Skip to contentSkip to conversation

powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory

PATCH Applied Gaurav Batra 5 participants 11 messages Jan 30, 2025 – May 9, 2025 PowerPC

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
 arch/powerpc/include/asm/mmzone.h      |  1 +
 arch/powerpc/mm/numa.c                 |  2 +-
 arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
 3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
 #ifdef CONFIG_MEMORY_HOTPLUG
 extern unsigned long max_pfn;
 u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
 #else
 #define memory_hotplug_max() memblock_end_of_DRAM()
 #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
 	return nid;
 }
 
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
 {
 	struct device_node *memory = NULL;
 	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +16 -13
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
 
 static phys_addr_t ddw_memory_hotplug_max(void)
 {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
 
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
 
 	return max_addr;
 }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
 
 	if (direct_mapping) {
 		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
 					    win64->value, tce_setrange_multi_pSeriesLP_walk);
 		if (ret) {
 			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
 	struct memory_notify *arg = data;
 	int ret = 0;
 
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
 	switch (action) {
 	case MEM_GOING_ONLINE:
 		spin_lock(&dma_win_list_lock);
 		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
 				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
 						arg->nr_pages, window->prop);
 			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
 	case MEM_OFFLINE:
 		spin_lock(&dma_win_list_lock);
 		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
 				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
 						arg->nr_pages, window->prop);
 			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Signature · 1 line
2.39.3 (Apple Git-146)
#1 lore raw
Quoted 97 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 +
  arch/powerpc/mm/numa.c                 |  2 +-
  arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
  #ifdef CONFIG_MEMORY_HOTPLUG
  extern unsigned long max_pfn;
  u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
  #else
  #define memory_hotplug_max() memblock_end_of_DRAM()
  #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
  	return nid;
  }
 
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
  {
  	struct device_node *memory = NULL;
  	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +9 -11
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
 
  static phys_addr_t ddw_memory_hotplug_max(void)
  {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
 
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
 
  	return max_addr;
  }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
 
  	if (direct_mapping) {
  		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
  					    win64->value, tce_setrange_multi_pSeriesLP_walk);
  		if (ret) {
  			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	struct memory_notify *arg = data;
  	int ret = 0;
 
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will

arch/powerpc/platforms/pseries/iommu.c static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action, · line 2355

Quoted 10 lines
arch/powerpc/platforms/pseries/iommu.c +5 -1
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
  	switch (action) {
  	case MEM_GOING_ONLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {

Hi Gaurav,

Since the pmem_start will be greater than ddw_memory_hotplug_max(), and we have not created DDW beyond ddw_memory_hotplug_max(), we are not adding TCE for this range, right?

I have tested this patch on my system, and daxctl reconfigure-device is able to reconfigure PMEM to system RAM.

~# daxctl reconfigure-device --mode=system-ram dax1.0 --force [
  {
    "chardev":"dax1.0",
    "size":5362417664,
    "target_node":4,
    "align":65536,
    "mode":"system-ram",
    "online_memblocks":4,
    "total_memblocks":4,
    "movable":true
  }
]
reconfigured 1 device
~#
~# lsmem
RANGE                                  SIZE  STATE REMOVABLE BLOCK 0x0000000000000000-0x000000697fffffff  422G online       yes 0-421 0x0000040380000000-0x000004047fffffff    4G online       yes 4110-4113

Memory block size:         1G
Total online memory:     426G
Total offline memory:      0B
root@ltcden14-lp2:~#

Thanks
Donet

Quoted 15 lines
arch/powerpc/platforms/pseries/iommu.c +2 -1
  				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	case MEM_OFFLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

#2 lore raw
Quoted 110 lines · Donet Tom
Quoted 98 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 + arch/powerpc/mm/numa.c                 |  2 +- arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------

