Re: [f2fs-dev] [PATCH] f2fs: prevent meta updates while checkpoint is in progress

From: Jaegeuk Kim
Date: Mon Apr 06 2020 - 23:00:18 EST

On 04/07, Chao Yu wrote:
> On 2020/4/4 1:27, Jaegeuk Kim wrote:
> > On 04/03, Jaegeuk Kim wrote:
> >> On 04/01, Sahitya Tummala wrote:
> >>> Hi Jaegeuk,
> >>>
> >>> Got it.
> >>> The diff below looks good to me.
> >>> Would you like me to test it and put a patch for this?
> >>
> >> Sahitya, Chao,
> >>
> >> Could you please take a look at this patch and test intensively?
> >>
> >> Thanks,
> >
> > v2:
> >
> >>From 6bf7d5b227d466b0fe90d4957af29bd184fb646e Mon Sep 17 00:00:00 2001
> > From: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > Date: Tue, 31 Mar 2020 11:43:07 -0700
> > Subject: [PATCH] f2fs: refactor resize_fs to avoid meta updates in progress
> >
> > Sahitya raised an issue:
> > - prevent meta updates while checkpoint is in progress
> >
> > allocate_segment_for_resize() can cause metapage updates if
> > it requires to change the current node/data segments for resizing.
> > Stop these meta updates when there is a checkpoint already
> > in progress to prevent inconsistent CP data.
> >
> > Signed-off-by: Sahitya Tummala <stummala@xxxxxxxxxxxxxx>
> > Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
> > ---
> > fs/f2fs/checkpoint.c | 6 +-
> > fs/f2fs/f2fs.h | 2 +-
> > fs/f2fs/file.c | 5 +-
> > fs/f2fs/gc.c | 107 +++++++++++++++++++-----------------
> > fs/f2fs/super.c | 1 -
> > include/trace/events/f2fs.h | 4 +-
> > 6 files changed, 67 insertions(+), 58 deletions(-)
> >
> > diff --git a/fs/f2fs/checkpoint.c b/fs/f2fs/checkpoint.c
> > index 852890b72d6ac..531995192b714 100644
> > --- a/fs/f2fs/checkpoint.c
> > +++ b/fs/f2fs/checkpoint.c
> > @@ -1553,7 +1553,8 @@ int f2fs_write_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
> > return 0;
> > f2fs_warn(sbi, "Start checkpoint disabled!");
> > }
> > - mutex_lock(&sbi->cp_mutex);
> > + if (cpc->reason != CP_RESIZE)
> > + mutex_lock(&sbi->cp_mutex);
> >
> > if (!is_sbi_flag_set(sbi, SBI_IS_DIRTY) &&
> > ((cpc->reason & CP_FASTBOOT) || (cpc->reason & CP_SYNC) ||
> > @@ -1622,7 +1623,8 @@ int f2fs_write_checkpoint(struct f2fs_sb_info *sbi, struct cp_control *cpc)
> > f2fs_update_time(sbi, CP_TIME);
> > trace_f2fs_write_checkpoint(sbi->sb, cpc->reason, "finish checkpoint");
> > out:
> > - mutex_unlock(&sbi->cp_mutex);
> > + if (cpc->reason != CP_RESIZE)
> > + mutex_unlock(&sbi->cp_mutex);
> > return err;
> > }
> >
> > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > index be02a5cadd944..f9b2caa2135bd 100644
> > --- a/fs/f2fs/f2fs.h
> > +++ b/fs/f2fs/f2fs.h
> > @@ -193,6 +193,7 @@ enum {
> > #define CP_DISCARD 0x00000010
> > #define CP_TRIMMED 0x00000020
> > #define CP_PAUSE 0x00000040
> > +#define CP_RESIZE 0x00000080
> >
> > #define MAX_DISCARD_BLOCKS(sbi) BLKS_PER_SEC(sbi)
> > #define DEF_MAX_DISCARD_REQUEST 8 /* issue 8 discards per round */
> > @@ -1421,7 +1422,6 @@ struct f2fs_sb_info {
> > unsigned int segs_per_sec; /* segments per section */
> > unsigned int secs_per_zone; /* sections per zone */
> > unsigned int total_sections; /* total section count */
> > - struct mutex resize_mutex; /* for resize exclusion */
> > unsigned int total_node_count; /* total node block count */
> > unsigned int total_valid_node_count; /* valid node block count */
> > loff_t max_file_blocks; /* max block index of file */
> > diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
> > index 257e61d0afffb..b4c12370bb3d6 100644
> > --- a/fs/f2fs/file.c
> > +++ b/fs/f2fs/file.c
> > @@ -3305,7 +3305,6 @@ static int f2fs_ioc_resize_fs(struct file *filp, unsigned long arg)
> > {
> > struct f2fs_sb_info *sbi = F2FS_I_SB(file_inode(filp));
> > __u64 block_count;
> > - int ret;
> >
> > if (!capable(CAP_SYS_ADMIN))
> > return -EPERM;
> > @@ -3317,9 +3316,7 @@ static int f2fs_ioc_resize_fs(struct file *filp, unsigned long arg)
> > sizeof(block_count)))
> > return -EFAULT;
