]> git.karo-electronics.de Git - linux-beck.git/commitdiff
xen-blkback: make the queue of free requests per backend
authorRoger Pau Monne <roger.pau@citrix.com>
Wed, 17 Apr 2013 18:18:59 +0000 (20:18 +0200)
committerKonrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Thu, 18 Apr 2013 13:29:25 +0000 (09:29 -0400)
Remove the last dependency from blkbk by moving the list of free
requests to blkif. This change reduces the contention on the list of
available requests.

Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: xen-devel@lists.xen.org
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
drivers/block/xen-blkback/blkback.c
drivers/block/xen-blkback/common.h
drivers/block/xen-blkback/xenbus.c

index ae7dc92ad3cfb359be5c2544ca4d310dc0799e49..90a57552f4b77f49b46933d2362b534221674137 100644 (file)
 #include <xen/balloon.h>
 #include "common.h"
 
-/*
- * These are rather arbitrary. They are fairly large because adjacent requests
- * pulled from a communication ring are quite likely to end up being part of
- * the same scatter/gather request at the disc.
- *
- * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
- *
- * This will increase the chances of being able to write whole tracks.
- * 64 should be enough to keep us competitive with Linux.
- */
-static int xen_blkif_reqs = 64;
-module_param_named(reqs, xen_blkif_reqs, int, 0);
-MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
-
 /*
  * Maximum number of unused free pages to keep in the internal buffer.
  * Setting this to a value too low will reduce memory used in each backend,
@@ -112,53 +98,11 @@ MODULE_PARM_DESC(max_persistent_grants,
 static unsigned int log_stats;
 module_param(log_stats, int, 0644);
 
-/*
- * Each outstanding request that we've passed to the lower device layers has a
- * 'pending_req' allocated to it. Each buffer_head that completes decrements
- * the pendcnt towards zero. When it hits zero, the specified domain has a
- * response queued for it, with the saved 'id' passed back.
- */
-struct pending_req {
-       struct xen_blkif        *blkif;
-       u64                     id;
-       int                     nr_pages;
-       atomic_t                pendcnt;
-       unsigned short          operation;
-       int                     status;
-       struct list_head        free_list;
-       struct page             *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
-       struct persistent_gnt   *persistent_gnts[BLKIF_MAX_SEGMENTS_PER_REQUEST];
-       grant_handle_t          grant_handles[BLKIF_MAX_SEGMENTS_PER_REQUEST];
-};
-
 #define BLKBACK_INVALID_HANDLE (~0)
 
 /* Number of free pages to remove on each call to free_xenballooned_pages */
 #define NUM_BATCH_FREE_PAGES 10
 
-struct xen_blkbk {
-       struct pending_req      *pending_reqs;
-       /* List of all 'pending_req' available */
-       struct list_head        pending_free;
-       /* And its spinlock. */
-       spinlock_t              pending_free_lock;
-       wait_queue_head_t       pending_free_wq;
-};
-
-static struct xen_blkbk *blkbk;
-
-/*
- * Little helpful macro to figure out the index and virtual address of the
- * pending_pages[..]. For each 'pending_req' we have have up to
- * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
- * 10 and would index in the pending_pages[..].
- */
-static inline int vaddr_pagenr(struct pending_req *req, int seg)
-{
-       return (req - blkbk->pending_reqs) *
-               BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
-}
-
 static inline int get_free_page(struct xen_blkif *blkif, struct page **page)
 {
        unsigned long flags;
@@ -485,18 +429,18 @@ finished:
 /*
  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
  */
-static struct pending_req *alloc_req(void)
+static struct pending_req *alloc_req(struct xen_blkif *blkif)
 {
        struct pending_req *req = NULL;
        unsigned long flags;
 
-       spin_lock_irqsave(&blkbk->pending_free_lock, flags);
-       if (!list_empty(&blkbk->pending_free)) {
-               req = list_entry(blkbk->pending_free.next, struct pending_req,
+       spin_lock_irqsave(&blkif->pending_free_lock, flags);
+       if (!list_empty(&blkif->pending_free)) {
+               req = list_entry(blkif->pending_free.next, struct pending_req,
                                 free_list);
                list_del(&req->free_list);
        }
-       spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
+       spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
        return req;
 }
 