  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
   #ifdef CONFIG_MEMORY_HOTPLUG
   extern unsigned long max_pfn;
   u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
   #else
   #define memory_hotplug_max() memblock_end_of_DRAM()
   #endif
arch/powerpc/mm/numa.c +1 -0
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
       return nid;
   }
   -static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
   {
       struct device_node *memory = NULL;
       struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +7 -10
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
     static phys_addr_t ddw_memory_hotplug_max(void)
   {
-    resource_size_t max_addr = memory_hotplug_max();
-    struct device_node *memory;
+    resource_size_t max_addr;
   -    for_each_node_by_type(memory, "memory") {
-        struct resource res;
-
-        if (of_address_to_resource(memory, 0, &res))
-            continue;
-
-        max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-    }
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+    max_addr = hot_add_drconf_memory_max();
+#else
+    max_addr = memblock_end_of_DRAM();
+#endif
         return max_addr;
   }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, 
 struct device_node *pdn)
         if (direct_mapping) {
           /* DDW maps the whole partition, so enable direct DMA mapping */
-        ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> 
 PAGE_SHIFT,
+        ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> 
 PAGE_SHIFT,
                           win64->value, tce_setrange_multi_pSeriesLP_walk);

          if (ret) {
              dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",

arch/powerpc/platforms/pseries/iommu.c +0 -0
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct 
 notifier_block *nb, unsigned long action,
       struct memory_notify *arg = data;
       int ret = 0;

  +    /* This notifier can get called when onlining persistent memory as well. +     * TCEs are not pre-mapped for persistent memory. Persistent memory will
+     * always be above ddw_memory_hotplug_max()
+     */

Quoted 8 lines

+
      switch (action) {
      case MEM_GOING_ONLINE:
          spin_lock(&dma_win_list_lock); list_for_each_entry(window, &dma_win_list, list) { -            if (window->direct) {
+            if (window->direct && (arg->start_pfn << PAGE_SHIFT) < +                ddw_memory_hotplug_max()) {

Hi Gaurav,

Quoted 2 lines

Since the pmem_start will be greater than ddw_memory_hotplug_max(), and we have not created DDW beyond ddw_memory_hotplug_max(), we are not adding TCE for this range, right?

That is correct

Quoted 45 lines

I have tested this patch on my system, and daxctl reconfigure-device is able to reconfigure PMEM to system RAM.

~# daxctl reconfigure-device --mode=system-ram dax1.0 --force [
  {
    "chardev":"dax1.0",
    "size":5362417664,
    "target_node":4,
    "align":65536,
    "mode":"system-ram",
    "online_memblocks":4,
    "total_memblocks":4,
    "movable":true
  }
]
reconfigured 1 device
~#
~# lsmem
RANGE                                  SIZE  STATE REMOVABLE BLOCK 0x0000000000000000-0x000000697fffffff  422G online       yes 0-421 0x0000040380000000-0x000004047fffffff    4G online       yes 4110-4113

Memory block size:         1G
Total online memory:     426G
Total offline memory:      0B
root@ltcden14-lp2:~#

Thanks
Donet

Quoted 16 lines

                  ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn, arg->nr_pages, window->prop);
              }

arch/powerpc/platforms/pseries/iommu.c +0 -0
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct 
 notifier_block *nb, unsigned long action,

      case MEM_OFFLINE:
          spin_lock(&dma_win_list_lock); list_for_each_entry(window, &dma_win_list, list) { -            if (window->direct) {
+            if (window->direct && (arg->start_pfn << PAGE_SHIFT) < +                ddw_memory_hotplug_max()) {
                  ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn, arg->nr_pages, window->prop);
              }

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Tested-by Reviewed-by
#3 lore raw
Quoted 113 lines · Gaurav Batra
Quoted 111 lines · Donet Tom
Quoted 99 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 +
  arch/powerpc/mm/numa.c                 |  2 +-
  arch/powerpc/platforms/pseries/iommu.c | 29
++++++++++++++------------

  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
   #ifdef CONFIG_MEMORY_HOTPLUG
   extern unsigned long max_pfn;
   u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
   #else
   #define memory_hotplug_max() memblock_end_of_DRAM()
   #endif
arch/powerpc/mm/numa.c +1 -0
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
       return nid;
   }
   -static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
   {
       struct device_node *memory = NULL;
       struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +7 -10
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
     static phys_addr_t ddw_memory_hotplug_max(void)
   {
-    resource_size_t max_addr = memory_hotplug_max();
-    struct device_node *memory;
+    resource_size_t max_addr;
   -    for_each_node_by_type(memory, "memory") {
-        struct resource res;
-
-        if (of_address_to_resource(memory, 0, &res))
-            continue;
-
-        max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-    }
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+    max_addr = hot_add_drconf_memory_max();
+#else
+    max_addr = memblock_end_of_DRAM();
+#endif
         return max_addr;
   }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, 
 struct device_node *pdn)
         if (direct_mapping) {
           /* DDW maps the whole partition, so enable direct DMA mapping */
-        ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> 
 PAGE_SHIFT,
+        ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> 
 PAGE_SHIFT,
                           win64->value, tce_setrange_multi_pSeriesLP_walk);

