Re: [f2fs-dev] [PATCH RESEND] f2fs: modify the procedure of scan free nid

From: Chao Yu
Date: Mon Nov 06 2017 - 06:08:38 EST


On 2017/11/6 18:42, Chao Yu wrote:
> On 2017/11/6 15:09, Fan Li wrote:
>>
>>
>>> -----Original Message-----
>>> From: Chao Yu [mailto:chao@xxxxxxxxxx]
>>> Sent: Friday, November 03, 2017 9:16 PM
>>> To: Fan Li; 'Chao Yu'; 'Jaegeuk Kim'
>>> Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
>>> Subject: Re: [f2fs-dev] [PATCH RESEND] f2fs: modify the procedure of scan free nid
>>>
>>> On 2017/11/3 18:29, Fan Li wrote:
>>>>
>>>>
>>>>> -----Original Message-----
>>>>> From: Chao Yu [mailto:yuchao0@xxxxxxxxxx]
>>>>> Sent: Friday, November 03, 2017 4:54 PM
>>>>> To: Fan Li; 'Chao Yu'; 'Jaegeuk Kim'
>>>>> Cc: linux-kernel@xxxxxxxxxxxxxxx;
>>>>> linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
>>>>> Subject: Re: [f2fs-dev] [PATCH RESEND] f2fs: modify the procedure of
>>>>> scan free nid
>>>>>
>>>>> On 2017/11/3 15:31, Fan Li wrote:
>>>>>> In current version, we preserve 8 pages of nat blocks as free nids,
>>>>>> we build bitmaps for it and use them to allocate nids until its
>>>>>> number drops below NAT_ENTRY_PER_BLOCK.
>>>>>>
>>>>>> After that, we have a problem, scan_free_nid_bits will scan the same
>>>>>> 8 pages trying to find more free nids, but in most cases the free
>>>>>> nids in these bitmaps are already in free list, scan them won't get
>>>>>> us any new nids.
>>>>>> Further more, after scan_free_nid_bits, the scan is over if
>>>>>> nid_cnt[FREE_NID] != 0.
>>>>>> It causes that we scan the same pages over and over again, and no
>>>>>> new free nids are found until nid_cnt[FREE_NID]==0. While the
>>>>>> scanned pages increase, the problem grows worse.
>>>>>>
>>>>>> This patch mark the range where new free nids could exist and keep
>>>>>> scan for free nids until nid_cnt[FREE_NID] >= NAT_ENTRY_PER_BLOCK.
>>>>>> The new vairable first_scan_block marks the start of the range, it's
>>>>>> initialized with NEW_ADDR, which means all free nids before
>>>>>> next_scan_nid are already in free list; and use next_scan_nid as the
>>>>>> end of the range since all free nids which are scanned in
>>>>>> scan_free_nid_bits must be smaller next_scan_nid.
>>>>>
>>>>> Think over again, IMO, we can add an variable for stating total count
>>>>> of free nids in bitamp, if there is no free nid, just
>>>> skipping scanning all
>>>>> existed bitmap.
>>>>>
>>>>> And if there is only few free nid scattered in bitmap, the cost will
>>>>> be limited because we will skip scanning
>>>> nm_i::free_nid_bitmap if
>>>>> nm_i::free_nid_count is zero. Once we find one free nid, let's skip out.
>>>>>
>>>>> Since there shouldn't be very heavy overhead for CPU during traveling
>>>>> nm_i::nat_block_bitmap, I expect below change could be more simple for maintaining and being with the same effect.
>>>>>
>>>>> How do you think?
>>>>>
>>>>
>>>> I think if you need this to work, check total_bitmap_free_nid may not be sufficient enough.
>>>> The problem this patch presents is that even all the free nids are
>>>> already in the free list, we still scan all the pages.
>>>> The scan proceeds once free nid count is below NAT_ENTRY_PER_BLOCK.
>>>> So in most cases, there are still free nids in the bitmap during the
>>>> scan, and current codes will check every one of them to see if they are actually in free list.
>>>> If only check total_bitmap_free_nid == 0 won't take this overhead away.
>>>
>>> Oh, you could see that, we have added free_nid_count in each NAT block's free_nid_bitmap bitmap, before scan the bitmap, we will
>> make
>>> sure there is at least one free nid.
>>>
>>> scan_free_nid_bits()
>>> for (i = 0; i < nm_i->nat_blocks; i++) {
>>> if (!test_bit_le(i, nm_i->nat_block_bitmap))
>>> continue;
>>> if (!nm_i->free_nid_count[i])
>>> continue;
>> Do you mean free_nid_count here?
>> I thought free_nid_count only represents how many nats are marked free in bitmap of one block.
>
> Right.
>
>>
>> To my understanding, even a nat is already in the free list, it will still have a bit marked as free in
>> free_nid_bitmap and a count in free_nid_count.
>> That means if free_nid_count != 0, and there are marked bits in the bitmap, the free nats in this
>> block could still be all in the free list.
>
> Yes.
>
>> The purpose of scan is to find new nats and add them to free list, go through the nats which are
>> already in the free list isn't what we want.
>> And in xfstest, under most cases scan_free_nid_bits runs, all free nats are indeed in the free list.
>
> Could you test that diff to check whether we will scan free_nid_bitmap
> which indicates there are free nids, but can not add any of them into
> free list due to they already are there.
>
> - if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
> + if (nm_i->nid_cnt[FREE_NID])
>
> Due to above change, I expect that will be eliminated...

