On 2024/8/10 0:20, Jan Kara wrote: > On Fri 09-08-24 11:35:49, Zhang Yi wrote: >> On 2024/8/9 2:36, Jan Kara wrote: >>> On Thu 08-08-24 19:18:30, Zhang Yi wrote: >>>> On 2024/8/8 1:41, Jan Kara wrote: >>>>> On Fri 02-08-24 19:51:16, Zhang Yi wrote: >>>>>> From: Zhang Yi <yi.zhang@xxxxxxxxxx> >>>>>> >>>>>> Now that we update data reserved space for delalloc after allocating >>>>>> new blocks in ext4_{ind|ext}_map_blocks(), and if bigalloc feature is >>>>>> enabled, we also need to query the extents_status tree to calculate the >>>>>> exact reserved clusters. This is complicated now and it appears that >>>>>> it's better to do this job in ext4_es_insert_extent(), because >>>>>> __es_remove_extent() have already count delalloc blocks when removing >>>>>> delalloc extents and __revise_pending() return new adding pending count, >>>>>> we could update the reserved blocks easily in ext4_es_insert_extent(). >>>>>> >>>>>> Thers is one special case needs to concern is the quota claiming, when >>>>>> bigalloc is enabled, if the delayed cluster allocation has been raced >>>>>> by another no-delayed allocation(e.g. from fallocate) which doesn't >>>>>> cover the delayed blocks: >>>>>> >>>>>> |< one cluster >| >>>>>> hhhhhhhhhhhhhhhhhhhdddddddddd >>>>>> ^ ^ >>>>>> |< >| < fallocate this range, don't claim quota again >>>>>> >>>>>> We can't claim quota as usual because the fallocate has already claimed >>>>>> it in ext4_mb_new_blocks(), we could notice this case through the >>>>>> removed delalloc blocks count. >>>>>> >>>>>> Signed-off-by: Zhang Yi <yi.zhang@xxxxxxxxxx> >>>>> ... >>>>>> @@ -926,9 +928,27 @@ void ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk, >>>>>> __free_pending(pr); >>>>>> pr = NULL; >>>>>> } >>>>>> + pending = err3; >>>>>> } >>>>>> error: >>>>>> write_unlock(&EXT4_I(inode)->i_es_lock); >>>>>> + /* >>>>>> + * Reduce the reserved cluster count to reflect successful deferred >>>>>> + * allocation of delayed allocated clusters or direct allocation of >>>>>> + * clusters discovered to be delayed allocated. Once allocated, a >>>>>> + * cluster is not included in the reserved count. >>>>>> + * >>>>>> + * When bigalloc is enabled, allocating non-delayed allocated blocks >>>>>> + * which belong to delayed allocated clusters (from fallocate, filemap, >>>>>> + * DIO, or clusters allocated when delalloc has been disabled by >>>>>> + * ext4_nonda_switch()). Quota has been claimed by ext4_mb_new_blocks(), >>>>>> + * so release the quota reservations made for any previously delayed >>>>>> + * allocated clusters. >>>>>> + */ >>>>>> + resv_used = rinfo.delonly_cluster + pending; >>>>>> + if (resv_used) >>>>>> + ext4_da_update_reserve_space(inode, resv_used, >>>>>> + rinfo.delonly_block); >>>>> >>>>> I'm not sure I understand here. We are inserting extent into extent status >>>>> tree. We are replacing resv_used clusters worth of space with delayed >>>>> allocation reservation with normally allocated clusters so we need to >>>>> release the reservation (mballoc already reduced freeclusters counter). >>>>> That I understand. In normal case we should also claim quota because we are >>>>> converting from reserved into allocated state. Now if we allocated blocks >>>>> under this range (e.g. from fallocate()) without >>>>> EXT4_GET_BLOCKS_DELALLOC_RESERVE, we need to release quota reservation here >>>>> instead of claiming it. But I fail to see how rinfo.delonly_block > 0 is >>>>> related to whether EXT4_GET_BLOCKS_DELALLOC_RESERVE was set when allocating >>>>> blocks for this extent or not. >>>> >>>> Oh, this is really complicated due to the bigalloc feature, please let me >>>> explain it more clearly by listing all related situations. >>>> >>>> There are 2 types of paths of allocating delayed/reserved cluster: >>>> 1. Normal case, normally allocate delayed clusters from the write back path. >>>> 2. Special case, allocate blocks under this delayed range, e.g. from >>>> fallocate(). >>>> >>>> There are 4 situations below: >>>> >>>> A. bigalloc is disabled. This case is simple, after path 2, we don't need >>>> to distinguish path 1 and 2, when calling ext4_es_insert_extent(), we >>>> set