          if (ret) {
              dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",

arch/powerpc/platforms/pseries/iommu.c +0 -0
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct 
 notifier_block *nb, unsigned long action,
       struct memory_notify *arg = data;
       int ret = 0;

  +    /* This notifier can get called when onlining persistent memory as well. +     * TCEs are not pre-mapped for persistent memory. Persistent memory will
+     * always be above ddw_memory_hotplug_max()
+     */

Quoted 8 lines

+
      switch (action) {
      case MEM_GOING_ONLINE:
          spin_lock(&dma_win_list_lock); list_for_each_entry(window, &dma_win_list, list) { -            if (window->direct) {
+            if (window->direct && (arg->start_pfn << PAGE_SHIFT) < +                ddw_memory_hotplug_max()) {

Hi Gaurav,

Quoted 4 lines
Quoted 2 lines

Since the pmem_start will be greater than ddw_memory_hotplug_max(), and we have not created DDW beyond ddw_memory_hotplug_max(), we are not adding TCE for this range, right?

That is correct

Thank you.
This looks good to me. feel free to add

Quoted 46 lines
Quoted 46 lines

I have tested this patch on my system, and daxctl reconfigure-device is able to reconfigure PMEM to system RAM.

~# daxctl reconfigure-device --mode=system-ram dax1.0 --force [
  {
    "chardev":"dax1.0",
    "size":5362417664,
    "target_node":4,
    "align":65536,
    "mode":"system-ram",
    "online_memblocks":4,
    "total_memblocks":4,
    "movable":true
  }
]
reconfigured 1 device
~#
~# lsmem
RANGE                                  SIZE  STATE REMOVABLE BLOCK 0x0000000000000000-0x000000697fffffff  422G online       yes 0-421 0x0000040380000000-0x000004047fffffff    4G online       yes 4110-4113

Memory block size:         1G
Total online memory:     426G
Total offline memory:      0B
root@ltcden14-lp2:~#

Thanks
Donet

Quoted 16 lines

                  ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn, arg->nr_pages, window->prop);
              }

arch/powerpc/platforms/pseries/iommu.c +0 -0
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct 
 notifier_block *nb, unsigned long action,

      case MEM_OFFLINE:
          spin_lock(&dma_win_list_lock); list_for_each_entry(window, &dma_win_list, list) { -            if (window->direct) {
+            if (window->direct && (arg->start_pfn << PAGE_SHIFT) < +                ddw_memory_hotplug_max()) {
                  ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn, arg->nr_pages, window->prop);
              }

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

maintainer Applied
#4 lore raw
In reply to Gaurav Batra
Quoted 10 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

[...]

Applied to powerpc/next.

[1/1] powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory

      https://git.kernel.org/powerpc/c/6aa989ab2bd0d37540c812b4270006ff794662e7

Thanks

#5 lore raw
In reply to Gaurav Batra

Hello,

looks like this upsets some assumption qemu has about these windows.

https://lists.nongnu.org/archive/html/qemu-devel/2025-03/msg05137.html

When Linux kernel that has this patch applied is running inside a qemu VM with a PCI device and the VM is rebooted qemu crashes shortly after the next Linux kernel starts.

This is quite curious since qemu does AFAIK not support pmemory at all.

Any idea what went wrong there?

