[PATCH] ext4: Change the handling of RAID stripe width

[PATCH] ext4: Change the handling of RAID stripe width

[Dan Ehrenberg]

Previously, the stripe width was blindly used for determining the size
of allocations. Now, the stripe width is used as a hint for the initial
mb_group_prealloc; if it is greater than 1, then we make sure that
mb_group_prealloc is some multiple of it, and otherwise it is ignored.
mb_group_prealloc is always usable to adjust the preallocation strategy,
not just when the stripe-width is 0 as before.

Signed-off-by: Dan Ehrenberg <dehrenberg@google.com>
---
 fs/ext4/mballoc.c |   40 +++++++++++++++++++++++++++++-----------
 1 files changed, 29 insertions(+), 11 deletions(-)

diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
index 6ed859d..710c27f 100644
--- a/fs/ext4/mballoc.c
+++ b/fs/ext4/mballoc.c
@@ -127,13 +127,14 @@
  * based on file size. This can be found in ext4_mb_normalize_request. If
  * we are doing a group prealloc we try to normalize the request to
  * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
- * 512 blocks. This can be tuned via
- * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
- * terms of number of blocks. If we have mounted the file system with -O
+ * 512 blocks. If we have mounted the file system with -O
  * stripe=<value> option the group prealloc request is normalized to the
- * stripe value (sbi->s_stripe)
+ * the smallest multiple of the stripe value (sbi->s_stripe) which is
+ * greater than the default mb_group_prealloc. This can be tuned via
+ * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
+ * terms of number of blocks.
  *
- * The regular allocator(using the buddy cache) supports few tunables.
+ * The regular allocator (using the buddy cache) supports a few tunables.
  *
  * /sys/fs/ext4/<partition>/mb_min_to_scan
  * /sys/fs/ext4/<partition>/mb_max_to_scan
@@ -2471,7 +2472,26 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
 	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;
+	/*
+	 * If the stripe width is 1, this makes no sense and
+	 * we set it to 0 to turn off stripe handling code.
+	 */
+	if (sbi->s_stripe == 1)
+		sbi->s_stripe = 0;
+
 	sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
+	/*
+	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
+	 * to the lowest multiple of s_stripe which is bigger than
+	 * the s_mb_group_prealloc as determined above. We want
+	 * the preallocation size to be an exact multiple of the
+	 * RAID stripe size so that preallocations don't fragment
+	 * the stripes.
+	 */
+	if (sbi->s_stripe > 1) {
+		sbi->s_mb_group_prealloc = roundup(
+			sbi->s_mb_group_prealloc, sbi->s_stripe);
+	}
 
 	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
 	if (sbi->s_locality_groups == NULL) {
@@ -2830,8 +2850,9 @@ out_err:
 
 /*
  * here we normalize request for locality group
- * Group request are normalized to s_strip size if we set the same via mount
- * option. If not we set it to s_mb_group_prealloc which can be configured via
+ * Group request are normalized to s_mb_group_prealloc, which goes to
+ * s_strip if we set the same via mount option.
+ * s_mb_group_prealloc can be configured via
  * /sys/fs/ext4/<partition>/mb_group_prealloc
  *
  * XXX: should we try to preallocate more than the group has now?
@@ -2842,10 +2863,7 @@ static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
 	struct ext4_locality_group *lg = ac->ac_lg;
 
 	BUG_ON(lg == NULL);
-	if (EXT4_SB(sb)->s_stripe)
-		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
-	else
-		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
+	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
 	mb_debug(1, "#%u: goal %u blocks for locality group\n",
 		current->pid, ac->ac_g_ex.fe_len);
 }
-- 
1.7.3.1

Re: [PATCH] ext4: Change the handling of RAID stripe width

[Eric Sandeen]

