As of c/s 21812:e382656e4dcc, IOMMU related allocations for Dom0
happen only after it got all of its memory allocated, and hence the
reserve (mainly for setting up its swiotlb) may get exhausted without
accounting for the necessary allocations up front.

While not precise, the estimate has been found to be within a couple
of pages for the systems it got tested on.

For the calculation to be reasonably correct, this depends on the
patch titled "x86/iommu: don't map RAM holes above 4G" sent out
yesterday.

Signed-off-by: Jan Beulich <jbeulich@xxxxxxxxxx>

--- a/xen/arch/x86/domain_build.c
+++ b/xen/arch/x86/domain_build.c
@@ -191,6 +191,15 @@ static unsigned long __init compute_dom0
     if ( is_pv_32on64_domain(d) )
         avail -= opt_dom0_max_vcpus - 1;
 
+    /* Reserve memory for iommu_dom0_init() (rough estimate). */
+    if ( iommu_enabled )
+    {
+        unsigned int s;
+
+        for ( s = 9; s < BITS_PER_LONG; s += 9 )
+            avail -= max_pdx >> s;
+    }
+
     /*
      * If domain 0 allocation isn't specified, reserve 1/16th of available
      * memory for things like DMA buffers. This reservation is clamped to