Re: [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async

[John Snow]

On 10/06/2015 05:20 AM, Peter Lieven wrote:
> Am 06.10.2015 um 10:57 schrieb Kevin Wolf:
>> Am 05.10.2015 um 23:15 hat John Snow geschrieben:
>>>
>>> On 09/21/2015 08:25 AM, Peter Lieven wrote:
>>>> PIO read requests on the ATAPI interface used to be sync blk requests.
>>>> This has to siginificant drawbacks. First the main loop hangs util an
>>>> I/O request is completed and secondly if the I/O request does not
>>>> complete (e.g. due to an unresponsive storage) Qemu hangs completely.
>>>>
>>>> Signed-off-by: Peter Lieven <address@hidden>
>>>> ---
>>>>   hw/ide/atapi.c | 69
>>>> ++++++++++++++++++++++++++++++++++++----------------------
>>>>   1 file changed, 43 insertions(+), 26 deletions(-)
>>>>
>>>> diff --git a/hw/ide/atapi.c b/hw/ide/atapi.c
>>>> index 747f466..9257e1c 100644
>>>> --- a/hw/ide/atapi.c
>>>> +++ b/hw/ide/atapi.c
>>>> @@ -105,31 +105,51 @@ static void cd_data_to_raw(uint8_t *buf, int lba)
>>>>       memset(buf, 0, 288);
>>>>   }
>>>>   -static int cd_read_sector(IDEState *s, int lba, uint8_t *buf, int
>>>> sector_size)
>>>> +static void cd_read_sector_cb(void *opaque, int ret)
>>>>   {
>>>> -    int ret;
>>>> +    IDEState *s = opaque;
>>>>   -    switch(sector_size) {
>>>> -    case 2048:
>>>> -        block_acct_start(blk_get_stats(s->blk), &s->acct,
>>>> -                         4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
>>>> -        ret = blk_read(s->blk, (int64_t)lba << 2, buf, 4);
>>>> -        block_acct_done(blk_get_stats(s->blk), &s->acct);
>>>> -        break;
>>>> -    case 2352:
>>>> -        block_acct_start(blk_get_stats(s->blk), &s->acct,
>>>> -                         4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
>>>> -        ret = blk_read(s->blk, (int64_t)lba << 2, buf + 16, 4);
>>>> -        block_acct_done(blk_get_stats(s->blk), &s->acct);
>>>> -        if (ret < 0)
>>>> -            return ret;
>>>> -        cd_data_to_raw(buf, lba);
>>>> -        break;
>>>> -    default:
>>>> -        ret = -EIO;
>>>> -        break;
>>>> +    block_acct_done(blk_get_stats(s->blk), &s->acct);
>>>> +
>>>> +    if (ret < 0) {
>>>> +        ide_atapi_io_error(s, ret);
>>>> +        return;
>>>> +    }
>>>> +
>>>> +    if (s->cd_sector_size == 2352) {
>>>> +        cd_data_to_raw(s->io_buffer, s->lba);
>>>>       }
>>>> -    return ret;
>>>> +
>>>> +    s->lba++;
>>>> +    s->io_buffer_index = 0;
>>>> +    s->status &= ~BUSY_STAT;
>>>> +
>>>> +    ide_atapi_cmd_reply_end(s);
>>>> +}
>>>> +
>>>> +static int cd_read_sector(IDEState *s, int lba, void *buf, int
>>>> sector_size)
>>>> +{
>>>> +    if (sector_size != 2048 && sector_size != 2352) {
>>>> +        return -EINVAL;
>>>> +    }
>>>> +
>>>> +    s->iov.iov_base = buf;
>>>> +    if (sector_size == 2352) {
>>>> +        buf += 4;
>>>> +    }
>> This doesn't look quite right, buf is never read after this.
>>
>> Also, why +=4 when it was originally buf + 16?
> 
> You are right. I mixed that up.
> 
>>
>>>> +
>>>> +    s->iov.iov_len = 4 * BDRV_SECTOR_SIZE;
>>>> +    qemu_iovec_init_external(&s->qiov, &s->iov, 1);
>>>> +
>>>> +    if (blk_aio_readv(s->blk, (int64_t)lba << 2, &s->qiov, 4,
>>>> +                      cd_read_sector_cb, s) == NULL) {
>>>> +        return -EIO;
>>>> +    }
>>>> +
>>>> +    block_acct_start(blk_get_stats(s->blk), &s->acct,
>>>> +                     4 * BDRV_SECTOR_SIZE, BLOCK_ACCT_READ);
>>>> +    s->status |= BUSY_STAT;
>>>> +    return 0;
>>>>   }
>>>>   
>>> We discussed this off-list a bit, but for upstream synchronization:
>>>
>>> Unfortunately, I believe making cd_read_sector here non-blocking makes
>>> ide_atapi_cmd_reply_end non-blocking, and as a result makes calls to
>>> s->end_transfer_func() nonblocking, which functions like ide_data_readw
>>> are not prepared to cope with.
>> I don't think that's a problem as long as BSY is set while the
>> asynchronous command is running and DRQ is cleared. The latter will
>> protect ide_data_readw(). ide_sector_read() does essentially the same
>> thing.
> 
>  I was thinking the same. Without the BSY its not working at all.
> 
>>
>> Or maybe I'm just missing what you're trying to say.
>>
>>> My suggestion is to buffer an entire DRQ block of data at once
>>> (byte_count_limit) to avoid the problem.
>> No matter whether there is a problem or not, buffering more data at once
>> (and therefore doing less requests) is better for performance anyway.
> 
> Its possible to do only one read in the backend and read the whole
> request into the IO buffer. I send a follow-up.
> 

Be cautious: we only have 128K (+4 bytes) to play with in the io_buffer
and the READ10 cdb can request up to 128MiB! For performance, it might
be nice to always buffer something like:

MIN(128K, nb_sectors * sector_size)

and then as the guest drains the DRQ block of size byte_count_limit
which can only be at largest 0xFFFE (we can fit in at least two of these
per io_buffer refill) we can just shift the data_ptr and data_end
pointers to utilize io_buffer like a ring buffer.

Because the guest can at most fetch 0xfffe bytes at a time, it will tend
to leave at least 4 bytes left over from a 64 block read. Luckily, we've
got 4 extra bytes in s->io_buffer, so with a ring buffer we can always
rebuffer *at least* two full DRQ blocks of data at a time.

The routine would basically look like this:

- No DRQ blocks buffered, so read up to 64 blocks or however many are
left for our transfer
- If we have at least one full DRQ block allocated, start the transfer
and send an interrupt
- If we ran out of DRQ blocks, go back to the top and buffer them.

This would eliminate the need for code stanza #3 in
ide_atapi_cmd_reply_end, which re-starts a transfer without signaling to
the guest. We'd only have:

ide_atapi_cmd_reply_end(...) {
  if (packet_transfer_size == 0) { end(...); return; }
  if (blocks_buffered < 1) { async_buffer_blocks(...); return; }
  ide_transfer_start(...)
  ide_set_irq(s->bus);
}


which is a good deal simpler than what we have now, though I need to
look into the formatting of raw CD data a little more to make sure my
numbers make sense... it may not be quite so easy to buffer multiple DRQ
blocks in some cases, but so it goes -- we should always be able to
buffer at least one.

> Maybe do you have a pointer to the test tool that John mentioned?
> 
> Peter
>