On 7/6/11 3:47 PM, Dan Ehrenberg wrote:
> Previously, the stripe width was blindly used for determining the size
> of allocations. Now, the stripe width is used as a hint for the initial
> mb_group_prealloc; if it is greater than 1, then we make sure that
> mb_group_prealloc is some multiple of it, and otherwise it is ignored.
> mb_group_prealloc is always usable to adjust the preallocation strategy,
> not just when the stripe-width is 0 as before.
> 
> Signed-off-by: Dan Ehrenberg <dehrenberg@google.com>
> ---
>  fs/ext4/mballoc.c |   40 +++++++++++++++++++++++++++++-----------
>  1 files changed, 29 insertions(+), 11 deletions(-)
> 
> diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
> index 6ed859d..710c27f 100644
> --- a/fs/ext4/mballoc.c
> +++ b/fs/ext4/mballoc.c
> @@ -127,13 +127,14 @@
>   * based on file size. This can be found in ext4_mb_normalize_request. If
>   * we are doing a group prealloc we try to normalize the request to
>   * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
> - * 512 blocks. This can be tuned via
> - * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
> - * terms of number of blocks. If we have mounted the file system with -O
> + * 512 blocks. If we have mounted the file system with -O
>   * stripe=<value> option the group prealloc request is normalized to the
> - * stripe value (sbi->s_stripe)
> + * the smallest multiple of the stripe value (sbi->s_stripe) which is
> + * greater than the default mb_group_prealloc. This can be tuned via
> + * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
> + * terms of number of blocks.
>   *
> - * The regular allocator(using the buddy cache) supports few tunables.
> + * The regular allocator (using the buddy cache) supports a few tunables.
>   *
>   * /sys/fs/ext4/<partition>/mb_min_to_scan
>   * /sys/fs/ext4/<partition>/mb_max_to_scan
> @@ -2471,7 +2472,26 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
>  	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;
> +	/*
> +	 * If the stripe width is 1, this makes no sense and
> +	 * we set it to 0 to turn off stripe handling code.
> +	 */
> +	if (sbi->s_stripe == 1)
> +		sbi->s_stripe = 0;

This strikes me as a weird band-aid-y place to fix this up.

Wouldn't it be better suited for the option-parsing code, and/or
in ext4_get_stripe_size()?  Why let a value of 1 get this far
only to override it here?

-Eric

>  	sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
> +	/*
> +	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
> +	 * to the lowest multiple of s_stripe which is bigger than
> +	 * the s_mb_group_prealloc as determined above. We want
> +	 * the preallocation size to be an exact multiple of the
> +	 * RAID stripe size so that preallocations don't fragment
> +	 * the stripes.
> +	 */
> +	if (sbi->s_stripe > 1) {
> +		sbi->s_mb_group_prealloc = roundup(
> +			sbi->s_mb_group_prealloc, sbi->s_stripe);
> +	}
>  
>  	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
>  	if (sbi->s_locality_groups == NULL) {
> @@ -2830,8 +2850,9 @@ out_err:
>  
>  /*
>   * here we normalize request for locality group
> - * Group request are normalized to s_strip size if we set the same via mount
> - * option. If not we set it to s_mb_group_prealloc which can be configured via
> + * Group request are normalized to s_mb_group_prealloc, which goes to
> + * s_strip if we set the same via mount option.
> + * s_mb_group_prealloc can be configured via
>   * /sys/fs/ext4/<partition>/mb_group_prealloc
>   *
>   * XXX: should we try to preallocate more than the group has now?
> @@ -2842,10 +2863,7 @@ static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
>  	struct ext4_locality_group *lg = ac->ac_lg;
>  
>  	BUG_ON(lg == NULL);
> -	if (EXT4_SB(sb)->s_stripe)
> -		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
> -	else
> -		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
> +	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
>  	mb_debug(1, "#%u: goal %u blocks for locality group\n",
>  		current->pid, ac->ac_g_ex.fe_len);
>  }

