tools/libxc: send page-in requests in batches in linux_privcmd_map_foreign_bulk

One of the bottlenecks with foreign page-in request is the poor retry
handling in linux_privcmd_map_foreign_bulk(). It sends one request per
paged gfn at a time and it waits until the gfn is accessible. This
causes long delays in mmap requests from qemu-dm and xc_save.

Instead of sending one request at a time, walk the entire gfn list and
send batches of mmap requests. They will eventually end up in the pager's
request ring (if it has room again), and will fill up this ring so that
in turn the pager can also process page-in in batches.

Signed-off-by: Olaf Hering <olaf@aepfle.de>
Committed-by: Ian Jackson <ian.jackson.citrix.com>

diff --git a/tools/libxc/xc_linux_osdep.c b/tools/libxc/xc_linux_osdep.c
index 896f5c1356391fb3b64c2b7bafd2446cffa7fc15..744dfea80d70bb52529e5136399a86dbeb3ee3e3 100644 (file)
--- a/tools/libxc/xc_linux_osdep.c
+++ b/tools/libxc/xc_linux_osdep.c
@@ -191,6 +191,59 @@ static void *linux_privcmd_map_foreign_batch(xc_interface *xch, xc_osdep_handle
     return addr;
 }
 
+/*
+ * Retry mmap of all paged gfns in batches
+ * retuns < 0 on fatal error
+ * returns 0 if all gfns left paging state
+ * returns > 0 if some gfns are still in paging state
+ *
+ * Walk all gfns and try to assemble blocks of gfns in paging state.
+ * This will keep the request ring full and avoids delays.
+ */
+static int retry_paged(int fd, uint32_t dom, void *addr,
+                       const xen_pfn_t *arr, int *err, unsigned int num)
+{
+    privcmd_mmapbatch_v2_t ioctlx;
+    int rc, paged = 0, i = 0;
+    
+    do
+    {
+        /* Skip gfns not in paging state */
+        if ( err[i] != -ENOENT )
+        {
+            i++;
+            continue;
+        }
+
+        paged++;
+
+        /* At least one gfn is still in paging state */
+        ioctlx.num = 1;
+        ioctlx.dom = dom;
+        ioctlx.addr = (unsigned long)addr + ((unsigned long)i<<XC_PAGE_SHIFT);
+        ioctlx.arr = arr + i;
+        ioctlx.err = err + i;
+        
+        /* Assemble a batch of requests */
+        while ( ++i < num )
+        {
+            if ( err[i] != -ENOENT )
+                break;
+            ioctlx.num++;
+        }
+        
+        /* Send request and abort on fatal error */
+        rc = ioctl(fd, IOCTL_PRIVCMD_MMAPBATCH_V2, &ioctlx);
+        if ( rc < 0 && errno != ENOENT )
+            goto out;
+
+    } while ( i < num );
+    
+    rc = paged;
+out:
+    return rc;
+}
+
 static void *linux_privcmd_map_foreign_bulk(xc_interface *xch, xc_osdep_handle h,
                                             uint32_t dom, int prot,
                                             const xen_pfn_t *arr, int *err, unsigned int num)
@@ -220,21 +273,10 @@ static void *linux_privcmd_map_foreign_bulk(xc_interface *xch, xc_osdep_handle h
     /* Command was recognized, some gfn in arr are in paging state */
     if ( rc < 0 && errno == ENOENT )
     {
-        for ( i = rc = 0; rc == 0 && i < num; i++ )
-        {
-            if ( err[i] != -ENOENT )
-                continue;
-
-            ioctlx.num = 1;
-            ioctlx.dom = dom;
-            ioctlx.addr = (unsigned long)addr + ((unsigned long)i<<XC_PAGE_SHIFT);
-            ioctlx.arr = arr + i;
-            ioctlx.err = err + i;
-            do {
-                usleep(100);
-                rc = ioctl(fd, IOCTL_PRIVCMD_MMAPBATCH_V2, &ioctlx);
-            } while ( rc < 0 && errno == ENOENT && err[i] == -ENOENT );
-        }
+        do {
+            usleep(100);
+            rc = retry_paged(fd, dom, addr, arr, err, num);
+        } while ( rc > 0 );
     }
     /* Command was not recognized, use fall back */
     else if ( rc < 0 && errno == EINVAL && (int)num > 0 )