arm64/mm: unmap the linear alias of module allocations

When CONFIG_STRICT_KERNEL_RWX=y [which is the default on arm64], we take great care to ensure that the mappings of modules in the vmalloc space are locked down as much as possible, i.e., executable code is mapped read-only, read-only data is mapped read-only and non-executable, and read-write data is mapped non-executable as well.

However, due to the way we map the linear region [aka the kernel direct mapping], those module regions are aliased by read-write mappings, and it is possible for an attacker to modify code or data that was assumed to be immutable in this configuration.

So let's ensure that the linear alias of module memory is remapped read-only upon allocation and remapped read-write when it is freed. The latter requires some special handling involving a workqueue due to the fact that it may be called in softirq context at which time calling find_vm_area() is unsafe.

Note that this requires the entire linear region to be mapped down to pages, which may result in a performance regression in some hardware configurations.

Signed-off-by Ard Biesheuvel <[email protected]>

---
Changes since RFC/v1:
- remap linear alias read-only rather than invalid
- honour rodata_enabled, i.e., 'rodata=off' will disable this functionality
- use VMA nr_pages field rather than size to obtain the number of pages to remap

 arch/arm64/include/asm/module.h |  6 +++
 arch/arm64/kernel/module.c      | 42 ++++++++++++++++++++
 arch/arm64/mm/mmu.c             |  2 +-
 arch/arm64/mm/pageattr.c        | 19 +++++++++
 4 files changed, 68 insertions(+), 1 deletion(-)

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
               a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +24 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+       module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+       WARN_ON(!module_free_wq);
+
+       return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+       remap_linear_module_alias(work, false);
+       vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+       struct work_struct *work;
+

arch/arm64/kernel/module.c +8 -0

+       work = module_region;
+       INIT_WORK(work, module_free_wq_worker);
+       queue_work(module_free_wq, work);
+}
+
+#endif
+
 void *module_alloc(unsigned long size)
 {
        gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
                return NULL;
        }
 
+       remap_linear_module_alias(p, true);
        return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644

arch/arm64/mm/mmu.c +0 -0

        /*

arch/arm64/mm/pageattr.c +19 -1

index a56359373d8b..cc04be572660 100644
@@ -138,6 +138,25 @@ int set_memory_valid(unsigned long addr, int numpages, int enable)
                                        __pgprot(PTE_VALID));
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro)
+{
+       struct vm_struct *vm;
+       int i;
+
+       if (!rodata_enabled)
+               return;
+
+       vm = find_vm_area(module_region);
+       WARN_ON(!vm || !vm->pages);
+
+       for (i = 0; i < vm->nr_pages; i++)
+               __change_memory_common((u64)page_address(vm->pages[i]), PAGE_SIZE,
+                               ro ? __pgprot(PTE_RDONLY) : __pgprot(PTE_WRITE),
+                               ro ? __pgprot(PTE_WRITE) : __pgprot(PTE_RDONLY));
+}
+#endif
+
 #ifdef CONFIG_DEBUG_PAGEALLOC
 void __kernel_map_pages(struct page *page, int numpages, int enable)
 {
--

2.17.1

When CONFIG_STRICT_KERNEL_RWX=y [which is the default on arm64], we take great care to ensure that the mappings of modules in the vmalloc space are locked down as much as possible, i.e., executable code is mapped read-only, read-only data is mapped read-only and non-executable, and read-write data is mapped non-executable as well.

However, due to the way we map the linear region [aka the kernel direct mapping], those module regions are aliased by read-write mappings, and it is possible for an attacker to modify code or data that was assumed to be immutable in this configuration.

So let's ensure that the linear alias of module memory is remapped read-only upon allocation and remapped read-write when it is freed. The latter requires some special handling involving a workqueue due to the fact that it may be called in softirq context at which time calling find_vm_area() is unsafe.

Note that this requires the entire linear region to be mapped down to pages, which may result in a performance regression in some hardware configurations.