EXT4_GET_BLOCKS_DELALLOC_RESERVE after EXT4_MAP_DELAYED bit is >>>> detected. If the flag is set, we must be replacing a delayed extent and >>>> rinfo.delonly_block must be > 0. So rinfo.delonly_block > 0 is equal >>>> to set EXT4_GET_BLOCKS_DELALLOC_RESERVE. >>> >>> Right. So fallocate() will call ext4_map_blocks() and >>> ext4_es_lookup_extent() will find delayed extent and set EXT4_MAP_DELAYED >>> which you (due to patch 2 of this series) transform into >>> EXT4_GET_BLOCKS_DELALLOC_RESERVE. We used to update the delalloc >>> accounting through in ext4_ext_map_blocks() but this patch moved the update >>> to ext4_es_insert_extent(). But there is one cornercase even here AFAICT: >>> >>> Suppose fallocate is called for range 0..16k, we have delalloc extent at >>> 8k..16k. In this case ext4_map_blocks() at block 0 will not find the >>> delalloc extent but ext4_ext_map_blocks() will allocate 16k from mballoc >>> without using delalloc reservation but then ext4_es_insert_extent() will >>> still have rinfo.delonly_block > 0 so we claim the quota reservation >>> instead of releasing it? >>> >> >> After commit 6430dea07e85 ("ext4: correct the hole length returned by >> ext4_map_blocks()"), the fallocate range 0-16K would be divided into two >> rounds. When we first calling ext4_map_blocks() with 0-16K, the map range >> will be corrected to 0-8k by ext4_ext_determine_insert_hole() and the >> allocating range should not cover any delayed range. > > Eww, subtle, subtle, subtle... And isn't this also racy? We drop i_data_sem > in ext4_map_blocks() after we do the initial lookup. So there can be some > changes to both the extent tree and extent status tree before we grab > i_data_sem again for the allocation. We hold inode_lock so there can be > only writeback and page faults racing with us but e.g. ext4_page_mkwrite() > -> block_page_mkwrite -> ext4_da_get_block_prep() -> ext4_da_map_blocks() > can add delayed extent into extent status tree in that window causing > breakage, can't it? Oh! you are totally right, I missed that current ext4_fallocate() doesn't hold invalidate_lock for the normal fallocate path, hence there's nothing could prevent this race now, thanks a lot for pointing this out. > >> Then >> ext4_alloc_file_blocks() will call ext4_map_blocks() again to allocate >> 8K-16K in the second round, in this round, we are allocating a real >> delayed range. Please below graph for details, >> >> ext4_alloc_file_blocks() //0-16K >> ext4_map_blocks() //0-16K >> ext4_es_lookup_extent() //find nothing >> ext4_ext_map_blocks(0) >> ext4_ext_determine_insert_hole() //change map range to 0-8K >> ext4_ext_map_blocks(EXT4_GET_BLOCKS_CREATE) //allocate blocks under hole >> ext4_map_blocks() //8-16K >> ext4_es_lookup_extent() //find delayed extent >> ext4_ext_map_blocks(EXT4_GET_BLOCKS_CREATE) >> //allocate blocks under a whole delayed range, >> //use rinfo.delonly_block > 0 is okay >> >> Hence the allocating range can't mixed with delayed and non-delayed extent >> at a time, and the rinfo.delonly_block > 0 should work. > > Besides the race above I agree. So either we need to trim mapping extent in > ext4_map_blocks() after re-acquiring i_data_sem Yeah, if we keep on using this solution, it looks like we have to add similar logic we've done in ext4_da_map_blocks() a few months ago into the begin of the new helper ext4_map_create_blocks(). I guess it may expensive and not worth now. if (ext4_es_lookup_extent(inode, map->m_lblk, NULL, &es)) { map->m_len = min_t(unsigned int, map->m_len, es.es_len - (map->m_lblk - es.es_lblk)); } else retval = ext4_map_query_blocks(NULL, inode, map); ... } > or we need to deal with > unwritten extents that are partially delalloc. I'm more and more leaning > towards just passing the information whether delalloc was used or not to > extent status tree insertion. Because that can deal with partial extents > just fine... > Yeah, I agree with you, passing the information to ext4_es_init_extent() is simple and looks fine. I will change to use this solution. > Thanks for your patience with me :). > Anytime! I appreciate your review and suggestions as well. :) Thanks, Yi.