Re: [PATCH v3 37/42] hpsa: use block layer tag for command allocation

[Webb Scales <webbnh@hp.com>]

At this point, I've turned this work over to Don; I presume he'll want 
to address the bitmask search. I'll leave it up to him to decide what to 
do about the refcount.


                 Webb


On 3/26/15 11:10 AM, Tomas Henzl wrote:
> On 03/26/2015 03:38 PM, Webb Scales wrote:
>> Ah!  Tomas, you are absolutely correct!  The loop should not be
>> restarting the search from the beginning of the bitfield, but rather
>> should be proceeding to check the next bit.  (Otherwise, there's no
>> point in having more than one bit!!)
> Most of the time it will work as expected, my comment was about a corner case.
> Are you going to repost this patch? Btw. I think that a local variable
> to restart the loop with an incremented value is enough.
>
>> (This code has been tweaked so many times that when I read it now I no
>> longer see what it actually does....)
>>
>> And, Tomas, you have a good point regarding the difference between
>> atomic_inc() and atomic_inc_return(), but, again, the difference is only
>> a couple of instructions in the context of a long code path, so I think
>> it's a difference without a distinction.  (And, I'm looking forward to
>> the day when all of the reference counting stuff can be removed...I
>> think it's not far off.)
> If i got it right the refcount exist only because of the error handler,
> i also think it might be rewritten so, that it could be removed.
> This the atomic_inc/and atomic_inc_return is a small detail, but if you repost
> the patch why not fix it too?
>
> tomash
>
>>
>>                   Webb
>>
>>
>>
>> On 3/26/15 8:47 AM, Tomas Henzl wrote:
>>> On 03/25/2015 07:33 PM, Webb Scales wrote:
>>>> Tomas,
>>>>
>>>> You are correct that the previous approach of using find_next_zero_bit()
>>>> with a persistent offset is more run-time efficient in the worst case;
>>>> however, given that "normal" operation doesn't call this allocation
>>>> routine, and given that, when this routine is called, the bit mask being
>>>> searched only has about 16 bits in it, I opted for simplicity over
>>>> efficiency -- that is, I doubt that the difference in efficiency is
>>>> discernible, while getting rid of the last_allocation field is a
>>>> worthwhile savings in both memory use and code.
>>> My comment is not about efficiency, I believe that when you measure it you wont be
>>> able to measure any significant difference.
>>> But if this '(unlikely(refcount > 1))' is true for let's say the first entry in the bitfield
>>> this and any other command submitted later will not pass the cmd-alloc until the command that
>>> caused '(unlikely(refcount > 1))' to be true is resolved. That might cause unexpected
>>> behavior.
>>>
>>>> Regarding your earlier comment on the refcount variable, I believe that
>>>> it was handy for debugging purposes.  The code has undergone numerous
>>>> revisions, and the variable certainly could now be removed from the
>>>> source per your suggestion.  (Of course, the compiler is already
>>>> removing it, I'm sure. ;-) )
>>> Not sure if the compiler is able to switch from 'atomic_inc_return' to 'atomic_inc' :) though.
>>> It is not important. (I'd not comment on this without the other cmd_alloc inaccuracy)
>>>
>>> tomash
>>>
>>>>                    Webb
>>>>
>>>>
>>>>
>>>> -----Original Message-----
>>>> From: Tomas Henzl [mailto:thenzl@redhat.com]
>>>> Sent: Monday, March 23, 2015 11:58 AM
>>>> To: Don Brace; Scott Teel; Kevin Barnett; james.bottomley@parallels.com; hch@infradead.org; Justin Lindley; brace
>>>> Cc: linux-scsi@vger.kernel.org
>>>> Subject: Re: [PATCH v3 37/42] hpsa: use block layer tag for command allocation
>>>>
>>>> On 03/17/2015 09:06 PM, Don Brace wrote:
>>>>
>>>>> From: Webb Scales <webbnh@hp.com>
>>>>>
>>>>> Rework slave allocation:
>>>>>      - separate the tagging support setup from the hostdata setup
>>>>>      - make the hostdata setup act consistently when the lookup fails
>>>>>      - make the hostdata setup act consistently when the device is not added
>>>>>      - set up the queue depth consistently across these scenarios
>>>>>      - if the block layer mq support is not available, explicitly enable and
>>>>>        activate the SCSI layer tcq support (and do this at allocation-time so
>>>>>        that the tags will be available for INQUIRY commands)
>>>>>
>>>>> Tweak slave configuration so that devices which are masked are also
>>>>> not attached.
