Skip to contentSkip to conversation

hv_balloon: Fallback to generic_online_page() for non-HV hot added mem

PATCH Applied Jacob Pan 2 participants 2 messages Jan 7, 2025 – Jan 8, 2025

Subsystems (2)
lore
Share & tools
lore

Jacob Pan

author
#0 lore raw

PATCH v2 1/1 hv_balloon: Fallback to generic_online_page() for non-HV hot added mem

The Hyper-V balloon driver installs a custom callback for handling page onlining operations performed by the memory hotplug subsystem. This custom callback is global, and overrides the default callback (generic_online_page) that Linux otherwise uses. The custom callback properly handles memory that is hot-added by the balloon driver as part of a Hyper-V hot-add region.

But memory can also be hot-added directly by a device driver for a vPCI device, particularly GPUs. In such a case, the custom callback installed by the balloon driver runs, but won't find the page in its hot-add region list and doesn't online it, which could cause driver initialization failures.

Fix this by having the balloon custom callback run generic_online_page() when the page isn't part of a Hyper-V hot-add region, thereby doing the default Linux behavior. This allows device driver hot-adds to work properly. Similar cases are handled the same way in the virtio-mem driver.

---
v2: Updated commit message suggested by Michael Kelley.

---
 drivers/hv/hv_balloon.c | 18 ++++++++++--------
 1 file changed, 10 insertions(+), 8 deletions(-)
drivers/hv/hv_balloon.c +10 -8
index a99112e6f0b8..c999daf34108 100644
@@ -766,16 +766,18 @@ static void hv_online_page(struct page *pg, unsigned int order)
 	struct hv_hotadd_state *has;
 	unsigned long pfn = page_to_pfn(pg);
 
-	guard(spinlock_irqsave)(&dm_device.ha_lock);
-	list_for_each_entry(has, &dm_device.ha_region_list, list) {
-		/* The page belongs to a different HAS. */
-		if (pfn < has->start_pfn ||
-		    (pfn + (1UL << order) > has->end_pfn))
-			continue;
+	scoped_guard(spinlock_irqsave, &dm_device.ha_lock) {
+		list_for_each_entry(has, &dm_device.ha_region_list, list) {
+			/* The page belongs to a different HAS. */
+			if (pfn < has->start_pfn ||
+				(pfn + (1UL << order) > has->end_pfn))
+				continue;
 
-		hv_bring_pgs_online(has, pfn, 1UL << order);
-		break;
+			hv_bring_pgs_online(has, pfn, 1UL << order);
+			return;
+		}
 	}
+	generic_online_page(pg, order);
 }
 
 static int pfn_covered(unsigned long start_pfn, unsigned long pfn_cnt)
Signature · 1 line
2.34.1
maintainer Applied
#1 lore raw
Quoted 17 lines · Jacob Pan

The Hyper-V balloon driver installs a custom callback for handling page onlining operations performed by the memory hotplug subsystem. This custom callback is global, and overrides the default callback (generic_online_page) that Linux otherwise uses. The custom callback properly handles memory that is hot-added by the balloon driver as part of a Hyper-V hot-add region.

But memory can also be hot-added directly by a device driver for a vPCI device, particularly GPUs. In such a case, the custom callback installed by the balloon driver runs, but won't find the page in its hot-add region list and doesn't online it, which could cause driver initialization failures.

Fix this by having the balloon custom callback run generic_online_page() when the page isn't part of a Hyper-V hot-add region, thereby doing the default Linux behavior. This allows device driver hot-adds to work properly. Similar cases are handled the same way in the virtio-mem driver.

Quoted 4 lines

Applied to hyperv-next. Thanks!