mm: kfence: export kfence_enabled as global variables

PATCH In review Zhenhua Huang 4 participants 7 messages Feb 7, 2023

Subsystems (4)

Memory management arm64 DMA buffer sharing framework Kfence

lore

Share & tools

lore

Zhenhua Huang

author

2023-02-07 06:15 UTC #0 lore raw

PATCH 1/1 mm: kfence: export kfence_enabled as global variables

Export the variable to ease the judgement of whether kfence enabled at runtime. It should be more precise than through kernel config "CONFIG_KFENCE".

For example We can disable kfence at runtime using bootargs "kfence.sample_interval=0" but CONFIG_KFENCE enabled. It was false positive.

Signed-off-by Zhenhua Huang <[email protected]>

---
 arch/arm64/mm/pageattr.c | 3 ++-
 include/linux/kfence.h   | 2 ++
 mm/kfence/core.c         | 2 +-
 3 files changed, 5 insertions(+), 2 deletions(-)

arch/arm64/mm/pageattr.c +2 -1

index 79dd201..208d780 100644
@@ -7,6 +7,7 @@
 #include <linux/module.h>
 #include <linux/sched.h>
 #include <linux/vmalloc.h>
+#include <linux/kfence.h>
 
 #include <asm/cacheflush.h>
 #include <asm/set_memory.h>

@@ -27,7 +28,7 @@ bool can_set_direct_map(void)
 	 * protect/unprotect single pages.
 	 */
 	return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() ||
-		IS_ENABLED(CONFIG_KFENCE);
+		kfence_enabled;
 }
 
 static int change_page_range(pte_t *ptep, unsigned long addr, void *data)

include/linux/kfence.h +2 -0

index 726857a..9fac824 100644
@@ -26,6 +26,7 @@ extern unsigned long kfence_sample_interval;
  */
 #define KFENCE_POOL_SIZE ((CONFIG_KFENCE_NUM_OBJECTS + 1) * 2 * PAGE_SIZE)
 extern char *__kfence_pool;
+extern bool kfence_enabled;
 
 DECLARE_STATIC_KEY_FALSE(kfence_allocation_key);
 extern atomic_t kfence_allocation_gate;

@@ -222,6 +223,7 @@ bool __kfence_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *sla
 
 #else /* CONFIG_KFENCE */
 
+#define kfence_enabled 0
 static inline bool is_kfence_address(const void *addr) { return false; }
 static inline void kfence_alloc_pool(void) { }
 static inline void kfence_init(void) { }

mm/kfence/core.c +1 -1

index 5349c37..e5bf93d 100644
@@ -48,7 +48,7 @@
 
 /* === Data ================================================================= */
 
-static bool kfence_enabled __read_mostly;
+bool kfence_enabled __read_mostly;
 static bool disabled_by_warn __read_mostly;
 
 unsigned long kfence_sample_interval __read_mostly = CONFIG_KFENCE_SAMPLE_INTERVAL;

Signature · 1 line

2.7.4

Marco Elver

maintainer

2023-02-07 07:19 UTC #1 lore raw

Quoted 29 lines · Zhenhua Huang

Export the variable to ease the judgement of whether kfence enabled at runtime. It should be more precise than through kernel config "CONFIG_KFENCE".
For example We can disable kfence at runtime using bootargs "kfence.sample_interval=0" but CONFIG_KFENCE enabled. It was false positive.
Signed-off-by Zhenhua Huang <[email protected]>
---
 arch/arm64/mm/pageattr.c | 3 ++-
 include/linux/kfence.h   | 2 ++
 mm/kfence/core.c         | 2 +-
 3 files changed, 5 insertions(+), 2 deletions(-)
arch/arm64/mm/pageattr.c +1 -0
index 79dd201..208d780 100644
@@ -7,6 +7,7 @@
 #include <linux/module.h>
 #include <linux/sched.h>
 #include <linux/vmalloc.h>
+#include <linux/kfence.h>
 
 #include <asm/cacheflush.h>
 #include <asm/set_memory.h>

arch/arm64/mm/pageattr.c bool can_set_direct_map(void) · line 31

Quoted 5 lines

arch/arm64/mm/pageattr.c +1 -1

@@ -27,7 +28,7 @@ bool can_set_direct_map(void)
         * protect/unprotect single pages.
         */
        return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() ||
-               IS_ENABLED(CONFIG_KFENCE);
+               kfence_enabled;

Unfortunately this won't work, because it's possible to enable KFENCE after the kernel has booted with e.g.: echo 100 > /sys/module/kfence/parameters/sample_interval

What is the problem you have encountered? Is the page-granular direct map causing issues?

Zhenhua Huang

author

2023-02-07 07:47 UTC #2 lore raw

Thanks Marco!

