Messages in this thread | | | Date | Tue, 8 Aug 2023 08:50:18 +0800 | Subject | Re: [f2fs-dev] [PATCH v2] f2fs: do not issue small discard commands during checkpoint | From | Chao Yu <> |
| |
On 2023/8/8 3:55, Jaegeuk Kim wrote: > On 08/07, Chao Yu wrote: >> On 2023/8/5 4:52, Jaegeuk Kim wrote: >>> On 07/25, Chao Yu wrote: >>>> On 2023/7/22 4:23, Jaegeuk Kim wrote: >>>>> On 07/13, Chao Yu wrote: >>>>>> On 2023/7/12 23:55, Jaegeuk Kim wrote: >>>>>>> On 07/12, Chao Yu wrote: >>>>>>>> On 2023/7/12 0:37, Jaegeuk Kim wrote: >>>>>>>>> On 07/06, Chao Yu wrote: >>>>>>>>>> On 2023/7/6 1:30, Jaegeuk Kim wrote: >>>>>>>>>>> On 07/04, Chao Yu wrote: >>>>>>>>>>>> On 2023/7/4 18:53, Jaegeuk Kim wrote: >>>>>>>>>>>>> On 07/03, Chao Yu wrote: >>>>>>>>>>>>>> On 2023/6/15 0:10, Jaegeuk Kim wrote: >>>>>>>>>>>>>>> If there're huge # of small discards, this will increase checkpoint latency >>>>>>>>>>>>>>> insanely. Let's issue small discards only by trim. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Signed-off-by: Jaegeuk Kim <jaegeuk@kernel.org> >>>>>>>>>>>>>>> --- >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Change log from v1: >>>>>>>>>>>>>>> - move the skip logic to avoid dangling objects >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> fs/f2fs/segment.c | 2 +- >>>>>>>>>>>>>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c >>>>>>>>>>>>>>> index 8c7af8b4fc47..0457d620011f 100644 >>>>>>>>>>>>>>> --- a/fs/f2fs/segment.c >>>>>>>>>>>>>>> +++ b/fs/f2fs/segment.c >>>>>>>>>>>>>>> @@ -2193,7 +2193,7 @@ void f2fs_clear_prefree_segments(struct f2fs_sb_info *sbi, >>>>>>>>>>>>>>> len = next_pos - cur_pos; >>>>>>>>>>>>>>> if (f2fs_sb_has_blkzoned(sbi) || >>>>>>>>>>>>>>> - (force && len < cpc->trim_minlen)) >>>>>>>>>>>>>>> + !force || len < cpc->trim_minlen) >>>>>>>>>>>>>>> goto skip; >>>>>>>>>>>>>> >>>>>>>>>>>>>> Sorry for late reply. >>>>>>>>>>>>>> >>>>>>>>>>>>>> We have a configuration for such case, what do you think of setting >>>>>>>>>>>>>> max_small_discards to zero? otherwise, w/ above change, max_small_discards >>>>>>>>>>>>>> logic may be broken? >>>>>>>>>>>>>> >>>>>>>>>>>>>> What: /sys/fs/f2fs/<disk>/max_small_discards >>>>>>>>>>>>>> Date: November 2013 >>>>>>>>>>>>>> Contact: "Jaegeuk Kim" <jaegeuk.kim@samsung.com> >>>>>>>>>>>>>> Description: Controls the issue rate of discard commands that consist of small >>>>>>>>>>>>>> blocks less than 2MB. The candidates to be discarded are cached until >>>>>>>>>>>>>> checkpoint is triggered, and issued during the checkpoint. >>>>>>>>>>>>>> By default, it is disabled with 0. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Or, if we prefer to disable small_discards by default, what about below change: >>>>>>>>>>>>> >>>>>>>>>>>>> I think small_discards is fine, but need to avoid long checkpoint latency only. >>>>>>>>>>>> >>>>>>>>>>>> I didn't get you, do you mean we can still issue small discard by >>>>>>>>>>>> fstrim, so small_discards functionality is fine? >>>>>>>>>>> >>>>>>>>>>> You got the point. >>>>>>>>>> >>>>>>>>>> Well, actually, what I mean is max_small_discards sysfs entry's functionality >>>>>>>>>> is broken. Now, the entry can not be used to control number of small discards >>>>>>>>>> committed by checkpoint. >>>>>>>>> >>>>>>>>> Could you descrbie this problem first? >>>>>>>> >>>>>>>> Oh, alright, actually, I've