]> xenbits.xensource.com Git - xen.git/commitdiff
x86/mem-sharing: mem-sharing a range of memory
authorTamas K Lengyel <tamas.lengyel@zentific.com>
Mon, 1 Aug 2016 17:14:27 +0000 (11:14 -0600)
committerWei Liu <wei.liu2@citrix.com>
Tue, 2 Aug 2016 16:03:10 +0000 (17:03 +0100)
Currently mem-sharing can be performed on a page-by-page basis from the control
domain. However, this process is quite wasteful when a range of pages have to
be deduplicated.

This patch introduces a new mem_sharing memop for range sharing where
the user doesn't have to separately nominate each page in both the source and
destination domain, and the looping over all pages happen in the hypervisor.
This significantly reduces the overhead of sharing a range of memory.

Signed-off-by: Tamas K Lengyel <tamas.lengyel@zentific.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Reviewed-by: Andrew Cooper <andrew.cooper3@citrix.com>
tools/libxc/include/xenctrl.h
tools/libxc/xc_memshr.c
tools/tests/mem-sharing/memshrtool.c
xen/arch/x86/mm/mem_sharing.c
xen/include/public/memory.h

index fdc148ac2d956f62aef998bcc9d5f57ce5cf21e2..560ce7b26e3aaaf9e9216d0455ca305ee7a47f42 100644 (file)
@@ -2334,6 +2334,21 @@ int xc_memshr_add_to_physmap(xc_interface *xch,
                     domid_t client_domain,
                     unsigned long client_gfn);
 
+/* Allows to deduplicate a range of memory of a client domain. Using
+ * this function is equivalent of calling xc_memshr_nominate_gfn for each gfn
+ * in the two domains followed by xc_memshr_share_gfns.
+ *
+ * May fail with -EINVAL if the source and client domain have different
+ * memory size or if memory sharing is not enabled on either of the domains.
+ * May also fail with -ENOMEM if there isn't enough memory available to store
+ * the sharing metadata before deduplication can happen.
+ */
+int xc_memshr_range_share(xc_interface *xch,
+                          domid_t source_domain,
+                          domid_t client_domain,
+                          uint64_t first_gfn,
+                          uint64_t last_gfn);
+
 /* Debug calls: return the number of pages referencing the shared frame backing
  * the input argument. Should be one or greater. 
  *
index deb0aa4c8bf934124e91a9f506c794c51045077c..f53a59a537d57f1b14876a75ae79cb4672b36c56 100644 (file)
@@ -181,6 +181,25 @@ int xc_memshr_add_to_physmap(xc_interface *xch,
     return xc_memshr_memop(xch, source_domain, &mso);
 }
 
+int xc_memshr_range_share(xc_interface *xch,
+                          domid_t source_domain,
+                          domid_t client_domain,
+                          uint64_t first_gfn,
+                          uint64_t last_gfn)
+{
+    xen_mem_sharing_op_t mso;
+
+    memset(&mso, 0, sizeof(mso));
+
+    mso.op = XENMEM_sharing_op_range_share;
+
+    mso.u.range.client_domain = client_domain;
+    mso.u.range.first_gfn = first_gfn;
+    mso.u.range.last_gfn = last_gfn;
+
+    return xc_memshr_memop(xch, source_domain, &mso);
+}
+
 int xc_memshr_domain_resume(xc_interface *xch,
                             domid_t domid)
 {
index 437c7c9c472c0eff29b60e54f3de90e878de6d34..8e5e22b9e95d91f1441d8eb226b64852eca075d5 100644 (file)
@@ -24,6 +24,8 @@ static int usage(const char* prog)
     printf("  nominate <domid> <gfn>  - Nominate a page for sharing.\n");
     printf("  share <domid> <gfn> <handle> <source> <source-gfn> <source-handle>\n");
     printf("                          - Share two pages.\n");
+    printf("  range <source-domid> <destination-domid> <first-gfn> <last-gfn>\n");
+    printf("                          - Share pages between domains in a range.\n");
     printf("  unshare <domid> <gfn>   - Unshare a page by grabbing a writable map.\n");
     printf("  add-to-physmap <domid> <gfn> <source> <source-gfn> <source-handle>\n");
     printf("                          - Populate a page in a domain with a shared page.\n");
@@ -180,6 +182,26 @@ int main(int argc, const char** argv)
         }
         printf("Audit returned %d errors.\n", rc);
     }
+    else if( !strcasecmp(cmd, "range") )
+    {
+        domid_t sdomid, cdomid;
+        int rc;
+        uint64_t first_gfn, last_gfn;
+
+        if ( argc != 6 )
+            return usage(argv[0]);
 