>>>>>
>>>>> Reviewed-by: Scott Teel <scott.teel@pmcs.com>
>>>>> Reviewed-by: Kevin Barnett <kevin.barnett@pmcs.com>
>>>>> Signed-off-by: Webb Scales <webbnh@hp.com>
>>>>> Signed-off-by: Don Brace <don.brace@pmcs.com>
>>>>> ---
>>>>>     drivers/scsi/hpsa.c |  153 +++++++++++++++++++++++++++++++++++++++++----------
>>>>>     drivers/scsi/hpsa.h |    1
>>>>>     2 files changed, 123 insertions(+), 31 deletions(-)
>>>>>
>>>>> diff --git a/drivers/scsi/hpsa.c b/drivers/scsi/hpsa.c
>>>>> index 34c178c..4e34a62 100644
>>>>> --- a/drivers/scsi/hpsa.c
>>>>> +++ b/drivers/scsi/hpsa.c
>>>>> @@ -44,6 +44,7 @@
>>>>>     #include <scsi/scsi_host.h>
>>>>>     #include <scsi/scsi_tcq.h>
>>>>>     #include <scsi/scsi_eh.h>
>>>>> +#include <scsi/scsi_dbg.h>
>>>>>     #include <linux/cciss_ioctl.h>
>>>>>     #include <linux/string.h>
>>>>>     #include <linux/bitmap.h>
>>>>> @@ -212,6 +213,9 @@ static int hpsa_compat_ioctl(struct scsi_device *dev, int cmd,
>>>>>     
>>>>>     static void cmd_free(struct ctlr_info *h, struct CommandList *c);
>>>>>     static struct CommandList *cmd_alloc(struct ctlr_info *h);
>>>>> +static void cmd_tagged_free(struct ctlr_info *h, struct CommandList *c);
>>>>> +static struct CommandList *cmd_tagged_alloc(struct ctlr_info *h,
>>>>> +					    struct scsi_cmnd *scmd);
>>>>>     static int fill_cmd(struct CommandList *c, u8 cmd, struct ctlr_info *h,
>>>>>     	void *buff, size_t size, u16 page_code, unsigned char *scsi3addr,
>>>>>     	int cmd_type);
>>>>> @@ -2047,11 +2051,17 @@ static void hpsa_cmd_resolve_events(struct ctlr_info *h,
>>>>>     	}
>>>>>     }
>>>>>     
>>>>> +static void hpsa_cmd_resolve_and_free(struct ctlr_info *h,
>>>>> +				      struct CommandList *c)
>>>>> +{
>>>>> +	hpsa_cmd_resolve_events(h, c);
>>>>> +	cmd_tagged_free(h, c);
>>>>> +}
>>>>> +
>>>>>     static void hpsa_cmd_free_and_done(struct ctlr_info *h,
>>>>>     		struct CommandList *c, struct scsi_cmnd *cmd)
>>>>>     {
>>>>> -	hpsa_cmd_resolve_events(h, c);
>>>>> -	cmd_free(h, c);
>>>>> +	hpsa_cmd_resolve_and_free(h, c);
>>>>>     	cmd->scsi_done(cmd);
>>>>>     }
>>>>>     
>>>>> @@ -2072,8 +2082,7 @@ static void hpsa_cmd_abort_and_free(struct ctlr_info *h, struct CommandList *c,
>>>>>     	hpsa_set_scsi_cmd_aborted(cmd);
