Re: [External Mail]Re: [PATCH v3] f2fs: avoid splitting bio when reading multiple pages

[Huang Jianan <huangjianan@xiaomi.com>]

On 2025/6/30 19:23, Chao Yu wrote:
> 
> On 6/25/25 17:50, Huang Jianan wrote:
>> On 2025/6/25 17:48, Jianan Huang wrote:
>>> On 2025/6/25 16:45, Chao Yu wrote:
>>>>
>>>> On 6/25/25 14:49, Jianan Huang wrote:
>>>>> When fewer pages are read, nr_pages may be smaller than nr_cpages. Due
>>>>> to the nr_vecs limit, the compressed pages will be split into multiple
>>>>> bios and then merged at the block level. In this case, nr_cpages should
>>>>> be used to pre-allocate bvecs.
>>>>> To handle this case, align max_nr_pages to cluster_size, which should be
>>>>> enough for all compressed pages.
>>>>>
>>>>> Signed-off-by: Jianan Huang <huangjianan@xiaomi.com>
>>>>> Signed-off-by: Sheng Yong <shengyong1@xiaomi.com>
>>>>> ---
>>>>> Changes since v2:
>>>>> - Initialize index only for compressed files.
>>>>> Changes since v1:
>>>>> - Use aligned nr_pages instead of nr_cpages to pre-allocate bvecs.
>>>>>
>>>>>    fs/f2fs/data.c | 12 ++++++++++--
>>>>>    1 file changed, 10 insertions(+), 2 deletions(-)
>>>>>
>>>>> diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
>>>>> index 31e892842625..d071d9f6a811 100644
>>>>> --- a/fs/f2fs/data.c
>>>>> +++ b/fs/f2fs/data.c
>>>>> @@ -2303,7 +2303,7 @@ int f2fs_read_multi_pages(struct compress_ctx
>>>>> *cc, struct bio **bio_ret,
>>>>>                 }
>>>>>
>>>>>                 if (!bio) {
>>>>> -                     bio = f2fs_grab_read_bio(inode, blkaddr, nr_pages,
>>>>> +                     bio = f2fs_grab_read_bio(inode, blkaddr,
>>>>> nr_pages - i,
>>>>
>>>> Jianan,
>>>>
>>>> Another case:
>>>>
>>>> read page #0,1,2,3 from block #1000,1001,1002, cluster_size=4.
>>>>
>>>> nr_pages=4
>>>> max_nr_pages=round_up(0+4,4)-round_down(0,4)=4
>>>>
>>>> f2fs_mpage_readpages() calls f2fs_read_multi_pages() when nr_pages=1, at
>>>> that time, max_nr_pages equals to 1 as well.
>>>>
>>>> f2fs_grab_read_bio(..., 1 - 0,...) allocate bio w/ 1 vec capacity,
>>>> however,
>>>> we need at least 3 vecs to merge all cpages, right?
>>>>
>>>
>>> Hi, chao,
>>>
>>> If we don't align nr_pages, then when entering f2fs_read_multi_pages,
>>> we have nr_pages pages left, which belong to other clusters.
>>> If this is the last page, we can simply pass nr_pages = 0.
>>>
>>> When allocating bio, we need:
>>> 1. The cpages remaining in the current cluster, which should be
>>> (nr_capges - i).
>>> 2. The maximum cpages remaining in other clusters, which should be
>>> max(nr_pages, cc->nr_cpages).
>>>
>>
>> align(nr_pages, cc->nr_cpages), sorry for this.
>>
>>> So (nr_capges - i) + max(nr_pages, nr_cpages), should be enough for all
>>> vecs?
> 
> Jianan,
> 
> What about getting rid of below change? and just passing max_nr_pages to
> f2fs_read_multi_pages? Maybe there is a little bit waste for bio vector space,
> but it will be more safe to reserve enough margin.
> 
> +       for (; nr_pages; nr_pages--, max_nr_pages--) {
> 

LGTM, updated in v4.

Thanks,

> Thanks,
> 
>>>
>>> Thanks,
>>>
>>>
>>>> Thanks,
>>>>
>>>>>                                         f2fs_ra_op_flags(rac),
>>>>>                                         folio->index, for_write);
>>>>>                         if (IS_ERR(bio)) {
>>>>> @@ -2376,6 +2376,14 @@ static int f2fs_mpage_readpages(struct inode
>>>>> *inode,
>>>>>         unsigned max_nr_pages = nr_pages;
>>>>>         int ret = 0;
>>>>>
>>>>> +#ifdef CONFIG_F2FS_FS_COMPRESSION
>>>>> +     if (f2fs_compressed_file(inode)) {
>>>>> +             index = rac ? readahead_index(rac) : folio->index;
>>>>> +             max_nr_pages = round_up(index + nr_pages,
>>>>> cc.cluster_size) -
>>>>> +                             round_down(index, cc.cluster_size);
>>>>> +     }
>>>>> +#endif
>>>>> +
>>>>>         map.m_pblk = 0;
>>>>>         map.m_lblk = 0;
>>>>>         map.m_len = 0;
>>>>> @@ -2385,7 +2393,7 @@ static int f2fs_mpage_readpages(struct inode
>>>>> *inode,
>>>>>         map.m_seg_type = NO_CHECK_TYPE;
>>>>>         map.m_may_create = false;
>>>>>
>>>>> -     for (; nr_pages; nr_pages--) {
>>>>> +     for (; nr_pages; nr_pages--, max_nr_pages--) {
>>>>>                 if (rac) {
>>>>>                         folio = readahead_folio(rac);
>>>>>                         prefetchw(&folio->flags);
>>>>
>>>
>>
>