> >
> > - ret = f2fs_resize_fs(sbi, block_count);
> > -
> > - return ret;
> > + return f2fs_resize_fs(sbi, block_count);
> > }
> >
> > static int f2fs_ioc_enable_verity(struct file *filp, unsigned long arg)
> > diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> > index 26248c8936db0..46c75ecb64a2e 100644
> > --- a/fs/f2fs/gc.c
> > +++ b/fs/f2fs/gc.c
> > @@ -1399,12 +1399,29 @@ void f2fs_build_gc_manager(struct f2fs_sb_info *sbi)
> > GET_SEGNO(sbi, FDEV(0).end_blk) + 1;
> > }
> >
> > -static int free_segment_range(struct f2fs_sb_info *sbi, unsigned int start,
> > - unsigned int end)
> > +static int free_segment_range(struct f2fs_sb_info *sbi,
> > + unsigned int secs, bool gc_only)
> > {
> > - int type;
> > - unsigned int segno, next_inuse;
> > + unsigned int segno, next_inuse, start, end;
> > + struct cp_control cpc = { CP_RESIZE, 0, 0, 0 };
> > + int gc_mode, gc_type;
> > int err = 0;
> > + int type;
> > +
> > + /* Force block allocation for GC */
> > + MAIN_SECS(sbi) -= secs;
> > + start = MAIN_SECS(sbi) * sbi->segs_per_sec;
> > + end = MAIN_SEGS(sbi) - 1;
> > +
> > + mutex_lock(&DIRTY_I(sbi)->seglist_lock);
> > + for (gc_mode = 0; gc_mode < MAX_GC_POLICY; gc_mode++)
> > + if (SIT_I(sbi)->last_victim[gc_mode] >= start)
> > + SIT_I(sbi)->last_victim[gc_mode] = 0;
> > +
> > + for (gc_type = BG_GC; gc_type <= FG_GC; gc_type++)
> > + if (sbi->next_victim_seg[gc_type] >= start)
> > + sbi->next_victim_seg[gc_type] = NULL_SEGNO;
> > + mutex_unlock(&DIRTY_I(sbi)->seglist_lock);
> >
> > /* Move out cursegs from the target range */
> > for (type = CURSEG_HOT_DATA; type < NR_CURSEG_TYPE; type++)
> > @@ -1417,18 +1434,20 @@ static int free_segment_range(struct f2fs_sb_info *sbi, unsigned int start,
> > .iroot = RADIX_TREE_INIT(gc_list.iroot, GFP_NOFS),
> > };
> >
> > - down_write(&sbi->gc_lock);
> > do_garbage_collect(sbi, segno, &gc_list, FG_GC);
> > - up_write(&sbi->gc_lock);
> > put_gc_inode(&gc_list);
> >
> > - if (get_valid_blocks(sbi, segno, true))
> > - return -EAGAIN;
> > + if (!gc_only && get_valid_blocks(sbi, segno, true)) {
> > + err = -EAGAIN;
> > + goto out;
> > + }
> > }
> > + if (gc_only)
> > + goto out;
> >
> > - err = f2fs_sync_fs(sbi->sb, 1);
> > + err = f2fs_write_checkpoint(sbi, &cpc);
> > if (err)
> > - return err;
> > + goto out;
> >
> > next_inuse = find_next_inuse(FREE_I(sbi), end + 1, start);
> > if (next_inuse <= end) {
> > @@ -1436,6 +1455,8 @@ static int free_segment_range(struct f2fs_sb_info *sbi, unsigned int start,
> > next_inuse);
> > f2fs_bug_on(sbi, 1);
> > }
> > +out:
> > + MAIN_SECS(sbi) -= secs;
> > return err;
> > }
> >
> > @@ -1481,6 +1502,7 @@ static void update_fs_metadata(struct f2fs_sb_info *sbi, int secs)
> >
> > SM_I(sbi)->segment_count = (int)SM_I(sbi)->segment_count + segs;
> > MAIN_SEGS(sbi) = (int)MAIN_SEGS(sbi) + segs;
> > + MAIN_SECS(sbi) += secs;
> > FREE_I(sbi)->free_sections = (int)FREE_I(sbi)->free_sections + secs;
> > FREE_I(sbi)->free_segments = (int)FREE_I(sbi)->free_segments + segs;
> > F2FS_CKPT(sbi)->user_block_count = cpu_to_le64(user_block_count + blks);
> > @@ -1502,8 +1524,8 @@ static void update_fs_metadata(struct f2fs_sb_info *sbi, int secs)
> > int f2fs_resize_fs(struct f2fs_sb_info *sbi, __u64 block_count)
> > {
> > __u64 old_block_count, shrunk_blocks;
> > + struct cp_control cpc = { CP_RESIZE, 0, 0, 0 };
> > unsigned int secs;
> > - int gc_mode, gc_type;
> > int err = 0;
> > __u32 rem;
> >
> > @@ -1538,10 +1560,22 @@ int f2fs_resize_fs(struct f2fs_sb_info *sbi, __u64 block_count)
> > return -EINVAL;
> > }
> >
> > - freeze_bdev(sbi->sb->s_bdev);
> > -
> > shrunk_blocks = old_block_count - block_count;
> > secs = div_u64(shrunk_blocks, BLKS_PER_SEC(sbi));
> > +
> > + /* protect MAIN_SEC in free_segment_range */
> > + f2fs_lock_op(sbi);
> > + err = free_segment_range(sbi, secs, true);
>
> For this path, we break the rule that we need hold gc_lock during
> do_garbage_collect().