>>>>>     	dev_warn(&h->pdev->dev, "CDB %16phN was aborted with status 0x%x\n",
>>>>>     			 c->Request.CDB, c->err_info->ScsiStatus);
>>>>> -	hpsa_cmd_resolve_events(h, c);
>>>>> -	cmd_free(h, c);		/* FIX-ME:  change to cmd_tagged_free(h, c) */
>>>>> +	hpsa_cmd_resolve_and_free(h, c);
>>>>>     }
>>>>>     
>>>>>     static void process_ioaccel2_completion(struct ctlr_info *h,
>>>>> @@ -4535,7 +4544,7 @@ static int hpsa_ciss_submit(struct ctlr_info *h,
>>>>>     	}
>>>>>     
>>>>>     	if (hpsa_scatter_gather(h, c, cmd) < 0) { /* Fill SG list */
>>>>> -		cmd_free(h, c);
>>>>> +		hpsa_cmd_resolve_and_free(h, c);
>>>>>     		return SCSI_MLQUEUE_HOST_BUSY;
>>>>>     	}
>>>>>     	enqueue_cmd_and_start_io(h, c);
>>>>> @@ -4581,6 +4590,8 @@ static inline void hpsa_cmd_partial_init(struct ctlr_info *h, int index,
>>>>>     {
>>>>>     	dma_addr_t cmd_dma_handle = h->cmd_pool_dhandle + index * sizeof(*c);
>>>>>     
>>>>> +	BUG_ON(c->cmdindex != index);
>>>>> +
>>>>>     	memset(c->Request.CDB, 0, sizeof(c->Request.CDB));
>>>>>     	memset(c->err_info, 0, sizeof(*c->err_info));
>>>>>     	c->busaddr = (u32) cmd_dma_handle;
>>>>> @@ -4675,27 +4686,24 @@ static int hpsa_scsi_queue_command(struct Scsi_Host *sh, struct scsi_cmnd *cmd)
>>>>>     
>>>>>     	/* Get the ptr to our adapter structure out of cmd->host. */
>>>>>     	h = sdev_to_hba(cmd->device);
>>>>> +
>>>>> +	BUG_ON(cmd->request->tag < 0);
>>>>> +
>>>>>     	dev = cmd->device->hostdata;
>>>>>     	if (!dev) {
>>>>>     		cmd->result = DID_NO_CONNECT << 16;
>>>>>     		cmd->scsi_done(cmd);
>>>>>     		return 0;
>>>>>     	}
>>>>> -	memcpy(scsi3addr, dev->scsi3addr, sizeof(scsi3addr));
>>>>>     
>>>>> -	if (unlikely(lockup_detected(h))) {
>>>>> -		cmd->result = DID_NO_CONNECT << 16;
>>>>> -		cmd->scsi_done(cmd);
>>>>> -		return 0;
>>>>> -	}
>>>>> -	c = cmd_alloc(h);
>>>>> +	memcpy(scsi3addr, dev->scsi3addr, sizeof(scsi3addr));
>>>>>     
>>>>>     	if (unlikely(lockup_detected(h))) {
>>>>>     		cmd->result = DID_NO_CONNECT << 16;
>>>>> -		cmd_free(h, c);
>>>>>     		cmd->scsi_done(cmd);
>>>>>     		return 0;
>>>>>     	}
>>>>> +	c = cmd_tagged_alloc(h, cmd);
>>>>>     
>>>>>     	/*
>>>>>     	 * Call alternate submit routine for I/O accelerated commands.
>>>>> @@ -4708,7 +4716,7 @@ static int hpsa_scsi_queue_command(struct Scsi_Host *sh, struct scsi_cmnd *cmd)
>>>>>     		if (rc == 0)
>>>>>     			return 0;
>>>>>     		if (rc == SCSI_MLQUEUE_HOST_BUSY) {
>>>>> -			cmd_free(h, c);
>>>>> +			hpsa_cmd_resolve_and_free(h, c);
>>>>>     			return SCSI_MLQUEUE_HOST_BUSY;
>>>>>     		}
>>>>>     	}
>>>>> @@ -4822,15 +4830,23 @@ static int hpsa_register_scsi(struct ctlr_info *h)
>>>>>     	sh->hostdata[0] = (unsigned long) h;
>>>>>     	sh->irq = h->intr[h->intr_mode];
>>>>>     	sh->unique_id = sh->irq;
>>>>> +	error = scsi_init_shared_tag_map(sh, sh->can_queue);
>>>>> +	if (error) {
>>>>> +		dev_err(&h->pdev->dev,
>>>>> +			"%s: scsi_init_shared_tag_map failed for controller %d\n",
>>>>> +			__func__, h->ctlr);
>>>>> +		goto fail_host_put;
>>>>> +	}
>>>>>     	error = scsi_add_host(sh, &h->pdev->dev);
>>>>> -	if (error)
>>>>> +	if (error) {
>>>>> +		dev_err(&h->pdev->dev, "%s: scsi_add_host failed for controller %d\n",
>>>>> +			__func__, h->ctlr);
>>>>>     		goto fail_host_put;
>>>>> +	}
>>>>>     	scsi_scan_host(sh);
>>>>>     	return 0;
>>>>>     
>>>>>      fail_host_put:
>>>>> -	dev_err(&h->pdev->dev, "%s: scsi_add_host"
>>>>> -		" failed for controller %d\n", __func__, h->ctlr);
>>>>>     	scsi_host_put(sh);
>>>>>     	return error;
>>>>>      fail:
>>>>> @@ -4840,6 +4856,23 @@ static int hpsa_register_scsi(struct ctlr_info *h)
>>>>>     }
>>>>>     
>>>>>     /*
>>>>> + * The block layer has already gone to the trouble of picking out a unique,
>>>>> + * small-integer tag for this request.  We use an offset from that value as
>>>>> + * an index to select our command block.  (The offset allows us to reserve the
>>>>> + * low-numbered entries for our own uses.)
>>>>> + */
>>>>> +static int hpsa_get_cmd_index(struct scsi_cmnd *scmd)
>>>>> +{
>>>>> +	int idx = scmd->request->tag;
>>>>> +
>>>>> +	if (idx < 0)
>>>>> +		return idx;
>>>>> +
>>>>> +	/* Offset to leave space for internal cmds. */
>>>>> +	return idx += HPSA_NRESERVED_CMDS;
>>>>> +}
>>>>> +
>>>>> +/*
>>>>>      * Send a TEST_UNIT_READY command to the specified LUN using the specified
>>>>>      * reply queue; returns zero if the unit is ready, and non-zero otherwise.
>>>>>      */
>>>>> @@ -4979,18 +5012,18 @@ static int hpsa_eh_device_reset_handler(struct scsi_cmnd *scsicmd)
>>>>>     	/* if controller locked up, we can guarantee command won't complete */
>>>>>     	if (lockup_detected(h)) {
>>>>>     		dev_warn(&h->pdev->dev,
>>>>> -			"scsi %d:%d:%d:%d RESET FAILED, lockup detected\n",
>>>>> -			h->scsi_host->host_no, dev->bus, dev->target,
>>>>> -			dev->lun);
>>>>> +			 "scsi %d:%d:%d:%u cmd %d RESET FAILED, lockup detected\n",
>>>>> +			 h->scsi_host->host_no, dev->bus, dev->target, dev->lun,
>>>>> +			 hpsa_get_cmd_index(scsicmd));
>>>>>     		return FAILED;
>>>>>     	}
>>>>>     
>>>>>     	/* this reset request might be the result of a lockup; check */
>>>>>     	if (detect_controller_lockup(h)) {
>>>>>     		dev_warn(&h->pdev->dev,
>>>>> -			 "scsi %d:%d:%d:%d RESET FAILED, new lockup detected\n",
>>>>> +			 "scsi %d:%d:%d:%u cmd %d RESET FAILED, new lockup detected\n",
>>>>>     			 h->scsi_host->host_no, dev->bus, dev->target,
>>>>> -			 dev->lun);
>>>>> +			 dev->lun, hpsa_get_cmd_index(scsicmd));
>>>>>     		return FAILED;
>>>>>     	}
>>>>>     
>>>>> @@ -5442,6 +5475,59 @@ static int hpsa_eh_abort_handler(struct scsi_cmnd *sc)
>>>>>     }
>>>>>     
>>>>>     /*
>>>>> + * For operations with an associated SCSI command, a command block is allocated
>>>>> + * at init, and managed by cmd_tagged_alloc() and cmd_tagged_free() using the
>>>>> + * block request tag as an index into a table of entries.  cmd_tagged_free() is
>>>>> + * the complement, although cmd_free() may be called instead.
