[sheepdog] [PATCH v2 02/17] block: use int64_t as bytes type in tracked requests

Vladimir Sementsov-Ogievskiy vsementsov at virtuozzo.com
Thu Apr 30 10:21:35 CEST 2020


29.04.2020 18:50, Eric Blake wrote:
> On 4/27/20 3:23 AM, Vladimir Sementsov-Ogievskiy wrote:
>> We are generally moving to int64_t for both offset and bytes parameters
>> on all io paths. Convert tracked requests now.
> 
> As mentioned elsewhere in the thread, this states 'what' but not 'why'; adding a bit more of the 'why' can be useful when revisiting this commit in the future.
> 
>>
>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov at virtuozzo.com>
>> Reviewed-by: Stefan Hajnoczi <stefanha at redhat.com>
>> ---
>>   include/block/block_int.h |  4 ++--
>>   block/io.c                | 11 ++++++-----
>>   2 files changed, 8 insertions(+), 7 deletions(-)
>>
>> diff --git a/include/block/block_int.h b/include/block/block_int.h
>> index 4c3587ea19..c8daba608b 100644
>> --- a/include/block/block_int.h
>> +++ b/include/block/block_int.h
>> @@ -70,12 +70,12 @@ enum BdrvTrackedRequestType {
>>   typedef struct BdrvTrackedRequest {
>>       BlockDriverState *bs;
>>       int64_t offset;
>> -    uint64_t bytes;
>> +    int64_t bytes;
>>       enum BdrvTrackedRequestType type;
>>       bool serialising;
>>       int64_t overlap_offset;
>> -    uint64_t overlap_bytes;
>> +    int64_t overlap_bytes;
> 
> unsigned values have defined wraparound semantics, signed values have a lower maximum and require careful handling to avoid undefined overflow. So we have to check all clients for any surprises.
> 
> block/file-posix.c:raw_do_pwrite_zeroes() -
>          assert(req->offset + req->bytes >= offset + bytes);
> pre-patch: assert(int64_t + uint64_t >= int64_t + int)
>             assert(uint64_t >= int64_t) - unsigned compare
> post-patch: assert(int64_t >= int64_t) - signed compare
> Risky if adding req->bytes could ever overflow 63 bits but still fit in 64 bits, but I couldn't find any way to trigger that.  I think we're safe because the block layer never calls a driver's .pwrite_zeroes with a bytes that would overflow 63 bits.
> 
> block/write-threshold.c:bdrv_write_threshold_exceeded() -
>          if ((req->offset + req->bytes) > bs->write_threshold_offset) {
> pre-patch: ((int64_t + uint64_t) > uint64_t) - unsigned compare
> post-patch: (int64_t > uint64_t) - still unsigned compare
> 
> Perhaps that function should be changed to return int64_t, but probably as a different patch in the series (I didn't read ahead yet to see if you already did).
> 
>>       QLIST_ENTRY(BdrvTrackedRequest) list;
>>       Coroutine *co; /* owner, used for deadlock detection */
>> diff --git a/block/io.c b/block/io.c
>> index aba67f66b9..7cbb80bd24 100644
>> --- a/block/io.c
>> +++ b/block/io.c
>> @@ -692,10 +692,11 @@ static void tracked_request_end(BdrvTrackedRequest *req)
>>   static void tracked_request_begin(BdrvTrackedRequest *req,
>>                                     BlockDriverState *bs,
>>                                     int64_t offset,
>> -                                  uint64_t bytes,
>> +                                  int64_t bytes,
>>                                     enum BdrvTrackedRequestType type)
>>   {
>> -    assert(bytes <= INT64_MAX && offset <= INT64_MAX - bytes);
>> +    assert(offset >= 0 && bytes >= 0 &&
>> +           bytes <= INT64_MAX && offset <= INT64_MAX - bytes);
> 
> This part is nice - it makes it very easy to prove all other uses of BdrvTrackedRequest.bytes were already in range (both pre- and post-patch).
> 
>>       *req = (BdrvTrackedRequest){
>>           .bs = bs,
>> @@ -716,7 +717,7 @@ static void tracked_request_begin(BdrvTrackedRequest *req,
>>   }
>>   static bool tracked_request_overlaps(BdrvTrackedRequest *req,
>> -                                     int64_t offset, uint64_t bytes)
>> +                                     int64_t offset, int64_t bytes)
>>   {
>>       /*        aaaa   bbbb */
>>       if (offset >= req->overlap_offset + req->overlap_bytes) {
>> @@ -773,8 +774,8 @@ bool bdrv_mark_request_serialising(BdrvTrackedRequest *req, uint64_t align)
>>   {
>>       BlockDriverState *bs = req->bs;
>>       int64_t overlap_offset = req->offset & ~(align - 1);
> 
> While here, should we use QEMU_ALIGN_DOWN instead of open-coding?

But then, also s/ROUND_UP/QEMU_ALIGN_UP/ for consistency? But ROUND_UP is faster. Still, we tend to generally use QEMU_ALIGN_UP everywhere.. So, may be better to refactor these thing alltogether in some good way. Maybe, add ROUND_DOWN macro for symmetry?

> 
>> -    uint64_t overlap_bytes = ROUND_UP(req->offset + req->bytes, align)
>> -                               - overlap_offset;
>> +    int64_t overlap_bytes =
>> +            ROUND_UP(req->offset + req->bytes, align) - overlap_offset;
>>       bool waited;
>>       qemu_co_mutex_lock(&bs->reqs_lock);
>>
> 
> Looking through uses of BdrvTrackedRequest in io.c, I couldn't find any other surprises (it seems everything starts with tracked_request_begin, and while you did switch between unsigned and signed, you did not switch width, so it's easier to reason about once we know things don't overflow).
> 
> Reviewed-by: Eric Blake <eblake at redhat.com>
> 
> 


-- 
Best regards,
Vladimir


More information about the sheepdog mailing list