fs_sb_info *sbi);
 int __init create_segment_manager_caches(void);
 void destroy_segment_manager_caches(void);
 int rw_hint_to_seg_type(enum rw_hint hint);
+enum rw_hint io_type_to_rw_hint(struct f2fs_sb_info *sbi, enum page_type type,
+                               enum temp_type temp);
 
 /*
  * checkpoint.c
index 205b0d9..c14d5a1 100644 (file)
@@ -2456,6 +2456,62 @@ int rw_hint_to_seg_type(enum rw_hint hint)
        }
 }
 
+/* This returns write hints for each segment type. This hints will be
+ * passed down to block layer. There are mapping tables which depend on
+ * the mount option 'whint_mode'.
+ *
+ * 1) whint_mode=off. F2FS only passes down WRITE_LIFE_NOT_SET.
+ *
+ * 2) whint_mode=user-based. F2FS tries to pass down hints given by users.
+ *
+ * User                  F2FS                     Block
+ * ----                  ----                     -----
+ *                       META                     WRITE_LIFE_NOT_SET
+ *                       HOT_NODE                 "
+ *                       WARM_NODE                "
+ *                       COLD_NODE                "
+ * ioctl(COLD)           COLD_DATA                WRITE_LIFE_EXTREME
+ * extension list        "                        "
+ *
+ * -- buffered io
+ * WRITE_LIFE_EXTREME    COLD_DATA                WRITE_LIFE_EXTREME
+ * WRITE_LIFE_SHORT      HOT_DATA                 WRITE_LIFE_SHORT
+ * WRITE_LIFE_NOT_SET    WARM_DATA                WRITE_LIFE_NOT_SET
+ * WRITE_LIFE_NONE       "                        "
+ * WRITE_LIFE_MEDIUM     "                        "
+ * WRITE_LIFE_LONG       "                        "
+ *
+ * -- direct io
+ * WRITE_LIFE_EXTREME    COLD_DATA                WRITE_LIFE_EXTREME
+ * WRITE_LIFE_SHORT      HOT_DATA                 WRITE_LIFE_SHORT
+ * WRITE_LIFE_NOT_SET    WARM_DATA                WRITE_LIFE_NOT_SET
+ * WRITE_LIFE_NONE       "                        WRITE_LIFE_NONE
+ * WRITE_LIFE_MEDIUM     "                        WRITE_LIFE_MEDIUM
+ * WRITE_LIFE_LONG       "                        WRITE_LIFE_LONG
+ *
+ */
+
+enum rw_hint io_type_to_rw_hint(struct f2fs_sb_info *sbi,
+                               enum page_type type, enum temp_type temp)
+{
+       if (sbi->whint_mode == WHINT_MODE_USER) {
+               if (type == DATA) {
+                       switch (temp) {
+                       case COLD:
+                               return WRITE_LIFE_EXTREME;
+                       case HOT:
+                               return WRITE_LIFE_SHORT;
+                       default:
+                               return WRITE_LIFE_NOT_SET;
+                       }
+               } else {
+                       return WRITE_LIFE_NOT_SET;
+               }
+       } else {
+               return WRITE_LIFE_NOT_SET;
+       }
+}
+
 static int __get_segment_type_2(struct f2fs_io_info *fio)
 {
        if (fio->type == DATA)
@@ -2643,6 +2699,7 @@ void write_meta_page(struct f2fs_sb_info *sbi, struct page *page,
        struct f2fs_io_info fio = {
                .sbi = sbi,
                .type = META,
+               .temp = HOT,
                .op = REQ_OP_WRITE,
                .op_flags = REQ_SYNC | REQ_META | REQ_PRIO,
                .old_blkaddr = page->index,
@@ -2691,6 +2748,8 @@ int rewrite_data_page(struct f2fs_io_info *fio)
        int err;
 
        fio->new_blkaddr = fio->old_blkaddr;
+       /* i/o temperature is needed for passing down write hints */
+       __get_segment_type(fio);
        stat_inc_inplace_blocks(fio->sbi);
 
        err = f2fs_submit_page_bio(fio);
index 9587ca0..28b9d60 100644 (file)
@@ -129,6 +129,7 @@ enum {
        Opt_jqfmt_vfsold,
        Opt_jqfmt_vfsv0,
        Opt_jqfmt_vfsv1,
+       Opt_whint,
        Opt_err,
 };
 