>>>>> + */
>>>>> +static struct CommandList *cmd_tagged_alloc(struct ctlr_info *h,
>>>>> +					    struct scsi_cmnd *scmd)
>>>>> +{
>>>>> +	int idx = hpsa_get_cmd_index(scmd);
>>>>> +	struct CommandList *c = h->cmd_pool + idx;
>>>>> +	int refcount = 0;
>>>>> +
>>>>> +	if (idx < HPSA_NRESERVED_CMDS || idx >= h->nr_cmds) {
>>>>> +		dev_err(&h->pdev->dev, "Bad block tag: %d not in [%d..%d]\n",
>>>>> +			idx, HPSA_NRESERVED_CMDS, h->nr_cmds - 1);
>>>>> +		/* The index value comes from the block layer, so if it's out of
>>>>> +		 * bounds, it's probably not our bug.
>>>>> +		 */
>>>>> +		BUG();
>>>>> +	}
>>>>> +
>>>>> +	refcount = atomic_inc_return(&c->refcount);
>>>> refcount is never used, use atomic_inc(&c->refcount); instead?
>>>>
>>>>> +	if (unlikely(!hpsa_is_cmd_idle(c))) {
>>>>> +		/*
>>>>> +		 * We expect that the SCSI layer will hand us a unique tag
>>>>> +		 * value.  Thus, there should never be a collision here between
>>>>> +		 * two requests...because if the selected command isn't idle
>>>>> +		 * then someone is going to be very disappointed.
>>>>> +		 */
>>>>> +		dev_err(&h->pdev->dev,
>>>>> +			"tag collision (tag=%d) in cmd_tagged_alloc().\n",
>>>>> +			idx);
>>>>> +		if (c->scsi_cmd != NULL)
>>>>> +			scsi_print_command(c->scsi_cmd);
>>>>> +		scsi_print_command(scmd);
>>>>> +	}
>>>>> +
>>>>> +	hpsa_cmd_partial_init(h, idx, c);
>>>>> +	return c;
>>>>> +}
>>>>> +
>>>>> +static void cmd_tagged_free(struct ctlr_info *h, struct CommandList *c)
>>>>> +{
>>>>> +	/*
>>>>> +	 * Release our reference to the block.  We don't need to do anything
>>>>> +	 * else to free it, because it is accessed by index.  (There's no point
>>>>> +	 * in checking the result of the decrement, since we cannot guarantee
>>>>> +	 * that there isn't a concurrent abort which is also accessing it.)
>>>>> +	 */
>>>>> +	(void)atomic_dec(&c->refcount);
>>>>> +}
>>>>> +
>>>>> +/*
>>>>>      * For operations that cannot sleep, a command block is allocated at init,
>>>>>      * and managed by cmd_alloc() and cmd_free() using a simple bitmap to track
>>>>>      * which ones are free or in use.  Lock must be held when calling this.
>>>>> @@ -5454,7 +5540,6 @@ static struct CommandList *cmd_alloc(struct ctlr_info *h)
>>>>>     {
>>>>>     	struct CommandList *c;
>>>>>     	int refcount, i;
>>>>> -	unsigned long offset;
>>>>>     
>>>>>     	/*
>>>>>     	 * There is some *extremely* small but non-zero chance that that
>>>>> @@ -5466,31 +5551,39 @@ static struct CommandList *cmd_alloc(struct ctlr_info *h)
>>>>>     	 * very unlucky thread might be starved anyway, never able to
>>>>>     	 * beat the other threads.  In reality, this happens so
>>>>>     	 * infrequently as to be indistinguishable from never.
