Skip to contentSkip to conversation

Introduce per-profile available space array to avoid over-confident can_overcommit()

PATCH Superseded Qu Wenruo 4 participants 11 messages Jan 9, 2020 – Jan 18, 2020 Btrfs file system

v5 of 9 Superseded Latest is v9 →

lore
Share & tools
lore

Qu Wenruo

author
#0 lore raw

v5 cover Introduce per-profile available space array to avoid over-confident can_overcommit()

There are several bug reports of ENOSPC error in
btrfs_run_delalloc_range().

With some extra info from one reporter, it turns out that can_overcommit() is using a wrong way to calculate allocatable metadata space.

The most typical case would look like:
devid 1 unallocated: 1G
devid 2 unallocated: 10G
metadata profile: RAID1

In above case, we can at most allocate 1G chunk for metadata, due to unbalanced disk free space.
But current can_overcommit() uses factor based calculation, which never consider the disk free space balance.

To address this problem, here comes the per-profile available space array, which gets updated every time a chunk get allocated/removed or a device get grown or shrunk.

This provides a quick way for hotter place like can_overcommit() to grab an estimation on how many bytes it can over-commit.

The per-profile available space calculation tries to keep the behavior of chunk allocator, thus it can handle uneven disks pretty well.

And statfs() can also grab that pre-calculated value for instance usage. For metadata over-commit, statfs() falls back to factor based educated guess method.
Since over-commit can only happen when we have unallocated space, the problem caused by over-commit should only be a first world problem.

Changelog:
v1:
- Fix a bug where we forgot to update per-profile array after allocating a chunk.
To avoid ABBA deadlock, this introduce a small windows at the end __btrfs_alloc_chunk(), it's not elegant but should be good enough before we rework chunk and device list mutex.

- Make statfs() to use virtual chunk allocator to do better estimation Now statfs() can report not only more accurate result, but can also handle RAID5/6 better.

v2:
- Fix a deadlock caused by acquiring device_list_mutex under __btrfs_alloc_chunk()
There is no need to acquire device_list_mutex when holding chunk_mutex.
Fix it and remove the lockdep assert.

v3:
- Use proper chunk_mutex instead of device_list_mutex
Since they are protecting two different things, and we only care about alloc_list, we should only use chunk_mutex.
With improved lock situation, it's easier to fold calc_per_profile_available() calls into the first patch.

- Add performance benchmark for statfs() modification
As Facebook seems to run into some problems with statfs() calls, add some basic ftrace results.

v4:
- Keep the lock-free design for statfs()
As extra sleeping in statfs() may not be a good idea, keep the old lock-free design, and use factor based calculation as fall back.

v5:
- Enhance btrfs_update_device() error handling in btrfs_grow_device() Revert device size to prevent possible mismatch.

- Ensure all failure caused by calc_per_profile_available() is the same with existing error handling
So no new failure pattern.

- Fix a bug where chunk_mutex is not released in btrfs_shrink_device()

Qu Wenruo (4):
btrfs: Reset device size when btrfs_update_device() failed in
btrfs_grow_device()
btrfs: Introduce per-profile available space facility
btrfs: space-info: Use per-profile available space in can_overcommit()
btrfs: statfs: Use pre-calculated per-profile available space

 fs/btrfs/space-info.c |  15 ++-
 fs/btrfs/super.c      | 182 +++++++++------------------------
 fs/btrfs/volumes.c    | 229 ++++++++++++++++++++++++++++++++++++++----
 fs/btrfs/volumes.h    |  11 ++
 4 files changed, 274 insertions(+), 163 deletions(-)
Signature · 1 line
2.24.1
author
#1 lore raw

PATCH v5 1/4 btrfs: Reset device size when btrfs_update_device() failed in btrfs_grow_device()

When btrfs_update_device() failed due to ENOMEM, we didn't reset device size back to its original size, causing the in-memory device size larger than original.

If somehow the memory pressure get solved, and the fs committed, since the device item is not updated, but super block total size get updated, it would cause mount failure due to size mismatch.

So here revert device size and super size to its original size when btrfs_update_device() failed, just like what we did in shrink_device().

---
 fs/btrfs/volumes.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
fs/btrfs/volumes.c +19 -1
index d8e5560db285..be638465f210 100644
@@ -2633,8 +2633,10 @@ int btrfs_grow_device(struct btrfs_trans_handle *trans,
 {
 	struct btrfs_fs_info *fs_info = device->fs_info;
 	struct btrfs_super_block *super_copy = fs_info->super_copy;
+	u64 old_device_size;
 	u64 old_total;
 	u64 diff;
+	int ret;
 
 	if (!test_bit(BTRFS_DEV_STATE_WRITEABLE, &device->dev_state))
 		return -EACCES;
@@ -2642,6 +2644,7 @@ int btrfs_grow_device(struct btrfs_trans_handle *trans,
 	new_size = round_down(new_size, fs_info->sectorsize);
 
 	mutex_lock(&fs_info->chunk_mutex);
+	old_device_size = device->total_bytes;
 	old_total = btrfs_super_total_bytes(super_copy);
 	diff = round_down(new_size - device->total_bytes, fs_info->sectorsize);
 
@@ -2663,7 +2666,22 @@ int btrfs_grow_device(struct btrfs_trans_handle *trans,
 			      &trans->transaction->dev_update_list);
 	mutex_unlock(&fs_info->chunk_mutex);
 
