Skip to contentSkip to conversation

ext4: Convert inode preallocation list to an rbtree

PATCH Superseded Ojaswin Mujoo 1 participant 10 messages Mar 17, 2023

Subsystems (3)
Documentation VFS ext4 file system

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Ojaswin Mujoo

author reviewer
#0 lore raw

v5 cover ext4: Convert inode preallocation list to an rbtree

This patch series aim to improve the performance and scalability of inode preallocation by changing inode preallocation linked list to an rbtree. I've ran xfstests quick on this series and plan to run auto group as well to confirm we have no regressions.

** Shortcomings of existing implementation **

Right now, we add all the inode preallocations(PAs) to a per inode linked list ei->i_prealloc_list. To prevent the list from growing infinitely during heavy sparse workloads, the length of this list was capped at 512 and a trimming logic was added to trim the list whenever it grew over this threshold, in patch 27bc446e2. This was discussed in detail in the following lore thread [1].

[1] https://lore.kernel.org/all/[email protected]/

But from our testing, we noticed that the current implementation still had issues with scalability as the performance degraded when the PAs stored in the list grew. Most of the degradation was seen in ext4_mb_normalize_request() and ext4_mb_use_preallocated() functions as they iterated the inode PA list.

** Improvements in this patchset **

To counter the above shortcomings, this patch series modifies the inode PA list to an rbtree, which:

- improves the performance of functions discussed above due to the improved lookup speed.

- improves scalability by changing lookup complexity from O(n) to O(logn). We no longer need the trimming logic as well.

As a result, the RCU implementation was needed to be changed since lockless lookups of rbtrees do have some issues like skipping subtrees. Hence, RCU was replaced with read write locks for inode PAs. More information can be found in Patch 7 (that has the core changes).

** Performance Numbers **

Performance numbers were collected with and without these patches, using an nvme device. Details of tests/benchmarks used are as follows:

Test 1: 200,000 1KiB sparse writes using (fio)
Test 2: Fill 5GiB w/ random writes, 1KiB burst size using (fio) Test 3: Test 2, but do 4 sequential writes before jumping to random
offset (fio)
Test 4: Fill 8GB FS w/ 2KiB files, 64 threads in parallel (fsmark)

+──────────+──────────────────+────────────────+──────────────────+──────────────────+
|          |            nodelalloc             |              delalloc               |
+──────────+──────────────────+────────────────+──────────────────+──────────────────+
|          | Unpatched        | Patched        | Unpatched        | Patched          |
+──────────+──────────────────+────────────────+──────────────────+──────────────────+
| Test 1   | 11.8 MB/s        | 23.3 MB/s      | 27.2 MB/s        | 63.7 MB/s        |
| Test 2   | 1617 MB/s        | 1740 MB/s      | 2223 MB/s        | 2208 MB/s        |
| Test 3   | 1715 MB/s        | 1823 MB/s      | 2346 MB/s        | 2364 MB/s        |
| Test 4   | 14284 files/sec  | 14347 files/s  | 13762 files/sec  | 13882 files/sec  |
+──────────+──────────────────+────────────────+──────────────────+──────────────────+

In test 1, we almost see 100 to 200% increase in performance due to the high number of sparse writes highlighting the bottleneck in the unpatched kernel. Further, on running "perf diff patched.data unpatched.data" for test 1, we see something as follows:

     2.83%    +29.67%  [kernel.vmlinux]          [k] _raw_spin_lock
												...
               +3.33%  [ext4]                    [k] ext4_mb_normalize_request.constprop.30
     0.25%     +2.81%  [ext4]                    [k] ext4_mb_use_preallocated

Here we can see that the biggest different is in the _raw_spin_lock() function of unpatched kernel, that is called from ext4_mb_normalize_request() as seen here:

    32.47%  fio              [kernel.vmlinux]            [k] _raw_spin_lock
            |
            ---_raw_spin_lock
               |          
                --32.22%--ext4_mb_normalize_request.constprop.30

This is coming from the spin_lock(&pa->pa_lock) that is called for each PA that we iterate over, in ext4_mb_normalize_request(). Since in rbtrees, we lookup log(n) PAs rather than n PAs, this spin lock is taken less frequently, as evident in the perf.

Furthermore, we see some improvements in other tests however since they don't exercise the PA traversal path as much as test 1, the improvements are relatively smaller.

** Summary of patches **

- Patch 1-5: Abstractions/Minor optimizations
- Patch 6: Change bex lstart adjustment logic to avoid overflow from goal
- Patch 7: Split common inode & locality group specific fields to a union
- Patch 8: Core changes to move inode PA logic from list to rbtree
- Patch 9: Remove the trim logic as it is not needed

** Changes since PATCH v3 [4] **
- ./checkpatch fixes
- Picked up RVBs by Ritesh and Jan
- Rebased over ted/dev which includes Kemeng's mballoc cleanup v3 patches [5]

** Changes since PATCH v3 [2] **
- Patch 6 fixes a bug in current best extent adjustment logic by introducing a new logic.
- Patch 8 - ext4_mb_pa_adjust_overlap functions uses a modified logic to trim the normalized range, which also takes care of presence deleted PAs in the preallocation tree.
- Ran xfstests -c all quick and everything seems good. Performance numbers seem similar to previous improvements as well.

Both the above changes are based on discussion here [3]

** Changes since PATCH v2 [1] **
- In patch 7, include a design change related to encountering deleted PAs in inode rbtree that overlap with to be inserted PA, when adjusting overlap. More details in the patch. (Removed Jan's RVB for this patch)

** Changes since PATCH v1 **
- fixed styling issue
- merged ext4_mb_rb_insert() and ext4_mb_pa_cmp()

** Changes since RFC v3 **
- Changed while loops to for loops in patch 7
- Fixed some data types
- Made rbtree comparison logic more intuitive. The rbtree insertion function still kept separate from comparison function for reusability.

** Changes since RFC v2 **
- Added a function definition that was deleted during v2 rebase

** Changes since RFC v1 **
- Rebased over ext4 dev branch which includes Jan's patchset that changed some code in mballoc.c

[1] https://lore.kernel.org/linux-ext4/[email protected]/
[2] https://lore.kernel.org/all/[email protected]/
[3]
https://lore.kernel.org/all/[email protected]/
[4]
https://lore.kernel.org/all/[email protected]/
[5]
https://lore.kernel.org/r/[email protected]

Ojaswin Mujoo (9):
ext4: Stop searching if PA doesn't satisfy non-extent file
ext4: Refactor code related to freeing PAs
ext4: Refactor code in ext4_mb_normalize_request() and
ext4_mb_use_preallocated()
ext4: Move overlap assert logic into a separate function
ext4: Abstract out overlap fix/check logic in
ext4_mb_normalize_request()
ext4: Fix best extent lstart adjustment logic in

    ext4_mb_new_inode_pa()

ext4: Convert pa->pa_inode_list and pa->pa_obj_lock into a union
ext4: Use rbtrees to manage PAs instead of inode i_prealloc_list
ext4: Remove the logic to trim inode PAs

 Documentation/admin-guide/ext4.rst |   3 -
 fs/ext4/ext4.h                     |   5 +-
 fs/ext4/mballoc.c                  | 536 ++++++++++++++++++++---------
 fs/ext4/mballoc.h                  |  17 +-
 fs/ext4/super.c                    |   4 +-
 fs/ext4/sysfs.c                    |   2 -
 6 files changed, 377 insertions(+), 190 deletions(-)
Signature · 1 line
2.31.1
author reviewer
#1 lore raw

PATCH v5 1/9 ext4: Stop searching if PA doesn't satisfy non-extent file

If we come across a PA that matches the logical offset but is unable to satisfy a non-extent file due to its physical start being higher than that supported by non extent files, then simply stop searching for another PA and break out of loop. This is because, since PAs don't overlap, we won't be able to find another inode PA which can satisfy the original request.