Signed-off-by Ard Biesheuvel <[email protected]>

---
Changes since RFC/v1:
- remap linear alias read-only rather than invalid
- honour rodata_enabled, i.e., 'rodata=off' will disable this functionality
- use VMA nr_pages field rather than size to obtain the number of pages to remap

  arch/arm64/include/asm/module.h |  6 +++
  arch/arm64/kernel/module.c      | 42 ++++++++++++++++++++
  arch/arm64/mm/mmu.c             |  2 +-
  arch/arm64/mm/pageattr.c        | 19 +++++++++
  4 files changed, 68 insertions(+), 1 deletion(-)

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
  	       a->mov2 == b->mov2;
  }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
  #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +42 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
  #include <linux/mm.h>
  #include <linux/moduleloader.h>
  #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
  #include <asm/alternative.h>
  #include <asm/insn.h>
  #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+	module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+	WARN_ON(!module_free_wq);
+
+	return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+	remap_linear_module_alias(work, false);
+	vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+	struct work_struct *work;
+
+	if (!module_region)
+		return;
+
+	/*
+	 * At this point, module_region is a pointer to an allocation of at
+	 * least PAGE_SIZE bytes that is mapped read-write. So instead of
+	 * allocating memory for a data structure containing a work_struct
+	 * instance and a copy of the value of module_region, just reuse the
+	 * allocation directly.
+	 */
+	work = module_region;
+	INIT_WORK(work, module_free_wq_worker);
+	queue_work(module_free_wq, work);
+}
+
+#endif
+
  void *module_alloc(unsigned long size)
  {
  	gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
  		return NULL;
  	}
 
+	remap_linear_module_alias(p, true);
  	return p;
  }
 

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

index 493ff75670ff..5492b691aafd 100644
@@ -432,7 +432,7 @@ static void __init map_mem(pgd_t *pgdp)
  	struct memblock_region *reg;
  	int flags = 0;
 
-	if (debug_pagealloc_enabled())
+	if (rodata_enabled || debug_pagealloc_enabled())
  		flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
 
  	/*

arch/arm64/mm/pageattr.c +19 -0

index a56359373d8b..cc04be572660 100644
@@ -138,6 +138,25 @@ int set_memory_valid(unsigned long addr, int numpages, int enable)
  					__pgprot(PTE_VALID));
  }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro)
+{
+	struct vm_struct *vm;
+	int i;
+
+	if (!rodata_enabled)
+		return;
+
+	vm = find_vm_area(module_region);
+	WARN_ON(!vm || !vm->pages);
+
+	for (i = 0; i < vm->nr_pages; i++)
+		__change_memory_common((u64)page_address(vm->pages[i]), PAGE_SIZE,
+				ro ? __pgprot(PTE_RDONLY) : __pgprot(PTE_WRITE),
+				ro ? __pgprot(PTE_WRITE) : __pgprot(PTE_RDONLY));
+}
+#endif
+
  #ifdef CONFIG_DEBUG_PAGEALLOC
  void __kernel_map_pages(struct page *page, int numpages, int enable)
  {

Quoted 18 lines · Ard Biesheuvel

When CONFIG_STRICT_KERNEL_RWX=y [which is the default on arm64], we take great care to ensure that the mappings of modules in the vmalloc space are locked down as much as possible, i.e., executable code is mapped read-only, read-only data is mapped read-only and non-executable, and read-write data is mapped non-executable as well.

However, due to the way we map the linear region [aka the kernel direct mapping], those module regions are aliased by read-write mappings, and it is possible for an attacker to modify code or data that was assumed to be immutable in this configuration.

So let's ensure that the linear alias of module memory is remapped read-only upon allocation and remapped read-write when it is freed. The latter requires some special handling involving a workqueue due to the fact that it may be called in softirq context at which time calling find_vm_area() is unsafe.

Quoted 4 lines

Note that this requires the entire linear region to be mapped down to pages, which may result in a performance regression in some hardware configurations.

I'm concerned any performance regression might just make people turn off rodata all together. It's not overly bad for the kernel mappings but modules are fully RWX so we might be trading a subtle threat for a more obvious one. Would putting this behind a Kconfig make sense for those who want the status quo (with a big caveat that this should have been present from the start)?

When CONFIG_STRICT_KERNEL_RWX=y [which is the default on arm64], we take great care to ensure that the mappings of modules in the vmalloc space are locked down as much as possible, i.e., executable code is mapped read-only, read-only data is mapped read-only and non-executable, and read-write data is mapped non-executable as well.

However, due to the way we map the linear region [aka the kernel direct mapping], those module regions are aliased by read-write mappings, and it is possible for an attacker to modify code or data that was assumed to be immutable in this configuration.

So let's ensure that the linear alias of module memory is remapped read-only upon allocation and remapped read-write when it is freed. The latter requires some special handling involving a workqueue due to the fact that it may be called in softirq context at which time calling find_vm_area() is unsafe.

Note that this requires the entire linear region to be mapped down to pages, which may result in a performance regression in some hardware configurations.

