Re: [Qemu-devel] [PATCH 01/17] cow: make reads go at a decent speed

[Fam Zheng <famz@redhat.com>]

On Wed, 07/03 16:34, Paolo Bonzini wrote:
> Do not do two reads for each sector; load each sector of the bitmap
> and use bitmap operations to process it.
> 
> Writes are still dog slow!
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  block/cow.c | 54 ++++++++++++++++++++++++++++++++----------------------
>  1 file changed, 32 insertions(+), 22 deletions(-)
> 
> diff --git a/block/cow.c b/block/cow.c
> index 1cc2e89..204451e 100644
> --- a/block/cow.c
> +++ b/block/cow.c
> @@ -126,18 +126,31 @@ static inline int cow_set_bit(BlockDriverState *bs, int64_t bitnum)
>      return 0;
>  }
>  
> -static inline int is_bit_set(BlockDriverState *bs, int64_t bitnum)
> +#define BITS_PER_BITMAP_SECTOR (512 * 8)
> +
> +/* Cannot use bitmap.c on big-endian machines.  */
> +static int cow_test_bit(int64_t bitnum, const uint8_t *bitmap)
>  {
> -    uint64_t offset = sizeof(struct cow_header_v2) + bitnum / 8;
> -    uint8_t bitmap;
> -    int ret;
> +    return (bitmap[bitnum / 8] & (1 << (bitnum & 7))) != 0;
> +}
>  
> -    ret = bdrv_pread(bs->file, offset, &bitmap, sizeof(bitmap));
> -    if (ret < 0) {
> -       return ret;
> +static int cow_find_streak(const uint8_t *bitmap, int value, int start, int nb_sectors)
I think type bool is better for 'value' as you don't booleanize it.  And
also int64_t for start?
> +{
> +    int streak_value = value ? 0xFF : 0;
> +    int last = MIN(start + nb_sectors, BITS_PER_BITMAP_SECTOR);
> +    int bitnum = start;
> +    while (bitnum < last) {
> +        if ((bitnum & 7) == 0 && bitmap[bitnum / 8] == streak_value) {
> +            bitnum += 8;
> +            continue;
> +        }
> +        if (cow_test_bit(bitnum, bitmap) == value) {
> +            bitnum++;
> +            continue;
> +        }
> +        break;
>      }
> -
> -    return !!(bitmap & (1 << (bitnum % 8)));
> +    return MIN(bitnum, last) - start;
>  }
>  
>  /* Return true if first block has been changed (ie. current version is
> @@ -146,23 +159,20 @@ static inline int is_bit_set(BlockDriverState *bs, int64_t bitnum)
>  static int coroutine_fn cow_co_is_allocated(BlockDriverState *bs,
>          int64_t sector_num, int nb_sectors, int *num_same)
>  {
> +    int64_t bitnum = sector_num + sizeof(struct cow_header_v2) * 8;
> +    uint64_t offset = (bitnum / 8) & -BDRV_SECTOR_SIZE;
> +    uint8_t bitmap[512];
> +    int ret;
>      int changed;
>  
> -    if (nb_sectors == 0) {
> -	*num_same = nb_sectors;
> -	return 0;
> -    }
> -
> -    changed = is_bit_set(bs, sector_num);
> -    if (changed < 0) {
> -        return 0; /* XXX: how to return I/O errors? */
> -    }
> -
> -    for (*num_same = 1; *num_same < nb_sectors; (*num_same)++) {
> -	if (is_bit_set(bs, sector_num + *num_same) != changed)
> -	    break;
> +    ret = bdrv_pread(bs->file, offset, &bitmap, sizeof(bitmap));
> +    if (ret < 0) {
> +        return ret;
>      }
>  
> +    bitnum &= BITS_PER_BITMAP_SECTOR - 1;
> +    changed = cow_test_bit(bitnum, bitmap);
> +    *num_same = cow_find_streak(bitmap, changed, bitnum, nb_sectors);
>      return changed;
>  }
>  
> -- 
> 1.8.2.1
> 
> 
> 

-- 
Fam