Re: [Qemu-devel] [PATCH 1/8] block: prepare bdrv_co_do_write_zeroes to deal with large bl.max_write_zeroes

["Denis V. Lunev" <den@openvz.org>]

On 05/01/15 14:23, Peter Lieven wrote:
> On 05.01.2015 12:06, Denis V. Lunev wrote:
>> On 05/01/15 10:34, Peter Lieven wrote:
>>> On 30.12.2014 10:20, Denis V. Lunev wrote:
>>>> bdrv_co_do_write_zeroes split writes using bl.max_write_zeroes or
>>>> 16 MiB as a chunk size. This is implemented in this way to tolerate
>>>> buggy block backends which do not accept too big requests.
>>>>
>>>> Though if the bdrv_co_write_zeroes callback is not good enough, we
>>>> fallback to write data explicitely using bdrv_co_writev and we
>>>> create buffer to accomodate zeroes inside. The size of this buffer
>>>> is the size of the chunk. Thus if the underlying layer will have
>>>> bl.max_write_zeroes high enough, f.e. 4 GiB, the allocation can fail.
>>>>
>>>> Actually, there is no need to allocate such a big amount of memory.
>>>> We could simply allocate 1 MiB buffer and create iovec, which will
>>>> point to the same memory.
>>>>
>>>> Signed-off-by: Denis V. Lunev <den@openvz.org>
>>>> CC: Kevin Wolf <kwolf@redhat.com>
>>>> CC: Stefan Hajnoczi <stefanha@redhat.com>
>>>> CC: Peter Lieven <pl@kamp.de>
>>>> ---
>>>>   block.c | 35 ++++++++++++++++++++++++-----------
>>>>   1 file changed, 24 insertions(+), 11 deletions(-)
>>>>
>>>> diff --git a/block.c b/block.c
>>>> index 4165d42..d69c121 100644
>>>> --- a/block.c
>>>> +++ b/block.c
>>>> @@ -3173,14 +3173,18 @@ int coroutine_fn 
>>>> bdrv_co_copy_on_readv(BlockDriverState *bs,
>>>>    * of 32768 512-byte sectors (16 MiB) per request.
>>>>    */
>>>>   #define MAX_WRITE_ZEROES_DEFAULT 32768
>>>> +/* allocate iovec with zeroes using 1 MiB chunks to avoid to big 
>>>> allocations */
>>>> +#define MAX_ZEROES_CHUNK (1024 * 1024)
>>>>     static int coroutine_fn 
>>>> bdrv_co_do_write_zeroes(BlockDriverState *bs,
>>>>       int64_t sector_num, int nb_sectors, BdrvRequestFlags flags)
>>>>   {
>>>>       BlockDriver *drv = bs->drv;
>>>>       QEMUIOVector qiov;
>>>> -    struct iovec iov = {0};
>>>>       int ret = 0;
>>>> +    void *chunk = NULL;
>>>> +
>>>> +    qemu_iovec_init(&qiov, 0);
>>>>         int max_write_zeroes = bs->bl.max_write_zeroes ?
>>>>                              bs->bl.max_write_zeroes : 
>>>> MAX_WRITE_ZEROES_DEFAULT;
>>>> @@ -3217,27 +3221,35 @@ static int coroutine_fn 
>>>> bdrv_co_do_write_zeroes(BlockDriverState *bs,
>>>>           }
>>>>             if (ret == -ENOTSUP) {
>>>> +            int64_t num_bytes = (int64_t)num << BDRV_SECTOR_BITS;
>>>> +            int chunk_size = MIN(MAX_ZEROES_CHUNK, num_bytes);
>>>> +
>>>>               /* Fall back to bounce buffer if write zeroes is 
>>>> unsupported */
>>>> -            iov.iov_len = num * BDRV_SECTOR_SIZE;
>>>> -            if (iov.iov_base == NULL) {
>>>> -                iov.iov_base = qemu_try_blockalign(bs, num * 
>>>> BDRV_SECTOR_SIZE);
>>>> -                if (iov.iov_base == NULL) {
>>>> +            if (chunk == NULL) {
>>>> +                chunk = qemu_try_blockalign(bs, chunk_size);
>>>> +                if (chunk == NULL) {
>>>>                       ret = -ENOMEM;
>>>>                       goto fail;
>>>>                   }
>>>> -                memset(iov.iov_base, 0, num * BDRV_SECTOR_SIZE);
>>>> +                memset(chunk, 0, chunk_size);
>>>> +            }
>>>> +
>>>> +            while (num_bytes > 0) {
>>>> +                int to_add = MIN(chunk_size, num_bytes);
>>>> +                qemu_iovec_add(&qiov, chunk, to_add);
>>>
>>> This can and likely will fail for big num_bytes if you exceed 
>>> IOV_MAX vectors.
>>>
>>> I would stick to the old method and limit the num to a reasonable 
>>> value e.g. MAX_WRITE_ZEROES_DEFAULT.
>>> This becomes necessary as you set INT_MAX for max_write_zeroes. That 
>>> hasn't been considered before in
>>> the original patch.
>>>
>>> Peter
>>>
>>
>> hmm. You are right, but I think that it would be better to limit 
>> iovec size
>> to 32 and this will solve the problem. Allocation of 32 Mb could be a 
>> real problem
>> on loaded system could be a problem.
>>
>> What do you think on this? May be we could consider 16 as a limit...
>
> I would do the following:
>
> ---8<---
>
> From 8c2a08baddbcd9e89bbb11fa83a42350bd7cc095 Mon Sep 17 00:00:00 2001
> From: Peter Lieven <pl@kamp.de>
> Date: Mon, 5 Jan 2015 12:14:52 +0100
> Subject: [PATCH] block: limited request size in write zeroes 
> unsupported path
>
> If bs->bl.max_write_zeroes is large and we end up in the unsupported
> path we might allocate a lot of memory for the iovector and/or even
> generate an oversized requests.
>
> Fix this by limiting the request by the minimum of the reported
> maximum transfer size or 16MB (32768 sectors).
>
> Reported-by: Denis V. Lunev <den@openvz.org>
> Signed-off-by: Peter Lieven <pl@kamp.de>
> ---
>  block.c |    5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/block.c b/block.c
> index a612594..8009478 100644
> --- a/block.c
> +++ b/block.c
> @@ -3203,6 +3203,9 @@ static int coroutine_fn 
> bdrv_co_do_write_zeroes(BlockDriverState *bs,
>
>          if (ret == -ENOTSUP) {
>              /* Fall back to bounce buffer if write zeroes is 
> unsupported */
> +            int max_xfer_len = MIN_NON_ZERO(bs->bl.max_transfer_length,
> + MAX_WRITE_ZEROES_DEFAULT);
> +            num = MIN(num, max_xfer_len);
this is not going to work IMHO. num is the number in sectors.
max_xfer_len is in bytes.

I will send my updated version using your approach in a
couple of minutes. Would like to test it a bit.

> iov.iov_len = num * BDRV_SECTOR_SIZE;
>              if (iov.iov_base == NULL) {
>                  iov.iov_base = qemu_try_blockalign(bs, num * 
> BDRV_SECTOR_SIZE);
> @@ -3219,7 +3222,7 @@ static int coroutine_fn 
> bdrv_co_do_write_zeroes(BlockDriverState *bs,
>              /* Keep bounce buffer around if it is big enough for all
>               * all future requests.
>               */
> -            if (num < max_write_zeroes) {
> +            if (num < max_xfer_len) {
>                  qemu_vfree(iov.iov_base);
>                  iov.iov_base = NULL;
>              }