Thanks

Michal

Quoted 126 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
 arch/powerpc/include/asm/mmzone.h      |  1 +
 arch/powerpc/mm/numa.c                 |  2 +-
 arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
 3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
 #ifdef CONFIG_MEMORY_HOTPLUG
 extern unsigned long max_pfn;
 u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
 #else
 #define memory_hotplug_max() memblock_end_of_DRAM()
 #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
 	return nid;
 }
 
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
 {
 	struct device_node *memory = NULL;
 	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +16 -13
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
 
 static phys_addr_t ddw_memory_hotplug_max(void)
 {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
 
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
 
 	return max_addr;
 }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
 
 	if (direct_mapping) {
 		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
 					    win64->value, tce_setrange_multi_pSeriesLP_walk);
 		if (ret) {
 			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
 	struct memory_notify *arg = data;
 	int ret = 0;
 
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
 	switch (action) {
 	case MEM_GOING_ONLINE:
 		spin_lock(&dma_win_list_lock);
 		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
 				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
 						arg->nr_pages, window->prop);
 			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
 	case MEM_OFFLINE:
 		spin_lock(&dma_win_list_lock);
 		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
 				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
 						arg->nr_pages, window->prop);
 			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Signature · 1 line
2.39.3 (Apple Git-146)
#6 lore raw

Hello Michal,

In the patch to fix the pmemory bug, I made some changes to the code that determines Max memory an LPAR can have (excluding pmemory). This information is needed while creating Dynamic DMA Window (DDW). These changes are in the main line code path of DDW creation. This might have irritated QEMU somehow, no idea yet on how.

Thanks,

Gaurav

Quoted 145 lines · Michal Suchánek

Hello,

looks like this upsets some assumption qemu has about these windows.

https://lists.nongnu.org/archive/html/qemu-devel/2025-03/msg05137.html

When Linux kernel that has this patch applied is running inside a qemu VM with a PCI device and the VM is rebooted qemu crashes shortly after the next Linux kernel starts.

This is quite curious since qemu does AFAIK not support pmemory at all.

Any idea what went wrong there?

Thanks

Michal

Quoted 126 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 +
  arch/powerpc/mm/numa.c                 |  2 +-
  arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
  #ifdef CONFIG_MEMORY_HOTPLUG
  extern unsigned long max_pfn;
  u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
  #else
  #define memory_hotplug_max() memblock_end_of_DRAM()
  #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
  	return nid;
  }
 
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
  {
  	struct device_node *memory = NULL;
  	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +16 -13
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
 
  static phys_addr_t ddw_memory_hotplug_max(void)
  {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
 
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
 
  	return max_addr;
  }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
 
  	if (direct_mapping) {
  		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
  					    win64->value, tce_setrange_multi_pSeriesLP_walk);
  		if (ret) {
  			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	struct memory_notify *arg = data;
  	int ret = 0;
 
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
  	switch (action) {
  	case MEM_GOING_ONLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	case MEM_OFFLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Signature · 1 line
2.39.3 (Apple Git-146)
#7 lore raw

Hello,

Quoted 7 lines · Gaurav Batra

Hello Michal,

In the patch to fix the pmemory bug, I made some changes to the code that determines Max memory an LPAR can have (excluding pmemory). This information is needed while creating Dynamic DMA Window (DDW). These changes are in the main line code path of DDW creation. This might have irritated QEMU somehow, no idea yet on how.

Yes, it's defeinitely something with the DDW code. Using the disable_ddw=1 kernel parameter avoids the qemu crash.

The kernels in
https://download.opensuse.org/repositories/Kernel:/SLE15-SP7/pool/ppc64le/

have the patch applied.

Booting the kernel inside qemu VM with a PCI device (such as the USB hub) and then rebooting the VM crashes qemu.

Thanks

Michal

Quoted 145 lines

Thanks,

Gaurav

Quoted 139 lines · Michal Suchánek

Hello,

looks like this upsets some assumption qemu has about these windows.

https://lists.nongnu.org/archive/html/qemu-devel/2025-03/msg05137.html

When Linux kernel that has this patch applied is running inside a qemu VM with a PCI device and the VM is rebooted qemu crashes shortly after the next Linux kernel starts.

This is quite curious since qemu does AFAIK not support pmemory at all.

Any idea what went wrong there?

