Message-ID: <20091116215417.GE3273@beardog.cce.hp.com>
Date:	Mon, 16 Nov 2009 15:54:17 -0600
From:	scameron@...rdog.cce.hp.com
To:	James Bottomley <James.Bottomley@...e.de>
Cc:	akpm@...ux-foundation.org, linux-kernel@...r.kernel.org,
	linux-scsi@...r.kernel.org, mikem@...rdog.cce.hp.com,
	scameron@...rdog.cce.hp.com
Subject: Re: [PATCH 00/17] hpsa driver updates

On Mon, Nov 16, 2009 at 03:06:49PM -0600, James Bottomley wrote:
> On Wed, 2009-11-11 at 10:50 -0600, Stephen M. Cameron wrote:
> > The following series implements hpsa scsi driver for HP Smart Arrays,
> > and some updates since the last time.
> > The first 5 patches in the series are already in Andrew Morton's tree.
> > 
> > ---
> > 
> > Andrew Morton (1):
> >       avoid helpful cleanup patches.
> > 
> > Stephen M. Cameron (16):
> >       hpsa: fix typo that causes scsi status to be lost
> >       hpsa:  Make fill_cmd() return void
> >       hpsa: Remove sendcmd, in no case are we required to poll for completions.
> >       hpsa: Flush cache with interrupts still enabled.
> >       hpsa: Retry driver initiated commands on unit attention
> >       hpsa: decode unit attention condition and retry commands.
> >       hpsa: Make hpsa_sdev_attrs static
> >       hpsa:  Allow device rescan to be triggered via sysfs.
> >       Add thread to allow controllers to register for rescan for new devices
> >       hpsa: Factor out some pci_unmap code
> >       hpsa: Factor out command submission sequence
> >       hpsa: Use shost_priv instead of accessing host->hostdata[0] directly.
> >       hpsa: Allocate the correct amount of extra space for the scsi host
> >       Fix use of unallocated memory for MSA2xxx enclosure device data.
> >       hpsa: Fix vendor id check
> >       Add hpsa driver for HP Smart Array controllers.
> 
> Actually, it's pretty difficult to review a 17 patch series like this
> because the human mind (or at least mine) doesn't retain sufficient
> context from patch to patch.  I ended up just pulling all 17 into a tree
> and reviewing the finished driver.
> 

I was trying to err on the safe side, in case anyone wanted the separate
patches, on the grounds that it's easier to bake a cake than to unbake one.
If it's preferred, I can pre-bake them into one big patch before sending
next time.  Also, not all the patches are authored by me, so I figured I
ought to preserve that.

> That said:
> 
> in hpsa.c:
> 
> > static struct device_attribute *hpsa_shost_attrs[] = {
> >	&dev_attr_rescan,
> >	NULL,
> > };
> 
> We already have a host scan attribute which (admittedly using the
> transport class logic) you can plug into ... can't you just use it?

Maybe.  I'll take a look at it. 

> It
> supplies user context, so you could dispense with all that scan thread
> stuff as well, I think.
> 
> > static DEFINE_MUTEX(scan_mutex);
> > static LIST_HEAD(scan_q);
> > static int scan_thread(void *data);
> 
> These names are too generic.  We already have a scan_mutex at least
> defined at the top level.  I know they're protected by static, but that
> doesn't necessarily help if they show up in a debug stack trace.

Ok.

> 
> All of this report luns stuff looks fairly identical to the report luns
> we do in scsi_scan.c ... barring the initial command, which could be
> translated.  Wouldn't it be easier to have the generic code parse and do
> all of this?
> 

Maybe.  There are complications.  We present some physical devices
(tape drives, changers, the RAID controller) some logical (logical
drives), the RAID controller is scsi revision 0, not 5, if we
present it first, the REPORT_LUNS is never sent by the upper layers
because of this, and only 7 devices get discovered, etc.  If we
present the first logical drive first, it sends the REPORT LUNS
to the the drive, etc.  It gets a little weird, as our hardware is
a little weird.  If it is possible, I doubt I would be able to have
such a change ready by 2.6.33 timeframe.

