Skip to contentSkip to conversation

handle init errors more gracefully in amd_iommu

PATCH In review Kevin Mitchell via iommu 2 participants 5 messages Apr 18, 2019 – May 7, 2019 1 message not in the archive IOMMU

lore
Share & tools
lore
#0 lore raw

cover handle init errors more gracefully in amd_iommu

This series makes error handling more robust in the amd_iommu init code. It was initially motivated by problematic firmware that does not set up the physical address of the iommu. This led to a NULL dereference panic when iommu_disable was called during cleanup.

While the first patch is sufficient to avoid the panic, the subsequent two move the cleanup closer to the actual error and avoid calling the cleanup code twice when amd_iommu=off is specified on the command line.

I have tested this series on a variety of AMD CPUs with firmware exhibiting the issue. I have additionally tested on platforms where the firmware has been fixed. I tried both with and without amd_iommu=off. I have also tested on older CPUs where no IOMMU is detected and even one where the GART driver ends up running.

Thanks,

Kevin

Kevin Mitchell (3):
iommu/amd: make iommu_disable safer iommu/amd: move gart fallback to amd_iommu_init iommu/amd: only free resources once on init error

 drivers/iommu/amd_iommu_init.c | 45 ++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 21 deletions(-)
Signature · 1 line
2.20.1

A message that is not in the archive (sent off-list or not yet received): 20190418175744.6229-1-kevmitch-nzgTgzXrdUbQT0dZR+AlfA@public.gmane.org

#2 lore raw

PATCH 1/3 iommu/amd: make iommu_disable safer

Make it safe to call iommu_disable during early init error conditions before mmio_base is set, but after the struct amd_iommu has been added to the amd_iommu_list. For example, this happens if firmware fails to fill in mmio_phys in the ACPI table leading to a NULL pointer dereference in iommu_feature_disable.

---
 drivers/iommu/amd_iommu_init.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/iommu/amd_iommu_init.c +3 -0
index f773792d77fd..3798d7303c99 100644
@@ -424,6 +424,9 @@ static void iommu_enable(struct amd_iommu *iommu)
 
 static void iommu_disable(struct amd_iommu *iommu)
 {
+	if (!iommu->mmio_base)
+		return;
+
 	/* Disable command buffer */
 	iommu_feature_disable(iommu, CONTROL_CMDBUF_EN);
 
Signature · 1 line
2.20.1
#3 lore raw

PATCH 2/3 iommu/amd: move gart fallback to amd_iommu_init

In reply to a message not in the archive

The fallback to the GART driver in the case amd_iommu doesn't work was executed in a function called free_iommu_resources, which didn't really make sense. This was even being called twice if amd_iommu=off was specified on the command line.

The only complication is that it needs to be verified that amd_iommu has fully relinquished control by calling free_iommu_resources and emptying the amd_iommu_list.

---
 drivers/iommu/amd_iommu_init.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)
drivers/iommu/amd_iommu_init.c +10 -9
index 3798d7303c99..5f3df5ae6ba8 100644
@@ -2345,15 +2345,6 @@ static void __init free_iommu_resources(void)
 	amd_iommu_dev_table = NULL;
 
 	free_iommu_all();
-
-#ifdef CONFIG_GART_IOMMU
-	/*
-	 * We failed to initialize the AMD IOMMU - try fallback to GART
-	 * if possible.
-	 */
-	gart_iommu_init();
-
-#endif
 }
 
 /* SB IOAPIC is always on this device in AMD systems */
@@ -2774,6 +2765,16 @@ static int __init amd_iommu_init(void)
 		}
 	}
 
+#ifdef CONFIG_GART_IOMMU
+	if (ret && list_empty(&amd_iommu_list)) {
+		/*
+		 * We failed to initialize the AMD IOMMU - try fallback
+		 * to GART if possible.
+		 */
+		gart_iommu_init();
+	}
+#endif
+
 	for_each_iommu(iommu)
 		amd_iommu_debugfs_setup(iommu);
 
Signature · 1 line
2.20.1
#4 lore raw

PATCH 3/3 iommu/amd: only free resources once on init error

In reply to a message not in the archive

When amd_iommu=off was specified on the command line, free_X_resources functions were called immediately after early_amd_iommu_init. They were then called again when amd_iommu_init also failed (as expected).

Instead, call them only once: at the end of state_next() whenever there's an error. These functions should be safe to call any time and any number of times. However, since state_next is never called again in an error state, the cleanup will only ever be run once.

This also ensures that cleanup code is run as soon as possible after an error is detected rather than waiting for amd_iommu_init() to be called.

---
 drivers/iommu/amd_iommu_init.c | 27 +++++++++++++--------------
 1 file changed, 13 insertions(+), 14 deletions(-)
drivers/iommu/amd_iommu_init.c +13 -14
index 5f3df5ae6ba8..24fc060fe596 100644
@@ -2638,8 +2638,6 @@ static int __init state_next(void)
 		init_state = ret ? IOMMU_INIT_ERROR : IOMMU_ACPI_FINISHED;
 		if (init_state == IOMMU_ACPI_FINISHED && amd_iommu_disabled) {
 			pr_info("AMD IOMMU disabled on kernel command-line\n");
-			free_dma_resources();
-			free_iommu_resources();
 			init_state = IOMMU_CMDLINE_DISABLED;
 			ret = -EINVAL;
 		}
@@ -2680,6 +2678,19 @@ static int __init state_next(void)
 		BUG();
 	}
 
+	if (ret) {
+		free_dma_resources();
+		if (!irq_remapping_enabled) {
+			disable_iommus();
+			free_iommu_resources();
+		} else {
+			struct amd_iommu *iommu;
+
+			uninit_device_table_dma();
+			for_each_iommu(iommu)
+				iommu_flush_all_caches(iommu);
+		}
+	}
 	return ret;
 }
 
@@ -2753,18 +2764,6 @@ static int __init amd_iommu_init(void)
 	int ret;
 
 	ret = iommu_go_to_state(IOMMU_INITIALIZED);
-	if (ret) {
-		free_dma_resources();
-		if (!irq_remapping_enabled) {
-			disable_iommus();
-			free_iommu_resources();
-		} else {
-			uninit_device_table_dma();
-			for_each_iommu(iommu)
-				iommu_flush_all_caches(iommu);
-		}
-	}
-
 #ifdef CONFIG_GART_IOMMU
 	if (ret && list_empty(&amd_iommu_list)) {
 		/*
Signature · 1 line
2.20.1
#5 lore raw
In reply to Kevin Mitchell via iommu

Sorry to nag, but did anyone have any comments about these patches? Is no news good news?

Thanks,
Kevin

Quoted 24 lines · Kevin Mitchell

This series makes error handling more robust in the amd_iommu init code. It was initially motivated by problematic firmware that does not set up the physical address of the iommu. This led to a NULL dereference panic when iommu_disable was called during cleanup.

While the first patch is sufficient to avoid the panic, the subsequent two move the cleanup closer to the actual error and avoid calling the cleanup code twice when amd_iommu=off is specified on the command line.

I have tested this series on a variety of AMD CPUs with firmware exhibiting the issue. I have additionally tested on platforms where the firmware has been fixed. I tried both with and without amd_iommu=off. I have also tested on older CPUs where no IOMMU is detected and even one where the GART driver ends up running.

Thanks,

Kevin

Kevin Mitchell (3):
iommu/amd: make iommu_disable safer iommu/amd: move gart fallback to amd_iommu_init iommu/amd: only free resources once on init error

Quoted 7 lines
 drivers/iommu/amd_iommu_init.c | 45 ++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 21 deletions(-)

--
2.20.1

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu