hvmloader: Correct bug in low mmio region accounting

When deciding whether to map a device in low MMIO space (<4GiB),
hvmloader compares it with "mmio_left", which is set to the size of
the low MMIO range (pci_mem_end - pci_mem_start).  However, even if it
does map a device in high MMIO space, it still removes the size of its
BAR from mmio_left.

In reality we don't need to do a separate accounting of the low memory
available -- this can be calculated from mem_resource.  Just get rid
of the variable and the duplicate accounting entirely.  This will make
the code more robust.

Note also that the calculation of whether to move a device to 64-bit
is fragile at the moment, depending on some unstated assumptions.
State those assumptions in a comment for future reference.

v5:
 - Add comment documenting fragility of the move-to-highmem check
v3:
 - Use mem_resource values directly instead of doing duplicate
   accounting

Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Acked-by: Ian Jackson <ian.jackson@eu.citrix.com>
CC: Ian Campbell <ian.campbell@citrix.com>
CC: Stefano Stabellini <stefano.stabellini@citrix.com>
CC: Hanweidong <hanweidong@huawei.com>
CC: Keir Fraser <keir@xen.org>

1 files changed, 15 insertions, 5 deletions

diff --git a/tools/firmware/hvmloader/pci.c b/tools/firmware/hvmloader/pci.c
index 6792ed47a4..1fe250ddc7 100644
--- a/tools/firmware/hvmloader/pci.c
+++ b/tools/firmware/hvmloader/pci.c
@@ -42,7 +42,6 @@ void pci_setup(void)
     uint32_t vga_devfn = 256;
     uint16_t class, vendor_id, device_id;
     unsigned int bar, pin, link, isa_irq;
-    int64_t mmio_left;
 
     /* Resources assignable to PCI devices via BARs. */
     struct resource {
@@ -264,8 +263,6 @@ void pci_setup(void)
     io_resource.base = 0xc000;
     io_resource.max = 0x10000;
 
-    mmio_left = pci_mem_end - pci_mem_start;
-
     /* Assign iomem and ioport resources in descending order of size. */
     for ( i = 0; i < nr_bars; i++ )
     {
@@ -273,7 +270,21 @@ void pci_setup(void)
         bar_reg = bars[i].bar_reg;
         bar_sz  = bars[i].bar_sz;
 
-        using_64bar = bars[i].is_64bar && bar64_relocate && (mmio_left < bar_sz);
+        /*
+         * NB: The code here is rather fragile, as the check here to see
+         * whether bar_sz will fit in the low MMIO region doesn't match the
+         * real check made below, which involves aligning the base offset of the
+         * bar with the size of the bar itself.  As it happens, this will always
+         * be satisfied because:
+         * - The first one will succeed because the MMIO hole can only start at
+         *   0x{f,e,c,8}00000000.  If it fits, it will be aligned properly.
+         * - All subsequent ones will be aligned because the list is ordered
+         *   large to small, and bar_sz is always a power of 2. (At least
+         *   the code here assumes it to be.)
+         * Should either of those two conditions change, this code will break.
+         */
+        using_64bar = bars[i].is_64bar && bar64_relocate
+            && (bar_sz > (mem_resource.max - mem_resource.base));
         bar_data = pci_readl(devfn, bar_reg);
 
         if ( (bar_data & PCI_BASE_ADDRESS_SPACE) ==
@@ -295,7 +306,6 @@ void pci_setup(void)
                 resource = &mem_resource;
                 bar_data &= ~PCI_BASE_ADDRESS_MEM_MASK;
             }
-            mmio_left -= bar_sz;
         }
         else
         {