> > static int hpsa_scsi_queue_command(struct scsi_cmnd *cmd,
> >	void (*done)(struct scsi_cmnd *))
> > {
> [...]
> >	c = cmd_alloc(h);
> >	spin_unlock_irqrestore(&h->lock, flags);
> >	if (c == NULL) {			/* trouble... */
> >		dev_err(&h->pdev->dev, "cmd_alloc returned NULL!\n");
> >		cmd->result = DID_NO_CONNECT << 16;
> >		done(cmd);
> >		return 0;
> >	}
> 
> I think you want to return SCSI_MLQUEUE_HOST_BUSY here, which will
> trigger a throttle and retry after either something frees or I/O
> pressure builds more.

Ok.

> 
> > static int hpsa_eh_device_reset_handler(struct scsi_cmnd *scsicmd)
> > {
> [...]
> >	rc = hpsa_send_reset(h, dev->scsi3addr);
> >	if (rc == 0 && wait_for_device_to_become_ready(h, dev->scsi3addr) == 0)
> >		return SUCCESS;
> 
> So the first thing we do after a device reset successful return is send
> a test unit ready to the failing device ... there's no real need for you
> to duplicate that, is there?

It didn't seem to work until I did that.  this was awhile ago that I 
wrote this so my memory of it isn't too clear, but as I recall, I tested
this (in cciss, which is what this code evolved from) with a tape drive
with the long timeout set rather short, and "mt -f /dev/st0 erase"   The
artificially shortened timeout would mean that the erase would time out,
and this reset code would get triggered.  I found that without my change,
the device didn't become ready in time, even though the upper layers sent
TURs, and the device would get kicked off.  So I put in that code to make
it work.  Maybe it was the mid layer that needed tweaking?  maybe the slow
to recover tape drives deserved to get kicked off? 

> 
> The ioctl stuff looks like you could do it all with SG_IO now rather
> than rolling your own versions ... or is there some backward
> compatibility problem here?
> 

Yeah, mainly backward compatibility.

HP's array config utility (ACU), snmp storage agents,
firmware flashing utilities, cciss_vol_status,
arrayprobe, to name a few things,  depend on the passthrough ioctl
and there is more complete error information returned by the passthrough.
I also have a vague recollection of some buffer size limitation with SG_IO
I'm not sure, might be wrong about that, but I know flashing firmware
takes some large buffers.

> 
> > static __devinit int hpsa_hard_reset_controller(struct pci_dev *pdev)
> > {
> [...]
> >	set_current_state(TASK_UNINTERRUPTIBLE);
> >	schedule_timeout(HZ >> 1);
> 
> msleep(500) please ..  This isn't the only place this occurs, could you
> replace all of them?

Ok.

> 
> in hpsa_cmd.h:
> 
> > /* Unit Attentions ASC's as defined for the MSA2012sa */
> > #define POWER_OR_RESET			0x29
> > #define STATE_CHANGED			0x2a
> > #define UNIT_ATTENTION_CLEARED		0x2f
> > #define LUN_FAILED			0x3e
> > #define REPORT_LUNS_CHANGED		0x3f
> > 
> > /* Unit Attentions ASCQ's as defined for the MSA2012sa */
> > 
> > 	/* These ASCQ's defined for ASC = POWER_OR_RESET */
> > #define POWER_ON_RESET			0x00
> > #define POWER_ON_REBOOT			0x01
> > #define SCSI_BUS_RESET			0x02
> > #define MSA_TARGET_RESET		0x03
> > #define CONTROLLER_FAILOVER		0x04
> > #define TRANSCEIVER_SE			0x05
> > #define TRANSCEIVER_LVD			0x06
> > 
> > 	/* These ASCQ's defined for ASC = STATE_CHANGED */
> > #define RESERVATION_PREEMPTED		0x03
> > #define ASYM_ACCESS_CHANGED		0x06
> > #define LUN_CAPACITY_CHANGED		0x09
> 
> Traditionally we've shied away from putting ASC/ASCQ values into
> defines ... but these all look to be global not hpsa local, so they
> should be in a common central file.
> 

Ok.  I think there are some in there we didn't even use.  We should
probably omit them in those cases.

> Otherwise looks OK to a cursory glance.
> 
> James
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists