Skip to contentSkip to conversation

nvdimm/pfn_dev: Prevent the creation of zero-sized namespaces

PATCH Superseded Aneesh Kumar K.V 2 participants 5 messages Aug 4, 2023 – Aug 5, 2023 libnvdimm: non-volatile memory device

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 nvdimm/pfn_dev: Prevent the creation of zero-sized namespaces

On architectures that have different page size values used for kernel direct mapping and userspace mappings, the user can end up creating zero-sized namespaces as shown below

:/sys/bus/nd/devices/region1# cat align
0x1000000
/sys/bus/nd/devices/region1# echo 0x200000 > align
/sys/bus/nd/devices/region1/dax1.0# cat supported_alignments
65536 16777216
 $ ndctl create-namespace -r region1 -m devdax -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"devdax",
  "map":"dev",
  "size":0,
  "uuid":"3094329a-0c66-4905-847e-357223e56ab0",
  "daxregion":{
    "id":1,
    "size":0,
    "align":65536
  },
  "align":65536
}
similarily for fsdax
 $ ndctl create-namespace -r region1 -m fsdax  -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"fsdax",
  "map":"dev",
  "size":0,
  "uuid":"45538a6f-dec7-405d-b1da-2a4075e06232",
  "sector_size":512,
  "align":65536,
  "blockdev":"pmem1"
}

In commit 9ffc1d19fc4a ("mm/memremap_pages: Introduce memremap_compat_align()") memremap_compat_align was added to make sure the kernel always creates namespaces with 16MB alignment. But the user can still override the region alignment and no input validation is done in the region alignment values to retain the flexibility user had before. However, the kernel ensures that only part of the namespace that can be mapped via kernel direct mapping page size is enabled. This is achieved by tracking the unmapped part of the namespace in pfn_sb->end_trunc. The kernel also ensures that the start address of the namespace is also aligned to the kernel direct mapping page size.

Depending on the user request, the kernel implements userspace mapping alignment by updating pfn device alignment attribute and this value is used to adjust the start address for userspace mappings. This is tracked in pfn_sb->dataoff. Hence the available size for userspace mapping is:

usermapping_size = size of the namespace - pfn_sb->end_trun - pfn_sb->dataoff

If the kernel finds the user mapping size zero then don't allow the creation of namespace.

After fix:

$ ndctl create-namespace -f  -r region1 -m devdax  -s 18M --align 64K
libndctl: ndctl_dax_enable: dax1.1: failed to enable
  Error: namespace1.2: failed to enable

failed to create namespace: No such device or address

And existing zero sized namespace will be marked disabled.

root@ltczz75-lp2:/home/kvaneesh# ndctl  list -N -r region1 -i
[
  {
    "dev":"namespace1.0",
    "mode":"raw",
    "size":18874368,
    "uuid":"94a90fb0-8e78-4fb6-a759-ffc62f9fa181",
    "sector_size":512,
    "state":"disabled"
  },
---
 drivers/nvdimm/pfn_devs.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/nvdimm/pfn_devs.c +8 -2
index af7d9301520c..36b904a129b9 100644
@@ -453,7 +453,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	struct resource *res;
 	enum nd_pfn_mode mode;
 	struct nd_namespace_io *nsio;
-	unsigned long align, start_pad;
+	unsigned long align, start_pad, end_trunc;
 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 	struct nd_namespace_common *ndns = nd_pfn->ndns;
 	const uuid_t *parent_uuid = nd_dev_to_uuid(&ndns->dev);
@@ -503,6 +503,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	align = le32_to_cpu(pfn_sb->align);
 	offset = le64_to_cpu(pfn_sb->dataoff);
 	start_pad = le32_to_cpu(pfn_sb->start_pad);
+	end_trunc = le32_to_cpu(pfn_sb->end_trunc);
 	if (align == 0)
 		align = 1UL << ilog2(offset);
 	mode = le32_to_cpu(pfn_sb->mode);
@@ -610,6 +611,10 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 		return -EOPNOTSUPP;
 	}
 
+	if (offset >= (res->end - res->start + 1 - start_pad - end_trunc)) {
+		dev_err(&nd_pfn->dev, "bad offset with small namespace\n");
+		return -EOPNOTSUPP;
+	}
 	return 0;
 }
 EXPORT_SYMBOL(nd_pfn_validate);
@@ -810,7 +815,8 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
 	else
 		return -ENXIO;
 
-	if (offset >= size) {
+	if (offset >= (size - end_trunc)) {
+		/* This implies we result in zero size devices */
 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
 				dev_name(&ndns->dev));
 		return -ENXIO;
Signature · 1 line
2.41.0
#1 lore raw

PATCH 2/2 nvdimm/pfn_dev: Avoid unnecessary endian conversion

use the local variable that already have the converted values.

