How can I identify the request queue for a linux block device

Question

I am working on this driver that connects the hard disk over the network. There is a bug that if I enable two or more hard disks on the computer, only the first one gets the partitions looked over and identified. The result is, if I have 1 partition on hda and 1 partitions on hdb, as soon as I connect hda there is a partition that can be mounted. So hda1 gets a blkid xyz123 as soon as it mounts. But when I go ahead and mount hdb1 it also comes up with the same blkid and in fact, the driver is reading it from hda, not hdb.

So I think I found the place where the driver is messing up. Below is a debug output including a dump_stack which I put at the first spot where it seems to be accessing the wrong device.

Here is the code section:

/*basically, this is just the request_queue processor. In the log output that
  follows, the second device, (hdb) has just been connected, right after hda
  was connected and hda1 was mounted to the system. */

void nblk_request_proc(struct request_queue *q)
{
struct request *req;
ndas_error_t err = NDAS_OK;

dump_stack();

while((req = NBLK_NEXT_REQUEST(q)) != NULL)
{
    dbgl_blk(8,"processing queue request from slot %d",SLOT_R(req));

    if (test_bit(NDAS_FLAG_QUEUE_SUSPENDED, &(NDAS_GET_SLOT_DEV(SLOT_R(req))->queue_flags)))  {
        printk ("ndas: Queue is suspended\n");
        /* Queue is suspended */
#if ( LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,31) )
        blk_start_request(req);
#else
        blkdev_dequeue_request(req);
#endif

Here is a log output. I have added some comments to help understand what is happening and where the bad call seems to come up.

  /* Just below here you can see "slot" mentioned many times. This is the 
     identification for the network case in which the hd is connected to the 
     network. So you will see slot 2 in this log because the first device has 
     already been connected and mounted. */

  kernel: [231644.155503] BL|4|slot_enable|/driver/block/ctrldev.c:281|adding disk: slot=2, first_minor=16, capacity=976769072|nd/dpcd1,64:15:44.38,3828:10
  kernel: [231644.155588] BL|3|ndop_open|/driver/block/ops.c:233|ing bdev=f6823400|nd/dpcd1,64:15:44.38,3720:10
  kernel: [231644.155598] BL|2|ndop_open|/driver/block/ops.c:247|slot =0x2|nd/dpcd1,64:15:44.38,3720:10
  kernel: [231644.155606] BL|2|ndop_open|/driver/block/ops.c:248|dev_t=0x3c00010|nd/dpcd1,64:15:44.38,3720:10
  kernel: [231644.155615] ND|3|ndas_query_slot|netdisk/nddev.c:791|slot=2 sdev=d33e2080|nd/dpcd1,64:15:44.38,3696:10
  kernel: [231644.155624] ND|3|ndas_query_slot|netdisk/nddev.c:817|ed|nd/dpcd1,64:15:44.38,3696:10
  kernel: [231644.155631] BL|3|ndop_open|/driver/block/ops.c:326|mode=1|nd/dpcd1,64:15:44.38,3720:10
  kernel: [231644.155640] BL|3|ndop_open|/driver/block/ops.c:365|ed open|nd/dpcd1,64:15:44.38,3724:10
  kernel: [231644.155653] BL|8|ndop_revalidate_disk|/driver/block/ops.c:2334|gendisk=c6afd800={major=60,first_minor=16,minors=0x10,disk_name=ndas-44700486-0,private_data=00000002,capacity=%lld}|nd/dpcd1,64:15:44.38,3660:10
  kernel: [231644.155668] BL|8|ndop_revalidate_disk|/driver/block/ops.c:2346|ed|nd/dpcd1,64:15:44.38,3652:10

  /* So at this point the hard disk is added (gendisk=c6...) and the identifications
     all match the network device. The driver is now about to begin scanning the 
     hard drive for existing partitions. the little 'ed', at the end of the previous
     line indicates that revalidate_disk has finished it's job. 

     Also, I think the request queue is indicated by the output dpcd1 near the very
     end of the line. 