Quoted 27 lines · Marco Elver

Quoted 26 lines · Zhenhua Huang
Export the variable to ease the judgement of whether kfence enabled at runtime. It should be more precise than through kernel config "CONFIG_KFENCE".
For example We can disable kfence at runtime using bootargs "kfence.sample_interval=0" but CONFIG_KFENCE enabled. It was false positive.
Signed-off-by Zhenhua Huang <[email protected]>
---
  arch/arm64/mm/pageattr.c | 3 ++-
  include/linux/kfence.h   | 2 ++
  mm/kfence/core.c         | 2 +-
  3 files changed, 5 insertions(+), 2 deletions(-)
arch/arm64/mm/pageattr.c +1 -0
index 79dd201..208d780 100644
@@ -7,6 +7,7 @@
  #include <linux/module.h>
  #include <linux/sched.h>
  #include <linux/vmalloc.h>
+#include <linux/kfence.h>
 

Quoted 12 lines

arch/arm64/mm/pageattr.c bool can_set_direct_map(void) · line 31
Quoted 8 lines
arch/arm64/mm/pageattr.c +1 -1
  #include <asm/cacheflush.h>
  #include <asm/set_memory.h>
@@ -27,7 +28,7 @@ bool can_set_direct_map(void)
          * protect/unprotect single pages.
          */
         return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() ||
-               IS_ENABLED(CONFIG_KFENCE);
+               kfence_enabled;
Unfortunately this won't work, because it's possible to enable KFENCE after the kernel has booted with e.g.: echo 100 > /sys/module/kfence/parameters/sample_interval

Yeah, got it. Thanks for catching it.

Quoted 3 lines

What is the problem you have encountered? Is the page-granular direct map causing issues?

We're working on a low memory target, page-granular mapping costed more (2M per 1GB) memory. Due to GKI constraints, it is not easy to disable CONFIG_KFENCE. So my intention was to move the judgement to runtime configurable w/ CONFIG_KFENCE on...

Do you have any further suggestion/proposal on this? Many Thanks!

Thanks,
Zhenhua

Alexander Potapenko

maintainer

2023-02-07 10:29 UTC #3 lore raw

Quoted 9 lines

Quoted 3 lines
What is the problem you have encountered? Is the page-granular direct map causing issues?
We're working on a low memory target, page-granular mapping costed more (2M per 1GB) memory. Due to GKI constraints, it is not easy to disable CONFIG_KFENCE. So my intention was to move the judgement to runtime configurable w/ CONFIG_KFENCE on...
Do you have any further suggestion/proposal on this? Many Thanks!

Right now CONFIG_KFENCE allocates 512Kb for the GKI kernel (only 63 objects instead of the default 255):
https://cs.android.com/android/kernel/superproject/+/common-android-mainline:common/arch/arm64/configs/gki_defconfig;l=686?q=CONFIG_KFENCE_NUM_OBJECTS
Where do 2M come from?

Zhenhua Huang

author

2023-02-07 10:43 UTC #4 lore raw

Quoted 8 lines · Alexander Potapenko

Quoted 8 lines
Quoted 3 lines
What is the problem you have encountered? Is the page-granular direct map causing issues?
We're working on a low memory target, page-granular mapping costed more (2M per 1GB) memory. Due to GKI constraints, it is not easy to disable CONFIG_KFENCE. So my intention was to move the judgement to runtime configurable w/ CONFIG_KFENCE on...

Quoted 6 lines

Quoted 1 line
Do you have any further suggestion/proposal on this? Many Thanks!
Right now CONFIG_KFENCE allocates 512Kb for the GKI kernel (only 63 objects instead of the default 255):
https://cs.android.com/android/kernel/superproject/+/common-android-mainline:common/arch/arm64/configs/gki_defconfig;l=686?q=CONFIG_KFENCE_NUM_OBJECTS
Where do 2M come from?

Hi Alexander,
It is not from KFENCE cost itself, but from the requirement it needs page-granular mapping.
2M per 1GB cost is from linear mapping for page granularity. You can see can_set_direct_map() is used for whether allowing section mapping or not.

Mark Rutland

reviewer

2023-02-07 10:48 UTC #5 lore raw

In reply to Zhenhua Huang

Quoted 44 lines · Zhenhua Huang

Thanks Marco!
Quoted 27 lines · Marco Elver
Quoted 26 lines · Zhenhua Huang
Export the variable to ease the judgement of whether kfence enabled at runtime. It should be more precise than through kernel config "CONFIG_KFENCE".
For example We can disable kfence at runtime using bootargs "kfence.sample_interval=0" but CONFIG_KFENCE enabled. It was false positive.
Signed-off-by Zhenhua Huang <[email protected]>
---
  arch/arm64/mm/pageattr.c | 3 ++-
  include/linux/kfence.h   | 2 ++
  mm/kfence/core.c         | 2 +-
  3 files changed, 5 insertions(+), 2 deletions(-)
arch/arm64/mm/pageattr.c +1 -0
index 79dd201..208d780 100644
@@ -7,6 +7,7 @@
  #include <linux/module.h>
  #include <linux/sched.h>
  #include <linux/vmalloc.h>
+#include <linux/kfence.h>
 