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)
fs/ext4/mballoc.c +7 -2
index 63a68cee36c6..ed58cefed7fa 100644
@@ -4379,8 +4379,13 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 		/* non-extent files can't have physical blocks past 2^32 */
 		if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
 		    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
-		     EXT4_MAX_BLOCK_FILE_PHYS))
-			continue;
+		     EXT4_MAX_BLOCK_FILE_PHYS)) {
+			/*
+			 * Since PAs don't overlap, we won't find any
+			 * other PA to satisfy this.
+			 */
+			break;
+		}
 
 		/* found preallocated blocks, use them */
 		spin_lock(&pa->pa_lock);
Signature · 1 line
2.31.1
author reviewer
#2 lore raw

PATCH v5 2/9 ext4: Refactor code related to freeing PAs

In reply to Ojaswin Mujoo

This patch makes the following changes:

* Rename ext4_mb_pa_free to ext4_mb_pa_put_free
to better reflect its purpose

* Add new ext4_mb_pa_free() which only handles freeing

* Refactor ext4_mb_pa_callback() to use ext4_mb_pa_free()

There are no functional changes in this patch

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 30 +++++++++++++++++++++---------
 1 file changed, 21 insertions(+), 9 deletions(-)
fs/ext4/mballoc.c +21 -9
index ed58cefed7fa..18b65e53abbb 100644
@@ -4527,16 +4527,22 @@ static void ext4_mb_mark_pa_deleted(struct super_block *sb,
 	}
 }
 
-static void ext4_mb_pa_callback(struct rcu_head *head)
+static inline void ext4_mb_pa_free(struct ext4_prealloc_space *pa)
 {
-	struct ext4_prealloc_space *pa;
-	pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
-
+	BUG_ON(!pa);
 	BUG_ON(atomic_read(&pa->pa_count));
 	BUG_ON(pa->pa_deleted == 0);
 	kmem_cache_free(ext4_pspace_cachep, pa);
 }
 
+static void ext4_mb_pa_callback(struct rcu_head *head)
+{
+	struct ext4_prealloc_space *pa;
+
+	pa = container_of(head, struct ext4_prealloc_space, u.pa_rcu);
+	ext4_mb_pa_free(pa);
+}
+
 /*
  * drops a reference to preallocated space descriptor
  * if this was the last reference and the space is consumed

@@ -5054,14 +5060,20 @@ static int ext4_mb_pa_alloc(struct ext4_allocation_context *ac)
 	return 0;
 }
 
-static void ext4_mb_pa_free(struct ext4_allocation_context *ac)
+static void ext4_mb_pa_put_free(struct ext4_allocation_context *ac)
 {
 	struct ext4_prealloc_space *pa = ac->ac_pa;
 
 	BUG_ON(!pa);
 	ac->ac_pa = NULL;
 	WARN_ON(!atomic_dec_and_test(&pa->pa_count));
-	kmem_cache_free(ext4_pspace_cachep, pa);
+	/*
+	 * current function is only called due to an error or due to
+	 * len of found blocks < len of requested blocks hence the PA has not
+	 * been added to grp->bb_prealloc_list. So we don't need to lock it
+	 */
+	pa->pa_deleted = 1;
+	ext4_mb_pa_free(pa);
 }
 
 #ifdef CONFIG_EXT4_DEBUG

@@ -5604,13 +5616,13 @@ ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
 		 * So we have to free this pa here itself.
 		 */
 		if (*errp) {
-			ext4_mb_pa_free(ac);
+			ext4_mb_pa_put_free(ac);
 			ext4_discard_allocated_blocks(ac);
 			goto errout;
 		}
 		if (ac->ac_status == AC_STATUS_FOUND &&
 			ac->ac_o_ex.fe_len >= ac->ac_f_ex.fe_len)
-			ext4_mb_pa_free(ac);
+			ext4_mb_pa_put_free(ac);
 	}
 	if (likely(ac->ac_status == AC_STATUS_FOUND)) {
 		*errp = ext4_mb_mark_diskspace_used(ac, handle, reserv_clstrs);

@@ -5629,7 +5641,7 @@ ext4_fsblk_t ext4_mb_new_blocks(handle_t *handle,
 		 * If block allocation fails then the pa allocated above
 		 * needs to be freed here itself.
 		 */
-		ext4_mb_pa_free(ac);
+		ext4_mb_pa_put_free(ac);
 		*errp = -ENOSPC;
 	}
Signature · 1 line
2.31.1
author reviewer
#3 lore raw

PATCH v5 3/9 ext4: Refactor code in ext4_mb_normalize_request() and ext4_mb_use_preallocated()

In reply to Ojaswin Mujoo

Change some variable names to be more consistent and refactor some of the code to make it easier to read.

There are no functional changes in this patch

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 96 ++++++++++++++++++++++++-----------------------
 1 file changed, 49 insertions(+), 47 deletions(-)
fs/ext4/mballoc.c +49 -47
index 18b65e53abbb..a1429d5a1197 100644
@@ -3994,7 +3994,8 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
 	loff_t orig_size __maybe_unused;
 	ext4_lblk_t start;
 	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
-	struct ext4_prealloc_space *pa;
+	struct ext4_prealloc_space *tmp_pa;
+	ext4_lblk_t tmp_pa_start, tmp_pa_end;
 
 	/* do normalize only data requests, metadata requests
 	   do not need preallocation */

@@ -4097,54 +4098,52 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
 
 	/* check we don't cross already preallocated blocks */
 	rcu_read_lock();
-	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
-		ext4_lblk_t pa_end;
-
-		if (pa->pa_deleted)
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+		if (tmp_pa->pa_deleted)
 			continue;
-		spin_lock(&pa->pa_lock);
-		if (pa->pa_deleted) {
-			spin_unlock(&pa->pa_lock);
+		spin_lock(&tmp_pa->pa_lock);
+		if (tmp_pa->pa_deleted) {
+			spin_unlock(&tmp_pa->pa_lock);
 			continue;
 		}
 
-		pa_end = pa->pa_lstart + EXT4_C2B(sbi, pa->pa_len);
+		tmp_pa_start = tmp_pa->pa_lstart;
+		tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
 
 		/* PA must not overlap original request */
-		BUG_ON(!(ac->ac_o_ex.fe_logical >= pa_end ||
-			ac->ac_o_ex.fe_logical < pa->pa_lstart));
+		BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
+			ac->ac_o_ex.fe_logical < tmp_pa_start));
 