-	return btrfs_update_device(trans, device);
+	ret = btrfs_update_device(trans, device);
+	if (ret < 0) {
+		/*
+		 * Although we dropped chunk_mutex halfway for
+		 * btrfs_update_device(), we have FS_EXCL_OP bit to prevent
+		 * shrinking/growing race.
+		 * So we're safe to use the old size directly.
+		 */
+		mutex_lock(&fs_info->chunk_mutex);
+		btrfs_set_super_total_bytes(super_copy, old_total);
+		device->fs_devices->total_rw_bytes -= diff;
+		btrfs_device_set_total_bytes(device, old_device_size);
+		btrfs_device_set_disk_total_bytes(device, old_device_size);
+		mutex_unlock(&fs_info->chunk_mutex);
+	}
+	return ret;
 }
 
 static int btrfs_free_chunk(struct btrfs_trans_handle *trans, u64 chunk_offset)
Signature · 1 line
2.24.1
#2 lore raw

Re: [PATCH v5 1/4] btrfs: Reset device size when btrfs_update_device() failed in btrfs_grow_device()

Quoted 4 lines · Qu Wenruo

When btrfs_update_device() failed due to ENOMEM, we didn't reset device size back to its original size, causing the in-memory device size larger than original.

Quoted 8 lines

If somehow the memory pressure get solved, and the fs committed, since the device item is not updated, but super block total size get updated, it would cause mount failure due to size mismatch.

So here revert device size and super size to its original size when btrfs_update_device() failed, just like what we did in shrink_device().

Did you test this with error injection to make sure nothing else wonky came out of this? If you are going to fix this I'd rather it be in a different series because it's not necessarily related to what you are doing, and isn't any more broken with your other patches. The thing you are fixing in this series is important and I'd rather not hold it up on some error handling shenanigans. Thanks,

Josef

#3 lore raw

Re: [PATCH v5 1/4] btrfs: Reset device size when btrfs_update_device() failed in btrfs_grow_device()

Quoted 12 lines · Josef Bacik
Quoted 11 lines · Qu Wenruo

When btrfs_update_device() failed due to ENOMEM, we didn't reset device size back to its original size, causing the in-memory device size larger than original.

If somehow the memory pressure get solved, and the fs committed, since the device item is not updated, but super block total size get updated, it would cause mount failure due to size mismatch.

So here revert device size and super size to its original size when btrfs_update_device() failed, just like what we did in shrink_device().

Quoted 8 lines
Quoted 1 line

Did you test this with error injection to make sure nothing else wonky came out of this?  If you are going to fix this I'd rather it be in a different series because it's not necessarily related to what you are doing, and isn't any more broken with your other patches.  The thing you are fixing in this series is important and I'd rather not hold it up on some error handling shenanigans.  Thanks,

Yes, I have the same feeling.

But sometimes I just can't stop addressing the comment that makes sense.

And you're right, I forgot the error injection test, and it detects one bug. In the error handling path, I forgot the re-update per-profile available, causing df showing the grown size, not the old size.

To David, what's your idea on this?
I guess the patchset can't be backported anyway due to new infrastructure. I'm OK solving the problem by either removing this patch, or fix the bug exposed by the error injection.

Thanks,
Qu

Quoted 2 lines

Josef

#4 lore raw

Re: [PATCH v5 1/4] btrfs: Reset device size when btrfs_update_device() failed in btrfs_grow_device()

Quoted 28 lines · Qu Wenruo
Quoted 12 lines · Josef Bacik
Quoted 11 lines · Qu Wenruo

When btrfs_update_device() failed due to ENOMEM, we didn't reset device size back to its original size, causing the in-memory device size larger than original.

If somehow the memory pressure get solved, and the fs committed, since the device item is not updated, but super block total size get updated, it would cause mount failure due to size mismatch.

So here revert device size and super size to its original size when btrfs_update_device() failed, just like what we did in shrink_device().

Quoted 8 lines
Quoted 1 line

Did you test this with error injection to make sure nothing else wonky came out of this?  If you are going to fix this I'd rather it be in a different series because it's not necessarily related to what you are doing, and isn't any more broken with your other patches.  The thing you are fixing in this series is important and I'd rather not hold it up on some error handling shenanigans.  Thanks,

Yes, I have the same feeling.

But sometimes I just can't stop addressing the comment that makes sense.

Quoted 8 lines

And you're right, I forgot the error injection test, and it detects one bug. In the error handling path, I forgot the re-update per-profile available, causing df showing the grown size, not the old size.

To David, what's your idea on this?
I guess the patchset can't be backported anyway due to new infrastructure. I'm OK solving the problem by either removing this patch, or fix the bug exposed by the error injection.

Gentle ping.

Any feedback, David?

Thanks,
Qu

Quoted 7 lines

Thanks,
Qu

Quoted 2 lines

Josef

author
#5 lore raw

PATCH v5 2/4 btrfs: Introduce per-profile available space facility

In reply to Qu Wenruo

[PROBLEM]
There are some locations in btrfs requiring accurate estimation on how many new bytes can be allocated on unallocated space.

We have two types of estimation:
- Factor based calculation
Just use all unallocated space, divide by the profile factor One obvious user is can_overcommit().

- Chunk allocator like calculation
This will emulate the chunk allocator behavior, to get a proper estimation.
The only user is btrfs_calc_avail_data_space(), utilized by btrfs_statfs().
The problem is, that function is not generic purposed enough, can't handle things like RAID5/6.