No functional change in this patch.

---
 drivers/nvdimm/pfn_devs.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/nvdimm/pfn_devs.c +2 -4
index 36b904a129b9..8b7342517895 100644
@@ -599,14 +599,12 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 		return -EOPNOTSUPP;
 	}
 
-	if (!IS_ALIGNED(res->start + le32_to_cpu(pfn_sb->start_pad),
-				memremap_compat_align())) {
+	if (!IS_ALIGNED(res->start + start_pad, memremap_compat_align())) {
 		dev_err(&nd_pfn->dev, "resource start misaligned\n");
 		return -EOPNOTSUPP;
 	}
 
-	if (!IS_ALIGNED(res->end + 1 - le32_to_cpu(pfn_sb->end_trunc),
-				memremap_compat_align())) {
+	if (!IS_ALIGNED(res->end + 1 - end_trunc, memremap_compat_align())) {
 		dev_err(&nd_pfn->dev, "resource end misaligned\n");
 		return -EOPNOTSUPP;
 	}
Signature · 1 line
2.41.0
Reviewed-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] nvdimm/pfn_dev: Avoid unnecessary endian conversion

Quoted 19 lines · Aneesh Kumar K.V

use the local variable that already have the converted values.

No functional change in this patch.

---
 drivers/nvdimm/pfn_devs.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)
drivers/nvdimm/pfn_devs.c +1 -2
index 36b904a129b9..8b7342517895 100644
@@ -599,14 +599,12 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 		return -EOPNOTSUPP;
 	}
 
-	if (!IS_ALIGNED(res->start + le32_to_cpu(pfn_sb->start_pad),
-				memremap_compat_align())) {
+	if (!IS_ALIGNED(res->start + start_pad, memremap_compat_align())) {

drivers/nvdimm/pfn_devs.c int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig) · line 610

Quoted 10 lines
drivers/nvdimm/pfn_devs.c +1 -2
 		dev_err(&nd_pfn->dev, "resource start misaligned\n");
 		return -EOPNOTSUPP;
 	}
 
-	if (!IS_ALIGNED(res->end + 1 - le32_to_cpu(pfn_sb->end_trunc),
-				memremap_compat_align())) {
+	if (!IS_ALIGNED(res->end + 1 - end_trunc, memremap_compat_align())) {
 		dev_err(&nd_pfn->dev, "resource end misaligned\n");
 		return -EOPNOTSUPP;
 	}
#3 lore raw
In reply to Aneesh Kumar K.V

Hi, Aneesh,

Quoted 25 lines · Aneesh Kumar K.V

On architectures that have different page size values used for kernel direct mapping and userspace mappings, the user can end up creating zero-sized namespaces as shown below

:/sys/bus/nd/devices/region1# cat align
0x1000000
/sys/bus/nd/devices/region1# echo 0x200000 > align
/sys/bus/nd/devices/region1/dax1.0# cat supported_alignments
65536 16777216
 $ ndctl create-namespace -r region1 -m devdax -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"devdax",
  "map":"dev",
  "size":0,
  "uuid":"3094329a-0c66-4905-847e-357223e56ab0",
  "daxregion":{
    "id":1,
    "size":0,
    "align":65536
  },
  "align":65536
}
similarily for fsdax
Quoted 11 lines
 $ ndctl create-namespace -r region1 -m fsdax  -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"fsdax",
  "map":"dev",
  "size":0,
  "uuid":"45538a6f-dec7-405d-b1da-2a4075e06232",
  "sector_size":512,
  "align":65536,
  "blockdev":"pmem1"
}

Just curious, but have you seen this in practice? It seems like an odd thing to do.

Quoted 29 lines

In commit 9ffc1d19fc4a ("mm/memremap_pages: Introduce memremap_compat_align()") memremap_compat_align was added to make sure the kernel always creates namespaces with 16MB alignment. But the user can still override the region alignment and no input validation is done in the region alignment values to retain the flexibility user had before. However, the kernel ensures that only part of the namespace that can be mapped via kernel direct mapping page size is enabled. This is achieved by tracking the unmapped part of the namespace in pfn_sb->end_trunc. The kernel also ensures that the start address of the namespace is also aligned to the kernel direct mapping page size.

Depending on the user request, the kernel implements userspace mapping alignment by updating pfn device alignment attribute and this value is used to adjust the start address for userspace mappings. This is tracked in pfn_sb->dataoff. Hence the available size for userspace mapping is:

usermapping_size = size of the namespace - pfn_sb->end_trun - pfn_sb->dataoff

If the kernel finds the user mapping size zero then don't allow the creation of namespace.