 		/* skip PAs this normalized request doesn't overlap with */
-		if (pa->pa_lstart >= end || pa_end <= start) {
-			spin_unlock(&pa->pa_lock);
+		if (tmp_pa_start >= end || tmp_pa_end <= start) {
+			spin_unlock(&tmp_pa->pa_lock);
 			continue;
 		}
-		BUG_ON(pa->pa_lstart <= start && pa_end >= end);
+		BUG_ON(tmp_pa_start <= start && tmp_pa_end >= end);
 
 		/* adjust start or end to be adjacent to this pa */
-		if (pa_end <= ac->ac_o_ex.fe_logical) {
-			BUG_ON(pa_end < start);
-			start = pa_end;
-		} else if (pa->pa_lstart > ac->ac_o_ex.fe_logical) {
-			BUG_ON(pa->pa_lstart > end);
-			end = pa->pa_lstart;
+		if (tmp_pa_end <= ac->ac_o_ex.fe_logical) {
+			BUG_ON(tmp_pa_end < start);
+			start = tmp_pa_end;
+		} else if (tmp_pa_start > ac->ac_o_ex.fe_logical) {
+			BUG_ON(tmp_pa_start > end);
+			end = tmp_pa_start;
 		}
-		spin_unlock(&pa->pa_lock);
+		spin_unlock(&tmp_pa->pa_lock);
 	}
 	rcu_read_unlock();
 	size = end - start;
 
 	/* XXX: extra loop to check we really don't overlap preallocations */
 	rcu_read_lock();
-	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
-		ext4_lblk_t pa_end;
-
-		spin_lock(&pa->pa_lock);
-		if (pa->pa_deleted == 0) {
-			pa_end = pa->pa_lstart + EXT4_C2B(sbi, pa->pa_len);
-			BUG_ON(!(start >= pa_end || end <= pa->pa_lstart));
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+		spin_lock(&tmp_pa->pa_lock);
+		if (tmp_pa->pa_deleted == 0) {
+			tmp_pa_start = tmp_pa->pa_lstart;
+			tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
+			BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
 		}
-		spin_unlock(&pa->pa_lock);
+		spin_unlock(&tmp_pa->pa_lock);
 	}
 	rcu_read_unlock();
 

@@ -4358,7 +4357,8 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 	int order, i;
 	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
 	struct ext4_locality_group *lg;
-	struct ext4_prealloc_space *pa, *cpa = NULL;
+	struct ext4_prealloc_space *tmp_pa, *cpa = NULL;
+	ext4_lblk_t tmp_pa_start, tmp_pa_end;
 	ext4_fsblk_t goal_block;
 
 	/* only data can be preallocated */

@@ -4367,18 +4367,20 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 
 	/* first, try per-file preallocation */
 	rcu_read_lock();
-	list_for_each_entry_rcu(pa, &ei->i_prealloc_list, pa_inode_list) {
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
 
 		/* all fields in this condition don't change,
 		 * so we can skip locking for them */
-		if (ac->ac_o_ex.fe_logical < pa->pa_lstart ||
-		    ac->ac_o_ex.fe_logical >= (pa->pa_lstart +
-					       EXT4_C2B(sbi, pa->pa_len)))
+		tmp_pa_start = tmp_pa->pa_lstart;
+		tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
+
+		if (ac->ac_o_ex.fe_logical < tmp_pa_start ||
+		    ac->ac_o_ex.fe_logical >= tmp_pa_end)
 			continue;
 
 		/* non-extent files can't have physical blocks past 2^32 */
 		if (!(ext4_test_inode_flag(ac->ac_inode, EXT4_INODE_EXTENTS)) &&
-		    (pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len) >
+		    (tmp_pa->pa_pstart + EXT4_C2B(sbi, tmp_pa->pa_len) >
 		     EXT4_MAX_BLOCK_FILE_PHYS)) {
 			/*
 			 * Since PAs don't overlap, we won't find any

@@ -4388,16 +4390,16 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 		}
 
 		/* found preallocated blocks, use them */
-		spin_lock(&pa->pa_lock);
-		if (pa->pa_deleted == 0 && pa->pa_free) {
-			atomic_inc(&pa->pa_count);
-			ext4_mb_use_inode_pa(ac, pa);
-			spin_unlock(&pa->pa_lock);
+		spin_lock(&tmp_pa->pa_lock);
+		if (tmp_pa->pa_deleted == 0 && tmp_pa->pa_free) {
+			atomic_inc(&tmp_pa->pa_count);
+			ext4_mb_use_inode_pa(ac, tmp_pa);
+			spin_unlock(&tmp_pa->pa_lock);
 			ac->ac_criteria = 10;
 			rcu_read_unlock();
 			return true;
 		}
-		spin_unlock(&pa->pa_lock);
+		spin_unlock(&tmp_pa->pa_lock);
 	}
 	rcu_read_unlock();
 

@@ -4421,16 +4423,16 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 	 */
 	for (i = order; i < PREALLOC_TB_SIZE; i++) {
 		rcu_read_lock();
-		list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[i],
+		list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
 					pa_inode_list) {
-			spin_lock(&pa->pa_lock);
-			if (pa->pa_deleted == 0 &&
-					pa->pa_free >= ac->ac_o_ex.fe_len) {
+			spin_lock(&tmp_pa->pa_lock);
+			if (tmp_pa->pa_deleted == 0 &&
+					tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {
 
 				cpa = ext4_mb_check_group_pa(goal_block,
-								pa, cpa);
+								tmp_pa, cpa);
 			}
-			spin_unlock(&pa->pa_lock);
+			spin_unlock(&tmp_pa->pa_lock);
 		}
 		rcu_read_unlock();
 	}
Signature · 1 line
2.31.1
author reviewer
#4 lore raw

PATCH v5 4/9 ext4: Move overlap assert logic into a separate function

In reply to Ojaswin Mujoo

Abstract out the logic to double check for overlaps in normalize_pa to a separate function. Since there has been no reports in past where we have seen any overlaps which hits this bug_on(), in future we can consider calling this function under "#ifdef AGGRESSIVE_CHECK" only.

There are no functional changes in this patch

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 35 ++++++++++++++++++++++++-----------
 1 file changed, 24 insertions(+), 11 deletions(-)
fs/ext4/mballoc.c +24 -11
index a1429d5a1197..daa8411d6b61 100644
@@ -3978,6 +3978,29 @@ static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
 	mb_debug(sb, "goal %u blocks for locality group\n", ac->ac_g_ex.fe_len);
 }
 
+static inline void
+ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
+			  ext4_lblk_t start, ext4_lblk_t end)
+{
+	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
+	struct ext4_prealloc_space *tmp_pa;
+	ext4_lblk_t tmp_pa_start, tmp_pa_end;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+		spin_lock(&tmp_pa->pa_lock);
+		if (tmp_pa->pa_deleted == 0) {
+			tmp_pa_start = tmp_pa->pa_lstart;
+			tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
+
+			BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
+		}
+		spin_unlock(&tmp_pa->pa_lock);
+	}
+	rcu_read_unlock();
+}
+
 /*
  * Normalization means making request better in terms of
  * size and alignment

@@ -4135,17 +4158,7 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
 	size = end - start;
 
 	/* XXX: extra loop to check we really don't overlap preallocations */
-	rcu_read_lock();
-	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
-		spin_lock(&tmp_pa->pa_lock);
-		if (tmp_pa->pa_deleted == 0) {
-			tmp_pa_start = tmp_pa->pa_lstart;
-			tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
-			BUG_ON(!(start >= tmp_pa_end || end <= tmp_pa_start));
-		}
-		spin_unlock(&tmp_pa->pa_lock);
-	}
-	rcu_read_unlock();
+	ext4_mb_pa_assert_overlap(ac, start, end);
 
 	/*
 	 * In this function "start" and "size" are normalized for better
Signature · 1 line
2.31.1
author reviewer
#5 lore raw

PATCH v5 5/9 ext4: Abstract out overlap fix/check logic in ext4_mb_normalize_request()

In reply to Ojaswin Mujoo

Abstract out the logic of fixing PA overlaps in ext4_mb_normalize_request to improve readability of code. This also makes it easier to make changes to the overlap logic in future.

