On Sep 29, 2014, at 2:59 PM, Darrick J. Wong <darrick.wong@xxxxxxxxxx> wrote:
>
> It'll end up recalculating stride for any flexbg FS with more than
> 12 BGs and more than 3 flexbgs. This piece is neither a part of nor
> is used for 32>64bit conversion.
>
> AFAICT, the point of determine_fs_stride() is to try to recover the
> RAID stride by inferring it from minor variations in the block/inode
> bitmap locations between successive block groups. This explodes when
> flexbg is turned on because bitmap blocks are stored in "other" bgs
> and there's a "big jump" between the bitmaps in the last bg of one
> flexbg and the bitmaps of the first bg of the next flexbg. Between
> bgs in a single flexbg the *_stride values are "negative" and don't
> contribute to the stride calculation.
>
> I /think/ the solution is to ignore first blockgroup when crossing a
> flexbg boundary when there are flexbgs. Can you give the following
> patch a spin? It shouldn't spit out "group XXX has stride..." messages
> after that. I'm not sure that "negative" stride ought to be ignored
> either, but....
>
> Honestly I'd rather just kill the whole thing, but someone must've had
> a reason to put it there? Ted?
I added this to try and preserve the RAID stride while doing the resize,
to avoid making one disk in a RAID be a hot-spot for bitmap updates.
With flex_bg the RAID stride becomes less critical, because the bitmaps
are contiguous and will naturally span the RAID stripes if the flex_bg
factor is large enough to have blocks on every stripe.
We normally specify the flex_bg factor to be 256 (== 1MB of contiguous
bitmaps) to exactly match the RAID stripe width when formatting Lustre,
so we don't need the stride for this, but still specify it to help the
mballoc align the IO blocks properly.
I haven't given much though about what happens if these two do not line
up evenly.
Cheers, Andreas
> --D
>
> diff --git a/resize/main.c b/resize/main.c
> index 060e67d..b993dfb 100644
> --- a/resize/main.c
> +++ b/resize/main.c
> @@ -105,6 +105,7 @@ static void determine_fs_stride(ext2_filsys fs)
> unsigned long long sum;
> unsigned int has_sb, prev_has_sb = 0, num;
> int i_stride, b_stride;
> + int flexbg_size = 1 << fs->super->s_log_groups_per_flex;
>
> if (fs->stride)
> return;
> @@ -120,10 +121,11 @@ static void determine_fs_stride(ext2_filsys fs)
> ext2fs_inode_bitmap_loc(fs, group - 1) -
> fs->super->s_blocks_per_group;
> if (b_stride != i_stride ||
> - b_stride < 0)
> + b_stride < 0 ||
> + (flexbg_size > 1 && (group % flexbg_size == 0)))
> goto next;
>
> - /* printf("group %d has stride %d\n", group, b_stride); */
> + printf("group %d has stride %d %d\n", group, b_stride, i_stride);
> sum += b_stride;
> num++;
>
> @@ -133,7 +135,7 @@ static void determine_fs_stride(ext2_filsys fs)
>
> if (fs->group_desc_count > 12 && num < 3)
> sum = 0;
> -
> +printf("sum is %llu %u\n", sum, num);
> if (num)
> fs->stride = sum / num;
> else
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
Cheers, Andreas
Attachment:
signature.asc
Description: Message signed with OpenPGP using GPGMail