Current factor based calculation can't handle the following case:
devid 1 unallocated: 1T
devid 2 unallocated: 10T
metadata type: RAID1

If using factor, we can use (1T + 10T) / 2 = 5.5T free space for metadata.
But in fact we can only get 1T free space, as we're limited by the smallest device for RAID1.

[SOLUTION]
This patch will introduce per-profile available space calculation, which can give an estimation based on chunk-allocator-like behavior.

The difference between it and chunk allocator is mostly on rounding and [0, 1M) reserved space handling, which shouldn't cause practical impact.

The newly introduced per-profile available space calculation will calculate available space for each type, using chunk-allocator like calculation.

With that facility, for above device layout we get the full available space array:
RAID10: 0 (not enough devices)
RAID1: 1T
RAID1C3: 0 (not enough devices)
RAID1C4: 0 (not enough devices)
DUP: 5.5T
RAID0: 2T
SINGLE: 11T
RAID5: 1T
RAID6: 0 (not enough devices)

Or for a more complex example:
devid 1 unallocated: 1T
devid 2 unallocated: 1T
devid 3 unallocated: 10T

We will get an array of:
RAID10: 0 (not enough devices)
RAID1: 2T
RAID1C3: 1T
RAID1C4: 0 (not enough devices)
DUP: 6T
RAID0: 3T
SINGLE: 12T
RAID5: 2T
RAID6: 0 (not enough devices)

And for the each profile , we go chunk allocator level calculation:
The pseudo code looks like:

  clear_virtual_used_space_of_all_rw_devices();
  do {
  	/*
  	 * The same as chunk allocator, despite used space,
  	 * we also take virtual used space into consideration.
  	 */
  	sort_device_with_virtual_free_space();

/*
* Unlike chunk allocator, we don't need to bother hole/stripe
* size, so we use the smallest device to make sure we can
* allocated as many stripes as regular chunk allocator */

  	stripe_size = device_with_smallest_free->avail_space;
	stripe_size = min(stripe_size, to_alloc / ndevs);

/*
* Allocate a virtual chunk, allocated virtual chunk will
* increase virtual used space, allow next iteration to
* properly emulate chunk allocator behavior. */
ret = alloc_virtual_chunk(stripe_size, &allocated_size); if (ret == 0)
avail += allocated_size;
} while (ret == 0)

As we always select the device with least free space, the device with the most space will be the first to be utilized, just like chunk allocator.
For above 1T + 10T device, we will allocate a 1T virtual chunk in the first iteration, then run out of device in next iteration.

Thus only get 1T free space for RAID1 type, just like what chunk allocator would do.

The patch will update such per-profile available space at the following timing:
- Mount time
- Chunk allocation
- Chunk removal
- Device grow
- Device shrink

Those timing are all protected by chunk_mutex, and what we do are only iterating in-memory only structures, no extra IO triggered, so the performance impact should be pretty small.

For the extra error handling, the principle is to keep the old behavior. That's to say, if old error handler would just return an error, then we follow it.
If the older error handler choose to abort transaction, then we follow it too.
So new failure mode is introduced.

---
 fs/btrfs/volumes.c | 209 ++++++++++++++++++++++++++++++++++++++++-----
 fs/btrfs/volumes.h |  11 +++
 2 files changed, 200 insertions(+), 20 deletions(-)
fs/btrfs/volumes.c +189 -20
index be638465f210..4e58a6b37848 100644
@@ -349,6 +349,7 @@ static struct btrfs_fs_devices *alloc_fs_devices(const u8 *fsid,
 	INIT_LIST_HEAD(&fs_devs->devices);
 	INIT_LIST_HEAD(&fs_devs->alloc_list);
 	INIT_LIST_HEAD(&fs_devs->fs_list);
+	spin_lock_init(&fs_devs->per_profile_lock);
 	if (fsid)
 		memcpy(fs_devs->fsid, fsid, BTRFS_FSID_SIZE);
 
@@ -2628,6 +2629,170 @@ static noinline int btrfs_update_device(struct btrfs_trans_handle *trans,
 	return ret;
 }
 
