Re: [PATCH] Chaining sg lists for big IO commands v5

[Benny Halevy <bhalevy@panasas.com>]

Jens Axboe wrote:
> On Mon, May 21 2007, Jens Axboe wrote:
>> On Fri, May 18 2007, Badari Pulavarty wrote:
>>> On Fri, 2007-05-18 at 09:35 +0200, Jens Axboe wrote:
>>>> On Thu, May 17 2007, Badari Pulavarty wrote:
>>>>> On Thu, 2007-05-17 at 08:27 +0200, Jens Axboe wrote:
>>>>>> On Wed, May 16 2007, Badari Pulavarty wrote:
>>>>>>> On Tue, 2007-05-15 at 19:50 +0200, Jens Axboe wrote:
>>>>>>>> On Tue, May 15 2007, Badari Pulavarty wrote:
>>>>>>>>> On Tue, 2007-05-15 at 19:20 +0200, Jens Axboe wrote:
>>>>>>>>>> On Tue, May 15 2007, Badari Pulavarty wrote:
>>>>>>>>>>> On Fri, 2007-05-11 at 15:51 +0200, Jens Axboe wrote:
>>>>>>>>>>>> Hi,
>>>>>>>>>>>>
>>>>>>>>>>>> Updated version of the patch - this time I'll just attach the patch
>>>>>>>>>>>> file...
>>>>>>>>>>> Missing scatterlist.h inclusions..
>>>>>>>>>>>
>>>>>>>>>>> drivers/scsi/sym53c8xx_2/sym_glue.c: In function ???sym_scatter???:
>>>>>>>>>>> drivers/scsi/sym53c8xx_2/sym_glue.c:385: warning: implicit declaration
>>>>>>>>>>> of function ???for_each_sg???
>>>>>>>>>>> drivers/scsi/sym53c8xx_2/sym_glue.c:385: error: expected ???;??? before ???{???
>>>>>>>>>>> token
>>>>>>>>>>> drivers/scsi/sym53c8xx_2/sym_glue.c:375: warning: unused variable ???tp???
>>>>>>>>>>> make[3]: *** [drivers/scsi/sym53c8xx_2/sym_glue.o] Error 1
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> drivers/scsi/qla2xxx/qla_iocb.c: In function ???qla24xx_build_scsi_iocbs???:
>>>>>>>>>>> drivers/scsi/qla2xxx/qla_iocb.c:678: warning: implicit declaration of
>>>>>>>>>>> function ???for_each_sg???
>>>>>>>>>>> drivers/scsi/qla2xxx/qla_iocb.c:678: error: expected ???;??? before ???{???
>>>>>>>>>>> token
>>>>>>>>>> Thanks, will fix those. What arch? I tested it here.
>>>>>>>>> I am playing with them on ppc64.
>>>>>>>> Ah ok, you need the updated patch series for ppc64 support. Builds fine
>>>>>>>> here on ppc64. See the #sglist branch of the block repo:
>>>>>>>>
>>>>>>>> git://git.kernel.dk/data/git/linux-2.6-block.git
>>>>>>>>
>>>>>>>> I can mail you an updated patch, if you want.
>>>>>>>
>>>>>>> Here is the whole panic stack..
>>>>>> Thanks will fix that up, the IDE part is totally untested. Can you try
>>>>>> and backout this patch and see if it boots?
>>>>> I increased max_segments to 1024 on my qla2200 attached disks and
>>>>> simple "dd" (direct read) resulted in following:
>>>>>
>>>>> elm3b29:/sys/block/sdd/queue # echo 1024 > max_segments
>>>>> elm3b29:/sys/block/sdd/queue # cat max_hw_sectors_kb > max_sectors_kb
>>>>> elm3b29:/mnt # dd iflag=direct if=./z of=/dev/null bs=512M
>>>>>
>>>>> Unable to handle kernel paging request at 0000000000001008 RIP:
>>>>>  [<ffffffff8025e7af>] __rmqueue+0x6f/0x120
>>>> Auch, that's a bug. I don't think the oom path has been tested yet,
>>>> perhaps this is hitting it.
>>>>
>>>> Can you try with this debug patch, plus enable the slab debugging
>>>> helpers (like poisoning)?
>>>>
>>>> diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
>>>> index 7456992..a479d1e 100644
>>>> --- a/drivers/scsi/scsi_lib.c
>>>> +++ b/drivers/scsi/scsi_lib.c
>>>> @@ -793,6 +793,7 @@ struct scatterlist *scsi_alloc_sgtable(struct scsi_cmnd *cmd, gfp_t gfp_mask)
>>>>  	return ret;
>>>>  enomem:
>>>>  	if (ret) {
>>>> +		printk(KERN_ERR "scsi: failed to allocate sg table\n");
>>>>  		/*
>>>>  		 * Free entries chained off ret. Since we were trying to
>>>>  		 * allocate another sglist, we know that all entries are of
>>>>
>>> Not much help. I get all kinds of weird panics.. This time I got (with
>>> the above debug).
>>>