After fix:

$ ndctl create-namespace -f  -r region1 -m devdax  -s 18M --align 64K
libndctl: ndctl_dax_enable: dax1.1: failed to enable
  Error: namespace1.2: failed to enable

failed to create namespace: No such device or address

And existing zero sized namespace will be marked disabled.

Quoted 10 lines
root@ltczz75-lp2:/home/kvaneesh# ndctl  list -N -r region1 -i
[
  {
    "dev":"namespace1.0",
    "mode":"raw",
    "size":18874368,
    "uuid":"94a90fb0-8e78-4fb6-a759-ffc62f9fa181",
    "sector_size":512,
    "state":"disabled"
  },

Thank you for providing examples of the command output before and after the change. I appreciate that.

Quoted 28 lines
---
 drivers/nvdimm/pfn_devs.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/nvdimm/pfn_devs.c +2 -1
index af7d9301520c..36b904a129b9 100644
@@ -453,7 +453,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	struct resource *res;
 	enum nd_pfn_mode mode;
 	struct nd_namespace_io *nsio;
-	unsigned long align, start_pad;
+	unsigned long align, start_pad, end_trunc;
 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 	struct nd_namespace_common *ndns = nd_pfn->ndns;
 	const uuid_t *parent_uuid = nd_dev_to_uuid(&ndns->dev);
@@ -503,6 +503,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	align = le32_to_cpu(pfn_sb->align);
 	offset = le64_to_cpu(pfn_sb->dataoff);
 	start_pad = le32_to_cpu(pfn_sb->start_pad);
+	end_trunc = le32_to_cpu(pfn_sb->end_trunc);
 	if (align == 0)
 		align = 1UL << ilog2(offset);
 	mode = le32_to_cpu(pfn_sb->mode);

drivers/nvdimm/pfn_devs.c int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig) · line 614

Quoted 4 lines
drivers/nvdimm/pfn_devs.c +1 -0
@@ -610,6 +611,10 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 		return -EOPNOTSUPP;
 	}
 
+	if (offset >= (res->end - res->start + 1 - start_pad - end_trunc)) {
                       ^^^^^^^^^^^^^^^^^^^^^^^^^ That's what

resource_size(res) does. It might be better to create a local variable 'size' to hold that, as there are now two instances of that in the function.

Quoted 7 lines
drivers/nvdimm/pfn_devs.c +3 -0
+		dev_err(&nd_pfn->dev, "bad offset with small namespace\n");
+		return -EOPNOTSUPP;
+	}
 	return 0;
 }
 EXPORT_SYMBOL(nd_pfn_validate);

drivers/nvdimm/pfn_devs.c static int nd_pfn_init(struct nd_pfn *nd_pfn) · line 822

Quoted 9 lines
drivers/nvdimm/pfn_devs.c +2 -1
@@ -810,7 +815,8 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
 	else
 		return -ENXIO;
 
-	if (offset >= size) {
+	if (offset >= (size - end_trunc)) {
+		/* This implies we result in zero size devices */
 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
 				dev_name(&ndns->dev));
 		return -ENXIO;

Functionally, this looks good to me.

Cheers,
Jeff

#4 lore raw
Quoted 41 lines · Jeff Moyer

Hi, Aneesh,

Quoted 25 lines · Aneesh Kumar K.V

On architectures that have different page size values used for kernel direct mapping and userspace mappings, the user can end up creating zero-sized namespaces as shown below

:/sys/bus/nd/devices/region1# cat align
0x1000000
/sys/bus/nd/devices/region1# echo 0x200000 > align
/sys/bus/nd/devices/region1/dax1.0# cat supported_alignments
65536 16777216
 $ ndctl create-namespace -r region1 -m devdax -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"devdax",
  "map":"dev",
  "size":0,
  "uuid":"3094329a-0c66-4905-847e-357223e56ab0",
  "daxregion":{
    "id":1,
    "size":0,
    "align":65536
  },
  "align":65536
}
similarily for fsdax
Quoted 11 lines
 $ ndctl create-namespace -r region1 -m fsdax  -s 18M --align 64K
{
  "dev":"namespace1.0",
  "mode":"fsdax",
  "map":"dev",
  "size":0,
  "uuid":"45538a6f-dec7-405d-b1da-2a4075e06232",
  "sector_size":512,
  "align":65536,
  "blockdev":"pmem1"
}
Quoted 3 lines

Just curious, but have you seen this in practice? It seems like an odd thing to do.

This was identified while writing new test cases for region alignment update.