described this problem literally, but maybe it's not >>>>>>>> clear, let me give some examples as below: >>>>>>>> >>>>>>>> echo 0 > /sys/fs/f2fs/vdb/max_small_discards >>>>>>>> xfs_io -f /mnt/f2fs/file -c "pwrite 0 2m" -c "fsync" >>>>>>>> xfs_io /mnt/f2fs/file -c "fpunch 0 4k" >>>>>>>> sync >>>>>>>> cat /proc/fs/f2fs/vdb/discard_plist_info |head -2 >>>>>>>> >>>>>>>> echo 100 > /sys/fs/f2fs/vdb/max_small_discards >>>>>>>> rm /mnt/f2fs/file >>>>>>>> xfs_io -f /mnt/f2fs/file -c "pwrite 0 2m" -c "fsync" >>>>>>>> xfs_io /mnt/f2fs/file -c "fpunch 0 4k" >>>>>>>> sync >>>>>>>> cat /proc/fs/f2fs/vdb/discard_plist_info |head -2 >>>>>>>> >>>>>>>> Before the patch: >>>>>>>> >>>>>>>> Discard pend list(Show diacrd_cmd count on each entry, .:not exist): >>>>>>>> 0 . . . . . . . . >>>>>>>> >>>>>>>> Discard pend list(Show diacrd_cmd count on each entry, .:not exist): >>>>>>>> 0 3 1 . . . . . . >>>>>>>> >>>>>>>> After the patch: >>>>>>>> Discard pend list(Show diacrd_cmd count on each entry, .:not exist): >>>>>>>> 0 . . . . . . . . >>>>>>>> >>>>>>>> Discard pend list(Show diacrd_cmd count on each entry, .:not exist): >>>>>>>> 0 . . . . . . . . >>>>>>>> >>>>>>>> So, now max_small_discards can not be used to control small discard number >>>>>>>> cached by checkpoint. >>>>>> >>>>>> Let me explain more: >>>>>> >>>>>> Previously, we have two mechanisms to cache & submit small discards: >>>>>> >>>>>> a) set max small discard number in /sys/fs/f2fs/vdb/max_small_discards, and checkpoint >>>>>> will cache small discard candidates w/ configured maximum number. >>>>>> >>>>>> b) call FITRIM ioctl, also, checkpoint in f2fs_trim_fs() will cache small discard >>>>>> candidates w/ configured discard granularity, but w/o limitation of number. FSTRIM >>>>>> interface is asynchronized, so it won't submit discard directly. >>>>>> >>>>>> Finally, discard thread will submit them in background periodically. >>>>>> >>>>>> So what I mean is the mechanism a) is broken, since no matter how we configure the >>>>>> sysfs entry /sys/fs/f2fs/vdb/max_small_discards, checkpoint will not cache small >>>>>> discard candidates any more. >>>>> >>>>> Ok, it seems what I encountered before was adding this small discard even >>>>> after issuing it by checkpoint. Thoughts? >>>> >>>> Do you mean: in f2fs_clear_prefree_segments(), small discard may overlap >>>> segment granularity discard? >>> >>> I didn't dig enough tho, don't think so. Somehow I got a loop as below which >>> said the same LBA was issued and added back repeatedly, not seen this short log >>> unfortunately. >>> >>>> >>>> e.g. >>>> - f2fs_clear_prefree_segments >>>> - f2fs_issue_discard(0, 512) --- segment granularity discard >>>> - f2fs_issue_discard(0, 1) --- small discard >>>> - f2fs_issue_discard(5, 1) --- small discard >>>> >>>> Thanks, >> >> [snip] >> >>>>> f2fs_discard-25-752 [003] ..... 9744.173111: f2fs_issue_discard: dev = (254,51), blkstart = 0x18c0ca, blklen = 0x1 >>>>> f2fs_discard-25-752 [004] ..... 9744.175348: f2fs_remove_discard: dev = (254,51), blkstart = 0x18c0ca, blklen = 0x1 >> >> I don't see any loop via above log, am I missing something? >> >> The traces were printed in below call paths, and it printed the same LBAs as >> expected? > > I said the log was cut to show the loop enoughly. I'll try to reproduce this
Oh, I misunderstood it...