@@ -504,17 +448,17 @@ static struct pending_req *alloc_req(void)
  * Return the 'pending_req' structure back to the freepool. We also
  * wake up the thread if it was waiting for a free page.
  */
-static void free_req(struct pending_req *req)
+static void free_req(struct xen_blkif *blkif, struct pending_req *req)
 {
        unsigned long flags;
        int was_empty;
 
-       spin_lock_irqsave(&blkbk->pending_free_lock, flags);
-       was_empty = list_empty(&blkbk->pending_free);
-       list_add(&req->free_list, &blkbk->pending_free);
-       spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
+       spin_lock_irqsave(&blkif->pending_free_lock, flags);
+       was_empty = list_empty(&blkif->pending_free);
+       list_add(&req->free_list, &blkif->pending_free);
+       spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
        if (was_empty)
-               wake_up(&blkbk->pending_free_wq);
+               wake_up(&blkif->pending_free_wq);
 }
 
 /*
@@ -649,8 +593,8 @@ int xen_blkif_schedule(void *arg)
                if (timeout == 0)
                        goto purge_gnt_list;
                timeout = wait_event_interruptible_timeout(
-                       blkbk->pending_free_wq,
-                       !list_empty(&blkbk->pending_free) ||
+                       blkif->pending_free_wq,
+                       !list_empty(&blkif->pending_free) ||
                        kthread_should_stop(),
                        timeout);
                if (timeout == 0)
@@ -907,7 +851,7 @@ static int dispatch_other_io(struct xen_blkif *blkif,
                             struct blkif_request *req,
                             struct pending_req *pending_req)
 {
-       free_req(pending_req);
+       free_req(blkif, pending_req);
        make_response(blkif, req->u.other.id, req->operation,
                      BLKIF_RSP_EOPNOTSUPP);
        return -EIO;
@@ -967,7 +911,7 @@ static void __end_block_io_op(struct pending_req *pending_req, int error)
                        if (atomic_read(&pending_req->blkif->drain))
                                complete(&pending_req->blkif->drain_complete);
                }
-               free_req(pending_req);
+               free_req(pending_req->blkif, pending_req);
        }
 }
 
@@ -1010,7 +954,7 @@ __do_block_io_op(struct xen_blkif *blkif)
                        break;
                }
 
-               pending_req = alloc_req();
+               pending_req = alloc_req(blkif);
                if (NULL == pending_req) {
                        blkif->st_oo_req++;
                        more_to_do = 1;
@@ -1044,7 +988,7 @@ __do_block_io_op(struct xen_blkif *blkif)
                                goto done;
                        break;
                case BLKIF_OP_DISCARD:
-                       free_req(pending_req);
+                       free_req(blkif, pending_req);
                        if (dispatch_discard_io(blkif, &req))
                                goto done;
                        break;
@@ -1246,7 +1190,7 @@ static int dispatch_rw_block_io(struct xen_blkif *blkif,
  fail_response:
        /* Haven't submitted any bio's yet. */
        make_response(blkif, req->u.rw.id, req->operation, BLKIF_RSP_ERROR);
-       free_req(pending_req);
+       free_req(blkif, pending_req);
        msleep(1); /* back off a bit */
        return -EIO;
 
@@ -1303,51 +1247,20 @@ static void make_response(struct xen_blkif *blkif, u64 id,
 
 static int __init xen_blkif_init(void)
 {
-       int i;
        int rc = 0;
 
        if (!xen_domain())
                return -ENODEV;
 
-       blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
-       if (!blkbk) {
-               pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
-               return -ENOMEM;
-       }
-
-
-       blkbk->pending_reqs          = kzalloc(sizeof(blkbk->pending_reqs[0]) *
-                                       xen_blkif_reqs, GFP_KERNEL);
-
-       if (!blkbk->pending_reqs) {
-               rc = -ENOMEM;
-               goto out_of_memory;
-       }
-
        rc = xen_blkif_interface_init();
        if (rc)
                goto failed_init;
 
-       INIT_LIST_HEAD(&blkbk->pending_free);
-       spin_lock_init(&blkbk->pending_free_lock);
-       init_waitqueue_head(&blkbk->pending_free_wq);
-
-       for (i = 0; i < xen_blkif_reqs; i++)
-               list_add_tail(&blkbk->pending_reqs[i].free_list,
-                             &blkbk->pending_free);
-
        rc = xen_blkif_xenbus_init();
        if (rc)
                goto failed_init;
 