     Now below we have entered the function that is pasted above. In the function
     you can see that the slot can be determined by the queue. And the log output
     after the stack dump shows it is from slot 1. (The first network drive that was
     already mounted.) */

        kernel: [231644.155677]  ndas-44700486-0:Pid: 467, comm: nd/dpcd1 Tainted: P           2.6.32-5-686 #1
  kernel: [231644.155711] Call Trace:
  kernel: [231644.155723]  [<fc5a7685>] ? nblk_request_proc+0x9/0x10c [ndas_block]
  kernel: [231644.155732]  [<c11298db>] ? __generic_unplug_device+0x23/0x25
  kernel: [231644.155737]  [<c1129afb>] ? generic_unplug_device+0x1e/0x2e
  kernel: [231644.155743]  [<c1123090>] ? blk_unplug+0x2e/0x31
  kernel: [231644.155750]  [<c10cceec>] ? block_sync_page+0x33/0x34
  kernel: [231644.155756]  [<c108770c>] ? sync_page+0x35/0x3d
  kernel: [231644.155763]  [<c126d568>] ? __wait_on_bit_lock+0x31/0x6a
  kernel: [231644.155768]  [<c10876d7>] ? sync_page+0x0/0x3d
  kernel: [231644.155773]  [<c10876aa>] ? __lock_page+0x76/0x7e
  kernel: [231644.155780]  [<c1043f1f>] ? wake_bit_function+0x0/0x3c
  kernel: [231644.155785]  [<c1087b76>] ? do_read_cache_page+0xdf/0xf8
  kernel: [231644.155791]  [<c10d21b9>] ? blkdev_readpage+0x0/0xc
  kernel: [231644.155796]  [<c1087bbc>] ? read_cache_page_async+0x14/0x18
  kernel: [231644.155801]  [<c1087bc9>] ? read_cache_page+0x9/0xf
  kernel: [231644.155808]  [<c10ed6fc>] ? read_dev_sector+0x26/0x60
  kernel: [231644.155813]  [<c10ee368>] ? adfspart_check_ICS+0x20/0x14c
  kernel: [231644.155819]  [<c10ee138>] ? rescan_partitions+0x17e/0x378
  kernel: [231644.155825]  [<c10ee348>] ? adfspart_check_ICS+0x0/0x14c
  kernel: [231644.155830]  [<c10d26a3>] ? __blkdev_get+0x225/0x2c7
  kernel: [231644.155836]  [<c10ed7e6>] ? register_disk+0xb0/0xfd
  kernel: [231644.155843]  [<c112e33b>] ? add_disk+0x9a/0xe8
  kernel: [231644.155848]  [<c112dafd>] ? exact_match+0x0/0x4
  kernel: [231644.155853]  [<c112deae>] ? exact_lock+0x0/0xd
  kernel: [231644.155861]  [<fc5a8b80>] ? slot_enable+0x405/0x4a5 [ndas_block]
  kernel: [231644.155868]  [<fc5a8c63>] ? ndcmd_enabled_handler+0x43/0x9e [ndas_block]
  kernel: [231644.155874]  [<fc5a8c20>] ? ndcmd_enabled_handler+0x0/0x9e [ndas_block]
  kernel: [231644.155891]  [<fc54b22b>] ? notify_func+0x38/0x4b [ndas_core]
  kernel: [231644.155906]  [<fc561cba>] ? _dpc_cancel+0x17c/0x626 [ndas_core]
  kernel: [231644.155919]  [<fc562005>] ? _dpc_cancel+0x4c7/0x626 [ndas_core]
  kernel: [231644.155933]  [<fc561cba>] ? _dpc_cancel+0x17c/0x626 [ndas_core]
  kernel: [231644.155941]  [<c1003d47>] ? kernel_thread_helper+0x7/0x10