+        sdomid = strtol(argv[2], NULL, 0);
+        cdomid = strtol(argv[3], NULL, 0);
+        first_gfn = strtoul(argv[4], NULL, 0);
+        last_gfn = strtoul(argv[5], NULL, 0);
+
+        rc = xc_memshr_range_share(xch, sdomid, cdomid, first_gfn, last_gfn);
+        if ( rc < 0 )
+        {
+            printf("error executing xc_memshr_range_share: %s\n", strerror(errno));
+            return rc;
+        }
+    }
     return 0;
 }
index 14952ce76ca65dff828953c45a5ed17923dc82e0..e7c6b74eedfd1029ebb9a063298d8f158f870d87 100644 (file)
@@ -1325,6 +1325,58 @@ int relinquish_shared_pages(struct domain *d)
     return rc;
 }
 
+static int range_share(struct domain *d, struct domain *cd,
+                       struct mem_sharing_op_range *range)
+{
+    int rc = 0;
+    shr_handle_t sh, ch;
+    unsigned long start = range->opaque ?: range->first_gfn;
+
+    while ( range->last_gfn >= start )
+    {
+        /*
+         * We only break out if we run out of memory as individual pages may
+         * legitimately be unsharable and we just want to skip over those.
+         */
+        rc = mem_sharing_nominate_page(d, start, 0, &sh);
+        if ( rc == -ENOMEM )
+            break;
+
+        if ( !rc )
+        {
+            rc = mem_sharing_nominate_page(cd, start, 0, &ch);
+            if ( rc == -ENOMEM )
+                break;
+
+            if ( !rc )
+            {
+                /* If we get here this should be guaranteed to succeed. */
+                rc = mem_sharing_share_pages(d, start, sh,
+                                             cd, start, ch);
+                ASSERT(!rc);
+            }
+        }
+
+        /* Check for continuation if it's not the last iteration. */
+        if ( range->last_gfn >= ++start && hypercall_preempt_check() )
+        {
+            rc = 1;
+            break;
+        }
+    }
+
+    range->opaque = start;
+
+    /*
+     * The last page may fail with -EINVAL, and for range sharing we don't
+     * care about that.
+     */
+    if ( range->last_gfn < start && rc == -EINVAL )
+        rc = 0;
+
+    return rc;
+}
+
 int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)
 {
     int rc;
@@ -1499,6 +1551,96 @@ int mem_sharing_memop(XEN_GUEST_HANDLE_PARAM(xen_mem_sharing_op_t) arg)
         }
         break;
 
