[sheepdog] [PATCH v3 06/17] block/io: support int64_t bytes in bdrv_aligned_pwritev()

[Eric Blake]

On 4/30/20 6:10 AM, Vladimir Sementsov-Ogievskiy wrote:
> We are generally moving to int64_t for both offset and bytes parameters
> on all io paths.
> 
> Main motivation is realization of 64-bit write_zeroes operation for
> fast zeroing large disk chunks, up to the whole disk.
> 
> We chose signed type, to be consistent with off_t (which is signed) and
> with possibility for signed return type (where negative value means
> error).
> 
> So, prepare bdrv_aligned_pwritev() now and convert the dependencies:
> bdrv_co_write_req_prepare() and bdrv_co_write_req_finish() to signed
> type bytes.
> 
> Series: 64bit-block-status
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov at virtuozzo.com>
> ---
>   block/io.c | 17 ++++++++++-------
>   1 file changed, 10 insertions(+), 7 deletions(-)
> 
> diff --git a/block/io.c b/block/io.c
> index b83749cc50..8bb4ea6285 100644
> --- a/block/io.c
> +++ b/block/io.c
> @@ -1686,12 +1686,11 @@ fail:
>   }
>   
>   static inline int coroutine_fn
> -bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
> +bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, int64_t bytes,
>                             BdrvTrackedRequest *req, int flags)

Changes from unsigned to signed.  Audit of callers:

bdrv_aligned_pwritev() - adjusted this patch, safe
bdrv_do_pdiscard() - passes int64_t, safe
bdrv_co_copy_range_internal() - passes int64_t, safe
bdrv_do_truncate() - passes int64_t, safe

Internal usage:

>   {
>       BlockDriverState *bs = child->bs;
>       bool waited;
> -    int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);

Drops an old sector calculation, and replaces it with:

>   
>       if (bs->read_only) {
>           return -EPERM;
> @@ -1716,8 +1715,10 @@ bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
>       }
>   
>       assert(req->overlap_offset <= offset);
> +    assert(offset <= INT64_MAX - bytes);
>       assert(offset + bytes <= req->overlap_offset + req->overlap_bytes);
> -    assert(end_sector <= bs->total_sectors || child->perm & BLK_PERM_RESIZE);
> +    assert(offset + bytes <= bs->total_sectors * BDRV_SECTOR_SIZE ||
> +           child->perm & BLK_PERM_RESIZE);

assertions that things fit within 63 bits.  Safe

[The req->overlap_offset+ req->overlap_bytes calculation used to be 
unsigned, but was changed to be signed earlier in this series]

>   
>       switch (req->type) {
>       case BDRV_TRACKED_WRITE:
> @@ -1738,7 +1739,7 @@ bdrv_co_write_req_prepare(BdrvChild *child, int64_t offset, uint64_t bytes,
>   }
>   
>   static inline void coroutine_fn
> -bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
> +bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, int64_t bytes,
>                            BdrvTrackedRequest *req, int ret)
>   {

Similar to the above; same four callers, all pass int64_t.


>       int64_t end_sector = DIV_ROUND_UP(offset + bytes, BDRV_SECTOR_SIZE);

This computation needs analysis.  Previously, we had:

DIV_ROUND_UP(int64_t + uint64_t, unsigned long long)
which expands to:
(((uint64_t) + (ull) - int) / (ull))
which simplifies to uint64_t.

Now we have:
DIV_ROUND_UP(int64_t + int64_t, ull)
Okay, in spite of our argument changing type, the macro still results in 
a 64-bit unsigned answer.  Either way, that answer fits within 63 bits, 
so it is safe when assigned to int64_t.

Also in this function:
             stat64_max(&bs->wr_highest_offset, offset + bytes);
in include/qemu/stats64.h, takes uint64_t parameter, but we're passing a 
positive 63-bit number - safe
             bdrv_set_dirty(bs, offset, bytes);
in block/dirty-bitmap.c, takes int64_t parameter - safe

> @@ -1780,14 +1781,14 @@ bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
>    * after possibly fragmenting it.
>    */
>   static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
> -    BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
> +    BdrvTrackedRequest *req, int64_t offset, int64_t bytes,
>       int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
>   {

changes signature from unsigned 32-bit to signed 64-bit.  callers:

bdrv_co_do_zero_pwritev() - passes int64_t, but that was clamped to 
either pad.buf_len [BdrvRequestPadding uses 'size_t buf_len', but 
initializes it in bdrv_init_padding() to at most 2*align] or align set 
from BlockLimits.request_alignment (naturally uint32_t, but documented 
as 'a power of 2 less than INT_MAX' which is at most 1G), so the old 
code never overflowed, and the new code introduces no change

Perhaps we should separately fix BdrvRequestPadding to use a saner type 
than size_t for continuity between 32- and 64-bit platforms (perhaps 
uint32_t rather than int64_t, since we know our padding is bounded by 
request_alignment), but it doesn't impact this patch

bdrv_do_pwritev_part() - still passes unsigned int at this point in the 
series, safe

Usage within the function:

>       BlockDriverState *bs = child->bs;
>       BlockDriver *drv = bs->drv;
>       int ret;
>   
> -    uint64_t bytes_remaining = bytes;
> +    int64_t bytes_remaining = bytes;

Previously we widened unsigned 32-bit into unsigned 64-bit; now we use 
signed 64-bit unchanged.

>       int max_transfer;
>   
>       if (!drv) {
> @@ -1799,6 +1800,8 @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
>       }
>   
>       assert(is_power_of_2(align));
> +    assert(offset >= 0);
> +    assert(bytes >= 0);
>       assert((offset & (align - 1)) == 0);
>       assert((bytes & (align - 1)) == 0);
>       assert(!qiov || qiov_offset + bytes <= qiov->size);

qiov->size is only size_t, while 'qiov_offset + bytes' changed from 
'size_t + unsigned int' to 'size_t + int64_t'.  The resulting type of 
the computation changes for some platforms, but the assertion is proving 
that things still fit (including in 32 bits, when size_t is constrained).

     ret = bdrv_co_write_req_prepare(child, offset, bytes, req, flags);
also touched in this patch, safe

         qemu_iovec_is_zero(qiov, qiov_offset, bytes)) {
Passes an 'int64_t' to a 'size_t' parameter, which is possibly 
narrowing.  Fortunately, the assertions just above prove that by this 
point, we are constrained by qiov->size, which is also size_t. Safe.

         ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
Passes to int64_t, safe

         ret = bdrv_driver_pwritev_compressed(bs, offset, bytes,
Passes to int64_t, safe

         ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, 
flags);
Passes to int64_t, safe

             ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
                                       num, qiov, bytes - bytes_remaining,
Passes int64_t to size_t parameter, but the previous assertion proved we 
did not overflow qiov->size which is size_t. Safe

     bdrv_co_write_req_finish(child, offset, bytes, req, ret);
also touched in this patch, safe

> @@ -1899,7 +1902,7 @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
>       assert(!bytes || (offset & (align - 1)) == 0);
>       if (bytes >= align) {
>           /* Write the aligned part in the middle. */
> -        uint64_t aligned_bytes = bytes & ~(align - 1);
> +        int64_t aligned_bytes = bytes & ~(align - 1);
>           ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
>                                      NULL, 0, flags);
>           if (ret < 0) {
> 

Reviewed-by: Eric Blake <eblake at redhat.com>

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org