I don't get the point.
In free_segment_range(), gc_lock is held before/after do_garbage_collect().

>
> One other concern is that still the granularity of lock_op is too large,
> to avoid potential hang if it triggers heavy gc migration, how about using
> a timeout mechanism in free_segment_range() like we did in
> f2fs_disable_checkpoint()?

We can do first round GC without f2fs_lock_op().

>
> > + f2fs_unlock_op(sbi);
> > + if (err)
> > + return err;
> > +
> > + set_sbi_flag(sbi, SBI_IS_RESIZEFS);
> > +
> > + freeze_super(sbi->sb);
> > + down_write(&sbi->gc_lock);
> > + mutex_lock(&sbi->cp_mutex);
> > +
> > spin_lock(&sbi->stat_lock);
> > if (shrunk_blocks + valid_user_blocks(sbi) +
> > sbi->current_reserved_blocks + sbi->unusable_block_count +
> > @@ -1550,69 +1584,44 @@ int f2fs_resize_fs(struct f2fs_sb_info *sbi, __u64 block_count)
> > else
> > sbi->user_block_count -= shrunk_blocks;
> > spin_unlock(&sbi->stat_lock);
> > - if (err) {
> > - thaw_bdev(sbi->sb->s_bdev, sbi->sb);
> > - return err;
> > - }
> > -
> > - mutex_lock(&sbi->resize_mutex);
> > - set_sbi_flag(sbi, SBI_IS_RESIZEFS);
> > -
> > - mutex_lock(&DIRTY_I(sbi)->seglist_lock);
> > -
> > - MAIN_SECS(sbi) -= secs;
> > -
> > - for (gc_mode = 0; gc_mode < MAX_GC_POLICY; gc_mode++)
> > - if (SIT_I(sbi)->last_victim[gc_mode] >=
> > - MAIN_SECS(sbi) * sbi->segs_per_sec)
> > - SIT_I(sbi)->last_victim[gc_mode] = 0;
> > -
> > - for (gc_type = BG_GC; gc_type <= FG_GC; gc_type++)
> > - if (sbi->next_victim_seg[gc_type] >=
> > - MAIN_SECS(sbi) * sbi->segs_per_sec)
> > - sbi->next_victim_seg[gc_type] = NULL_SEGNO;
> > -
> > - mutex_unlock(&DIRTY_I(sbi)->seglist_lock);
> > + if (err)
> > + goto out_err;
> >
> > - err = free_segment_range(sbi, MAIN_SECS(sbi) * sbi->segs_per_sec,
> > - MAIN_SEGS(sbi) - 1);
> > + err = free_segment_range(sbi, secs, false);
>
> Lock coverage is still large here, what about just checking the resize condition
> with find_next_inuse(, end + 1, start), if the migration finished, then let's call
> write_checkpoint(), otherwise, returning -EAGAIN.