Please check and test this diff:

diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
index 8df2ce9a1356..ff0ce3a4ceab 100644
--- a/fs/f2fs/f2fs.h
+++ b/fs/f2fs/f2fs.h
@@ -729,6 +729,7 @@ struct f2fs_nm_info {
unsigned char (*free_nid_bitmap)[NAT_ENTRY_BITMAP_SIZE];
unsigned char *nat_block_bitmap;
unsigned short *free_nid_count; /* free nid count of NAT block */
+ unsigned int total_bitmap_free_nid; /* total free nid count in bitmap */

/* for checkpoint */
char *nat_bitmap; /* NAT bitmap pointer */
diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c
index fef5c68886b1..33ee1302dbe1 100644
--- a/fs/f2fs/node.c
+++ b/fs/f2fs/node.c
@@ -1911,10 +1911,13 @@ static void update_free_nid_bitmap(struct f2fs_sb_info *sbi, nid_t nid,
else
__clear_bit_le(nid_ofs, nm_i->free_nid_bitmap[nat_ofs]);

- if (set)
+ if (set) {
nm_i->free_nid_count[nat_ofs]++;
- else if (!build)
+ nm_i->total_bitmap_free_nid++;
+ } else if (!build) {
nm_i->free_nid_count[nat_ofs]--;
+ nm_i->total_bitmap_free_nid--;
+ }
}

static void scan_nat_page(struct f2fs_sb_info *sbi,
@@ -1958,6 +1961,9 @@ static void scan_free_nid_bits(struct f2fs_sb_info *sbi)

down_read(&nm_i->nat_tree_lock);

+ if (!nm_i->total_bitmap_free_nid)
+ goto out;
+
for (i = 0; i < nm_i->nat_blocks; i++) {
if (!test_bit_le(i, nm_i->nat_block_bitmap))
continue;
@@ -1972,7 +1978,7 @@ static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
nid = i * NAT_ENTRY_PER_BLOCK + idx;
add_free_nid(sbi, nid, true);

- if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
+ if (nm_i->nid_cnt[FREE_NID])
goto out;
}
}
@@ -2012,6 +2018,10 @@ static void __build_free_nids(struct f2fs_sb_info *sbi, bool sync, bool mount)
return;

if (!mount) {
+ /* if there are free nids in list, allocate them in prior */
+ if (sync && nm_i->nid_cnt[FREE_NID])
+ return;
+
/* try to find free nids in free_nid_bitmap */
scan_free_nid_bits(sbi);