There are no functional changes in this patch

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 109 +++++++++++++++++++++++++++++-----------------
 1 file changed, 68 insertions(+), 41 deletions(-)
fs/ext4/mballoc.c +68 -41
index daa8411d6b61..37bf6507cbfd 100644
@@ -4001,6 +4001,73 @@ ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
 	rcu_read_unlock();
 }
 
+/*
+ * Given an allocation context "ac" and a range "start", "end", check
+ * and adjust boundaries if the range overlaps with any of the existing
+ * preallocatoins stored in the corresponding inode of the allocation context.
+ *
+ *Parameters:
+ *	ac			allocation context
+ *	start			start of the new range
+ *	end			end of the new range
+ */
+static inline void
+ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
+			 ext4_lblk_t *start, ext4_lblk_t *end)
+{
+	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
+	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
+	struct ext4_prealloc_space *tmp_pa;
+	ext4_lblk_t new_start, new_end;
+	ext4_lblk_t tmp_pa_start, tmp_pa_end;
+
+	new_start = *start;
+	new_end = *end;
+
+	/* check we don't cross already preallocated blocks */
+	rcu_read_lock();
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+		if (tmp_pa->pa_deleted)
+			continue;
+		spin_lock(&tmp_pa->pa_lock);
+		if (tmp_pa->pa_deleted) {
+			spin_unlock(&tmp_pa->pa_lock);
+			continue;
+		}
+
+		tmp_pa_start = tmp_pa->pa_lstart;
+		tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
+
+		/* PA must not overlap original request */
+		BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
+			ac->ac_o_ex.fe_logical < tmp_pa_start));
+
+		/* skip PAs this normalized request doesn't overlap with */
+		if (tmp_pa_start >= new_end || tmp_pa_end <= new_start) {
+			spin_unlock(&tmp_pa->pa_lock);
+			continue;
+		}
+		BUG_ON(tmp_pa_start <= new_start && tmp_pa_end >= new_end);
+
+		/* adjust start or end to be adjacent to this pa */
+		if (tmp_pa_end <= ac->ac_o_ex.fe_logical) {
+			BUG_ON(tmp_pa_end < new_start);
+			new_start = tmp_pa_end;
+		} else if (tmp_pa_start > ac->ac_o_ex.fe_logical) {
+			BUG_ON(tmp_pa_start > new_end);
+			new_end = tmp_pa_start;
+		}
+		spin_unlock(&tmp_pa->pa_lock);
+	}
+	rcu_read_unlock();
+
+	/* XXX: extra loop to check we really don't overlap preallocations */
+	ext4_mb_pa_assert_overlap(ac, new_start, new_end);
+
+	*start = new_start;
+	*end = new_end;
+}
+
 /*
  * Normalization means making request better in terms of
  * size and alignment

@@ -4016,9 +4083,6 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
 	loff_t size, start_off;
 	loff_t orig_size __maybe_unused;
 	ext4_lblk_t start;
-	struct ext4_inode_info *ei = EXT4_I(ac->ac_inode);
-	struct ext4_prealloc_space *tmp_pa;
-	ext4_lblk_t tmp_pa_start, tmp_pa_end;
 
 	/* do normalize only data requests, metadata requests
 	   do not need preallocation */

@@ -4119,47 +4183,10 @@ ext4_mb_normalize_request(struct ext4_allocation_context *ac,
 
 	end = start + size;
 
-	/* check we don't cross already preallocated blocks */
-	rcu_read_lock();
-	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
-		if (tmp_pa->pa_deleted)
-			continue;
-		spin_lock(&tmp_pa->pa_lock);
-		if (tmp_pa->pa_deleted) {
-			spin_unlock(&tmp_pa->pa_lock);
-			continue;
-		}
-
-		tmp_pa_start = tmp_pa->pa_lstart;
-		tmp_pa_end = tmp_pa->pa_lstart + EXT4_C2B(sbi, tmp_pa->pa_len);
-
-		/* PA must not overlap original request */
-		BUG_ON(!(ac->ac_o_ex.fe_logical >= tmp_pa_end ||
-			ac->ac_o_ex.fe_logical < tmp_pa_start));
-
-		/* skip PAs this normalized request doesn't overlap with */
-		if (tmp_pa_start >= end || tmp_pa_end <= start) {
-			spin_unlock(&tmp_pa->pa_lock);
-			continue;
-		}
-		BUG_ON(tmp_pa_start <= start && tmp_pa_end >= end);
+	ext4_mb_pa_adjust_overlap(ac, &start, &end);
 
-		/* adjust start or end to be adjacent to this pa */
-		if (tmp_pa_end <= ac->ac_o_ex.fe_logical) {
-			BUG_ON(tmp_pa_end < start);
-			start = tmp_pa_end;
-		} else if (tmp_pa_start > ac->ac_o_ex.fe_logical) {
-			BUG_ON(tmp_pa_start > end);
-			end = tmp_pa_start;
-		}
-		spin_unlock(&tmp_pa->pa_lock);
-	}
-	rcu_read_unlock();
 	size = end - start;
 
-	/* XXX: extra loop to check we really don't overlap preallocations */
-	ext4_mb_pa_assert_overlap(ac, start, end);
-
 	/*
 	 * In this function "start" and "size" are normalized for better
 	 * alignment and length such that we could preallocate more blocks.
Signature · 1 line
2.31.1
author reviewer
#6 lore raw

PATCH v5 6/9 ext4: Fix best extent lstart adjustment logic in ext4_mb_new_inode_pa()

In reply to Ojaswin Mujoo

When the length of best extent found is less than the length of goal extent we need to make sure that the best extent atleast covers the start of the original request. This is done by adjusting the ac_b_ex.fe_logical (logical start) of the extent.

While doing so, the current logic sometimes results in the best extent's logical range overflowing the goal extent. Since this best extent is later added to the inode preallocation list, we have a possibility of introducing overlapping preallocations. This is discussed in detail here [1].

To fix this, replace the existing logic with the below logic for adjusting best extent as it keeps fragmentation in check while ensuring logical range of best extent doesn't overflow out of goal extent:

1. Check if best extent can be kept at end of goal range and still cover original start.
2. Else, check if best extent can be kept at start of goal range and still cover original start.
3. Else, keep the best extent at start of original request.

Also, add a few extra BUG_ONs that might help catch errors faster.