@@ -182,6 +183,7 @@ static match_table_t f2fs_tokens = {
        {Opt_jqfmt_vfsold, "jqfmt=vfsold"},
        {Opt_jqfmt_vfsv0, "jqfmt=vfsv0"},
        {Opt_jqfmt_vfsv1, "jqfmt=vfsv1"},
+       {Opt_whint, "whint_mode=%s"},
        {Opt_err, NULL},
 };
 
@@ -679,6 +681,22 @@ static int parse_options(struct super_block *sb, char *options)
                                        "quota operations not supported");
                        break;
 #endif
+               case Opt_whint:
+                       name = match_strdup(&args[0]);
+                       if (!name)
+                               return -ENOMEM;
+                       if (strlen(name) == 10 &&
+                                       !strncmp(name, "user-based", 10)) {
+                               sbi->whint_mode = WHINT_MODE_USER;
+                       } else if (strlen(name) == 3 &&
+                                       !strncmp(name, "off", 3)) {
+                               sbi->whint_mode = WHINT_MODE_OFF;
+                       } else {
+                               kfree(name);
+                               return -EINVAL;
+                       }
+                       kfree(name);
+                       break;
                default:
                        f2fs_msg(sb, KERN_ERR,
                                "Unrecognized mount option \"%s\" or missing value",
@@ -722,6 +740,12 @@ static int parse_options(struct super_block *sb, char *options)
                        return -EINVAL;
                }
        }
+
+       /* Not pass down write hints if the number of active logs is lesser
+        * than NR_CURSEG_TYPE.
+        */
+       if (sbi->active_logs != NR_CURSEG_TYPE)
+               sbi->whint_mode = WHINT_MODE_OFF;
        return 0;
 }
 
@@ -1232,6 +1256,8 @@ static int f2fs_show_options(struct seq_file *seq, struct dentry *root)
                seq_puts(seq, ",prjquota");
 #endif
        f2fs_show_quota_options(seq, sbi->sb);
+       if (sbi->whint_mode == WHINT_MODE_USER)
+               seq_printf(seq, ",whint_mode=%s", "user-based");
 
        return 0;
 }
@@ -1241,6 +1267,7 @@ static void default_options(struct f2fs_sb_info *sbi)
        /* init some FS parameters */
        sbi->active_logs = NR_CURSEG_TYPE;
        sbi->inline_xattr_size = DEFAULT_INLINE_XATTR_ADDRS;
+       sbi->whint_mode = WHINT_MODE_OFF;
 
        set_opt(sbi, BG_GC);
        set_opt(sbi, INLINE_XATTR);
@@ -1281,6 +1308,7 @@ static int f2fs_remount(struct super_block *sb, int *flags, char *data)
        bool need_restart_gc = false;
        bool need_stop_gc = false;
        bool no_extent_cache = !test_opt(sbi, EXTENT_CACHE);
+       int old_whint_mode = sbi->whint_mode;
 #ifdef CONFIG_F2FS_FAULT_INJECTION
        struct f2fs_fault_info ffi = sbi->fault_info;
 #endif
@@ -1380,7 +1408,7 @@ static int f2fs_remount(struct super_block *sb, int *flags, char *data)
                need_stop_gc = true;
        }
 
-       if (*flags & SB_RDONLY) {
+       if (*flags & SB_RDONLY || sbi->whint_mode != old_whint_mode) {
                writeback_inodes_sb(sb, WB_REASON_SYNC);
                sync_inodes_sb(sb);
 
@@ -1430,6 +1458,7 @@ restore_opts:
                sbi->s_qf_names[i] = s_qf_names[i];
        }
 #endif
+       sbi->whint_mode = old_whint_mode;
        sbi->mount_opt = org_mount_opt;
        sbi->active_logs = active_logs;
        sb->s_flags = old_sb_flags;