> later locally.
Alright.
Thanks,
> >> >> - issue_discard_thread >> - __issue_discard_cmd >> - __submit_discard_cmd >> - trace_f2fs_issue_discard >> 9744.173111: f2fs_issue_discard: dev = (254,51), blkstart = 0x18c0ca, blklen = 0x1 >> - __wait_all_discard_cmd >> - __wait_discard_cmd_range >> - __remove_discard_cmd >> - trace_f2fs_remove_discard >> 9744.175348: f2fs_remove_discard: dev = (254,51), blkstart = 0x18c0ca, blklen = 0x1 >> >> Thanks, >> >>>>> >>>>>> >>>>>> So, it needs to fix max_small_discards sysfs functionality? or just drop the >>>>>> functionality? >>>>>> >>>>>>> >>>>>>> Since we do not submit small discards anymore during checkpoint. Why not relying >>>>>>> on the discard thread to issue them? >>>>>> >>>>>> Sorry, I'm not sure I get your point, do you mean max_small_discards functionality >>>>>> is obsoleted, so it recommended to use fstrim to cache & submit small discards? >>>>>> >>>>>> Let me know, if I'm missing something or misunderstanding the point. >>>>>> >>>>>> Thanks, >>>>>> >>>>>>> >>>>>>>> >>>>>>>> Thanks, >>>>>>>> >>>>>>>>> >>>>>>>>>> >>>>>>>>>> I think there is another way to achieve "avoid long checkpoint latency caused >>>>>>>>>> by committing huge # of small discards", the way is we can set max_small_discards >>>>>>>>>> to small value or zero, w/ such configuration, it will take checkpoint much less >>>>>>>>>> time or no time to committing small discard due to below control logic: >>>>>>>>>> >>>>>>>>>> f2fs_flush_sit_entries() >>>>>>>>>> { >>>>>>>>>> ... >>>>>>>>>> if (!(cpc->reason & CP_DISCARD)) { >>>>>>>>>> cpc->trim_start = segno; >>>>>>>>>> add_discard_addrs(sbi, cpc, false); >>>>>>>>>> } >>>>>>>>>> ... >>>>>>>>>> } >>>>>>>>>> >>>>>>>>>> add_discard_addrs() >>>>>>>>>> { >>>>>>>>>> ... >>>>>>>>>> while (force || SM_I(sbi)->dcc_info->nr_discards <= >>>>>>>>>> SM_I(sbi)->dcc_info->max_discards) { >>>>>>>>>> >>>>>>>>>> It will break the loop once nr_discards is larger than max_discards, if >>>>>>>>>> max_discards is set to zero, checkpoint won't take time to handle small discards. >>>>>>>>>> >>>>>>>>>> ... >>>>>>>>>> if (!de) { >>>>>>>>>> de = f2fs_kmem_cache_alloc(discard_entry_slab, >>>>>>>>>> GFP_F2FS_ZERO, true, NULL); >>>>>>>>>> de->start_blkaddr = START_BLOCK(sbi, cpc->trim_start); >>>>>>>>>> list_add_tail(&de->list, head); >>>>>>>>>> } >>>>>>>>>> ... >>>>>>>>>> } >>>>>>>>>> ... >>>>>>>>>> >>>>>>>>>> Thanks, >>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> Thanks, >>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>> From