Quoted 12 lines
arch/arm64/mm/pageattr.c bool can_set_direct_map(void) · line 31
Quoted 8 lines
arch/arm64/mm/pageattr.c +1 -1
  #include <asm/cacheflush.h>
  #include <asm/set_memory.h>
@@ -27,7 +28,7 @@ bool can_set_direct_map(void)
          * protect/unprotect single pages.
          */
         return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() ||
-               IS_ENABLED(CONFIG_KFENCE);
+               kfence_enabled;
Unfortunately this won't work, because it's possible to enable KFENCE after the kernel has booted with e.g.: echo 100 > /sys/module/kfence/parameters/sample_interval
Yeah, got it. Thanks for catching it.
Quoted 1 line

Quoted 8 lines

Quoted 2 lines
What is the problem you have encountered? Is the page-granular direct map causing issues?
We're working on a low memory target, page-granular mapping costed more (2M per 1GB) memory. Due to GKI constraints, it is not easy to disable CONFIG_KFENCE. So my intention was to move the judgement to runtime configurable w/ CONFIG_KFENCE on...
Do you have any further suggestion/proposal on this? Many Thanks!

Just to check, the cost is because we're mapping *all* of memory at page granulatrity, right? If we were to just map the KFENCE region a page granularity, would that be a sufficient saving?

We didn't do that so far because it was simpler to just map everything at page granularity (and that's also required by rodata_full, which I though android used?).

If it's really important (and rodata_full isn't being used), we could try to do that.

To do that we'd need to choose the KFENCE region *before* arm64 creates the fine-grain translation tables, which probable needs an arch_ hook.

Thanks,
Mark.

Zhenhua Huang

author

2023-02-07 12:36 UTC #6 lore raw

Thanks Mark!

Quoted 52 lines · Mark Rutland

Quoted 51 lines · Zhenhua Huang
Thanks Marco!
Quoted 27 lines · Marco Elver
Quoted 26 lines · Zhenhua Huang
Export the variable to ease the judgement of whether kfence enabled at runtime. It should be more precise than through kernel config "CONFIG_KFENCE".
For example We can disable kfence at runtime using bootargs "kfence.sample_interval=0" but CONFIG_KFENCE enabled. It was false positive.
Signed-off-by Zhenhua Huang <[email protected]>
---
   arch/arm64/mm/pageattr.c | 3 ++-
   include/linux/kfence.h   | 2 ++
   mm/kfence/core.c         | 2 +-
   3 files changed, 5 insertions(+), 2 deletions(-)
arch/arm64/mm/pageattr.c +1 -0
index 79dd201..208d780 100644
@@ -7,6 +7,7 @@
   #include <linux/module.h>
   #include <linux/sched.h>
   #include <linux/vmalloc.h>
+#include <linux/kfence.h>
 
Quoted 12 lines
arch/arm64/mm/pageattr.c bool can_set_direct_map(void) · line 31
Quoted 8 lines
arch/arm64/mm/pageattr.c +1 -1
   #include <asm/cacheflush.h>
   #include <asm/set_memory.h>
@@ -27,7 +28,7 @@ bool can_set_direct_map(void)
           * protect/unprotect single pages.
           */
          return (rodata_enabled && rodata_full) || debug_pagealloc_enabled() ||
-               IS_ENABLED(CONFIG_KFENCE);
+               kfence_enabled;
Unfortunately this won't work, because it's possible to enable KFENCE after the kernel has booted with e.g.: echo 100 > /sys/module/kfence/parameters/sample_interval
Yeah, got it. Thanks for catching it.
Quoted 3 lines
What is the problem you have encountered? Is the page-granular direct map causing issues?
We're working on a low memory target, page-granular mapping costed more (2M per 1GB) memory. Due to GKI constraints, it is not easy to disable CONFIG_KFENCE. So my intention was to move the judgement to runtime configurable w/ CONFIG_KFENCE on...

Quoted 5 lines

Quoted 1 line
Do you have any further suggestion/proposal on this? Many Thanks!
Just to check, the cost is because we're mapping *all* of memory at page granulatrity, right? If we were to just map the KFENCE region a page granularity, would that be a sufficient saving?

Yes,that's expected saving which I would like to address here.

Quoted 4 lines

We didn't do that so far because it was simpler to just map everything at page granularity (and that's also required by rodata_full, which I though android used?).

Yes, by default we're setting rodata_full. While for low-ram target, we also can lower the security requirement. This is why I am pursuing.. BTW, rodata_full can be easily runtime configured through bootargs.

Quoted 6 lines

If it's really important (and rodata_full isn't being used), we could try to do that.
To do that we'd need to choose the KFENCE region *before* arm64 creates the fine-grain translation tables, which probable needs an arch_ hook.

It sounds a good solution. Let me also dig into this. Many Thanks!

Quoted 4 lines

Thanks,
Mark.