>>>>> +	 *
>>>>> +	 * Note that we start allocating commands before the SCSI host structure
>>>>> +	 * is initialized.  Since the search starts at bit zero, this
>>>>> +	 * all works, since we have at least one command structure available;
>>>>> +	 * however, it means that the structures with the low indexes have to be
>>>>> +	 * reserved for driver-initiated requests, while requests from the block
>>>>> +	 * layer will use the higher indexes.
>>>>>     	 */
>>>>>     
>>>>> -	offset = h->last_allocation; /* benignly racy */
>>>>>     	for (;;) {
>>>>> -		i = find_next_zero_bit(h->cmd_pool_bits, h->nr_cmds, offset);
>>>>> -		if (unlikely(i == h->nr_cmds)) {
>>>>> -			offset = 0;
>>>>> +		i = find_first_zero_bit(h->cmd_pool_bits, HPSA_NRESERVED_CMDS);
>>>>> +		if (unlikely(i >= HPSA_NRESERVED_CMDS))
>>>>>     			continue;
>>>>> -		}
>>>>>     		c = h->cmd_pool + i;
>>>>>     		refcount = atomic_inc_return(&c->refcount);
>>>>>     		if (unlikely(refcount > 1)) {
>>>>>     			cmd_free(h, c); /* already in use */
>>>>> -			offset = (i + 1) % h->nr_cmds;
>>>> Hi Don,
>>>> when this happens - a command has its bitfield flag cleared and, but it's taken - refcount is > 1
>>>> it will be so likely for next several thousands of tests in this function until the it is freed.
>>>> When it is the first bit in the bitfield it will block all following commands sent to the card for that time.
>>>> The previous variant  'find_next_zero_bit + offset = (i + 1) % h->nr_cmds' seems to handle this better.
>>>> Cheers, Tomas
>>>>
>>>>>     			continue;
>>>>>     		}
>>>>>     		set_bit(i & (BITS_PER_LONG - 1),
>>>>>     			h->cmd_pool_bits + (i / BITS_PER_LONG));
>>>>>     		break; /* it's ours now. */
>>>>>     	}
>>>>> -	h->last_allocation = i; /* benignly racy */
>>>>>     	hpsa_cmd_partial_init(h, i, c);
>>>>>     	return c;
>>>>>     }
>>>>>     
>>>>> +/*
>>>>> + * This is the complementary operation to cmd_alloc().  Note, however, in some
>>>>> + * corner cases it may also be used to free blocks allocated by
>>>>> + * cmd_tagged_alloc() in which case the ref-count decrement does the trick and
>>>>> + * the clear-bit is harmless.
>>>>> + */
>>>>>     static void cmd_free(struct ctlr_info *h, struct CommandList *c)
>>>>>     {
>>>>>     	if (atomic_dec_and_test(&c->refcount)) {
>>>>> diff --git a/drivers/scsi/hpsa.h b/drivers/scsi/hpsa.h
>>>>> index 3ec8934..2536b67 100644
>>>>> --- a/drivers/scsi/hpsa.h
>>>>> +++ b/drivers/scsi/hpsa.h
>>>>> @@ -141,7 +141,6 @@ struct ctlr_info {
>>>>>     	struct CfgTable __iomem *cfgtable;
>>>>>     	int	interrupts_enabled;
>>>>>     	int 	max_commands;
>>>>> -	int last_allocation;
>>>>>     	atomic_t commands_outstanding;
>>>>>     #	define PERF_MODE_INT	0
>>>>>     #	define DOORBELL_INT	1
>>>>>
>>>>> --
>>>>> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
>>>>> the body of a message to majordomo@vger.kernel.org
>>>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>>> --
>>>> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
>>>> the body of a message to majordomo@vger.kernel.org
>>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html