+/*
+ * sort the devices in descending order by max_avail, total_avail
+ */
+static int btrfs_cmp_device_info(const void *a, const void *b)
+{
+	const struct btrfs_device_info *di_a = a;
+	const struct btrfs_device_info *di_b = b;
+
+	if (di_a->max_avail > di_b->max_avail)
+		return -1;
+	if (di_a->max_avail < di_b->max_avail)
+		return 1;
+	if (di_a->total_avail > di_b->total_avail)
+		return -1;
+	if (di_a->total_avail < di_b->total_avail)
+		return 1;
+	return 0;
+}
+
+/*
+ * Return 0 if we allocated any virtual(*) chunk, and restore the size to
+ * @allocated_size
+ * Return -ENOSPC if we have no more space to allocate virtual chunk
+ *
+ * *: virtual chunk is a space holder for per-profile available space
+ *    calculator.
+ *    Such virtual chunks won't take on-disk space, thus called virtual, and
+ *    only affects per-profile available space calulation.
+ */
+static int alloc_virtual_chunk(struct btrfs_fs_info *fs_info,
+			       struct btrfs_device_info *devices_info,
+			       enum btrfs_raid_types type,
+			       u64 *allocated)
+{
+	const struct btrfs_raid_attr *raid_attr = &btrfs_raid_array[type];
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u64 stripe_size;
+	int i;
+	int ndevs = 0;
+
+	lockdep_assert_held(&fs_info->chunk_mutex);
+
+	/* Go through devices to collect their unallocated space */
+	list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
+		u64 avail;
+		if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
+					&device->dev_state) ||
+		    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state))
+			continue;
+
+		if (device->total_bytes > device->bytes_used +
+				device->virtual_allocated)
+			avail = device->total_bytes - device->bytes_used -
+				device->virtual_allocated;
+		else
+			avail = 0;
+
+		/* And exclude the [0, 1M) reserved space */
+		if (avail > SZ_1M)
+			avail -= SZ_1M;
+		else
+			avail = 0;
+
+		if (avail < fs_info->sectorsize)
+			continue;
+		/*
+		 * Unlike chunk allocator, we don't care about stripe or hole
+		 * size, so here we use @avail directly
+		 */
+		devices_info[ndevs].dev_offset = 0;
+		devices_info[ndevs].total_avail = avail;
+		devices_info[ndevs].max_avail = avail;
+		devices_info[ndevs].dev = device;
+		++ndevs;
+	}
+	sort(devices_info, ndevs, sizeof(struct btrfs_device_info),
+	     btrfs_cmp_device_info, NULL);
+	ndevs -= ndevs % raid_attr->devs_increment;
+	if (ndevs < raid_attr->devs_min)
+		return -ENOSPC;
+	if (raid_attr->devs_max)
+		ndevs = min(ndevs, (int)raid_attr->devs_max);
+	else
+		ndevs = min(ndevs, (int)BTRFS_MAX_DEVS(fs_info));
+
+	/*
+	 * Now allocate a virtual chunk using the unallocate space of the
+	 * device with the least unallocated space.
+	 */
+	stripe_size = round_down(devices_info[ndevs - 1].total_avail,
+				 fs_info->sectorsize);
+	if (stripe_size == 0)
+		return -ENOSPC;
+
+	for (i = 0; i < ndevs; i++)
+		devices_info[i].dev->virtual_allocated += stripe_size;
+	*allocated = stripe_size * (ndevs - raid_attr->nparity) /
+		     raid_attr->ncopies;
+	return 0;
+}
+
+static int calc_one_profile_avail(struct btrfs_fs_info *fs_info,
+				  enum btrfs_raid_types type)
+{
+	struct btrfs_device_info *devices_info = NULL;
+	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
+	struct btrfs_device *device;
+	u64 allocated;
+	u64 result = 0;
+	int ret = 0;
+
+	ASSERT(type >= 0 && type < BTRFS_NR_RAID_TYPES);
+
+	/* Not enough devices, quick exit, just update the result */
+	if (fs_devices->rw_devices < btrfs_raid_array[type].devs_min)
+		goto out;
+
+	devices_info = kcalloc(fs_devices->rw_devices, sizeof(*devices_info),
+			       GFP_NOFS);
+	if (!devices_info) {
+		ret = -ENOMEM;
+		goto out;
+	}
+	/* Clear virtual chunk used space for each device */
+	list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list)
+		device->virtual_allocated = 0;
+	while (ret == 0) {
+		ret = alloc_virtual_chunk(fs_info, devices_info, type,
+					  &allocated);
+		if (ret == 0)
+			result += allocated;
+	}
+	list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list)
+		device->virtual_allocated = 0;
+out:
+	kfree(devices_info);
+	if (ret < 0 && ret != -ENOSPC)
+		return ret;
+	spin_lock(&fs_devices->per_profile_lock);
+	fs_devices->per_profile_avail[type] = result;
+	spin_unlock(&fs_devices->per_profile_lock);
+	return 0;
+}
+
+/*
+ * Calculate the per-profile available space array.
+ *
+ * Return 0 if we succeeded updating the array.
+ * Return <0 if something went wrong. (ENOMEM)
+ */
+static int calc_per_profile_avail(struct btrfs_fs_info *fs_info)
+{
+	int i;
+	int ret;
+
+	for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
+		ret = calc_one_profile_avail(fs_info, i);
+		if (ret < 0)
+			break;
+	}
+	return ret;
+}
+
 int btrfs_grow_device(struct btrfs_trans_handle *trans,
 		      struct btrfs_device *device, u64 new_size)
 {
@@ -2664,9 +2829,13 @@ int btrfs_grow_device(struct btrfs_trans_handle *trans,
 	if (list_empty(&device->post_commit_list))
 		list_add_tail(&device->post_commit_list,
 			      &trans->transaction->dev_update_list);
+	ret = calc_per_profile_avail(fs_info);
 	mutex_unlock(&fs_info->chunk_mutex);
+	if (ret < 0)
+		goto out;
 
 	ret = btrfs_update_device(trans, device);
+out:
 	if (ret < 0) {
 		/*
 		 * Although we dropped chunk_mutex halfway for
@@ -2849,7 +3018,13 @@ int btrfs_remove_chunk(struct btrfs_trans_handle *trans, u64 chunk_offset)
 					device->bytes_used - dev_extent_len);
 			atomic64_add(dev_extent_len, &fs_info->free_chunk_space);
 			btrfs_clear_space_info_full(fs_info);
+			ret = calc_per_profile_avail(fs_info);
 			mutex_unlock(&fs_info->chunk_mutex);
+			if (ret < 0) {
+				mutex_unlock(&fs_devices->device_list_mutex);
+				btrfs_abort_transaction(trans, ret);
+				goto out;
+			}
 		}
 
 		ret = btrfs_update_device(trans, device);
@@ -4544,6 +4719,11 @@ int btrfs_shrink_device(struct btrfs_device *device, u64 new_size)
 		atomic64_sub(diff, &fs_info->free_chunk_space);
 	}
 
+	ret = calc_per_profile_avail(fs_info);
+	if (ret < 0) {
+		mutex_unlock(&fs_info->chunk_mutex);
+		goto done;
+	}
 	/*
 	 * Once the device's size has been set to the new size, ensure all
 	 * in-memory chunks are synced to disk so that the loop below sees them
@@ -4708,25 +4888,6 @@ static int btrfs_add_system_chunk(struct btrfs_fs_info *fs_info,
 	return 0;
 }
 
-/*
- * sort the devices in descending order by max_avail, total_avail
- */
-static int btrfs_cmp_device_info(const void *a, const void *b)
-{
-	const struct btrfs_device_info *di_a = a;
-	const struct btrfs_device_info *di_b = b;
-
-	if (di_a->max_avail > di_b->max_avail)
-		return -1;
-	if (di_a->max_avail < di_b->max_avail)
-		return 1;
-	if (di_a->total_avail > di_b->total_avail)
-		return -1;
-	if (di_a->total_avail < di_b->total_avail)
-		return 1;
-	return 0;
-}
-
 static void check_raid56_incompat_flag(struct btrfs_fs_info *info, u64 type)
 {
 	if (!(type & BTRFS_BLOCK_GROUP_RAID56_MASK))
@@ -5010,9 +5171,10 @@ static int __btrfs_alloc_chunk(struct btrfs_trans_handle *trans,
 	free_extent_map(em);
 	check_raid56_incompat_flag(info, type);
 	check_raid1c34_incompat_flag(info, type);
+	ret = calc_per_profile_avail(info);
 
 	kfree(devices_info);
-	return 0;
+	return ret;
 
 error_del_extent:
 	write_lock(&em_tree->lock);
@@ -7647,6 +7809,13 @@ int btrfs_verify_dev_extents(struct btrfs_fs_info *fs_info)
 
 	/* Ensure all chunks have corresponding dev extents */
 	ret = verify_chunk_dev_extent_mapping(fs_info);
+	if (ret < 0)
+		goto out;
+
+	/* All dev extents are verified, update per-profile available space */
+	mutex_lock(&fs_info->chunk_mutex);
+	ret = calc_per_profile_avail(fs_info);
+	mutex_unlock(&fs_info->chunk_mutex);
 out:
 	btrfs_free_path(path);
 	return ret;
fs/btrfs/volumes.h +11 -0
index fc1b564b9cfe..5cddfe7cfee8 100644
@@ -138,6 +138,13 @@ struct btrfs_device {
 	atomic_t dev_stat_values[BTRFS_DEV_STAT_VALUES_MAX];
 
 	struct extent_io_tree alloc_state;
+
+	/*
+	 * the "virtual" allocated space by virtual chunk allocator, which
+	 * is used to do accurate estimation on available space.
+	 * Doesn't affect real chunk allocator.
+	 */
+	u64 virtual_allocated;
 };
 
 /*
@@ -257,6 +264,10 @@ struct btrfs_fs_devices {
 	struct kobject fsid_kobj;
 	struct kobject *device_dir_kobj;
 	struct completion kobj_unregister;
+
+	/* Records per-type available space */
+	spinlock_t per_profile_lock;
+	u64 per_profile_avail[BTRFS_NR_RAID_TYPES];
 };
 
 #define BTRFS_BIO_INLINE_CSUM_SIZE	64
Signature · 1 line
2.24.1
Reviewed-by on v5 2/4
#6 lore raw

Re: [PATCH v5 2/4] btrfs: Introduce per-profile available space facility

Quoted 123 lines · Qu Wenruo

[PROBLEM]
There are some locations in btrfs requiring accurate estimation on how many new bytes can be allocated on unallocated space.

We have two types of estimation:
- Factor based calculation
Just use all unallocated space, divide by the profile factor One obvious user is can_overcommit().

- Chunk allocator like calculation
This will emulate the chunk allocator behavior, to get a proper estimation.
The only user is btrfs_calc_avail_data_space(), utilized by btrfs_statfs().
The problem is, that function is not generic purposed enough, can't handle things like RAID5/6.

Current factor based calculation can't handle the following case:
devid 1 unallocated: 1T
devid 2 unallocated: 10T
metadata type: RAID1

If using factor, we can use (1T + 10T) / 2 = 5.5T free space for metadata.
But in fact we can only get 1T free space, as we're limited by the smallest device for RAID1.

[SOLUTION]
This patch will introduce per-profile available space calculation, which can give an estimation based on chunk-allocator-like behavior.

The difference between it and chunk allocator is mostly on rounding and [0, 1M) reserved space handling, which shouldn't cause practical impact.

The newly introduced per-profile available space calculation will calculate available space for each type, using chunk-allocator like calculation.