Quoted 74 lines
Quoted 29 lines

In commit 9ffc1d19fc4a ("mm/memremap_pages: Introduce memremap_compat_align()") memremap_compat_align was added to make sure the kernel always creates namespaces with 16MB alignment. But the user can still override the region alignment and no input validation is done in the region alignment values to retain the flexibility user had before. However, the kernel ensures that only part of the namespace that can be mapped via kernel direct mapping page size is enabled. This is achieved by tracking the unmapped part of the namespace in pfn_sb->end_trunc. The kernel also ensures that the start address of the namespace is also aligned to the kernel direct mapping page size.

Depending on the user request, the kernel implements userspace mapping alignment by updating pfn device alignment attribute and this value is used to adjust the start address for userspace mappings. This is tracked in pfn_sb->dataoff. Hence the available size for userspace mapping is:

usermapping_size = size of the namespace - pfn_sb->end_trun - pfn_sb->dataoff

If the kernel finds the user mapping size zero then don't allow the creation of namespace.

After fix:

$ ndctl create-namespace -f  -r region1 -m devdax  -s 18M --align 64K
libndctl: ndctl_dax_enable: dax1.1: failed to enable
  Error: namespace1.2: failed to enable

failed to create namespace: No such device or address

And existing zero sized namespace will be marked disabled.

Quoted 10 lines
root@ltczz75-lp2:/home/kvaneesh# ndctl  list -N -r region1 -i
[
  {
    "dev":"namespace1.0",
    "mode":"raw",
    "size":18874368,
    "uuid":"94a90fb0-8e78-4fb6-a759-ffc62f9fa181",
    "sector_size":512,
    "state":"disabled"
  },

Thank you for providing examples of the command output before and after the change. I appreciate that.

Quoted 31 lines
---
 drivers/nvdimm/pfn_devs.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
drivers/nvdimm/pfn_devs.c +2 -1
index af7d9301520c..36b904a129b9 100644
@@ -453,7 +453,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	struct resource *res;
 	enum nd_pfn_mode mode;
 	struct nd_namespace_io *nsio;
-	unsigned long align, start_pad;
+	unsigned long align, start_pad, end_trunc;
 	struct nd_pfn_sb *pfn_sb = nd_pfn->pfn_sb;
 	struct nd_namespace_common *ndns = nd_pfn->ndns;
 	const uuid_t *parent_uuid = nd_dev_to_uuid(&ndns->dev);
@@ -503,6 +503,7 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 	align = le32_to_cpu(pfn_sb->align);
 	offset = le64_to_cpu(pfn_sb->dataoff);
 	start_pad = le32_to_cpu(pfn_sb->start_pad);
+	end_trunc = le32_to_cpu(pfn_sb->end_trunc);
 	if (align == 0)
 		align = 1UL << ilog2(offset);
 	mode = le32_to_cpu(pfn_sb->mode);
@@ -610,6 +611,10 @@ int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig)
 		return -EOPNOTSUPP;
 	}
 
Quoted 5 lines

drivers/nvdimm/pfn_devs.c int nd_pfn_validate(struct nd_pfn *nd_pfn, const char *sig) · line 614

Quoted 1 line
drivers/nvdimm/pfn_devs.c +1 -0
+	if (offset >= (res->end - res->start + 1 - start_pad - end_trunc)) {
                       ^^^^^^^^^^^^^^^^^^^^^^^^^ That's what

resource_size(res) does. It might be better to create a local variable 'size' to hold that, as there are now two instances of that in the function.

Will update.

Quoted 20 lines
Quoted 8 lines
drivers/nvdimm/pfn_devs.c +3 -0
+		dev_err(&nd_pfn->dev, "bad offset with small namespace\n");
+		return -EOPNOTSUPP;
+	}
 	return 0;
 }
 EXPORT_SYMBOL(nd_pfn_validate);

drivers/nvdimm/pfn_devs.c static int nd_pfn_init(struct nd_pfn *nd_pfn) · line 822

Quoted 9 lines
drivers/nvdimm/pfn_devs.c +2 -1
@@ -810,7 +815,8 @@ static int nd_pfn_init(struct nd_pfn *nd_pfn)
 	else
 		return -ENXIO;
 
-	if (offset >= size) {
+	if (offset >= (size - end_trunc)) {
+		/* This implies we result in zero size devices */
 		dev_err(&nd_pfn->dev, "%s unable to satisfy requested alignment\n",
 				dev_name(&ndns->dev));
 		return -ENXIO;

Functionally, this looks good to me.

Quoted 3 lines

Cheers,
Jeff

Thanks for reviewing the patch.
-aneesh