linux: support 32 bit PCI domains (v3)

The PCI domain may be larger than 16 bits on Microsoft Azure and other
virtual environments. PCI busses reported by ACPI are limited to 16
bits, but in Azure the domain value for pass through devices is
intentionally larger than 16 bits to avoid clashing with local devices.
This is needed to support pass through of GPU devices.

v3: (ajax)
Update FreeBSD and Solaris backends to preserve the full 32-bit domain
number, since on those OSes it stands a chance of working already.
Update NetBSD and OpenBSD backends to initialize domain_16 compatibly
with older libpciaccess; neither backend appears to support more than a
handful of domains to begin with though. Trivially update the generic
x86 backend for source compatibility, though it still only supports one
domain and will never be better.

Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=101744
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Reviewed-by: Eric Anholt <eric@anholt.net>

6 files changed, 40 insertions, 15 deletions

diff --git a/include/pciaccess.h b/include/pciaccess.h
index 1d7aa4b..8167be6 100644
--- a/include/pciaccess.h
+++ b/include/pciaccess.h
@@ -311,6 +311,10 @@ struct pci_mem_region {
  * PCI device.
  *
  * Contains all of the information about a particular PCI device.
+ *
+ * This structure - like everything else in libpciaccess - is allocated
+ * by the library itself. Do not embed this structure in other structs,
+ * or otherwise allocate them yourself.
  */
 struct pci_device {
     /**
@@ -319,9 +323,12 @@ struct pci_device {
      * Complete bus identification, including domain, of the device.  On
      * platforms that do not support PCI domains (e.g., 32-bit x86 hardware),
      * the domain will always be zero.
+     *
+     * The domain_16 field is provided for binary compatibility with older
+     * libpciaccess.
      */
     /*@{*/
-    uint16_t    domain;
+    uint16_t    domain_16;
     uint8_t     bus;
     uint8_t     dev;
     uint8_t     func;
@@ -385,6 +392,12 @@ struct pci_device {
       * Used by the VGA arbiter. Type of resource decoded by the device and
       * the file descriptor (/dev/vga_arbiter). */
     int vgaarb_rsrc;
+
+
+    /**
+     * PCI domain value (full 32 bits)
+     */
+    uint32_t    domain;
 };
 
 
diff --git a/src/linux_sysfs.c b/src/linux_sysfs.c
index dd8ef3e..a8bc2e1 100644
--- a/src/linux_sysfs.c
+++ b/src/linux_sysfs.c
@@ -118,28 +118,18 @@ pci_system_linux_sysfs_create( void )
 
 
 /**
- * Filter out the names "." and ".." from the scanned sysfs entries, and
- * domains requiring 32-bits.
+ * Filter out the names "." and ".." from the scanned sysfs entries.
  *
  * \param d  Directory entry being processed by \c scandir.
  *
  * \return
- * Zero if the entry name matches either "." or "..", or the domain requires
- * 32 bits, non-zero otherwise.
+ * Zero if the entry name matches either "." or ".."
  *
  * \sa scandir, populate_entries
  */
 static int
 scan_sys_pci_filter( const struct dirent * d )
 {
-    if (d->d_name[0] != '.') {
-        unsigned dom = 0;
-
-        sscanf(d->d_name, "%x:", &dom);
-        if (dom > USHRT_MAX)
-            return 0;
-    }
-
     return !((strcmp( d->d_name, "." ) == 0)
 	     || (strcmp( d->d_name, ".." ) == 0));
 }
@@ -218,10 +208,19 @@ populate_entries( struct pci_system * p )
 			(struct pci_device_private *) &p->devices[i];
 
 
-		sscanf(devices[i]->d_name, "%04x:%02x:%02x.%1u",
+		sscanf(devices[i]->d_name, "%x:%02x:%02x.%1u",
 		       & dom, & bus, & dev, & func);
 
 		device->base.domain = dom;
+		/*
+		 * Applications compiled with older versions  do not expect
+		 * 32-bit domain numbers. To keep them working, we keep a 16-bit
+		 * version of the domain number at the previous location.
+		 */
+		if (dom > 0xffff)
+		     device->base.domain_16 = 0xffff;
+		else
+		     device->base.domain_16 = dom;
 		device->base.bus = bus;
 		device->base.dev = dev;
 		device->base.func = func;
diff --git a/src/netbsd_pci.c b/src/netbsd_pci.c
index f972f94..1f3bcea 100644
--- a/src/netbsd_pci.c
+++ b/src/netbsd_pci.c
@@ -959,6 +959,10 @@ pci_system_netbsd_create(void)
 					continue;
 
 				device->base.domain = domain;
+				if (domain > 0xffff)
+				    device->base.domain_16 = 0xffff;
+				else
+				    device->base.domain_16 = domain & 0xffff;
 				device->base.bus = bus;
 				device->base.dev = dev;
 				device->base.func = func;
diff --git a/src/openbsd_pci.c b/src/openbsd_pci.c
index b8ce318..c061fd8 100644
--- a/src/openbsd_pci.c
+++ b/src/openbsd_pci.c
@@ -656,6 +656,10 @@ pci_system_openbsd_create(void)
 						continue;
 
 					device->base.domain = domain;
+					if (domain > 0xffff)
+					    device->base.domain_16 = 0xffff;
+					else
+					    device->base.domain_16 = domain & 0xffff;
 					device->base.bus = bus;
 					device->base.dev = dev;
 					device->base.func = func;
diff --git a/src/solx_devfs.c b/src/solx_devfs.c
index 46fc301..dc1464d 100644
--- a/src/solx_devfs.c
+++ b/src/solx_devfs.c
@@ -213,6 +213,11 @@ probe_device_node(di_node_t node, void *arg)
     pci_base->dev = PCI_REG_DEV_G(retbuf[0]);
     pci_base->func  = PCI_REG_FUNC_G(retbuf[0]);
 
+    if (nexus->domain > 0xffff)
+	pci_base->domain_16 = 0xffff;
+    else
+	pci_base->domain_16 = nexus->domain;
+
     /* Get property values */
     for (i = 0; i < NUM_PROPERTIES; i++) {
 	len = di_prop_lookup_ints(DDI_DEV_T_ANY, node,
diff --git a/src/x86_pci.c b/src/x86_pci.c
index 32daa04..6b6a026 100644
--- a/src/x86_pci.c
+++ b/src/x86_pci.c
@@ -915,7 +915,7 @@ pci_system_x86_create(void)
 		if (PCI_VENDOR(reg) == PCI_VENDOR_INVALID ||
 		    PCI_VENDOR(reg) == 0)
 		    continue;
-		device->base.domain = 0;
+		device->base.domain = device->base.domain_16 = 0;
 		device->base.bus = bus;
 		device->base.dev = dev;
 		device->base.func = func;