Re: [PATCH 06/18] scsi: Make scsi_ioctl_reset() pass the request queue pointer to blk_rq_init()

From: "hch@lst.de" <hch@lst.de>
To: Hannes Reinecke <hare@suse.com>
Cc: "hch@lst.de" <hch@lst.de>,
	Bart Van Assche <Bart.VanAssche@sandisk.com>,
	"linux-scsi@vger.kernel.org" <linux-scsi@vger.kernel.org>,
	"James.Bottomley@HansenPartnership.com"
	<James.Bottomley@HansenPartnership.com>,
	"martin.petersen@oracle.com" <martin.petersen@oracle.com>
Subject: Re: [PATCH 06/18] scsi: Make scsi_ioctl_reset() pass the request queue pointer to blk_rq_init()
Date: Mon, 22 May 2017 15:00:40 +0200	[thread overview]
Message-ID: <20170522130040.GA12773@lst.de> (raw)
In-Reply-To: <95b0aba4-9463-eb8e-5c27-4a6951c3112a@suse.com>

On Mon, May 22, 2017 at 02:56:18PM +0200, Hannes Reinecke wrote:
> > No, we don't.  The driver simply sets a tag aside and doesn't expose
> > it to the block layer.  Similar to what smartpqi already does for LUN
> > resets and AENs, mpt3sas does for the ioctl tags and NVMe does for AERs.
> > Personally I feel a bit uncomfortable by setting aside just one tag for
> TMFs; this assumes we'll never be sending LUN resets to devices in parallel.
> 
> But maybe this is a discussion for another time, if and when we finally
> move to that.

For midlayer initiated EH that's the case - they are serialized by
being issued from the eh thread.  For the ioctl we have the
scsi_block_when_processing_errors / scsi_host_in_recovery magic,
although that hand crafted primitive is a bit of a mess, we'd
be much better off with a real EH mutex.