With that facility, for above device layout we get the full available space array:

   RAID10:	0  (not enough devices)
   RAID1:	1T
   RAID1C3:	0  (not enough devices)
   RAID1C4:	0  (not enough devices)
   DUP:		5.5T
   RAID0:	2T
   SINGLE:	11T
   RAID5:	1T
   RAID6:	0  (not enough devices)

Or for a more complex example:
devid 1 unallocated: 1T
devid 2 unallocated: 1T
devid 3 unallocated: 10T

We will get an array of:

   RAID10:	0  (not enough devices)
   RAID1:	2T
   RAID1C3:	1T
   RAID1C4:	0  (not enough devices)
   DUP:		6T
   RAID0:	3T
   SINGLE:	12T
   RAID5:	2T
   RAID6:	0  (not enough devices)

And for the each profile , we go chunk allocator level calculation:
The pseudo code looks like:

   clear_virtual_used_space_of_all_rw_devices();
   do {
   	/*
   	 * The same as chunk allocator, despite used space,
   	 * we also take virtual used space into consideration.
   	 */
   	sort_device_with_virtual_free_space();
   	/*
   	 * Unlike chunk allocator, we don't need to bother hole/stripe
   	 * size, so we use the smallest device to make sure we can
   	 * allocated as many stripes as regular chunk allocator
   	 */
   	stripe_size = device_with_smallest_free->avail_space;
	stripe_size = min(stripe_size, to_alloc / ndevs);

   	/*
   	 * Allocate a virtual chunk, allocated virtual chunk will
   	 * increase virtual used space, allow next iteration to
   	 * properly emulate chunk allocator behavior.
   	 */
   	ret = alloc_virtual_chunk(stripe_size, &allocated_size);
   	if (ret == 0)
   		avail += allocated_size;
   } while (ret == 0)

As we always select the device with least free space, the device with the most space will be the first to be utilized, just like chunk allocator.
For above 1T + 10T device, we will allocate a 1T virtual chunk in the first iteration, then run out of device in next iteration.

Thus only get 1T free space for RAID1 type, just like what chunk allocator would do.

The patch will update such per-profile available space at the following timing:
- Mount time
- Chunk allocation
- Chunk removal
- Device grow
- Device shrink

Those timing are all protected by chunk_mutex, and what we do are only iterating in-memory only structures, no extra IO triggered, so the performance impact should be pretty small.

For the extra error handling, the principle is to keep the old behavior. That's to say, if old error handler would just return an error, then we follow it.
If the older error handler choose to abort transaction, then we follow it too.
So new failure mode is introduced.

Quoted 2 lines

Thanks,

Josef

#7 lore raw

Re: [PATCH v5 2/4] btrfs: Introduce per-profile available space facility

In reply to Qu Wenruo
Thank you for the patch! Yet something to improve:

Hi Qu,

Thank you for the patch! Yet something to improve:

[auto build test ERROR on v5.5-rc5]
[also build test ERROR on next-20200117]
[cannot apply to btrfs/next]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system. BTW, we also suggest to use '--base' option to specify the base tree in git format-patch, please see https://stackoverflow.com/a/37406982]

url:    https://github.com/0day-ci/linux/commits/Qu-Wenruo/Introduce-per-profile-available-space-array-to-avoid-over-confident-can_overcommit/20200110-082114
base:    c79f46a282390e0f5b306007bf7b11a46d529538
config: i386-randconfig-e003-20200118 (attached as .config)
compiler: gcc-7 (Debian 7.5.0-3) 7.5.0
reproduce:
        # save the attached .config to linux build tree
        make ARCH=i386 

If you fix the issue, kindly add following tag

All errors (new ones prefixed by >>):

>> ERROR: "__udivdi3" [fs/btrfs/btrfs.ko] undefined!

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org Intel Corporation

author
#8 lore raw

PATCH v5 3/4 btrfs: space-info: Use per-profile available space in can_overcommit()

In reply to Qu Wenruo

For the following disk layout, can_overcommit() can cause false confidence in available space:

devid 1 unallocated: 1T
devid 2 unallocated: 10T
metadata type: RAID1

As can_overcommit() simply uses unallocated space with factor to calculate the allocatable metadata chunk size.

can_overcommit() believes we still have 5.5T for metadata chunks, while the truth is, we only have 1T available for metadata chunks. This can lead to ENOSPC at run_delalloc_range() and cause transaction abort.

Since factor based calculation can't distinguish RAID1/RAID10 and DUP at all, we need proper chunk-allocator level awareness to do such estimation.

Thankfully, we have per-profile available space already calculated, just use that facility to avoid such false confidence.

---
 fs/btrfs/space-info.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
fs/btrfs/space-info.c +7 -8
index f09aa6ee9113..c26aba9e7124 100644
@@ -164,10 +164,10 @@ static int can_overcommit(struct btrfs_fs_info *fs_info,
 			  enum btrfs_reserve_flush_enum flush,
 			  bool system_chunk)
 {
+	enum btrfs_raid_types index;
 	u64 profile;
 	u64 avail;
 	u64 used;
-	int factor;
 
 	/* Don't overcommit when in mixed mode. */
 	if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
@@ -179,16 +179,15 @@ static int can_overcommit(struct btrfs_fs_info *fs_info,
 		profile = btrfs_metadata_alloc_profile(fs_info);
 
 	used = btrfs_space_info_used(space_info, true);
-	avail = atomic64_read(&fs_info->free_chunk_space);
 
 	/*
-	 * If we have dup, raid1 or raid10 then only half of the free
-	 * space is actually usable.  For raid56, the space info used
-	 * doesn't include the parity drive, so we don't have to
-	 * change the math
+	 * Grab avail space from per-profile array which should be as accurate
+	 * as chunk allocator.
 	 */
-	factor = btrfs_bg_type_to_factor(profile);
-	avail = div_u64(avail, factor);
+	index = btrfs_bg_flags_to_raid_index(profile);
+	spin_lock(&fs_info->fs_devices->per_profile_lock);
+	avail = fs_info->fs_devices->per_profile_avail[index];
+	spin_unlock(&fs_info->fs_devices->per_profile_lock);
 
 	/*
 	 * If we aren't flushing all things, let us overcommit up to
Signature · 1 line
2.24.1
author
#9 lore raw

PATCH v5 4/4 btrfs: statfs: Use pre-calculated per-profile available space

In reply to Qu Wenruo

Although btrfs_calc_avail_data_space() is trying to do an estimation on how many data chunks it can allocate, the estimation is far from perfect:

- Metadata over-commit is not considered at all
- Chunk allocation doesn't take RAID5/6 into consideration

This patch will change btrfs_calc_avail_data_space() to use pre-calculated per-profile available space.

