Re: [dm-devel] [Lsf] Notes from the four separate IO track sessions at LSF/MM

[Bart Van Assche <bart.vanassche@sandisk.com>]

On 04/29/2016 05:47 PM, Laurence Oberman wrote:
> From: "Bart Van Assche" <bart.vanassche@sandisk.com>
> To: "Laurence Oberman" <loberman@redhat.com>
> Cc: "James Bottomley" <James.Bottomley@HansenPartnership.com>, "linux-scsi" <linux-scsi@vger.kernel.org>, "Mike Snitzer" <snitzer@redhat.com>, linux-block@vger.kernel.org, "device-mapper development" <dm-devel@redhat.com>, lsf@lists.linux-foundation.org
> Sent: Friday, April 29, 2016 8:36:22 PM
> Subject: Re: [dm-devel] [Lsf] Notes from the four separate IO track sessions at LSF/MM
>
>> On 04/29/2016 02:47 PM, Laurence Oberman wrote:
>>> Recovery with 21 LUNS is 300s that have in-flights to abort.
>>> [ ... ]
>>> eh_deadline is set to 10 on the 2 qlogic ports, eh_timeout is set
>>> to 10 for all devices. In multipath fast_io_fail_tmo=5
>>>
>>> I jam one of the target array ports and discard the commands
>>> effectively black-holing the commands and leave it that way until
>>> we recover and I watch the I/O. The recovery takes around 300s even
>>> with all the tuning and this effectively lands up in Oracle cluster
>>> evictions.
>>
>> This discussion started as a discussion about the time needed to fail
>> over from one path to another. How long did it take in your test before
>> I/O failed over from the jammed port to another port?
 >
 > Around 300s before the paths were declared hard failed and the
 > devices offlined. This is when I/O restarts.
 > The remaining paths on the second Qlogic port (that are not jammed)
 > will not be used until the error handler activity completes.
 >
 > Until we get these for example, and device-mapper starts declaring
 > paths down we are blocked.
 > Apr 29 17:20:51 localhost kernel: sd 1:0:1:0: Device offlined - not
 > ready after error recovery
 > Apr 29 17:20:51 localhost kernel: sd 1:0:1:13: Device offlined - not
 > ready after error recovery

Hello Laurence,

Everyone else on all mailing lists to which this message has been posted 
replies below the message. Please follow this convention.

Regarding the fail-over time: the ib_srp driver guarantees that 
scsi_done() is invoked from inside its terminate_rport_io() function. 
Apparently the lpfc and the qla2xxx drivers behave differently. Please 
work with the maintainers of these drivers to reduce fail-over time.

Bart.