[1] https://lore.kernel.org/r/[email protected]

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 49 ++++++++++++++++++++++++++++++-----------------
 1 file changed, 31 insertions(+), 18 deletions(-)
fs/ext4/mballoc.c +31 -18
index 37bf6507cbfd..1304c95d8c59 100644
@@ -4328,6 +4328,7 @@ static void ext4_mb_use_inode_pa(struct ext4_allocation_context *ac,
 	BUG_ON(start < pa->pa_pstart);
 	BUG_ON(end > pa->pa_pstart + EXT4_C2B(sbi, pa->pa_len));
 	BUG_ON(pa->pa_free < len);
+	BUG_ON(ac->ac_b_ex.fe_len <= 0);
 	pa->pa_free -= len;
 
 	mb_debug(ac->ac_sb, "use %llu/%d from inode pa %p\n", start, len, pa);

@@ -4666,10 +4667,8 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
 	pa = ac->ac_pa;
 
 	if (ac->ac_b_ex.fe_len < ac->ac_g_ex.fe_len) {
-		int winl;
-		int wins;
-		int win;
-		int offs;
+		int new_bex_start;
+		int new_bex_end;
 
 		/* we can't allocate as much as normalizer wants.
 		 * so, found space must get proper lstart

@@ -4677,26 +4676,40 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
 		BUG_ON(ac->ac_g_ex.fe_logical > ac->ac_o_ex.fe_logical);
 		BUG_ON(ac->ac_g_ex.fe_len < ac->ac_o_ex.fe_len);
 
-		/* we're limited by original request in that
-		 * logical block must be covered any way
-		 * winl is window we can move our chunk within */
-		winl = ac->ac_o_ex.fe_logical - ac->ac_g_ex.fe_logical;
+		/*
+		 * Use the below logic for adjusting best extent as it keeps
+		 * fragmentation in check while ensuring logical range of best
+		 * extent doesn't overflow out of goal extent:
+		 *
+		 * 1. Check if best ex can be kept at end of goal and still
+		 *    cover original start
+		 * 2. Else, check if best ex can be kept at start of goal and
+		 *    still cover original start
+		 * 3. Else, keep the best ex at start of original request.
+		 */
+		new_bex_end = ac->ac_g_ex.fe_logical +
+			EXT4_C2B(sbi, ac->ac_g_ex.fe_len);
+		new_bex_start = new_bex_end - EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
+		if (ac->ac_o_ex.fe_logical >= new_bex_start)
+			goto adjust_bex;
 
-		/* also, we should cover whole original request */
-		wins = EXT4_C2B(sbi, ac->ac_b_ex.fe_len - ac->ac_o_ex.fe_len);
+		new_bex_start = ac->ac_g_ex.fe_logical;
+		new_bex_end =
+			new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
+		if (ac->ac_o_ex.fe_logical < new_bex_end)
+			goto adjust_bex;
 
-		/* the smallest one defines real window */
-		win = min(winl, wins);
+		new_bex_start = ac->ac_o_ex.fe_logical;
+		new_bex_end =
+			new_bex_start + EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
 
-		offs = ac->ac_o_ex.fe_logical %
-			EXT4_C2B(sbi, ac->ac_b_ex.fe_len);
-		if (offs && offs < win)
-			win = offs;
+adjust_bex:
+		ac->ac_b_ex.fe_logical = new_bex_start;
 
-		ac->ac_b_ex.fe_logical = ac->ac_o_ex.fe_logical -
-			EXT4_NUM_B2C(sbi, win);
 		BUG_ON(ac->ac_o_ex.fe_logical < ac->ac_b_ex.fe_logical);
 		BUG_ON(ac->ac_o_ex.fe_len > ac->ac_b_ex.fe_len);
+		BUG_ON(new_bex_end > (ac->ac_g_ex.fe_logical +
+				      EXT4_C2B(sbi, ac->ac_g_ex.fe_len)));
 	}
 
 	pa->pa_lstart = ac->ac_b_ex.fe_logical;
Signature · 1 line
2.31.1
author reviewer
#7 lore raw

PATCH v5 7/9 ext4: Convert pa->pa_inode_list and pa->pa_obj_lock into a union

In reply to Ojaswin Mujoo

** Splitting pa->pa_inode_list **

Currently, we use the same pa->pa_inode_list to add a pa to either the inode preallocation list or the locality group preallocation list. For better clarity, split this list into a union of 2 list_heads and use either of the them based on the type of pa.

** Splitting pa->pa_obj_lock **

Currently, pa->pa_obj_lock is either assigned &ei->i_prealloc_lock for inode PAs or lg_prealloc_lock for lg PAs, and is then used to lock the lists containing these PAs. Make the distinction between the 2 PA types clear by changing this lock to a union of 2 locks. Explicitly use the pa_lock_node.inode_lock for inode PAs and pa_lock_node.lg_lock for lg PAs.

This patch is required so that the locality group preallocation code remains the same as in upcoming patches we are going to make changes to inode preallocation code to move from list to rbtree based implementation. This patch also makes it easier to review the upcoming patches.

There are no functional changes in this patch.

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/mballoc.c | 76 +++++++++++++++++++++++++++--------------------
 fs/ext4/mballoc.h | 10 +++++--
 2 files changed, 52 insertions(+), 34 deletions(-)
fs/ext4/mballoc.c +44 -32
index 1304c95d8c59..711661219473 100644
@@ -3988,7 +3988,7 @@ ext4_mb_pa_assert_overlap(struct ext4_allocation_context *ac,
 	ext4_lblk_t tmp_pa_start, tmp_pa_end;
 
 	rcu_read_lock();
-	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_node.inode_list) {
 		spin_lock(&tmp_pa->pa_lock);
 		if (tmp_pa->pa_deleted == 0) {
 			tmp_pa_start = tmp_pa->pa_lstart;

@@ -4026,7 +4026,7 @@ ext4_mb_pa_adjust_overlap(struct ext4_allocation_context *ac,
 
 	/* check we don't cross already preallocated blocks */
 	rcu_read_lock();
-	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_node.inode_list) {
 		if (tmp_pa->pa_deleted)
 			continue;
 		spin_lock(&tmp_pa->pa_lock);

@@ -4408,7 +4408,7 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 
 	/* first, try per-file preallocation */
 	rcu_read_lock();
-	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_inode_list) {
+	list_for_each_entry_rcu(tmp_pa, &ei->i_prealloc_list, pa_node.inode_list) {
 
 		/* all fields in this condition don't change,
 		 * so we can skip locking for them */

@@ -4465,7 +4465,7 @@ ext4_mb_use_preallocated(struct ext4_allocation_context *ac)
 	for (i = order; i < PREALLOC_TB_SIZE; i++) {
 		rcu_read_lock();
 		list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[i],
-					pa_inode_list) {
+					pa_node.lg_list) {
 			spin_lock(&tmp_pa->pa_lock);
 			if (tmp_pa->pa_deleted == 0 &&
 					tmp_pa->pa_free >= ac->ac_o_ex.fe_len) {

@@ -4639,9 +4639,15 @@ static void ext4_mb_put_pa(struct ext4_allocation_context *ac,
 	list_del(&pa->pa_group_list);
 	ext4_unlock_group(sb, grp);
 