This provides the following benefits:
- Accurate unallocated data space estimation, including RAID5/6 It's as accurate as chunk allocator, and can handle RAID5/6.

Although it still can't handle metadata over-commit that accurately, we still have fallback method for over-commit, by using factor based estimation.

The good news is, over-commit can only happen when we have enough unallocated space, so even we may not report byte accurate result when the fs is empty, the result will get more and more accurate when unallocated space is reducing.

So the metadata over-commit shouldn't cause too many problem.

Since we're keeping the old lock-free design, statfs should not experience any extra delay.

---
 fs/btrfs/super.c | 182 +++++++++++++----------------------------------
 1 file changed, 48 insertions(+), 134 deletions(-)
fs/btrfs/super.c +48 -134
index f452a94abdc3..b5a32339e544 100644
@@ -1894,118 +1894,53 @@ static inline void btrfs_descending_sort_devices(
  * The helper to calc the free space on the devices that can be used to store
  * file data.
  */
-static inline int btrfs_calc_avail_data_space(struct btrfs_fs_info *fs_info,
-					      u64 *free_bytes)
+static u64 btrfs_calc_avail_data_space(struct btrfs_fs_info *fs_info,
+				       u64 free_meta)
 {
-	struct btrfs_device_info *devices_info;
 	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
-	struct btrfs_device *device;
-	u64 type;
-	u64 avail_space;
-	u64 min_stripe_size;
-	int num_stripes = 1;
-	int i = 0, nr_devices;
-	const struct btrfs_raid_attr *rattr;
+	enum btrfs_raid_types data_type;
+	u64 data_profile = btrfs_data_alloc_profile(fs_info);
+	u64 data_avail;
+	u64 meta_rsv;
 
-	/*
-	 * We aren't under the device list lock, so this is racy-ish, but good
-	 * enough for our purposes.
-	 */
-	nr_devices = fs_info->fs_devices->open_devices;
-	if (!nr_devices) {
-		smp_mb();
-		nr_devices = fs_info->fs_devices->open_devices;
-		ASSERT(nr_devices);
-		if (!nr_devices) {
-			*free_bytes = 0;
-			return 0;
-		}
-	}
-
-	devices_info = kmalloc_array(nr_devices, sizeof(*devices_info),
-			       GFP_KERNEL);
-	if (!devices_info)
-		return -ENOMEM;
-
-	/* calc min stripe number for data space allocation */
-	type = btrfs_data_alloc_profile(fs_info);
-	rattr = &btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)];
-
-	if (type & BTRFS_BLOCK_GROUP_RAID0)
-		num_stripes = nr_devices;
-	else if (type & BTRFS_BLOCK_GROUP_RAID1)
-		num_stripes = 2;
-	else if (type & BTRFS_BLOCK_GROUP_RAID1C3)
-		num_stripes = 3;
-	else if (type & BTRFS_BLOCK_GROUP_RAID1C4)
-		num_stripes = 4;
-	else if (type & BTRFS_BLOCK_GROUP_RAID10)
-		num_stripes = 4;
+	spin_lock(&fs_info->global_block_rsv.lock);
+	meta_rsv = fs_info->global_block_rsv.size;
+	spin_unlock(&fs_info->global_block_rsv.lock);
 
-	/* Adjust for more than 1 stripe per device */
-	min_stripe_size = rattr->dev_stripes * BTRFS_STRIPE_LEN;
+	data_type = btrfs_bg_flags_to_raid_index(data_profile);
 
-	rcu_read_lock();
-	list_for_each_entry_rcu(device, &fs_devices->devices, dev_list) {
-		if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA,
-						&device->dev_state) ||
-		    !device->bdev ||
-		    test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state))
-			continue;
+	spin_lock(&fs_devices->per_profile_lock);
+	data_avail = fs_devices->per_profile_avail[data_type];
+	spin_unlock(&fs_devices->per_profile_lock);
 