Re: [PATCH] ext4: Change the handling of RAID stripe width

[Daniel Ehrenberg]

On Wed, Jul 6, 2011 at 2:18 PM, Eric Sandeen <sandeen@redhat.com> wrote:
> On 7/6/11 3:47 PM, Dan Ehrenberg wrote:
>> Previously, the stripe width was blindly used for determining the size
>> of allocations. Now, the stripe width is used as a hint for the initial
>> mb_group_prealloc; if it is greater than 1, then we make sure that
>> mb_group_prealloc is some multiple of it, and otherwise it is ignored.
>> mb_group_prealloc is always usable to adjust the preallocation strategy,
>> not just when the stripe-width is 0 as before.
>>
>> Signed-off-by: Dan Ehrenberg <dehrenberg@google.com>
>> ---
>>  fs/ext4/mballoc.c |   40 +++++++++++++++++++++++++++++-----------
>>  1 files changed, 29 insertions(+), 11 deletions(-)
>>
>> diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
>> index 6ed859d..710c27f 100644
>> --- a/fs/ext4/mballoc.c
>> +++ b/fs/ext4/mballoc.c
>> @@ -127,13 +127,14 @@
>>   * based on file size. This can be found in ext4_mb_normalize_request. If
>>   * we are doing a group prealloc we try to normalize the request to
>>   * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
>> - * 512 blocks. This can be tuned via
>> - * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
>> - * terms of number of blocks. If we have mounted the file system with -O
>> + * 512 blocks. If we have mounted the file system with -O
>>   * stripe=<value> option the group prealloc request is normalized to the
>> - * stripe value (sbi->s_stripe)
>> + * the smallest multiple of the stripe value (sbi->s_stripe) which is
>> + * greater than the default mb_group_prealloc. This can be tuned via
>> + * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
>> + * terms of number of blocks.
>>   *
>> - * The regular allocator(using the buddy cache) supports few tunables.
>> + * The regular allocator (using the buddy cache) supports a few tunables.
>>   *
>>   * /sys/fs/ext4/<partition>/mb_min_to_scan
>>   * /sys/fs/ext4/<partition>/mb_max_to_scan
>> @@ -2471,7 +2472,26 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
>>       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;
>> +     /*
>> +      * If the stripe width is 1, this makes no sense and
>> +      * we set it to 0 to turn off stripe handling code.
>> +      */
>> +     if (sbi->s_stripe == 1)
>> +             sbi->s_stripe = 0;
>
> This strikes me as a weird band-aid-y place to fix this up.
>
> Wouldn't it be better suited for the option-parsing code, and/or
> in ext4_get_stripe_size()?  Why let a value of 1 get this far
> only to override it here?
>
> -Eric

The mount option parsing code wouldn't be a working place to do it,
since it can be specified on-disk what the stripe size is. But
ext4_get_stripe_size might be a good place to put it instead. I guess
there are two unrelated parts to this patch: handling where the stripe
width was set to 1, and handling where it's more than 1 but much less
than what it should be.

Dan

Dan

Re: [PATCH] ext4: Change the handling of RAID stripe width

[Eric Sandeen]