-	spin_lock(pa->pa_obj_lock);
-	list_del_rcu(&pa->pa_inode_list);
-	spin_unlock(pa->pa_obj_lock);
+	if (pa->pa_type == MB_INODE_PA) {
+		spin_lock(pa->pa_node_lock.inode_lock);
+		list_del_rcu(&pa->pa_node.inode_list);
+		spin_unlock(pa->pa_node_lock.inode_lock);
+	} else {
+		spin_lock(pa->pa_node_lock.lg_lock);
+		list_del_rcu(&pa->pa_node.lg_list);
+		spin_unlock(pa->pa_node_lock.lg_lock);
+	}
 
 	call_rcu(&(pa)->u.pa_rcu, ext4_mb_pa_callback);
 }

@@ -4717,7 +4723,7 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
 	pa->pa_len = ac->ac_b_ex.fe_len;
 	pa->pa_free = pa->pa_len;
 	spin_lock_init(&pa->pa_lock);
-	INIT_LIST_HEAD(&pa->pa_inode_list);
+	INIT_LIST_HEAD(&pa->pa_node.inode_list);
 	INIT_LIST_HEAD(&pa->pa_group_list);
 	pa->pa_deleted = 0;
 	pa->pa_type = MB_INODE_PA;

@@ -4732,14 +4738,14 @@ ext4_mb_new_inode_pa(struct ext4_allocation_context *ac)
 	ei = EXT4_I(ac->ac_inode);
 	grp = ext4_get_group_info(sb, ac->ac_b_ex.fe_group);
 
-	pa->pa_obj_lock = &ei->i_prealloc_lock;
+	pa->pa_node_lock.inode_lock = &ei->i_prealloc_lock;
 	pa->pa_inode = ac->ac_inode;
 
 	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);
 
-	spin_lock(pa->pa_obj_lock);
-	list_add_rcu(&pa->pa_inode_list, &ei->i_prealloc_list);
-	spin_unlock(pa->pa_obj_lock);
+	spin_lock(pa->pa_node_lock.inode_lock);
+	list_add_rcu(&pa->pa_node.inode_list, &ei->i_prealloc_list);
+	spin_unlock(pa->pa_node_lock.inode_lock);
 	atomic_inc(&ei->i_prealloc_active);
 }
 

@@ -4767,7 +4773,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
 	pa->pa_len = ac->ac_b_ex.fe_len;
 	pa->pa_free = pa->pa_len;
 	spin_lock_init(&pa->pa_lock);
-	INIT_LIST_HEAD(&pa->pa_inode_list);
+	INIT_LIST_HEAD(&pa->pa_node.lg_list);
 	INIT_LIST_HEAD(&pa->pa_group_list);
 	pa->pa_deleted = 0;
 	pa->pa_type = MB_GROUP_PA;

@@ -4783,7 +4789,7 @@ ext4_mb_new_group_pa(struct ext4_allocation_context *ac)
 	lg = ac->ac_lg;
 	BUG_ON(lg == NULL);
 
-	pa->pa_obj_lock = &lg->lg_prealloc_lock;
+	pa->pa_node_lock.lg_lock = &lg->lg_prealloc_lock;
 	pa->pa_inode = NULL;
 
 	list_add(&pa->pa_group_list, &grp->bb_prealloc_list);

@@ -4959,9 +4965,15 @@ ext4_mb_discard_group_preallocations(struct super_block *sb,
 	list_for_each_entry_safe(pa, tmp, &list, u.pa_tmp_list) {
 
 		/* remove from object (inode or locality group) */
-		spin_lock(pa->pa_obj_lock);
-		list_del_rcu(&pa->pa_inode_list);
-		spin_unlock(pa->pa_obj_lock);
+		if (pa->pa_type == MB_GROUP_PA) {
+			spin_lock(pa->pa_node_lock.lg_lock);
+			list_del_rcu(&pa->pa_node.lg_list);
+			spin_unlock(pa->pa_node_lock.lg_lock);
+		} else {
+			spin_lock(pa->pa_node_lock.inode_lock);
+			list_del_rcu(&pa->pa_node.inode_list);
+			spin_unlock(pa->pa_node_lock.inode_lock);
+		}
 
 		if (pa->pa_type == MB_GROUP_PA)
 			ext4_mb_release_group_pa(&e4b, pa);

@@ -5023,8 +5035,8 @@ void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
 	spin_lock(&ei->i_prealloc_lock);
 	while (!list_empty(&ei->i_prealloc_list) && needed) {
 		pa = list_entry(ei->i_prealloc_list.prev,
-				struct ext4_prealloc_space, pa_inode_list);
-		BUG_ON(pa->pa_obj_lock != &ei->i_prealloc_lock);
+				struct ext4_prealloc_space, pa_node.inode_list);
+		BUG_ON(pa->pa_node_lock.inode_lock != &ei->i_prealloc_lock);
 		spin_lock(&pa->pa_lock);
 		if (atomic_read(&pa->pa_count)) {
 			/* this shouldn't happen often - nobody should

@@ -5041,7 +5053,7 @@ void ext4_discard_preallocations(struct inode *inode, unsigned int needed)
 		if (pa->pa_deleted == 0) {
 			ext4_mb_mark_pa_deleted(sb, pa);
 			spin_unlock(&pa->pa_lock);
-			list_del_rcu(&pa->pa_inode_list);
+			list_del_rcu(&pa->pa_node.inode_list);
 			list_add(&pa->u.pa_tmp_list, &list);
 			needed--;
 			continue;

@@ -5331,7 +5343,7 @@ ext4_mb_discard_lg_preallocations(struct super_block *sb,
 
 	spin_lock(&lg->lg_prealloc_lock);
 	list_for_each_entry_rcu(pa, &lg->lg_prealloc_list[order],
-				pa_inode_list,
+				pa_node.lg_list,
 				lockdep_is_held(&lg->lg_prealloc_lock)) {
 		spin_lock(&pa->pa_lock);
 		if (atomic_read(&pa->pa_count)) {

@@ -5354,7 +5366,7 @@ ext4_mb_discard_lg_preallocations(struct super_block *sb,
 		ext4_mb_mark_pa_deleted(sb, pa);
 		spin_unlock(&pa->pa_lock);
 
-		list_del_rcu(&pa->pa_inode_list);
+		list_del_rcu(&pa->pa_node.lg_list);
 		list_add(&pa->u.pa_tmp_list, &discard_list);
 
 		total_entries--;

@@ -5415,7 +5427,7 @@ static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
 	/* Add the prealloc space to lg */
 	spin_lock(&lg->lg_prealloc_lock);
 	list_for_each_entry_rcu(tmp_pa, &lg->lg_prealloc_list[order],
-				pa_inode_list,
+				pa_node.lg_list,
 				lockdep_is_held(&lg->lg_prealloc_lock)) {
 		spin_lock(&tmp_pa->pa_lock);
 		if (tmp_pa->pa_deleted) {

@@ -5424,8 +5436,8 @@ static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
 		}
 		if (!added && pa->pa_free < tmp_pa->pa_free) {
 			/* Add to the tail of the previous entry */
-			list_add_tail_rcu(&pa->pa_inode_list,
-						&tmp_pa->pa_inode_list);
+			list_add_tail_rcu(&pa->pa_node.lg_list,
+						&tmp_pa->pa_node.lg_list);
 			added = 1;
 			/*
 			 * we want to count the total