-       return 0;
-
- out_of_memory:
-       pr_alert(DRV_PFX "%s: out of memory\n", __func__);
  failed_init:
-       kfree(blkbk->pending_reqs);
-       kfree(blkbk);
-       blkbk = NULL;
        return rc;
 }
 
index af9bed48f7736563e7fb440a143ab6f448739541..e33fafa0facd4a0b4f66843bb6914569e42b64a3 100644 (file)
@@ -192,6 +192,9 @@ struct backend_info;
  */
 #define PERSISTENT_GNT_WAS_ACTIVE      1
 
+/* Number of requests that we can fit in a ring */
+#define XEN_BLKIF_REQS                 32
+
 struct persistent_gnt {
        struct page *page;
        grant_ref_t gnt;
@@ -242,6 +245,14 @@ struct xen_blkif {
        int                     free_pages_num;
        struct list_head        free_pages;
 
+       /* Allocation of pending_reqs */
+       struct pending_req      *pending_reqs;
+       /* List of all 'pending_req' available */
+       struct list_head        pending_free;
+       /* And its spinlock. */
+       spinlock_t              pending_free_lock;
+       wait_queue_head_t       pending_free_wq;
+
        /* statistics */
        unsigned long           st_print;
        unsigned long long                      st_rd_req;
@@ -255,6 +266,25 @@ struct xen_blkif {
        wait_queue_head_t       waiting_to_free;
 };
 
+/*
+ * Each outstanding request that we've passed to the lower device layers has a
+ * 'pending_req' allocated to it. Each buffer_head that completes decrements
+ * the pendcnt towards zero. When it hits zero, the specified domain has a
+ * response queued for it, with the saved 'id' passed back.
+ */
+struct pending_req {
+       struct xen_blkif        *blkif;
+       u64                     id;
+       int                     nr_pages;
+       atomic_t                pendcnt;
+       unsigned short          operation;
+       int                     status;
+       struct list_head        free_list;
+       struct page             *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
+       struct persistent_gnt   *persistent_gnts[BLKIF_MAX_SEGMENTS_PER_REQUEST];
+       grant_handle_t          grant_handles[BLKIF_MAX_SEGMENTS_PER_REQUEST];
+};
+
 
 #define vbd_sz(_v)     ((_v)->bdev->bd_part ? \
                         (_v)->bdev->bd_part->nr_sects : \
index e0fd92a2a4cd651a0f6e61db6aba6d160feaa434..1f1ade6d6e09f26c2636049d0c5876b60c95a52f 100644 (file)
@@ -105,6 +105,7 @@ static void xen_update_blkif_status(struct xen_blkif *blkif)
 static struct xen_blkif *xen_blkif_alloc(domid_t domid)
 {
        struct xen_blkif *blkif;
+       int i;
 
        blkif = kmem_cache_zalloc(xen_blkif_cachep, GFP_KERNEL);
        if (!blkif)
@@ -124,6 +125,21 @@ static struct xen_blkif *xen_blkif_alloc(domid_t domid)
        blkif->free_pages_num = 0;
        atomic_set(&blkif->persistent_gnt_in_use, 0);
 
+       blkif->pending_reqs = kcalloc(XEN_BLKIF_REQS,
+                                     sizeof(blkif->pending_reqs[0]),
+                                     GFP_KERNEL);
+       if (!blkif->pending_reqs) {
+               kmem_cache_free(xen_blkif_cachep, blkif);
+               return ERR_PTR(-ENOMEM);
+       }
+       INIT_LIST_HEAD(&blkif->pending_free);
+       spin_lock_init(&blkif->pending_free_lock);
+       init_waitqueue_head(&blkif->pending_free_wq);
+
+       for (i = 0; i < XEN_BLKIF_REQS; i++)
+               list_add_tail(&blkif->pending_reqs[i].free_list,
+                             &blkif->pending_free);
+
        return blkif;
 }
 
@@ -203,8 +219,18 @@ static void xen_blkif_disconnect(struct xen_blkif *blkif)
 
 static void xen_blkif_free(struct xen_blkif *blkif)
 {
+       struct pending_req *req;
+       int i = 0;
+
        if (!atomic_dec_and_test(&blkif->refcnt))
                BUG();
+
+       /* Check that there is no request in use */
+       list_for_each_entry(req, &blkif->pending_free, free_list)
+               i++;
+       BUG_ON(i != XEN_BLKIF_REQS);
+
+       kfree(blkif->pending_reqs);
        kmem_cache_free(xen_blkif_cachep, blkif);
 }