xen/blkfront: negotiate number of queues/rings to be used with backend
authorBob Liu <bob.liu@oracle.com>
Sat, 14 Nov 2015 03:12:14 +0000 (11:12 +0800)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Mon, 4 Jan 2016 17:21:02 +0000 (12:21 -0500)
The max number of hardware queues for xen/blkfront is set by parameter
'max_queues'(default 4), while it is also capped by the max value that the
xen/blkback exposes through XenStore key 'multi-queue-max-queues'.

The negotiated number is the smaller one and would be written back to xenstore
as "multi-queue-num-queues", blkback needs to read this negotiated number.

Signed-off-by: Bob Liu <bob.liu@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
drivers/block/xen-blkfront.c

index a9058bb..81c87e6 100644 (file)
@@ -99,6 +99,10 @@ static unsigned int xen_blkif_max_segments = 32;
 module_param_named(max, xen_blkif_max_segments, int, S_IRUGO);
 MODULE_PARM_DESC(max, "Maximum amount of segments in indirect requests (default is 32)");
 
+static unsigned int xen_blkif_max_queues = 4;
+module_param_named(max_queues, xen_blkif_max_queues, uint, S_IRUGO);
+MODULE_PARM_DESC(max_queues, "Maximum number of hardware queues/rings used per virtual disk");
+
 /*
  * Maximum order of pages to be used for the shared ring between front and
  * backend, 4KB page granularity is used.
@@ -118,6 +122,10 @@ MODULE_PARM_DESC(max_ring_page_order, "Maximum order of pages to be used for the
  * characters are enough. Define to 20 to keep consist with backend.
  */
 #define RINGREF_NAME_LEN (20)
+/*
+ * queue-%u would take 7 + 10(UINT_MAX) = 17 characters.
+ */
+#define QUEUE_NAME_LEN (17)
 
 /*
  *  Per-ring info.
@@ -823,7 +831,7 @@ static int xlvbd_init_blk_queue(struct gendisk *gd, u16 sector_size,
 
        memset(&info->tag_set, 0, sizeof(info->tag_set));
        info->tag_set.ops = &blkfront_mq_ops;
-       info->tag_set.nr_hw_queues = 1;
+       info->tag_set.nr_hw_queues = info->nr_rings;
        info->tag_set.queue_depth =  BLK_RING_SIZE(info);
        info->tag_set.numa_node = NUMA_NO_NODE;
        info->tag_set.flags = BLK_MQ_F_SHOULD_MERGE | BLK_MQ_F_SG_MERGE;
@@ -1522,6 +1530,53 @@ fail:
        return err;
 }
 
+/*
+ * Write out per-ring/queue nodes including ring-ref and event-channel, and each
+ * ring buffer may have multi pages depending on ->nr_ring_pages.
+ */
+static int write_per_ring_nodes(struct xenbus_transaction xbt,
+                               struct blkfront_ring_info *rinfo, const char *dir)
+{
+       int err;
+       unsigned int i;
+       const char *message = NULL;
+       struct blkfront_info *info = rinfo->dev_info;
+
+       if (info->nr_ring_pages == 1) {
+               err = xenbus_printf(xbt, dir, "ring-ref", "%u", rinfo->ring_ref[0]);
+               if (err) {
+                       message = "writing ring-ref";
+                       goto abort_transaction;
+               }
+       } else {
+               for (i = 0; i < info->nr_ring_pages; i++) {
+                       char ring_ref_name[RINGREF_NAME_LEN];
+
+                       snprintf(ring_ref_name, RINGREF_NAME_LEN, "ring-ref%u", i);
+                       err = xenbus_printf(xbt, dir, ring_ref_name,
+                                           "%u", rinfo->ring_ref[i]);
+                       if (err) {
+                               message = "writing ring-ref";
+                               goto abort_transaction;
+                       }
+               }
+       }
+
+       err = xenbus_printf(xbt, dir, "event-channel", "%u", rinfo->evtchn);
+       if (err) {
+               message = "writing event-channel";
+               goto abort_transaction;
+       }
+
+       return 0;
+
+abort_transaction:
+       xenbus_transaction_end(xbt, 1);
+       if (message)
+               xenbus_dev_fatal(info->xbdev, err, "%s", message);
+
+       return err;
+}
 
 /* Common code used when first setting up, and when resuming. */
 static int talk_to_blkback(struct xenbus_device *dev,
@@ -1529,10 +1584,9 @@ static int talk_to_blkback(struct xenbus_device *dev,
 {
        const char *message = NULL;
        struct xenbus_transaction xbt;
-       int err, i;
-       unsigned int max_page_order = 0;
+       int err;
+       unsigned int i, max_page_order = 0;
        unsigned int ring_page_order = 0;
-       struct blkfront_ring_info *rinfo;
 
        err = xenbus_scanf(XBT_NIL, info->xbdev->otherend,
                           "max-ring-page-order", "%u", &max_page_order);
@@ -1544,7 +1598,8 @@ static int talk_to_blkback(struct xenbus_device *dev,
        }
 
        for (i = 0; i < info->nr_rings; i++) {
-               rinfo = &info->rinfo[i];
+               struct blkfront_ring_info *rinfo = &info->rinfo[i];
+
                /* Create shared ring, alloc event channel. */
                err = setup_blkring(dev, rinfo);
                if (err)
@@ -1558,44 +1613,49 @@ again:
                goto destroy_blkring;
        }
 