We did GC above, so how much time do you expect to complete here? Basically
I hesitate to return EAGAIN, since there's no context with respect to how
many times user need to retry to get succeess. Disabling checkpoint has
some ways to get a sense tho. Nevertheless, if we want to return EAGAIN,
it'd be better to give # of bytes that it requires to migrate further?

>
> > if (err)
> > - goto out;
> > + goto recover_out;
> >
> > update_sb_metadata(sbi, -secs);
> >
> > err = f2fs_commit_super(sbi, false);
> > if (err) {
> > update_sb_metadata(sbi, secs);
> > - goto out;
> > + goto recover_out;
> > }
> >
> > - mutex_lock(&sbi->cp_mutex);
> > update_fs_metadata(sbi, -secs);
> > clear_sbi_flag(sbi, SBI_IS_RESIZEFS);
> > set_sbi_flag(sbi, SBI_IS_DIRTY);
> > - mutex_unlock(&sbi->cp_mutex);
> >
> > - err = f2fs_sync_fs(sbi->sb, 1);
> > + err = f2fs_write_checkpoint(sbi, &cpc);
> > if (err) {
> > - mutex_lock(&sbi->cp_mutex);
> > update_fs_metadata(sbi, secs);
> > - mutex_unlock(&sbi->cp_mutex);
> > update_sb_metadata(sbi, secs);
> > f2fs_commit_super(sbi, false);
> > }
> > -out:
> > +recover_out:
> > if (err) {
> > set_sbi_flag(sbi, SBI_NEED_FSCK);
> > f2fs_err(sbi, "resize_fs failed, should run fsck to repair!");
> >
> > - MAIN_SECS(sbi) += secs;
> > spin_lock(&sbi->stat_lock);
> > sbi->user_block_count += shrunk_blocks;
> > spin_unlock(&sbi->stat_lock);
> > }
> > +out_err:
> > + mutex_unlock(&sbi->cp_mutex);
> > + up_write(&sbi->gc_lock);
> > + thaw_super(sbi->sb);
> > clear_sbi_flag(sbi, SBI_IS_RESIZEFS);
> > - mutex_unlock(&sbi->resize_mutex);
> > - thaw_bdev(sbi->sb->s_bdev, sbi->sb);
> > return err;
> > }
> > diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> > index b83b17b54a0a6..1e7b1d21d0177 100644
> > --- a/fs/f2fs/super.c
> > +++ b/fs/f2fs/super.c
> > @@ -3412,7 +3412,6 @@ static int f2fs_fill_super(struct super_block *sb, void *data, int silent)
> > init_rwsem(&sbi->gc_lock);
> > mutex_init(&sbi->writepages);
> > mutex_init(&sbi->cp_mutex);
> > - mutex_init(&sbi->resize_mutex);
> > init_rwsem(&sbi->node_write);
> > init_rwsem(&sbi->node_change);
> >
> > diff --git a/include/trace/events/f2fs.h b/include/trace/events/f2fs.h
> > index 4d7d4c391879d..5d1a72001fdb4 100644
> > --- a/include/trace/events/f2fs.h
> > +++ b/include/trace/events/f2fs.h
> > @@ -50,6 +50,7 @@ TRACE_DEFINE_ENUM(CP_RECOVERY);
> > TRACE_DEFINE_ENUM(CP_DISCARD);
> > TRACE_DEFINE_ENUM(CP_TRIMMED);
> > TRACE_DEFINE_ENUM(CP_PAUSE);
> > +TRACE_DEFINE_ENUM(CP_RESIZE);
> >
> > #define show_block_type(type) \
> > __print_symbolic(type, \
> > @@ -126,7 +127,8 @@ TRACE_DEFINE_ENUM(CP_PAUSE);
> > { CP_RECOVERY, "Recovery" }, \
> > { CP_DISCARD, "Discard" }, \
> > { CP_PAUSE, "Pause" }, \
> > - { CP_TRIMMED, "Trimmed" })
> > + { CP_TRIMMED, "Trimmed" }, \
> > + { CP_RESIZE, "Resize" })
> >
> > #define show_fsync_cpreason(type) \
> > __print_symbolic(type, \
> >