Thanks

Michal

Quoted 120 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 +
  arch/powerpc/mm/numa.c                 |  2 +-
  arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
  #ifdef CONFIG_MEMORY_HOTPLUG
  extern unsigned long max_pfn;
  u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
  #else
  #define memory_hotplug_max() memblock_end_of_DRAM()
  #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
  	return nid;
  }
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
  {
  	struct device_node *memory = NULL;
  	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +16 -13
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
  static phys_addr_t ddw_memory_hotplug_max(void)
  {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
  	return max_addr;
  }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
  	if (direct_mapping) {
  		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
  					    win64->value, tce_setrange_multi_pSeriesLP_walk);
  		if (ret) {
  			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	struct memory_notify *arg = data;
  	int ret = 0;
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
  	switch (action) {
  	case MEM_GOING_ONLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	case MEM_OFFLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Signature · 1 line
2.39.3 (Apple Git-146)
#8 lore raw

Hi Michal,

I can recreate this issue on sles16 distro kernel but I don't observe this issue with upstream Linux 6.15-rc5 on the **same** sles16 guest.

Note: the commit 6aa989ab2bd0 ("powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory") was included since Linux 6.15-rc1.

I think there's something more to this crash.

Thanks,
Amit

Quoted 172 lines · Michal Suchánek

Hello,

Quoted 7 lines · Gaurav Batra

Hello Michal,

In the patch to fix the pmemory bug, I made some changes to the code that determines Max memory an LPAR can have (excluding pmemory). This information is needed while creating Dynamic DMA Window (DDW). These changes are in the main line code path of DDW creation. This might have irritated QEMU somehow, no idea yet on how.

Yes, it's defeinitely something with the DDW code. Using the disable_ddw=1 kernel parameter avoids the qemu crash.

The kernels in
https://download.opensuse.org/repositories/Kernel:/SLE15-SP7/pool/ppc64le/

have the patch applied.

Booting the kernel inside qemu VM with a PCI device (such as the USB hub) and then rebooting the VM crashes qemu.

Thanks

Michal

Quoted 145 lines

Thanks,

Gaurav

Quoted 139 lines · Michal Suchánek

Hello,

looks like this upsets some assumption qemu has about these windows.

https://lists.nongnu.org/archive/html/qemu-devel/2025-03/msg05137.html

When Linux kernel that has this patch applied is running inside a qemu VM with a PCI device and the VM is rebooted qemu crashes shortly after the next Linux kernel starts.

This is quite curious since qemu does AFAIK not support pmemory at all.

Any idea what went wrong there?

Thanks

Michal

Quoted 120 lines · Gaurav Batra

iommu_mem_notifier() is invoked when RAM is dynamically added/removed. This notifier call is responsible to add/remove TCEs from the Dynamic DMA Window (DDW) when TCEs are pre-mapped. TCEs are pre-mapped only for RAM and not for persistent memory (pmemory). For DMA buffers in pmemory, TCEs are dynamically mapped when the device driver instructs to do so.

The issue is 'daxctl' command is capable of adding pmemory as "System RAM" after LPAR boot. The command to do so is -

daxctl reconfigure-device --mode=system-ram dax0.0 --force

This will dynamically add pmemory range to LPAR RAM eventually invoking iommu_mem_notifier(). The address range of pmemory is way beyond the Max RAM that the LPAR can have. Which means, this range is beyond the DDW created for the device, at device initialization time.

As a result when TCEs are pre-mapped for the pmemory range, by iommu_mem_notifier(), PHYP HCALL returns H_PARAMETER. This failed the command, daxctl, to add pmemory as RAM.

The solution is to not pre-map TCEs for pmemory.