+        case XENMEM_sharing_op_range_share:
+        {
+            unsigned long max_sgfn, max_cgfn;
+            struct domain *cd;
+
+            rc = -EINVAL;
+            if ( mso.u.range._pad[0] || mso.u.range._pad[1] ||
+                 mso.u.range._pad[2] )
+                 goto out;
+
+            /*
+             * We use opaque for the hypercall continuation value.
+             * Ideally the user sets this to 0 in the beginning but
+             * there is no good way of enforcing that here, so we just check
+             * that it's at least in range.
+             */
+            if ( mso.u.range.opaque &&
+                 (mso.u.range.opaque < mso.u.range.first_gfn ||
+                  mso.u.range.opaque > mso.u.range.last_gfn) )
+                goto out;
+
+            if ( !mem_sharing_enabled(d) )
+                goto out;
+
+            rc = rcu_lock_live_remote_domain_by_id(mso.u.range.client_domain,
+                                                   &cd);
+            if ( rc )
+                goto out;
+
+            /*
+             * We reuse XENMEM_sharing_op_share XSM check here as this is
+             * essentially the same concept repeated over multiple pages.
+             */
+            rc = xsm_mem_sharing_op(XSM_DM_PRIV, d, cd,
+                                    XENMEM_sharing_op_share);
+            if ( rc )
+            {
+                rcu_unlock_domain(cd);
+                goto out;
+            }
+
+            if ( !mem_sharing_enabled(cd) )
+            {
+                rcu_unlock_domain(cd);
+                rc = -EINVAL;
+                goto out;
+            }
+
+            /*
+             * Sanity check only, the client should keep the domains paused for
+             * the duration of this op.
+             */
+            if ( !atomic_read(&d->pause_count) ||
+                 !atomic_read(&cd->pause_count) )
+            {
+                rcu_unlock_domain(cd);
+                rc = -EINVAL;
+                goto out;
+            }
+
+            max_sgfn = domain_get_maximum_gpfn(d);
+            max_cgfn = domain_get_maximum_gpfn(cd);
+
+            if ( max_sgfn < mso.u.range.first_gfn ||
+                 max_sgfn < mso.u.range.last_gfn ||
+                 max_cgfn < mso.u.range.first_gfn ||
+                 max_cgfn < mso.u.range.last_gfn )
+            {
+                rcu_unlock_domain(cd);
+                rc = -EINVAL;
+                goto out;
+            }
+
+            rc = range_share(d, cd, &mso.u.range);
+            rcu_unlock_domain(cd);
+
+            if ( rc > 0 )
+            {
+                if ( __copy_to_guest(arg, &mso, 1) )
+                    rc = -EFAULT;
+                else
+                    rc = hypercall_create_continuation(__HYPERVISOR_memory_op,
+                                                       "lh", XENMEM_sharing_op,
+                                                       arg);
+            }
+            else
+                mso.u.range.opaque = 0;
+        }
+        break;
+
         case XENMEM_sharing_op_debug_gfn:
         {
             unsigned long gfn = mso.u.debug.u.gfn;
index 29ec571d892e722affeeac4f3ead5699ca6c1a60..3badfb977e261d84a433a856542b37fd0dcadc4d 100644 (file)
@@ -465,6 +465,7 @@ DEFINE_XEN_GUEST_HANDLE(xen_mem_access_op_t);
 #define XENMEM_sharing_op_debug_gref        5
 #define XENMEM_sharing_op_add_physmap       6
 #define XENMEM_sharing_op_audit             7
+#define XENMEM_sharing_op_range_share       8
 
 #define XENMEM_SHARING_OP_S_HANDLE_INVALID  (-10)
 #define XENMEM_SHARING_OP_C_HANDLE_INVALID  (-9)
@@ -500,7 +501,14 @@ struct xen_mem_sharing_op {
             uint64_aligned_t client_gfn;    /* IN: the client gfn */
             uint64_aligned_t client_handle; /* IN: handle to the client page */
             domid_t  client_domain; /* IN: the client domain id */
-        } share; 
+        } share;
+        struct mem_sharing_op_range {         /* OP_RANGE_SHARE */
+            uint64_aligned_t first_gfn;      /* IN: the first gfn */
+            uint64_aligned_t last_gfn;       /* IN: the last gfn */
+            uint64_aligned_t opaque;         /* Must be set to 0 */
+            domid_t client_domain;           /* IN: the client domain id */
+            uint16_t _pad[3];                /* Must be set to 0 */
+        } range;
         struct mem_sharing_op_debug {     /* OP_DEBUG_xxx */
             union {
                 uint64_aligned_t gfn;      /* IN: gfn to debug          */