[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read reque
From: |
John Snow |
Subject: |
Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async |
Date: |
Tue, 6 Oct 2015 13:56:43 -0400 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.2.0 |
On 10/06/2015 01:12 PM, Peter Lieven wrote:
>
>> Am 06.10.2015 um 19:07 schrieb John Snow <address@hidden>:
>>
>>
>>
>>> 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)
>
> isnt nb_sectors limited to CD_MAX_SECTORS (32)?
>
> Peter
>
CD_MAX_SECTORS is... (80 * 60 * 75 * 2048) / 512 --> 1440000, and
describes the maximum sector size for a CD medium, not the request size.
Where'd you get the 32 number?
>
>>
>> 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
>>>
--
—js
Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, Kevin Wolf, 2015/10/08
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, Peter Lieven, 2015/10/08
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, John Snow, 2015/10/08
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, Peter Lieven, 2015/10/08
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async,
John Snow <=
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, Peter Lieven, 2015/10/08
- Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, John Snow, 2015/10/08
Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, John Snow, 2015/10/08
Re: [Qemu-block] [Qemu-devel] [PATCH 1/5] ide/atapi: make PIO read requests async, Kevin Wolf, 2015/10/08