Signed-off-by Ard Biesheuvel <[email protected]>

---
Changes since RFC/v1:
- remap linear alias read-only rather than invalid
- honour rodata_enabled, i.e., 'rodata=off' will disable this functionality
- use VMA nr_pages field rather than size to obtain the number of pages to remap

 arch/arm64/include/asm/module.h |  6 +++
 arch/arm64/kernel/module.c      | 42 ++++++++++++++++++++
 arch/arm64/mm/mmu.c             |  2 +-
 arch/arm64/mm/pageattr.c        | 19 +++++++++
 4 files changed, 68 insertions(+), 1 deletion(-)

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
 	       a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +28 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+	module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+	WARN_ON(!module_free_wq);
+
+	return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+	remap_linear_module_alias(work, false);
+	vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+	struct work_struct *work;
+
+	if (!module_region)
+		return;
+
+	/*

arch/arm64/kernel/module.c +3 -0

+#endif
+
 void *module_alloc(unsigned long size)
 {
 	gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
 		return NULL;
 	}
 
+	remap_linear_module_alias(p, true);
 	return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644

arch/arm64/mm/pageattr.c +7 -0

+void remap_linear_module_alias(void *module_region, bool ro)
+{
+	struct vm_struct *vm;
+	int i;
+
+	if (!rodata_enabled)
+		return;

Hi Ard,

Remember writing this patch? ;)

Quoted 97 lines · Ard Biesheuvel

When CONFIG_STRICT_KERNEL_RWX=y [which is the default on arm64], we take great care to ensure that the mappings of modules in the vmalloc space are locked down as much as possible, i.e., executable code is mapped read-only, read-only data is mapped read-only and non-executable, and read-write data is mapped non-executable as well.

However, due to the way we map the linear region [aka the kernel direct mapping], those module regions are aliased by read-write mappings, and it is possible for an attacker to modify code or data that was assumed to be immutable in this configuration.

So let's ensure that the linear alias of module memory is remapped read-only upon allocation and remapped read-write when it is freed. The latter requires some special handling involving a workqueue due to the fact that it may be called in softirq context at which time calling find_vm_area() is unsafe.

Note that this requires the entire linear region to be mapped down to pages, which may result in a performance regression in some hardware configurations.

Signed-off-by Ard Biesheuvel <[email protected]>

---
Changes since RFC/v1:
- remap linear alias read-only rather than invalid
- honour rodata_enabled, i.e., 'rodata=off' will disable this functionality
- use VMA nr_pages field rather than size to obtain the number of pages to remap

 arch/arm64/include/asm/module.h |  6 +++
 arch/arm64/kernel/module.c      | 42 ++++++++++++++++++++
 arch/arm64/mm/mmu.c             |  2 +-
 arch/arm64/mm/pageattr.c        | 19 +++++++++
 4 files changed, 68 insertions(+), 1 deletion(-)

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
             a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +35 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+     module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+     WARN_ON(!module_free_wq);
+
+     return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+     remap_linear_module_alias(work, false);
+     vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+     struct work_struct *work;
+
+     if (!module_region)
+             return;
+
+     /*
+      * At this point, module_region is a pointer to an allocation of at
+      * least PAGE_SIZE bytes that is mapped read-write. So instead of
+      * allocating memory for a data structure containing a work_struct
+      * instance and a copy of the value of module_region, just reuse the
+      * allocation directly.
+      */
+     work = module_region;

arch/arm64/kernel/module.c line 70

Quoted 3 lines

arch/arm64/kernel/module.c +3 -0

+     INIT_WORK(work, module_free_wq_worker);
+     queue_work(module_free_wq, work);
+}

Whilst I can see that this works, it's a bit grotty because we're duplicating much of the hoop-jumping that vfree() also has to go through. Given that we allocate the module memory in the arch code, could we grab the pages at alloc time and stash them in the mod_arch_specific field?

Quoted 21 lines

arch/arm64/kernel/module.c +3 -0

+#endif
+
 void *module_alloc(unsigned long size)
 {
      gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
              return NULL;
      }
 
+     remap_linear_module_alias(p, true);
      return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644
@@ -432,7 +432,7 @@ static void __init map_mem(pgd_t *pgdp)
      struct memblock_region *reg;
      int flags = 0;
 

arch/arm64/mm/mmu.c static void __init map_mem(pgd_t *pgdp) · line 436

Quoted 3 lines

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

-     if (debug_pagealloc_enabled())
+     if (rodata_enabled || debug_pagealloc_enabled())
              flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;

I suppose this is the most contentious part of the patch. Perhaps a better solution would be to tweak CONFIG_STRICT_MODULE_RWX so that it's structured more like CONFIG_STACKPROTECTOR. That way, your patch here could be the "strong" version, which comes with more of a performance cost.