On 7/6/11 5:17 PM, Daniel Ehrenberg wrote:
> On Wed, Jul 6, 2011 at 2:18 PM, Eric Sandeen <sandeen@redhat.com> wrote:
>> On 7/6/11 3:47 PM, Dan Ehrenberg wrote:
>>> Previously, the stripe width was blindly used for determining the size
>>> of allocations. Now, the stripe width is used as a hint for the initial
>>> mb_group_prealloc; if it is greater than 1, then we make sure that
>>> mb_group_prealloc is some multiple of it, and otherwise it is ignored.
>>> mb_group_prealloc is always usable to adjust the preallocation strategy,
>>> not just when the stripe-width is 0 as before.
>>>
>>> Signed-off-by: Dan Ehrenberg <dehrenberg@google.com>
>>> ---
>>>  fs/ext4/mballoc.c |   40 +++++++++++++++++++++++++++++-----------
>>>  1 files changed, 29 insertions(+), 11 deletions(-)
>>>
>>> diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
>>> index 6ed859d..710c27f 100644
>>> --- a/fs/ext4/mballoc.c
>>> +++ b/fs/ext4/mballoc.c
>>> @@ -127,13 +127,14 @@
>>>   * based on file size. This can be found in ext4_mb_normalize_request. If
>>>   * we are doing a group prealloc we try to normalize the request to
>>>   * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
>>> - * 512 blocks. This can be tuned via
>>> - * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
>>> - * terms of number of blocks. If we have mounted the file system with -O
>>> + * 512 blocks. If we have mounted the file system with -O
>>>   * stripe=<value> option the group prealloc request is normalized to the
>>> - * stripe value (sbi->s_stripe)
>>> + * the smallest multiple of the stripe value (sbi->s_stripe) which is
>>> + * greater than the default mb_group_prealloc. This can be tuned via
>>> + * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
>>> + * terms of number of blocks.
>>>   *
>>> - * The regular allocator(using the buddy cache) supports few tunables.
>>> + * The regular allocator (using the buddy cache) supports a few tunables.
>>>   *
>>>   * /sys/fs/ext4/<partition>/mb_min_to_scan
>>>   * /sys/fs/ext4/<partition>/mb_max_to_scan
>>> @@ -2471,7 +2472,26 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
>>>       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;
>>> +     /*
>>> +      * If the stripe width is 1, this makes no sense and
>>> +      * we set it to 0 to turn off stripe handling code.
>>> +      */
>>> +     if (sbi->s_stripe == 1)
>>> +             sbi->s_stripe = 0;
>>
>> This strikes me as a weird band-aid-y place to fix this up.
>>
>> Wouldn't it be better suited for the option-parsing code, and/or
>> in ext4_get_stripe_size()?  Why let a value of 1 get this far
>> only to override it here?
>>
>> -Eric
> 
> The mount option parsing code wouldn't be a working place to do it,
> since it can be specified on-disk what the stripe size is. But

well I think that the mount option overrides the on disk value, since the
first thing that ext4_get_stripe_size() does is look for s_stripe
and return it if it's already set.

I just think it's good to keep this stuff consolidated, and not
sprinkle special-cases and error checking around in the code 
too much... complicated enough already IMHO.

-Eric

> ext4_get_stripe_size might be a good place to put it instead. I guess
> there are two unrelated parts to this patch: handling where the stripe
> width was set to 1, and handling where it's more than 1 but much less
> than what it should be.
> 
> Dan
> 
> Dan

Re: [PATCH] ext4: Change the handling of RAID stripe width

[Andreas Dilger]

On 2011-07-06, at 2:47 PM, Dan Ehrenberg wrote:
> Previously, the stripe width was blindly used for determining the size
> of allocations. Now, the stripe width is used as a hint for the initial
> mb_group_prealloc; if it is greater than 1, then we make sure that
> mb_group_prealloc is some multiple of it, and otherwise it is ignored.
> mb_group_prealloc is always usable to adjust the preallocation strategy,
> not just when the stripe-width is 0 as before.

In general I like this patch.  In the past, stripe width was only set
manually by people who understood their storage and would pick a suitable
value.  With modern mke2fs and disks, it is managing to get set to a wide
variety of strange values, so it makes sense to have better sanity checking
in the kernel.  In particular, I like that the preallocation is a multiple
of the underlying stripe size, since this was the original intent.