-		if (i >= nr_devices)
-			break;
-
-		avail_space = device->total_bytes - device->bytes_used;
-
-		/* align with stripe_len */
-		avail_space = rounddown(avail_space, BTRFS_STRIPE_LEN);
-
-		/*
-		 * In order to avoid overwriting the superblock on the drive,
-		 * btrfs starts at an offset of at least 1MB when doing chunk
-		 * allocation.
-		 *
-		 * This ensures we have at least min_stripe_size free space
-		 * after excluding 1MB.
-		 */
-		if (avail_space <= SZ_1M + min_stripe_size)
-			continue;
-
-		avail_space -= SZ_1M;
-
-		devices_info[i].dev = device;
-		devices_info[i].max_avail = avail_space;
-
-		i++;
-	}
-	rcu_read_unlock();
-
-	nr_devices = i;
-
-	btrfs_descending_sort_devices(devices_info, nr_devices);
-
-	i = nr_devices - 1;
-	avail_space = 0;
-	while (nr_devices >= rattr->devs_min) {
-		num_stripes = min(num_stripes, nr_devices);
-
-		if (devices_info[i].max_avail >= min_stripe_size) {
-			int j;
-			u64 alloc_size;
-
-			avail_space += devices_info[i].max_avail * num_stripes;
-			alloc_size = devices_info[i].max_avail;
-			for (j = i + 1 - num_stripes; j <= i; j++)
-				devices_info[j].max_avail -= alloc_size;
-		}
-		i--;
-		nr_devices--;
+	/*
+	 * We have meta over-committed, do some wild guess using factor.
+	 *
+	 * To get an accurate result, we should allocate a metadata virtual
+	 * chunk first, then allocate data virtual chunks to get real
+	 * estimation.
+	 * But that needs chunk_mutex, which could be very slow to accquire.
+	 *
+	 * So here we trade for non-blocking statfs. And meta over-committing is
+	 * less a problem because:
+	 * - Meta over-commit only happens when we have unallocated space
+	 *   So no over-commit if we're low on available space.
+	 *
+	 * This may not be as accurate as virtual chunk based one, but it
+	 * should be good enough for statfs usage.
+	 */
+	if (free_meta < meta_rsv) {
+		u64 meta_needed = meta_rsv - free_meta;
+		int data_factor = btrfs_bg_type_to_factor(data_profile);
+		int meta_factor = btrfs_bg_type_to_factor(
+				btrfs_metadata_alloc_profile(fs_info));
+
+		if (data_avail < meta_needed * meta_factor / data_factor)
+			data_avail = 0;
+		else
+			data_avail -= meta_needed * meta_factor / data_factor;
 	}
-
-	kfree(devices_info);
-	*free_bytes = avail_space;
-	return 0;
+	return data_avail;
 }
 
 /*
@@ -2016,10 +1951,13 @@ static inline int btrfs_calc_avail_data_space(struct btrfs_fs_info *fs_info,
  *
  * Unused device space usage is based on simulating the chunk allocator
  * algorithm that respects the device sizes and order of allocations.  This is
- * a close approximation of the actual use but there are other factors that may
- * change the result (like a new metadata chunk).
+ * a very close approximation of the actual use.
+ * There are some inaccurate accounting for metadata over-commit, but it
+ * shouldn't cause big difference.
  *
- * If metadata is exhausted, f_bavail will be 0.
+ * There should be no way to exhaust metadata so that we can't even delete
+ * files. Thus the old "exhaused metadata means 0 f_bavail" behavior will be
+ * discard.
  */
 static int btrfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 {
@@ -2033,8 +1971,6 @@ static int btrfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 	__be32 *fsid = (__be32 *)fs_info->fs_devices->fsid;
 	unsigned factor = 1;
 	struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
-	int ret;
-	u64 thresh = 0;
 	int mixed = 0;
 
 	rcu_read_lock();
@@ -2082,31 +2018,9 @@ static int btrfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 		buf->f_bfree = 0;
 	spin_unlock(&block_rsv->lock);
 
-	buf->f_bavail = div_u64(total_free_data, factor);
-	ret = btrfs_calc_avail_data_space(fs_info, &total_free_data);
-	if (ret)
-		return ret;
-	buf->f_bavail += div_u64(total_free_data, factor);
+	buf->f_bavail = btrfs_calc_avail_data_space(fs_info, total_free_meta);
+	buf->f_bavail += total_free_data;
 	buf->f_bavail = buf->f_bavail >> bits;
-
-	/*
-	 * We calculate the remaining metadata space minus global reserve. If
-	 * this is (supposedly) smaller than zero, there's no space. But this
-	 * does not hold in practice, the exhausted state happens where's still
-	 * some positive delta. So we apply some guesswork and compare the
-	 * delta to a 4M threshold.  (Practically observed delta was ~2M.)
-	 *
-	 * We probably cannot calculate the exact threshold value because this
-	 * depends on the internal reservations requested by various
-	 * operations, so some operations that consume a few metadata will
-	 * succeed even if the Avail is zero. But this is better than the other
-	 * way around.
-	 */
-	thresh = SZ_4M;
-
-	if (!mixed && total_free_meta - thresh < block_rsv->size)
-		buf->f_bavail = 0;
-
 	buf->f_type = BTRFS_SUPER_MAGIC;
 	buf->f_bsize = dentry->d_sb->s_blocksize;
 	buf->f_namelen = BTRFS_NAME_LEN;
Signature · 1 line
2.24.1
Reviewed-by on v5 4/4
#10 lore raw

Re: [PATCH v5 4/4] btrfs: statfs: Use pre-calculated per-profile available space

Quoted 23 lines · Qu Wenruo

Although btrfs_calc_avail_data_space() is trying to do an estimation on how many data chunks it can allocate, the estimation is far from perfect:

- Metadata over-commit is not considered at all
- Chunk allocation doesn't take RAID5/6 into consideration

This patch will change btrfs_calc_avail_data_space() to use pre-calculated per-profile available space.

This provides the following benefits:
- Accurate unallocated data space estimation, including RAID5/6 It's as accurate as chunk allocator, and can handle RAID5/6.

Although it still can't handle metadata over-commit that accurately, we still have fallback method for over-commit, by using factor based estimation.

The good news is, over-commit can only happen when we have enough unallocated space, so even we may not report byte accurate result when the fs is empty, the result will get more and more accurate when unallocated space is reducing.

Quoted 6 lines

So the metadata over-commit shouldn't cause too many problem.

Since we're keeping the old lock-free design, statfs should not experience any extra delay.

Thanks,

Josef