Thanks,

>
> Thanks,
>
>>
>>> for (idx = 0; idx < NAT_ENTRY_PER_BLOCK; idx++) {
>>> nid_t nid;
>>>
>>> if (!test_bit_le(idx, nm_i->free_nid_bitmap[i]))
>>> continue;
>>>
>>> nid = i * NAT_ENTRY_PER_BLOCK + idx;
>>> add_free_nid(sbi, nid, true);
>>>
>>> if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
>>> goto out;
>>> }
>>> }
>>>
>>> And In that diff, we have changed the exiting condition, once we have grabbed one free nid, stop building.
>>>
>>>>> - if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
>>>>> + if (nm_i->nid_cnt[FREE_NID])
>>>>> goto out;
>>>
>>>
>>> So with that simple change, only overhead here is we need to travel nat_block_bitmap all the time when total_bitmap_free_nid is
>> nonzero,
>>> but I think that would not be an critical issue here.
>>>
>>>>
>>>> I considered a lot of ways to fix this problem before I submit this
>>>> patch, One of my idea is quite similar to yours, but I use "if
>>>> (total_bitmap_free_nid == nm_i->nid_cnt[FREE_NID])" to decide whether
>>>> skip or not.
>>>
>>> Hmm.. can we confirm that if there is no free nid in all bitmap, we can skip the unneeded scanning? Anyway, I think you can write
>> a patch to
>>> fix that first?
>>> More like that diff.
>>>
>>>> If you insist, I can submit this simpler one instead, but some follow
>>>> upgrade would be unavailable, for example, use smaller granularity for
>>>> tracking last-scanned-position that we talked about.> I know sometimes
>>>> I can be obsessed with the performance, I usually choose the faster
>>>> way over simpler ones. If you think it's too much, please tell me, I'm
>>>> sure we can find some middle ground.
>>>
>>> Yup, I think that's why you're the expert of algorithm, I have no doubt about that. :)
>>>
>>> IMO, instead of reducing cpu overhead without simple change, I prefer the one can reducing IO, e.g. if NAT block contains maximum
>> count
>>> free nids, we can load these nids first, after they were been allocated, in checkpoint, we can write these nat entries into one
>> NAT block. On
>>> the contrary, if we load free nids with same count from different NAT blocks, in checkpoint, maybe we will write them into more
>> NAT blocks.
>>>
>>> Thanks,
>>>
>>>>
>>>> Thank you
>>>>
>>>>
>>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h index
>>>>> cb3f10bc8723..238d95e89dec 100644
>>>>> --- a/fs/f2fs/f2fs.h
>>>>> +++ b/fs/f2fs/f2fs.h
>>>>> @@ -729,6 +729,7 @@ struct f2fs_nm_info {
>>>>> unsigned char (*free_nid_bitmap)[NAT_ENTRY_BITMAP_SIZE];
>>>>> unsigned char *nat_block_bitmap;
>>>>> unsigned short *free_nid_count; /* free nid count of NAT block */
>>>>> + unsigned int total_bitmap_free_nid; /* total free nid count in bitmap */
>>>>>
>>>>> /* for checkpoint */
>>>>> char *nat_bitmap; /* NAT bitmap pointer */
>>>>> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c index
>>>>> fef5c68886b1..e4861908a396 100644
>>>>> --- a/fs/f2fs/node.c
>>>>> +++ b/fs/f2fs/node.c
>>>>> @@ -1911,10 +1911,13 @@ static void update_free_nid_bitmap(struct f2fs_sb_info *sbi, nid_t nid,
>>>>> else
>>>>> __clear_bit_le(nid_ofs, nm_i->free_nid_bitmap[nat_ofs]);
>>>>>
>>>>> - if (set)