-       if (info->nr_rings == 1) {
-               rinfo = &info->rinfo[0];
-               if (info->nr_ring_pages == 1) {
-                       err = xenbus_printf(xbt, dev->nodename,
-                                           "ring-ref", "%u", rinfo->ring_ref[0]);
-                       if (err) {
-                               message = "writing ring-ref";
-                               goto abort_transaction;
-                       }
-               } else {
-                       err = xenbus_printf(xbt, dev->nodename,
-                                           "ring-page-order", "%u", ring_page_order);
-                       if (err) {
-                               message = "writing ring-page-order";
-                               goto abort_transaction;
-                       }
+       if (info->nr_ring_pages > 1) {
+               err = xenbus_printf(xbt, dev->nodename, "ring-page-order", "%u",
+                                   ring_page_order);
+               if (err) {
+                       message = "writing ring-page-order";
+                       goto abort_transaction;
+               }
+       }
 
-                       for (i = 0; i < info->nr_ring_pages; i++) {
-                               char ring_ref_name[RINGREF_NAME_LEN];
+       /* We already got the number of queues/rings in _probe */
+       if (info->nr_rings == 1) {
+               err = write_per_ring_nodes(xbt, &info->rinfo[0], dev->nodename);
+               if (err)
+                       goto destroy_blkring;
+       } else {
+               char *path;
+               size_t pathsize;
 
-                               snprintf(ring_ref_name, RINGREF_NAME_LEN, "ring-ref%u", i);
-                               err = xenbus_printf(xbt, dev->nodename, ring_ref_name,
-                                                   "%u", rinfo->ring_ref[i]);
-                               if (err) {
-                                       message = "writing ring-ref";
-                                       goto abort_transaction;
-                               }
-                       }
-               }
-               err = xenbus_printf(xbt, dev->nodename,
-                                   "event-channel", "%u", rinfo->evtchn);
+               err = xenbus_printf(xbt, dev->nodename, "multi-queue-num-queues", "%u",
+                                   info->nr_rings);
                if (err) {
-                       message = "writing event-channel";
+                       message = "writing multi-queue-num-queues";
                        goto abort_transaction;
                }
-       } else {
-               /* Not supported at this stage. */
-               goto abort_transaction;
+
+               pathsize = strlen(dev->nodename) + QUEUE_NAME_LEN;
+               path = kmalloc(pathsize, GFP_KERNEL);
+               if (!path) {
+                       err = -ENOMEM;
+                       message = "ENOMEM while writing ring references";
+                       goto abort_transaction;
+               }
+
+               for (i = 0; i < info->nr_rings; i++) {
+                       memset(path, 0, pathsize);
+                       snprintf(path, pathsize, "%s/queue-%u", dev->nodename, i);
+                       err = write_per_ring_nodes(xbt, &info->rinfo[i], path);
+                       if (err) {
+                               kfree(path);
+                               goto destroy_blkring;
+                       }
+               }
+               kfree(path);
        }
        err = xenbus_printf(xbt, dev->nodename, "protocol", "%s",
                            XEN_IO_PROTO_ABI_NATIVE);
@@ -1619,8 +1679,7 @@ again:
 
        for (i = 0; i < info->nr_rings; i++) {
                unsigned int j;
-
-               rinfo = &info->rinfo[i];
+               struct blkfront_ring_info *rinfo = &info->rinfo[i];
 
                for (j = 0; j < BLK_RING_SIZE(info); j++)
                        rinfo->shadow[j].req.u.rw.id = j + 1;
@@ -1652,6 +1711,7 @@ static int blkfront_probe(struct xenbus_device *dev,
        int err, vdevice;
        unsigned int r_index;
        struct blkfront_info *info;
+       unsigned int backend_max_queues = 0;
 
        /* FIXME: Use dynamic device id if this is not set. */
        err = xenbus_scanf(XBT_NIL, dev->nodename,
@@ -1701,7 +1761,18 @@ static int blkfront_probe(struct xenbus_device *dev,
                return -ENOMEM;
        }
 
-       info->nr_rings = 1;
+       info->xbdev = dev;
+       /* Check if backend supports multiple queues. */
+       err = xenbus_scanf(XBT_NIL, info->xbdev->otherend,
+                          "multi-queue-max-queues", "%u", &backend_max_queues);
+       if (err < 0)
+               backend_max_queues = 1;
+
+       info->nr_rings = min(backend_max_queues, xen_blkif_max_queues);
+       /* We need at least one ring. */
+       if (!info->nr_rings)
+               info->nr_rings = 1;
+
        info->rinfo = kzalloc(sizeof(struct blkfront_ring_info) * info->nr_rings, GFP_KERNEL);
        if (!info->rinfo) {
                xenbus_dev_fatal(dev, -ENOMEM, "allocating ring_info structure");
@@ -2390,6 +2461,7 @@ static struct xenbus_driver blkfront_driver = {
 static int __init xlblk_init(void)
 {
        int ret;
+       int nr_cpus = num_online_cpus();
 
        if (!xen_domain())
                return -ENODEV;
@@ -2400,6 +2472,12 @@ static int __init xlblk_init(void)
                xen_blkif_max_ring_order = 0;
        }
 
+       if (xen_blkif_max_queues > nr_cpus) {
+               pr_info("Invalid max_queues (%d), will use default max: %d.\n",
+                       xen_blkif_max_queues, nr_cpus);
+               xen_blkif_max_queues = nr_cpus;
+       }
+
        if (!xen_has_pv_disk_devices())
                return -ENODEV;