Quoted 8 lines

arch/arm64/mm/pageattr.c +0 -0

index a56359373d8b..cc04be572660 100644
@@ -138,6 +138,25 @@ int set_memory_valid(unsigned long addr, int numpages, int enable)
                                      __pgprot(PTE_VALID));
 }
 

arch/arm64/mm/pageattr.c int set_memory_valid(unsigned long addr, int numpages, int enable) · line 141

Quoted 1 line

arch/arm64/mm/pageattr.c +1 -0

+#ifdef CONFIG_STRICT_KERNEL_RWX

CONFIG_STRICT_MODULE_RWX instead?

Quoted 7 lines

arch/arm64/mm/pageattr.c +7 -0

+void remap_linear_module_alias(void *module_region, bool ro)
+{
+     struct vm_struct *vm;
+     int i;
+
+     if (!rodata_enabled)
+             return;

arch/arm64/mm/pageattr.c int set_memory_valid(unsigned long addr, int numpages, int enable) · line 158

Quoted 10 lines

arch/arm64/mm/pageattr.c +10 -0

+
+     vm = find_vm_area(module_region);
+     WARN_ON(!vm || !vm->pages);
+
+     for (i = 0; i < vm->nr_pages; i++)
+             __change_memory_common((u64)page_address(vm->pages[i]), PAGE_SIZE,
+                             ro ? __pgprot(PTE_RDONLY) : __pgprot(PTE_WRITE),
+                             ro ? __pgprot(PTE_WRITE) : __pgprot(PTE_RDONLY));
+}
+#endif

I think it would be cleaner just to have two wrapper functions for this:

remap_linear_module_alias_ro
remap_linear_module_alias_rw

so we can drop the bool parameter from the API and just pass the set/clear attributes directly to something like __remap_linear_module_alias.

Quoted 63 lines · Will Deacon

Quoted 62 lines · Ard Biesheuvel

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
             a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +35 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+     module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+     WARN_ON(!module_free_wq);
+
+     return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+     remap_linear_module_alias(work, false);
+     vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+     struct work_struct *work;
+
+     if (!module_region)
+             return;
+
+     /*
+      * At this point, module_region is a pointer to an allocation of at
+      * least PAGE_SIZE bytes that is mapped read-write. So instead of
+      * allocating memory for a data structure containing a work_struct
+      * instance and a copy of the value of module_region, just reuse the
+      * allocation directly.
+      */
+     work = module_region;

Quoted 9 lines

arch/arm64/kernel/module.c line 70

Quoted 3 lines

arch/arm64/kernel/module.c +3 -0

+     INIT_WORK(work, module_free_wq_worker);
+     queue_work(module_free_wq, work);
+}

Whilst I can see that this works, it's a bit grotty because we're duplicating much of the hoop-jumping that vfree() also has to go through. Given that we allocate the module memory in the arch code, could we grab the pages at alloc time and stash them in the mod_arch_specific field?

Yeah that should work

Quoted 21 lines

Quoted 21 lines

arch/arm64/kernel/module.c +3 -0

+#endif
+
 void *module_alloc(unsigned long size)
 {
      gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
              return NULL;
      }
 
+     remap_linear_module_alias(p, true);
      return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644
@@ -432,7 +432,7 @@ static void __init map_mem(pgd_t *pgdp)
      struct memblock_region *reg;
      int flags = 0;
 

Quoted 10 lines

arch/arm64/mm/mmu.c static void __init map_mem(pgd_t *pgdp) · line 436

Quoted 3 lines

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

-     if (debug_pagealloc_enabled())
+     if (rodata_enabled || debug_pagealloc_enabled())
              flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;

I suppose this is the most contentious part of the patch. Perhaps a better solution would be to tweak CONFIG_STRICT_MODULE_RWX so that it's structured more like CONFIG_STACKPROTECTOR. That way, your patch here could be the "strong" version, which comes with more of a performance cost.

I'd like to understand whether this performance hit is theoretical or not. It is obviously less efficient to map the linear region at page granularity, but which workloads will actually notice the difference?

Also, I think this should be a runtime setting, perhaps rodata=full? With a CONFIG_ option to set the default.