---
  arch/powerpc/include/asm/mmzone.h      |  1 +
  arch/powerpc/mm/numa.c                 |  2 +-
  arch/powerpc/platforms/pseries/iommu.c | 29 ++++++++++++++------------
  3 files changed, 18 insertions(+), 14 deletions(-)
arch/powerpc/include/asm/mmzone.h +1 -0
index d99863cd6cde..049152f8d597 100644
@@ -29,6 +29,7 @@ extern cpumask_var_t node_to_cpumask_map[];
  #ifdef CONFIG_MEMORY_HOTPLUG
  extern unsigned long max_pfn;
  u64 memory_hotplug_max(void);
+u64 hot_add_drconf_memory_max(void);
  #else
  #define memory_hotplug_max() memblock_end_of_DRAM()
  #endif
arch/powerpc/mm/numa.c +1 -1
index 3c1da08304d0..603a0f652ba6 100644
@@ -1336,7 +1336,7 @@ int hot_add_scn_to_nid(unsigned long scn_addr)
  	return nid;
  }
-static u64 hot_add_drconf_memory_max(void)
+u64 hot_add_drconf_memory_max(void)
  {
  	struct device_node *memory = NULL;
  	struct device_node *dn = NULL;
arch/powerpc/platforms/pseries/iommu.c +16 -13
index 29f1a0cc59cd..abd9529a8f41 100644
@@ -1284,17 +1284,13 @@ static LIST_HEAD(failed_ddw_pdn_list);
  static phys_addr_t ddw_memory_hotplug_max(void)
  {
-	resource_size_t max_addr = memory_hotplug_max();
-	struct device_node *memory;
+	resource_size_t max_addr;
-	for_each_node_by_type(memory, "memory") {
-		struct resource res;
-
-		if (of_address_to_resource(memory, 0, &res))
-			continue;
-
-		max_addr = max_t(resource_size_t, max_addr, res.end + 1);
-	}
+#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
+	max_addr = hot_add_drconf_memory_max();
+#else
+	max_addr = memblock_end_of_DRAM();
+#endif
  	return max_addr;
  }
@@ -1600,7 +1596,7 @@ static bool enable_ddw(struct pci_dev *dev, struct device_node *pdn)
  	if (direct_mapping) {
  		/* DDW maps the whole partition, so enable direct DMA mapping */
-		ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
+		ret = walk_system_ram_range(0, ddw_memory_hotplug_max() >> PAGE_SHIFT,
  					    win64->value, tce_setrange_multi_pSeriesLP_walk);
  		if (ret) {
  			dev_info(&dev->dev, "failed to map DMA window for %pOF: %d\n",
@@ -2346,11 +2342,17 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	struct memory_notify *arg = data;
  	int ret = 0;
+	/* This notifier can get called when onlining persistent memory as well.
+	 * TCEs are not pre-mapped for persistent memory. Persistent memory will
+	 * always be above ddw_memory_hotplug_max()
+	 */
+
  	switch (action) {
  	case MEM_GOING_ONLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}
@@ -2362,7 +2364,8 @@ static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
  	case MEM_OFFLINE:
  		spin_lock(&dma_win_list_lock);
  		list_for_each_entry(window, &dma_win_list, list) {
-			if (window->direct) {
+			if (window->direct && (arg->start_pfn << PAGE_SHIFT) <
+				ddw_memory_hotplug_max()) {
  				ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
  						arg->nr_pages, window->prop);
  			}

base-commit: 95ec54a420b8f445e04a7ca0ea8deb72c51fe1d3

Signature · 1 line
2.39.3 (Apple Git-146)
#9 lore raw

Hello,

Quoted 7 lines · Amit Machhiwal

Hi Michal,

I can recreate this issue on sles16 distro kernel but I don't observe this issue with upstream Linux 6.15-rc5 on the **same** sles16 guest.

Note: the commit 6aa989ab2bd0 ("powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory") was included since Linux 6.15-rc1.

Indeed, it's not reproducible with 6.15-rc5.

It's not clear what changed in TCE handling between 6.12 and 6.15 that avoids this problem.

Thanks

Michal

#10 lore raw
Quoted 8 lines · Michal Suchánek

Hello,

Quoted 5 lines · Amit Machhiwal

Hi Michal,

I can recreate this issue on sles16 distro kernel but I don't observe this issue with upstream Linux 6.15-rc5 on the **same** sles16 guest.

Quoted 7 lines
Quoted 2 lines

Note: the commit 6aa989ab2bd0 ("powerpc/pseries/iommu: memory notifier incorrectly adds TCEs for pmemory") was included since Linux 6.15-rc1.

Indeed, it's not reproducible with 6.15-rc5.

It's not clear what changed in TCE handling between 6.12 and 6.15 that avoids this problem.

Seems backporting commit 67dfc11982f7 ("powerpc/pseries/iommu: create DDW for devices with DMA mask less than 64-bits") avoids the problem but it's not clear why.

Thanks

Michal