e2fsck: improve performance of region_allocate()

Hi Ted, Doug,

Ted, I already emailed you the patch for the latter discussion here, 
including my rudimentary benchmarks.

Unfortunately I'm having trouble with inline formatting of the patch, so 
I have attached it instead of providing inline (sorry - I know that 
makes discussion difficult).  Was originally emailed to you as a series 
of three independent patches, with the below as 0001.  We still need to 
discuss the other optimization.

The attached improves CPU performance from O(e*h) to O(e) and memory 
from O(h) to O(1).  The patch below does similar for CPU but nothing for 
memory (In my case it took fsck down by a significant margin).

Previously my fsck got stuck on 0.5% (we both know it got stuck on a 
single 340GB file with numerous holes in it, of which I have multiple 
such files on my filesystem) and stayed there for a few hours.  With 
this (and the memory map link-count for pass2) I managed to finish a 
fsck on a 40TB filesystem in 508 minutes.

Ted - the provided patch by Doug may still improve performance for the 
other uses of region.c as well, but the impact will probably not be as 
severe since (as far as I know) there are usually not a great many 
number of EAs for each file.

Kind Regards,
Jaco

On 22/08/2017 04:29, Theodore Ts'o wrote:
> On Fri, Aug 18, 2017 at 06:16:35PM -0700, Doug Porter wrote:
>> Use a red-black tree to track allocations instead of a linked list.
>> This provides a substantial performance boost when the number of
>> allocations in a region is large.  This change resulted in a reduction
>> in runtime from 4821s to 6s on one filesystem.
>>
>> Signed-off-by: Doug Porter <dsp@fb.com>
> Hi Doug, as it turns out, Jaco Kroon and I had been debugging the same
> problem as oen you were working on.  We came up with a different way
> of solving this problem (see below).  It works by observing that
> unless the extent tree is terribly corrupted, region_allocate() will
> always be appending to the very end of the list.
>
> However, it has since occurred to me that since we are doing an
> breadth-first traversal of the extent tree, the starting lba for each
> leaf node *must* always be monotonically increasing, and we already
> check to make sure this is true within an extent tree block.  So I
> think it might be possible to dispense with using any kind of data
> structure, whether it's an rbtree or a linked list, and instead just
> simply make sure we enforce the start+len of the last entry in an
> extent tree block is < the starting lba of the first entry in the next
> extent tree block.
>
> We are already checking all of the necessary other conditions in
> scan_extent_node, so with this additional check, I believe that using
> the region tracking code in scan_extent_node (which was originally
> written to make sure that extended attribute block did not have any
> parts of a string shared between more than one EA key or value pair)
> can be made entirely unnecessary for scan_extent_node().
>
> I haven't had a chance to code this alternate fix, but I believe it
> should be superior to either your patch or the one which I had drafted
> below.  Does this make sense to you?
>
> 							- Ted
>
> commit 8a48ce07a5923242fecc5dc04d6e30dd59a8f07d
> Author: Theodore Ts'o <tytso@mit.edu>
> Date:   Mon Aug 14 19:52:39 2017 -0400
>
>      e2fsck: add optimization for large, fragmented sparse files
>      
>      The code which checks for overlapping logical blocks in an extent tree
>      is O(h*e) in time, where h is the number of holes in the file, and e
>      is the number of extents in the file.  So a file with a large number
>      of holes can take e2fsck a long time process.  Optimize this taking
>      advantage of the fact the vast majority of the time, region_allocate()
>      is called with increasing logical block numbers, so we are almost
>      always append onto the end of the region list.
>      
>      Signed-off-by: Theodore Ts'o <tytso@mit.edu>
>
> diff --git a/e2fsck/region.c b/e2fsck/region.c
> index e32f89db0..95d23be4f 100644
> --- a/e2fsck/region.c
> +++ b/e2fsck/region.c
> @@ -30,6 +30,7 @@ struct region_struct {
>   	region_addr_t	min;
>   	region_addr_t	max;
>   	struct region_el *allocated;
> +	struct region_el *last;
>   };
>   
>   region_t region_create(region_addr_t min, region_addr_t max)
> @@ -42,6 +43,7 @@ region_t region_create(region_addr_t min, region_addr_t max)
>   	memset(region, 0, sizeof(struct region_struct));
>   	region->min = min;
>   	region->max = max;
> +	region->last = NULL;
>   	return region;
>   }
>   
> @@ -68,6 +70,18 @@ int region_allocate(region_t region, region_addr_t start, int n)
>   	if (n == 0)
>   		return 1;
>   
> +	if (region->last && region->last->end == start &&
> +	    !region->last->next) {
> +		region->last->end = end;
> +		return 0;
> +	}
> +	if (region->last && start > region->last->end &&
> +	    !region->last->next) {
> +		r = NULL;
> +		prev = region->last;
> +		goto append_to_list;
> +	}
> +
>   	/*
>   	 * Search through the linked list.  If we find that it
>   	 * conflicts witih something that's already allocated, return
> @@ -92,6 +106,8 @@ int region_allocate(region_t region, region_addr_t start, int n)
>   					r->end = next->end;
>   					r->next = next->next;
>   					free(next);
> +					if (!r->next)
> +						region->last = r;
>   					return 0;
>   				}
>   			}
> @@ -104,12 +120,15 @@ int region_allocate(region_t region, region_addr_t start, int n)
>   	/*
>   	 * Insert a new region element structure into the linked list
>   	 */
> +append_to_list:
>   	new_region = malloc(sizeof(struct region_el));
>   	if (!new_region)
>   		return -1;
>   	new_region->start = start;
>   	new_region->end = start + n;
>   	new_region->next = r;
> +	if (!new_region->next)
> +		region->last = new_region;
>   	if (prev)
>   		prev->next = new_region;
>   	else

e2fsck: improve performance of region_allocate()

Commit Message

Comments

Patch