  /* here are the output of the driver debugs. They show that this operation is
     being performed on the first devices request queue. */

  kernel: [231644.155948] BL|8|nblk_request_proc|/driver/block/block26.c:494|processing queue request from slot 1|nd/dpcd1,64:15:44.38,3408:10
  kernel: [231644.155959] BL|8|nblk_handle_io|/driver/block/block26.c:374|struct ndas_slot sd = NDAS GET SLOT DEV(slot 1)
  kernel: [231644.155966] |nd/dpcd1,64:15:44.38,3328:10
  kernel: [231644.155970] BL|8|nblk_handle_io|/driver/block/block26.c:458|case READA call ndas_read(slot=1, ndas_req)|nd/dpcd1,64:15:44.38,3328:10
  kernel: [231644.155979] ND|8|ndas_read|netdisk/nddev.c:824|read io: slot=1, cmd=0, req=x00|nd/dpcd1,64:15:44.38,3320:10

I hope this is enough background information. Maybe an obvious question at this moment is "When and where are the request_queues assigned?"

Well that is handled a little bit before the add_disk function. adding disk, is the first line on the log output.

slot->disk = NULL;
spin_lock_init(&slot->lock);
slot->queue = blk_init_queue(
    nblk_request_proc, 
    &slot->lock
);

As far as I know, this is the standard operation. So back to my original question. Can I find the request queue somewhere and make sure it is incremented or unique for each new device or does the Linux kernel only use one queue for each Major number? I want to discover why this driver is loading the same queue on two different block storages, and determine if that is causing the duplicate blkid during the initial registration process.

Thanks for looking at this situation for me.

Sumikosumma

You might want to try the Kernel Newbies mailing list. – Application

@ndasusers: It may make sense to check the 'type' of request before starting a new xx_request(req). Maybe an incoming request is only about a read access to your already mounted hda and you handle that as something related to fresh plugged hdb. – Grandpa

You wrote and mount hdab1 it also comes up I guess you mean hdb1 right ? – Kruse

I don't think you can have two devices with the same major and minor numbers. What are your minor and major numbers for the drive? – Anoxia

nice question, but are you trying to re-implement nbd? – Pape

Answer 1

Queue = blk_init_queue(sbd_request, &Device.lock);

Jurisdiction

Answer 2

I share the solution to the bug that led me to post this question. Though it does not actually answer the question of how to identify the device request queue.

In the code above is the following:

if (test_bit(NDAS_FLAG_QUEUE_SUSPENDED, 
       &(NDAS_GET_SLOT_DEV(SLOT_R(req))->queue_flags))) 

Well, that "SLOT_R(req)" was causing the trouble. That is defined else where to return the gendisk device.

#define SLOT_R(_request_) SLOT((_request_)->rq_disk)

This returned the disk, but not the proper value for various operations later on. So as the extra block devices were loaded, this function basically kept returning 1. (I think it was processing as a boolean.) Therefore, all requests were piled on the to the request queue for disk 1.

The fix was to access the correct disk identification value that was already stored in the disk's private_data when the it was added to the system.

Correct identifier definition:
   #define SLOT_R(_request_) ( (int) _request_->rq_disk->private_data )

How the correct disk number was stored.
   slot->disk->queue = slot->queue;
   slot->disk->private_data = (void*) (long) s;  <-- 's' is the disk id
   slot->queue_flags = 0;

Now the correct disk id is returned from private data, so all requests to the correct queue.

As mentioned, this does not show how to identify the queue though. An un-educated guess might be:

 x = (int) _request_->rq_disk->queue->id;

Ref. the request_queue function in linux http://lxr.free-electrons.com/source/include/linux/blkdev.h#L270 & 321

Thanks everyone for helping!

Sumikosumma

In reply to previous comments: -boto: The requests are tested down further in this function. the error was actually happening in another place were SLOT_R was being used. -Yves: Thanks for correction. It was my mistake. -Ethan: The disk major and minors would come up correct, in almost all cases that they were discovered. -Ian: This is NDAS. It is quite like nbd, providing the block storage on the lan. NDAS is actual hardware though. The other close implementation is AoE. – Sumikosumma