@@ -5436,7 +5448,7 @@ static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
 		lg_prealloc_count++;
 	}
 	if (!added)
-		list_add_tail_rcu(&pa->pa_inode_list,
+		list_add_tail_rcu(&pa->pa_node.lg_list,
 					&lg->lg_prealloc_list[order]);
 	spin_unlock(&lg->lg_prealloc_lock);
 

@@ -5492,9 +5504,9 @@ static int ext4_mb_release_context(struct ext4_allocation_context *ac)
 			 * doesn't grow big.
 			 */
 			if (likely(pa->pa_free)) {
-				spin_lock(pa->pa_obj_lock);
-				list_del_rcu(&pa->pa_inode_list);
-				spin_unlock(pa->pa_obj_lock);
+				spin_lock(pa->pa_node_lock.lg_lock);
+				list_del_rcu(&pa->pa_node.lg_list);
+				spin_unlock(pa->pa_node_lock.lg_lock);
 				ext4_mb_add_n_trim(ac);
 			}
 		}

@@ -5504,9 +5516,9 @@ static int ext4_mb_release_context(struct ext4_allocation_context *ac)
 			 * treat per-inode prealloc list as a lru list, then try
 			 * to trim the least recently used PA.
 			 */
-			spin_lock(pa->pa_obj_lock);
-			list_move(&pa->pa_inode_list, &ei->i_prealloc_list);
-			spin_unlock(pa->pa_obj_lock);
+			spin_lock(pa->pa_node_lock.inode_lock);
+			list_move(&pa->pa_node.inode_list, &ei->i_prealloc_list);
+			spin_unlock(pa->pa_node_lock.inode_lock);
 		}
 
 		ext4_mb_put_pa(ac, ac->ac_sb, pa);
fs/ext4/mballoc.h +8 -2
index dcda2a943cee..398a6688c341 100644
@@ -114,7 +114,10 @@ struct ext4_free_data {
 };
 
 struct ext4_prealloc_space {
-	struct list_head	pa_inode_list;
+	union {
+		struct list_head	inode_list; /* for inode PAs */
+		struct list_head	lg_list;	/* for lg PAs */
+	} pa_node;
 	struct list_head	pa_group_list;
 	union {
 		struct list_head pa_tmp_list;

@@ -128,7 +131,10 @@ struct ext4_prealloc_space {
 	ext4_grpblk_t		pa_len;		/* len of preallocated chunk */
 	ext4_grpblk_t		pa_free;	/* how many blocks are free */
 	unsigned short		pa_type;	/* pa type. inode or group */
-	spinlock_t		*pa_obj_lock;
+	union {
+		spinlock_t		*inode_lock;	/* locks the inode list holding this PA */
+		spinlock_t		*lg_lock;	/* locks the lg list holding this PA */
+	} pa_node_lock;
 	struct inode		*pa_inode;	/* hack, for history only */
 };
Signature · 1 line
2.31.1
author reviewer
#8 lore raw

PATCH v5 8/9 ext4: Use rbtrees to manage PAs instead of inode i_prealloc_list

In reply to Ojaswin Mujoo

Currently, the kernel uses i_prealloc_list to hold all the inode preallocations. This is known to cause degradation in performance in workloads which perform large number of sparse writes on a single file. This is mainly because functions like ext4_mb_normalize_request() and ext4_mb_use_preallocated() iterate over this complete list, resulting in slowdowns when large number of PAs are present.

Patch 27bc446e2 partially fixed this by enforcing a limit of 512 for the inode preallocation list and adding logic to continually trim the list if it grows above the threshold, however our testing revealed that a hardcoded value is not suitable for all kinds of workloads.

To optimize this, add an rbtree to the inode and hold the inode preallocations in this rbtree. This will make iterating over inode PAs faster and scale much better than a linked list. Additionally, we also had to remove the LRU logic that was added during trimming of the list (in ext4_mb_release_context()) as it will add extra overhead in rbtree. The discards now happen in the lowest-logical-offset-first order.

** Locking notes **

With the introduction of rbtree to maintain inode PAs, we can't use RCU to walk the tree for searching since it can result in partial traversals which might miss some nodes(or entire subtrees) while discards happen in parallel (which happens under a lock). Hence this patch converts the ei->i_prealloc_lock spin_lock to rw_lock.

Almost all the codepaths that read/modify the PA rbtrees are protected by the higher level inode->i_data_sem (except ext4_mb_discard_group_preallocations() and ext4_clear_inode()) IIUC, the only place we need lock protection is when one thread is reading "searching" the PA rbtree (earlier protected under rcu_read_lock()) and another is "deleting" the PAs in ext4_mb_discard_group_preallocations() function (which iterates all the PAs using the grp->bb_prealloc_list and deletes PAs from the tree without taking any inode lock (i_data_sem)).

So, this patch converts all rcu_read_lock/unlock() paths for inode list PA to use read_lock() and all places where we were using ei->i_prealloc_lock spinlock will now be using write_lock().

Note that this makes the fast path (searching of the right PA e.g. ext4_mb_use_preallocated() or ext4_mb_normalize_request()), now use read_lock() instead of rcu_read_lock/unlock(). Ths also will now block due to slow discard path (ext4_mb_discard_group_preallocations()) which uses write_lock().

But this is not as bad as it looks. This is because -

1. The slow path only occurs when the normal allocation failed and we can say that we are low on disk space. One can argue this scenario won't be much frequent.

2. ext4_mb_discard_group_preallocations(), locks and unlocks the rwlock for deleting every individual PA. This gives enough opportunity for the fast path to acquire the read_lock for searching the PA inode list.

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 fs/ext4/ext4.h    |   4 +-
 fs/ext4/mballoc.c | 286 +++++++++++++++++++++++++++++++++-------------
 fs/ext4/mballoc.h |   6 +-
 fs/ext4/super.c   |   4 +-
 4 files changed, 214 insertions(+), 86 deletions(-)
fs/ext4/ext4.h +2 -2
index 9b2cfc32cf78..8cf955b1dca5 100644
@@ -1120,8 +1120,8 @@ struct ext4_inode_info {
 
 	/* mballoc */
 	atomic_t i_prealloc_active;
-	struct list_head i_prealloc_list;
-	spinlock_t i_prealloc_lock;
+	struct rb_root i_prealloc_node;
+	rwlock_t i_prealloc_lock;
 
 	/* extents status tree */
 	struct ext4_es_tree i_es_tree;
fs/ext4/mballoc.c +207 -79

460 lines not on this page · raw message

fs/ext4/mballoc.h +3 -3
index 398a6688c341..f8e8ee493867 100644
@@ -115,7 +115,7 @@ struct ext4_free_data {
 
 struct ext4_prealloc_space {
 	union {
-		struct list_head	inode_list; /* for inode PAs */
+		struct rb_node	inode_node;		/* for inode PA rbtree */
 		struct list_head	lg_list;	/* for lg PAs */
 	} pa_node;
 	struct list_head	pa_group_list;

