Message-ID: <ff0c2ab7-8e82-40d9-1adf-78ee12846e1f@opensource.wdc.com>
Date:   Wed, 2 Nov 2022 19:07:06 +0900
From:   Damien Le Moal <damien.lemoal@...nsource.wdc.com>
To:     John Garry <john.g.garry@...cle.com>,
        John Garry <john.garry@...wei.com>, jejb@...ux.ibm.com,
        martin.petersen@...cle.com, hare@...e.de, bvanassche@....org,
        hch@....de, ming.lei@...hat.com, niklas.cassel@....com
Cc:     axboe@...nel.dk, jinpu.wang@...ud.ionos.com,
        linux-block@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-ide@...r.kernel.org, linux-scsi@...r.kernel.org,
        linuxarm@...wei.com, john.garry2@...l.dcu.ie
Subject: Re: [PATCH RFC v3 2/7] ata: libata-scsi: Add
 ata_internal_queuecommand()

On 11/2/22 18:52, John Garry wrote:
> Hi Damien,
> 
>>>>>
>>>>> Please also note that for AHCI, I make reserved depth =1, while for SAS
>>>>> controllers it is greater. This means that in theory we could alloc > 1x
>>>>> reserved command for SATA disk, but I don't think it matters.
>>>> Yes, 1 is enough. However, is 1 reserved out of 32 total, meaning that
>>>> the
>>>> user can only use 31 tags ? or is it 32+1 reserved ? which we can do
>>>> since
>>>> when using the reserved request, we will not use a hw tag (all reserved
>>>> requests will be non-ncq).
>>>>
>>>> The 32 + 1 scheme will work.
>>>
>>> Yes, 32 + 1 is what we want. I now think that there is a mistake in my
>>> code in this series for ahci.
>>>
>>> So I think we need for ahci:
>>>
>>> can_queue = 33 >
>> Hmm.. For ATA, can_queue should be only 32. nr_tags is going to be 33
>> though as we include one tag for a reserved command. No ? (may be I
>> misunderstand can_queue though).
> 
> Yes, we want nr_tags=33. But according to semantics of can_queue, it 
> includes total of regular and reserved tags. This is because tag_set 
> depth is total of regular and reserved tags, and we subtract reserved 
> tags from total depth in blk_mq_init_bitmaps():
> 
> int blk_mq_init_bitmaps(.., unsigned int queue_depth, unsigned int 
> reserved, ..)
> {
> 	unsigned int depth = queue_depth - reserved;
> 	...
> 
> 	if (bt_alloc(bitmap_tags, depth, round_robin, node
> 
> 
> So we want a change like this as well:
> 
> diff --git a/drivers/ata/ahci.h b/drivers/ata/ahci.h
> index da7ee8bec165..cbcc337055a7 100644
> --- a/drivers/ata/ahci.h
> +++ b/drivers/ata/ahci.h
> @@ -390,7 +390,7 @@ extern const struct attribute_group 
> *ahci_sdev_groups[];
>   */
> #define AHCI_SHT(drv_name)                                             \
>         __ATA_BASE_SHT(drv_name),                                       \
> -       .can_queue              = AHCI_MAX_CMDS,                        \
> +       .can_queue              = AHCI_MAX_CMDS + 1,                    \
>         .sg_tablesize           = AHCI_MAX_SG,                          \
>         .dma_boundary           = AHCI_DMA_BOUNDARY,                    \
>         .shost_groups           = ahci_shost_groups,
> 
> I know it seems strange, but still 32 tags will only ever be available 
> for non-internal commands (as it is today) and 1 for ata internal command.
> 
>>
>>> nr_reserved_cmds = 1
>>>
>>> while I only have can_queue = 32
>>
>> Which seems right to me.
>>
>>>
>>> I need to check that again for ahci driver and AHCI SHT...
>>>
>>>> But for CDL command completion handling, we
>>>> will need a NCQ command to do a read log, to avoid forcing a queue drain.
>>>> For that to reliably work, we'll need a 31+1+1 setup...
>>>>
>>>
>>> So is your idea to permanently reserve 1 more command from 32 commands ?
>>
>> Yes. Command Duration Limits has this weird case were commands may be
>> failed when exceeding their duration limit with a "good status" and
>> "sense data available bit" set. This case was defined to avoid the queue
>> stall that happens with any NCQ error. The way to handle this without
>> relying on EH (as that would also cause a queue drain) is to issue an
>> NCQ read log command to fetch the "sense data for successful NCQ
>> commands" log, to retrieve the sense data for the completed command and
>> check if it really failed or not. So we absolutely need a reserved
>> command for this, Without a reserved command, it is a nightmare to
>> support (tag reuse would be another solution, but it is horrible).
>>
>>> Or re-use 1 from 32 (and still also have 1 separate internal command)?
>>
>> I am not yet 100% sure if we can treat that internal NCQ read log like
>> any other read/write request... If we can, then the 1-out-of-32
>> reservation would not be needed. Need to revisit all the cases we need
>> to take care of (because in the middle of this CDL completion handling,
>> regular NCQ errors can happen, resulting in a drive reset that could
>> wreck everything as we lose the sense data for the completed requests).
>>
>> In any case, I think that we can deal with that extra reserved command
>> on top of you current series. No need to worry about it for now I think.
>>
> 
> So are you saying that you are basing current CDL support on libata 
> internally managing this extra reserved tag (and so do not need this 
> SCSI midlayer reserved tag support yet)?

Not really. For now, it is using libata EH, that is, when we need the
internal command for the read log, we know the device is idle and no
command is on-going. So we send a non-NCQ command which does not have a tag.

Ideally, all of this should use a real reserved tag to allow for an NCQ
read log outside of EH, avoiding the drive queue drain.

> 
> Thanks,
> John

-- 
Damien Le Moal
Western Digital Research

Powered by blists - more mailing lists