You can add my:
Reviewed-by: Andreas Dilger <adilger.kernel@dilger.ca>

> Signed-off-by: Dan Ehrenberg <dehrenberg@google.com>
> ---
> fs/ext4/mballoc.c |   40 +++++++++++++++++++++++++++++-----------
> 1 files changed, 29 insertions(+), 11 deletions(-)
> 
> diff --git a/fs/ext4/mballoc.c b/fs/ext4/mballoc.c
> index 6ed859d..710c27f 100644
> --- a/fs/ext4/mballoc.c
> +++ b/fs/ext4/mballoc.c
> @@ -127,13 +127,14 @@
> * based on file size. This can be found in ext4_mb_normalize_request. If
> * we are doing a group prealloc we try to normalize the request to
> * sbi->s_mb_group_prealloc. Default value of s_mb_group_prealloc is
> - * 512 blocks. This can be tuned via
> - * /sys/fs/ext4/<partition/mb_group_prealloc. The value is represented in
> - * terms of number of blocks. If we have mounted the file system with -O
> + * 512 blocks. If we have mounted the file system with -O
> * stripe=<value> option the group prealloc request is normalized to the
> - * stripe value (sbi->s_stripe)
> + * the smallest multiple of the stripe value (sbi->s_stripe) which is
> + * greater than the default mb_group_prealloc. This can be tuned via
> + * /sys/fs/ext4/<partition>/mb_group_prealloc. The value is represented in
> + * terms of number of blocks.
> *
> - * The regular allocator(using the buddy cache) supports few tunables.
> + * The regular allocator (using the buddy cache) supports a few tunables.
> *
> * /sys/fs/ext4/<partition>/mb_min_to_scan
> * /sys/fs/ext4/<partition>/mb_max_to_scan
> @@ -2471,7 +2472,26 @@ int ext4_mb_init(struct super_block *sb, int needs_recovery)
> 	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;
> +	/*
> +	 * If the stripe width is 1, this makes no sense and
> +	 * we set it to 0 to turn off stripe handling code.
> +	 */
> +	if (sbi->s_stripe == 1)
> +		sbi->s_stripe = 0;
> +
> 	sbi->s_mb_group_prealloc = MB_DEFAULT_GROUP_PREALLOC;
> +	/*
> +	 * If there is a s_stripe > 1, then we set the s_mb_group_prealloc
> +	 * to the lowest multiple of s_stripe which is bigger than
> +	 * the s_mb_group_prealloc as determined above. We want
> +	 * the preallocation size to be an exact multiple of the
> +	 * RAID stripe size so that preallocations don't fragment
> +	 * the stripes.
> +	 */
> +	if (sbi->s_stripe > 1) {
> +		sbi->s_mb_group_prealloc = roundup(
> +			sbi->s_mb_group_prealloc, sbi->s_stripe);
> +	}
> 
> 	sbi->s_locality_groups = alloc_percpu(struct ext4_locality_group);
> 	if (sbi->s_locality_groups == NULL) {
> @@ -2830,8 +2850,9 @@ out_err:
> 
> /*
> * here we normalize request for locality group
> - * Group request are normalized to s_strip size if we set the same via mount
> - * option. If not we set it to s_mb_group_prealloc which can be configured via
> + * Group request are normalized to s_mb_group_prealloc, which goes to
> + * s_strip if we set the same via mount option.
> + * s_mb_group_prealloc can be configured via
> * /sys/fs/ext4/<partition>/mb_group_prealloc
> *
> * XXX: should we try to preallocate more than the group has now?
> @@ -2842,10 +2863,7 @@ static void ext4_mb_normalize_group_request(struct ext4_allocation_context *ac)
> 	struct ext4_locality_group *lg = ac->ac_lg;
> 
> 	BUG_ON(lg == NULL);
> -	if (EXT4_SB(sb)->s_stripe)
> -		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_stripe;
> -	else
> -		ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
> +	ac->ac_g_ex.fe_len = EXT4_SB(sb)->s_mb_group_prealloc;
> 	mb_debug(1, "#%u: goal %u blocks for locality group\n",
> 		current->pid, ac->ac_g_ex.fe_len);
> }
> -- 
> 1.7.3.1
>