>>>>> + if (set) {
>>>>> nm_i->free_nid_count[nat_ofs]++;
>>>>> - else if (!build)
>>>>> + nm_i->total_bitmap_free_nid++;
>>>>> + } else if (!build) {
>>>>> nm_i->free_nid_count[nat_ofs]--;
>>>>> + nm_i->total_bitmap_free_nid--;
>>>>> + }
>>>>> }
>>>>>
>>>>> static void scan_nat_page(struct f2fs_sb_info *sbi, @@ -1958,6
>>>>> +1961,9 @@ static void scan_free_nid_bits(struct f2fs_sb_info
>>>> *sbi)
>>>>>
>>>>> down_read(&nm_i->nat_tree_lock);
>>>>>
>>>>> + if (!nm_i->total_bitmap_free_nid)
>>>>> + goto out;
>>>>> +
>>>>> for (i = 0; i < nm_i->nat_blocks; i++) {
>>>>> if (!test_bit_le(i, nm_i->nat_block_bitmap))
>>>>> continue;
>>>>> @@ -1972,7 +1978,7 @@ static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
>>>>> nid = i * NAT_ENTRY_PER_BLOCK + idx;
>>>>> add_free_nid(sbi, nid, true);
>>>>>
>>>>> - if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
>>>>> + if (nm_i->nid_cnt[FREE_NID])
>>>>> goto out;
>>>>> }
>>>>> }
>>>>>
>>>>> Thanks,
>>>>>
>>>>>>
>>>>>> Signed-off-by: Fan li <fanofcode.li@xxxxxxxxxxx>
>>>>>> ---
>>>>>> fs/f2fs/f2fs.h | 1 +
>>>>>> fs/f2fs/node.c | 42 +++++++++++++++++++++++++++++++++++-------
>>>>>> 2 files changed, 36 insertions(+), 7 deletions(-)
>>>>>>
>>>>>> diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h index e0ef31c..ae1cf91
>>>>>> 100644
>>>>>> --- a/fs/f2fs/f2fs.h
>>>>>> +++ b/fs/f2fs/f2fs.h
>>>>>> @@ -705,6 +705,7 @@ struct f2fs_nm_info {
>>>>>> nid_t max_nid; /* maximum possible node ids */
>>>>>> nid_t available_nids; /* # of available node ids */
>>>>>> nid_t next_scan_nid; /* the next nid to be scanned */
>>>>>> + block_t first_scan_block; /* the first NAT block to be scanned */
>>>>>> unsigned int ram_thresh; /* control the memory footprint */
>>>>>> unsigned int ra_nid_pages; /* # of nid pages to be readaheaded */
>>>>>> unsigned int dirty_nats_ratio; /* control dirty nats ratio threshold */
>>>>>> diff --git a/fs/f2fs/node.c b/fs/f2fs/node.c index 3d0d1be..f921e0c
>>>>>> 100644
>>>>>> --- a/fs/f2fs/node.c
>>>>>> +++ b/fs/f2fs/node.c
>>>>>> @@ -1812,7 +1812,7 @@ static bool add_free_nid(struct f2fs_sb_info *sbi, nid_t nid, bool build)
>>>>>> struct f2fs_nm_info *nm_i = NM_I(sbi);
>>>>>> struct free_nid *i, *e;
>>>>>> struct nat_entry *ne;
>>>>>> - int err = -EINVAL;
>>>>>> + int need_free = 1;
>>>>>> bool ret = false;
>>>>>>
>>>>>> /* 0 nid should not be used */
>>>>>> @@ -1863,13 +1863,25 @@ static bool add_free_nid(struct f2fs_sb_info *sbi, nid_t nid, bool build)
>>>>>> }
>>>>>> }
>>>>>> ret = true;
>>>>>> - err = __insert_free_nid(sbi, i, FREE_NID);
>>>>>> + need_free = __insert_free_nid(sbi, i, FREE_NID);
>>>>>> err_out:
>>>>>> spin_unlock(&nm_i->nid_list_lock);
>>>>>> radix_tree_preload_end();
>>>>>> err:
>>>>>> - if (err)
>>>>>> + if (need_free)
>>>>>> kmem_cache_free(free_nid_slab, i);
>>>>>> + /*
>>>>>> + * For nid that should be free but not in the free
>>>>>> + * structure, update the scan range in hope of adding
>>>>>> + * it in the next scan.
>>>>>> + */
>>>>>> + if (!ret || need_free < 0) {
>>>>>> + block_t tmp_block = NAT_BLOCK_OFFSET(nid);
>>>>>> +
>>>>>> + if (tmp_block < nm_i->first_scan_block)