>>> general protection fault: 0000 [1] SMP
>>> CPU 1
>>> Modules linked in: jfs sg sd_mod qla2xxx firmware_class
>>> scsi_transport_fc scsi_mod vfat fat ipv6 thermal processor fan button
>>> battery ac dm_mod floppy parport_pc lp parport
>>> Pid: 56, comm: kblockd/1 Not tainted 2.6.22-rc1-sg #8
>>> RIP: 0010:[<ffffffff802816b6>]  [<ffffffff802816b6>] kmem_cache_alloc
>>> +0x36/0x70
>>> RSP: 0018:ffff81017abbfc10  EFLAGS: 00010002
>>> RAX: 0000000000000000 RBX: 0000000000000082 RCX: 0600000000000064
>>> RDX: ffff81019ff2b8a0 RSI: 0000000000011220 RDI: ffffffff8068d120
>>> RBP: ffff81017abbfc20 R08: 00000000000039f8 R09: 0000000000000000
>>> R10: ffff81019cbee700 R11: 0000000000000188 R12: ffff8101df2a64e0
>>> R13: 0000000000011220 R14: ffff8101df2a6510 R15: ffff81017abbfc50
>>> FS:  00002b505b027f20(0000) GS:ffff81018021f300(0000)
>>> knlGS:00000000f7da26b0
>>> CS:  0010 DS: 0018 ES: 0018 CR0: 000000008005003b
>>> CR2: 00002b505b029008 CR3: 000000019af73000 CR4: 00000000000006e0
>>> Process kblockd/1 (pid: 56, threadinfo ffff81017abbe000, task
>>> ffff81017a571440)
>>> Stack:  000000007abbfc30 0000000000000000 ffff81017abbfc30
>>> ffffffff8025d001
>>>  ffff81017abbfcb0 ffffffff8025d122 ffff81017abbfc60 ffffffff80219dc0
>>>  ffffffff880e5da6 00000000000000ad ffff81017abbfcd0 ffffffff8021a366
>>> Call Trace:
>>>  [<ffffffff8025d001>] mempool_alloc_slab+0x11/0x20
>>>  [<ffffffff8025d122>] mempool_alloc+0x42/0x110
>>>  [<ffffffff80219dc0>] flush_gart+0x40/0x50
>>>  [<ffffffff880e5da6>] :scsi_mod:__scsi_get_command+0x26/0x90
>>>  [<ffffffff8021a366>] gart_map_sg+0x2d6/0x3e0
>> Smells like a bug in the gart modifications, I'll double check them.
>> Does it work if you boot with iommu=off?
> 
> If iommu=off works, can you try a normal boot but with this applied on
> top of the sglist patches? That should fix gart mapping.
> 
> diff --git a/arch/x86_64/kernel/pci-gart.c b/arch/x86_64/kernel/pci-gart.c
> index 2e22a3a..b16384f 100644
> --- a/arch/x86_64/kernel/pci-gart.c
> +++ b/arch/x86_64/kernel/pci-gart.c
> @@ -381,7 +381,7 @@ int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
>  	int start;
>  	unsigned long pages = 0;
>  	int need = 0, nextneed;
> -	struct scatterlist *s, *ps, *start_sg;
> +	struct scatterlist *s, *ps, *start_sg, *sgmap;
>  
>  	if (nents == 0) 
>  		return 0;
> @@ -391,7 +391,7 @@ int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
>  
>  	out = 0;
>  	start = 0;
> -	start_sg = sg;
> +	start_sg = sgmap = sg;
>  	ps = NULL; /* shut up gcc */
>  	for_each_sg(sg, s, nents, i) {
>  		dma_addr_t addr = page_to_phys(s->page) + s->offset;
> @@ -405,11 +405,12 @@ int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
>  			/* Can only merge when the last chunk ends on a page 
>  			   boundary and the new one doesn't have an offset. */
>  			if (!iommu_merge || !nextneed || !need || s->offset ||
> -			    (ps->offset + ps->length) % PAGE_SIZE) { 
> -				if (dma_map_cont(start_sg, i - start, sg+out,
> +			    (ps->offset + ps->length) % PAGE_SIZE) {
> +				if (dma_map_cont(start_sg, i - start, sgmap,
>  						  pages, need) < 0)
>  					goto error;
>  				out++;
> +				sgmap = sg_next(sgmap);
>  				pages = 0;
>  				start = i;
>  				start_sg = s;
> @@ -420,7 +421,7 @@ int gart_map_sg(struct device *dev, struct scatterlist *sg, int nents, int dir)
>  		pages += to_pages(s->offset, s->length);
>  		ps = s;
>  	}
> -	if (dma_map_cont(start_sg, i - start, sg+out, pages, need) < 0)
> +	if (dma_map_cont(start_sg, i - start, sgmap, pages, need) < 0)
>  		goto error;
>  	out++;
>  	flush_gart();
> 

Looks good.
Thanks for picking this up.

Benny