eb89d9b56e817e3046d7fa17165b12416f09d456 Mon Sep 17 00:00:00 2001 >>>>>>>>>>>>>> From: Chao Yu <chao@kernel.org> >>>>>>>>>>>>>> Date: Mon, 3 Jul 2023 09:06:53 +0800 >>>>>>>>>>>>>> Subject: [PATCH] Revert "f2fs: enable small discard by default" >>>>>>>>>>>>>> >>>>>>>>>>>>>> This reverts commit d618ebaf0aa83d175658aea5291e0c459d471d39 in order >>>>>>>>>>>>>> to disable small discard by default, so that if there're huge number of >>>>>>>>>>>>>> small discards, it will decrease checkpoint's latency obviously. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Also, this patch reverts 9ac00e7cef10 ("f2fs: do not issue small discard >>>>>>>>>>>>>> commands during checkpoint"), due to it breaks small discard feature which >>>>>>>>>>>>>> may be configured via sysfs entry max_small_discards. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Fixes: 9ac00e7cef10 ("f2fs: do not issue small discard commands during checkpoint") >>>>>>>>>>>>>> Signed-off-by: Chao Yu <chao@kernel.org> >>>>>>>>>>>>>> --- >>>>>>>>>>>>>> fs/f2fs/segment.c | 4 ++-- >>>>>>>>>>>>>> 1 file changed, 2 insertions(+), 2 deletions(-) >>>>>>>>>>>>>> >>>>>>>>>>>>>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c >>>>>>>>>>>>>> index 14c822e5c9c9..0a313368f18b 100644 >>>>>>>>>>>>>> --- a/fs/f2fs/segment.c >>>>>>>>>>>>>> +++ b/fs/f2fs/segment.c >>>>>>>>>>>>>> @@ -2193,7 +2193,7 @@ void f2fs_clear_prefree_segments(struct f2fs_sb_info *sbi, >>>>>>>>>>>>>> len = next_pos - cur_pos; >>>>>>>>>>>>>> >>>>>>>>>>>>>> if (f2fs_sb_has_blkzoned(sbi) || >>>>>>>>>>>>>> - !force || len < cpc->trim_minlen) >>>>>>>>>>>>>> + (force && len < cpc->trim_minlen)) >>>>>>>>>>>>>> goto skip; >>>>>>>>>>>>>> >>>>>>>>>>>>>> f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos, >>>>>>>>>>>>>> @@ -2269,7 +2269,7 @@ static int create_discard_cmd_control(struct f2fs_sb_info *sbi) >>>>>>>>>>>>>> atomic_set(&dcc->queued_discard, 0); >>>>>>>>>>>>>> atomic_set(&dcc->discard_cmd_cnt, 0); >>>>>>>>>>>>>> dcc->nr_discards = 0; >>>>>>>>>>>>>> - dcc->max_discards = MAIN_SEGS(sbi) << sbi->log_blocks_per_seg; >>>>>>>>>>>>>> + dcc->max_discards = 0; >>>>>>>>>>>>>> dcc->max_discard_request = DEF_MAX_DISCARD_REQUEST; >>>>>>>>>>>>>> dcc->min_discard_issue_time = DEF_MIN_DISCARD_ISSUE_TIME; >>>>>>>>>>>>>> dcc->mid_discard_issue_time = DEF_MID_DISCARD_ISSUE_TIME; >>>>>>>>>>>>>> -- >>>>>>>>>>>>>> 2.40.1 >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>>> f2fs_issue_discard(sbi, entry->start_blkaddr + cur_pos,
| |