@@ -132,10 +132,10 @@ struct ext4_prealloc_space {
 	ext4_grpblk_t		pa_free;	/* how many blocks are free */
 	unsigned short		pa_type;	/* pa type. inode or group */
 	union {
-		spinlock_t		*inode_lock;	/* locks the inode list holding this PA */
+		rwlock_t		*inode_lock;	/* locks the rbtree holding this PA */
 		spinlock_t		*lg_lock;	/* locks the lg list holding this PA */
 	} pa_node_lock;
-	struct inode		*pa_inode;	/* hack, for history only */
+	struct inode		*pa_inode;	/* used to get the inode during group discard */
 };
 
 enum {
fs/ext4/super.c +2 -2
index f226f8ab469b..54fb600bf51f 100644
@@ -1325,9 +1325,9 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 	inode_set_iversion(&ei->vfs_inode, 1);
 	ei->i_flags = 0;
 	spin_lock_init(&ei->i_raw_lock);
-	INIT_LIST_HEAD(&ei->i_prealloc_list);
+	ei->i_prealloc_node = RB_ROOT;
 	atomic_set(&ei->i_prealloc_active, 0);
-	spin_lock_init(&ei->i_prealloc_lock);
+	rwlock_init(&ei->i_prealloc_lock);
 	ext4_es_init_tree(&ei->i_es_tree);
 	rwlock_init(&ei->i_es_lock);
 	INIT_LIST_HEAD(&ei->i_es_list);
Signature · 1 line
2.31.1
author reviewer
#9 lore raw

PATCH v5 9/9 ext4: Remove the logic to trim inode PAs

In reply to Ojaswin Mujoo

Earlier, inode PAs were stored in a linked list. This caused a need to periodically trim the list down inorder to avoid growing it to a very large size, as this would severly affect performance during list iteration.

Recent patches changed this list to an rbtree, and since the tree scales up much better, we no longer need to have the trim functionality, hence remove it.

Signed-off-by Ojaswin Mujoo <[email protected]>
Reviewed-by Ritesh Harjani (IBM) <[email protected]>
Reviewed-by Jan Kara <[email protected]>
---
 Documentation/admin-guide/ext4.rst |  3 ---
 fs/ext4/ext4.h                     |  1 -
 fs/ext4/mballoc.c                  | 20 --------------------
 fs/ext4/mballoc.h                  |  5 -----
 fs/ext4/sysfs.c                    |  2 --
 5 files changed, 31 deletions(-)
Documentation/admin-guide/ext4.rst +0 -3
index 4c559e08d11e..5740d85439ff 100644
@@ -489,9 +489,6 @@ Files in /sys/fs/ext4/<devname>:
         multiple of this tuning parameter if the stripe size is not set in the
         ext4 superblock
 
-  mb_max_inode_prealloc
-        The maximum length of per-inode ext4_prealloc_space list.
-
   mb_max_to_scan
         The maximum number of extents the multiblock allocator will search to
         find the best extent.
fs/ext4/ext4.h +0 -1
index 8cf955b1dca5..993d3284b430 100644
@@ -1613,7 +1613,6 @@ struct ext4_sb_info {
 	unsigned int s_mb_stats;
 	unsigned int s_mb_order2_reqs;
 	unsigned int s_mb_group_prealloc;
-	unsigned int s_mb_max_inode_prealloc;
 	unsigned int s_max_dir_size_kb;
 	/* where last allocation was done - for stream allocation */
 	unsigned long s_mb_last_group;
fs/ext4/mballoc.c +0 -20
index f7603dafb672..d5837c5acbc7 100644
@@ -3417,7 +3417,6 @@ int ext4_mb_init(struct super_block *sb)
 	sbi->s_mb_stats = MB_DEFAULT_STATS;
 	sbi->s_mb_stream_request = MB_DEFAULT_STREAM_THRESHOLD;
 	sbi->s_mb_order2_reqs = MB_DEFAULT_ORDER2_REQS;
-	sbi->s_mb_max_inode_prealloc = MB_DEFAULT_MAX_INODE_PREALLOC;
 	/*
 	 * The default group preallocation is 512, which for 4k block
 	 * sizes translates to 2 megabytes.  However for bigalloc file

@@ -5600,29 +5599,11 @@ static void ext4_mb_add_n_trim(struct ext4_allocation_context *ac)
 	return ;
 }
 
-/*
- * if per-inode prealloc list is too long, trim some PA
- */
-static void ext4_mb_trim_inode_pa(struct inode *inode)
-{
-	struct ext4_inode_info *ei = EXT4_I(inode);
-	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
-	int count, delta;
-
-	count = atomic_read(&ei->i_prealloc_active);
-	delta = (sbi->s_mb_max_inode_prealloc >> 2) + 1;
-	if (count > sbi->s_mb_max_inode_prealloc + delta) {
-		count -= sbi->s_mb_max_inode_prealloc;
-		ext4_discard_preallocations(inode, count);
-	}
-}
-
 /*
  * release all resource we used in allocation
  */
 static int ext4_mb_release_context(struct ext4_allocation_context *ac)
 {
-	struct inode *inode = ac->ac_inode;
 	struct ext4_sb_info *sbi = EXT4_SB(ac->ac_sb);
 	struct ext4_prealloc_space *pa = ac->ac_pa;
 	if (pa) {

@@ -5658,7 +5639,6 @@ static int ext4_mb_release_context(struct ext4_allocation_context *ac)
 	if (ac->ac_flags & EXT4_MB_HINT_GROUP_ALLOC)
 		mutex_unlock(&ac->ac_lg->lg_mutex);
 	ext4_mb_collect_stats(ac);
-	ext4_mb_trim_inode_pa(inode);
 	return 0;
 }
fs/ext4/mballoc.h +0 -5
index f8e8ee493867..6d85ee8674a6 100644
@@ -73,11 +73,6 @@
  */
 #define MB_DEFAULT_GROUP_PREALLOC	512
 
-/*
- * maximum length of inode prealloc list
- */
-#define MB_DEFAULT_MAX_INODE_PREALLOC	512
-
 /*
  * Number of groups to search linearly before performing group scanning
  * optimization.
fs/ext4/sysfs.c +0 -2
index 12d6252e3e22..3042bc605bbf 100644
@@ -214,7 +214,6 @@ EXT4_RW_ATTR_SBI_UI(mb_min_to_scan, s_mb_min_to_scan);
 EXT4_RW_ATTR_SBI_UI(mb_order2_req, s_mb_order2_reqs);
 EXT4_RW_ATTR_SBI_UI(mb_stream_req, s_mb_stream_request);
 EXT4_RW_ATTR_SBI_UI(mb_group_prealloc, s_mb_group_prealloc);
-EXT4_RW_ATTR_SBI_UI(mb_max_inode_prealloc, s_mb_max_inode_prealloc);
 EXT4_RW_ATTR_SBI_UI(mb_max_linear_groups, s_mb_max_linear_groups);
 EXT4_RW_ATTR_SBI_UI(extent_max_zeroout_kb, s_extent_max_zeroout_kb);
 EXT4_ATTR(trigger_fs_error, 0200, trigger_test_error);

@@ -264,7 +263,6 @@ static struct attribute *ext4_attrs[] = {
 	ATTR_LIST(mb_order2_req),
 	ATTR_LIST(mb_stream_req),
 	ATTR_LIST(mb_group_prealloc),
-	ATTR_LIST(mb_max_inode_prealloc),
 	ATTR_LIST(mb_max_linear_groups),
 	ATTR_LIST(max_writeback_mb_bump),
 	ATTR_LIST(extent_max_zeroout_kb),
Signature · 1 line
2.31.1