xen-netback: remove page tracking facility
authorWei Liu <wei.liu2@citrix.com>
Mon, 26 Aug 2013 11:59:37 +0000 (12:59 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 29 Aug 2013 05:18:04 +0000 (01:18 -0400)
The data flow from DomU to DomU on the same host in current copying
scheme with tracking facility:

       copy
DomU --------> Dom0          DomU
 |                            ^
 |____________________________|
             copy

The page in Dom0 is a page with valid MFN. So we can always copy from
page Dom0, thus removing the need for a tracking facility.

       copy           copy
DomU --------> Dom0 -------> DomU

Simple iperf test shows no performance regression (obviously we copy
twice either way):

  W/  tracking: ~5.3Gb/s
  W/o tracking: ~5.4Gb/s

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Acked-by: Matt Wilson <msw@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/xen-netback/netback.c

index 64828de25d9a252119a16bb805bfe4fb3ed4b809..91f163d03a49da0dfa76c832701bfee6e6d59dc7 100644 (file)
@@ -95,21 +95,6 @@ struct netbk_rx_meta {
 
 #define MAX_BUFFER_OFFSET PAGE_SIZE
 
-/* extra field used in struct page */
-union page_ext {
-       struct {
-#if BITS_PER_LONG < 64
-#define IDX_WIDTH   8
-#define GROUP_WIDTH (BITS_PER_LONG - IDX_WIDTH)
-               unsigned int group:GROUP_WIDTH;
-               unsigned int idx:IDX_WIDTH;
-#else
-               unsigned int group, idx;
-#endif
-       } e;
-       void *mapping;
-};
-
 struct xen_netbk {
        wait_queue_head_t wq;
        struct task_struct *task;
@@ -214,45 +199,6 @@ static inline unsigned long idx_to_kaddr(struct xen_netbk *netbk,
        return (unsigned long)pfn_to_kaddr(idx_to_pfn(netbk, idx));
 }
 
-/* extra field used in struct page */
-static inline void set_page_ext(struct page *pg, struct xen_netbk *netbk,
-                               unsigned int idx)
-{
-       unsigned int group = netbk - xen_netbk;
-       union page_ext ext = { .e = { .group = group + 1, .idx = idx } };
-
-       BUILD_BUG_ON(sizeof(ext) > sizeof(ext.mapping));
-       pg->mapping = ext.mapping;
-}
-
-static int get_page_ext(struct page *pg,
-                       unsigned int *pgroup, unsigned int *pidx)
-{
-       union page_ext ext = { .mapping = pg->mapping };
-       struct xen_netbk *netbk;
-       unsigned int group, idx;
-
-       group = ext.e.group - 1;
-
-       if (group < 0 || group >= xen_netbk_group_nr)
-               return 0;
-
-       netbk = &xen_netbk[group];
-
-       idx = ext.e.idx;
-
-       if ((idx < 0) || (idx >= MAX_PENDING_REQS))
-               return 0;
-
-       if (netbk->mmap_pages[idx] != pg)
-               return 0;
-
-       *pgroup = group;
-       *pidx = idx;
-
-       return 1;
-}
-
 /*
  * This is the amount of packet we copy rather than map, so that the
  * guest can't fiddle with the contents of the headers while we do
@@ -453,12 +399,6 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 {
        struct gnttab_copy *copy_gop;
        struct netbk_rx_meta *meta;
-       /*
-        * These variables are used iff get_page_ext returns true,
-        * in which case they are guaranteed to be initialized.
-        */
-       unsigned int uninitialized_var(group), uninitialized_var(idx);
-       int foreign = get_page_ext(page, &group, &idx);
        unsigned long bytes;
 
        /* Data must not cross a page boundary. */
@@ -494,20 +434,9 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
 
                copy_gop = npo->copy + npo->copy_prod++;
                copy_gop->flags = GNTCOPY_dest_gref;
-               if (foreign) {
-                       struct xen_netbk *netbk = &xen_netbk[group];
-                       struct pending_tx_info *src_pend;
+               copy_gop->source.domid = DOMID_SELF;
+               copy_gop->source.u.gmfn = virt_to_mfn(page_address(page));
 
-                       src_pend = &netbk->pending_tx_info[idx];
-
-                       copy_gop->source.domid = src_pend->vif->domid;
-                       copy_gop->source.u.ref = src_pend->req.gref;
-                       copy_gop->flags |= GNTCOPY_source_gref;
-               } else {
-                       void *vaddr = page_address(page);
-                       copy_gop->source.domid = DOMID_SELF;
-                       copy_gop->source.u.gmfn = virt_to_mfn(vaddr);
-               }
                copy_gop->source.offset = offset;
                copy_gop->dest.domid = vif->domid;
 
@@ -1047,7 +976,6 @@ static struct page *xen_netbk_alloc_page(struct xen_netbk *netbk,
        page = alloc_page(GFP_KERNEL|__GFP_COLD);
        if (!page)
                return NULL;
-       set_page_ext(page, netbk, pending_idx);
        netbk->mmap_pages[pending_idx] = page;
        return page;
 }
@@ -1155,7 +1083,6 @@ static struct gnttab_copy *xen_netbk_get_requests(struct xen_netbk *netbk,
                first->req.offset = 0;
                first->req.size = dst_offset;
                first->head = start_idx;
-               set_page_ext(page, netbk, head_idx);
                netbk->mmap_pages[head_idx] = page;
                frag_set_pending_idx(&frags[shinfo->nr_frags], head_idx);
        }