>>>>>> + nm_i->first_scan_block = tmp_block;
>>>>>> + }
>>>>>> +
>>>>>> return ret;
>>>>>> }
>>>>>>
>>>>>> @@ -1950,10 +1962,17 @@ static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
>>>>>> struct curseg_info *curseg = CURSEG_I(sbi, CURSEG_HOT_DATA);
>>>>>> struct f2fs_journal *journal = curseg->journal;
>>>>>> unsigned int i, idx;
>>>>>> + unsigned int max_blocks = NAT_BLOCK_OFFSET(nm_i->next_scan_nid);
>>>>>>
>>>>>> - down_read(&nm_i->nat_tree_lock);
>>>>>> + /* every free nid in blocks scanned previously is in the free list */
>>>>>> + if (nm_i->first_scan_block == NEW_ADDR)
>>>>>> + return;
>>>>>>
>>>>>> - for (i = 0; i < nm_i->nat_blocks; i++) {
>>>>>> + if (max_blocks == 0)
>>>>>> + max_blocks = nm_i->nat_blocks;
>>>>>> +
>>>>>> + down_read(&nm_i->nat_tree_lock);
>>>>>> + for (i = nm_i->first_scan_block; i < max_blocks; i++) {
>>>>>> if (!test_bit_le(i, nm_i->nat_block_bitmap))
>>>>>> continue;
>>>>>> if (!nm_i->free_nid_count[i])
>>>>>> @@ -1967,10 +1986,13 @@ static void scan_free_nid_bits(struct f2fs_sb_info *sbi)
>>>>>> nid = i * NAT_ENTRY_PER_BLOCK + idx;
>>>>>> add_free_nid(sbi, nid, true);
>>>>>>
>>>>>> - if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS)
>>>>>> + if (nm_i->nid_cnt[FREE_NID] >= MAX_FREE_NIDS) {
>>>>>> + nm_i->first_scan_block = i;
>>>>>> goto out;
>>>>>> + }
>>>>>> }
>>>>>> }
>>>>>> + nm_i->first_scan_block = NEW_ADDR;
>>>>>> out:
>>>>>> down_read(&curseg->journal_rwsem);
>>>>>> for (i = 0; i < nats_in_cursum(journal); i++) { @@ -2010,7 +2032,7
>>>>>> @@ static void __build_free_nids(struct f2fs_sb_info *sbi, bool sync, bool mount)
>>>>>> /* try to find free nids in free_nid_bitmap */
>>>>>> scan_free_nid_bits(sbi);
>>>>>>
>>>>>> - if (nm_i->nid_cnt[FREE_NID])
>>>>>> + if (nm_i->nid_cnt[FREE_NID] >= NAT_ENTRY_PER_BLOCK)
>>>>>> return;
>>>>>> }
>>>>>>
>>>>>> @@ -2163,6 +2185,7 @@ int try_to_free_nids(struct f2fs_sb_info *sbi, int nr_shrink)
>>>>>> struct f2fs_nm_info *nm_i = NM_I(sbi);
>>>>>> struct free_nid *i, *next;
>>>>>> int nr = nr_shrink;
>>>>>> + nid_t min_nid = nm_i->max_nid;
>>>>>>
>>>>>> if (nm_i->nid_cnt[FREE_NID] <= MAX_FREE_NIDS)
>>>>>> return 0;
>>>>>> @@ -2176,11 +2199,15 @@ int try_to_free_nids(struct f2fs_sb_info *sbi, int nr_shrink)
>>>>>> nm_i->nid_cnt[FREE_NID] <= MAX_FREE_NIDS)
>>>>>> break;
>>>>>>
>>>>>> + if (i->nid < min_nid)
>>>>>> + min_nid = i->nid;
>>>>>> __remove_free_nid(sbi, i, FREE_NID);
>>>>>> kmem_cache_free(free_nid_slab, i);
>>>>>> nr_shrink--;
>>>>>> }
>>>>>> spin_unlock(&nm_i->nid_list_lock);
>>>>>> + if (min_nid != nm_i->max_nid)
>>>>>> + nm_i->first_scan_block = NAT_BLOCK_OFFSET(min_nid);
>>>>>> mutex_unlock(&nm_i->build_lock);
>>>>>>
>>>>>> return nr - nr_shrink;
>>>>>> @@ -2674,6 +2701,7 @@ static int init_node_manager(struct f2fs_sb_info *sbi)
>>>>>> init_rwsem(&nm_i->nat_tree_lock);
>>>>>>
>>>>>> nm_i->next_scan_nid = le32_to_cpu(sbi->ckpt->next_free_nid);
>>>>>> + nm_i->first_scan_block = NEW_ADDR;
>>>>>> nm_i->bitmap_size = __bitmap_size(sbi, NAT_BITMAP);
>>>>>> version_bitmap = __bitmap_ptr(sbi, NAT_BITMAP);
>>>>>> if (!version_bitmap)
>>>>>>
>>>>>
>>>>
>>>>
>>
>>
>>
>> .
>>
>
>
> .
>