Quoted 74 lines · Ard Biesheuvel

Quoted 72 lines · Will Deacon

Quoted 55 lines · Ard Biesheuvel

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
             a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +28 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+     module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+     WARN_ON(!module_free_wq);
+
+     return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+     remap_linear_module_alias(work, false);
+     vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+     struct work_struct *work;
+
+     if (!module_region)
+             return;
+
+     /*

arch/arm64/kernel/module.c line 70

Quoted 10 lines

arch/arm64/kernel/module.c +10 -0

+      * At this point, module_region is a pointer to an allocation of at
+      * least PAGE_SIZE bytes that is mapped read-write. So instead of
+      * allocating memory for a data structure containing a work_struct
+      * instance and a copy of the value of module_region, just reuse the
+      * allocation directly.
+      */
+     work = module_region;
+     INIT_WORK(work, module_free_wq_worker);
+     queue_work(module_free_wq, work);
+}

Whilst I can see that this works, it's a bit grotty because we're duplicating much of the hoop-jumping that vfree() also has to go through. Given that we allocate the module memory in the arch code, could we grab the pages at alloc time and stash them in the mod_arch_specific field?

Quoted 1 line

Yeah that should work

Great!

Quoted 36 lines

Quoted 21 lines

Quoted 21 lines

arch/arm64/kernel/module.c +3 -0

+#endif
+
 void *module_alloc(unsigned long size)
 {
      gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
              return NULL;
      }
 
+     remap_linear_module_alias(p, true);
      return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644
@@ -432,7 +432,7 @@ static void __init map_mem(pgd_t *pgdp)
      struct memblock_region *reg;
      int flags = 0;
 

Quoted 10 lines

arch/arm64/mm/mmu.c static void __init map_mem(pgd_t *pgdp) · line 436

Quoted 3 lines

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

-     if (debug_pagealloc_enabled())
+     if (rodata_enabled || debug_pagealloc_enabled())
              flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;

I suppose this is the most contentious part of the patch. Perhaps a better solution would be to tweak CONFIG_STRICT_MODULE_RWX so that it's structured more like CONFIG_STACKPROTECTOR. That way, your patch here could be the "strong" version, which comes with more of a performance cost.

I'd like to understand whether this performance hit is theoretical or not. It is obviously less efficient to map the linear region at page granularity, but which workloads will actually notice the difference?

Quoted 2 lines

Also, I think this should be a runtime setting, perhaps rodata=full? With a CONFIG_ option to set the default.

Yes, that's a better idea. We could even default to "full" if we don't have any performance data (if people start shouting then we'll know it's an issue!).

Do you plan to spin a v2?

Quoted 75 lines · Will Deacon

Quoted 74 lines · Ard Biesheuvel

Quoted 72 lines · Will Deacon

Quoted 55 lines · Ard Biesheuvel

arch/arm64/include/asm/module.h +6 -0

index 97d0ef12e2ff..f16f0073642b 100644
@@ -91,4 +91,10 @@ static inline bool plt_entries_equal(const struct plt_entry *a,
             a->mov2 == b->mov2;
 }
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+void remap_linear_module_alias(void *module_region, bool ro);
+#else
+static inline void remap_linear_module_alias(void *module_region, bool ro) {}
+#endif
+
 #endif /* __ASM_MODULE_H */

arch/arm64/kernel/module.c +28 -0

index f0f27aeefb73..2b61ca0285eb 100644
@@ -26,10 +26,51 @@
 #include <linux/mm.h>
 #include <linux/moduleloader.h>
 #include <linux/vmalloc.h>
+#include <linux/workqueue.h>
 #include <asm/alternative.h>
 #include <asm/insn.h>
 #include <asm/sections.h>
 
+#ifdef CONFIG_STRICT_KERNEL_RWX
+
+static struct workqueue_struct *module_free_wq;
+
+static int init_workqueue(void)
+{
+     module_free_wq = alloc_ordered_workqueue("module_free_wq", 0);
+     WARN_ON(!module_free_wq);
+
+     return 0;
+}
+pure_initcall(init_workqueue);
+
+static void module_free_wq_worker(struct work_struct *work)
+{
+     remap_linear_module_alias(work, false);
+     vfree(work);
+}
+
+void module_memfree(void *module_region)
+{
+     struct work_struct *work;
+
+     if (!module_region)
+             return;
+
+     /*

arch/arm64/kernel/module.c line 70

Quoted 10 lines

arch/arm64/kernel/module.c +10 -0

+      * At this point, module_region is a pointer to an allocation of at
+      * least PAGE_SIZE bytes that is mapped read-write. So instead of
+      * allocating memory for a data structure containing a work_struct
+      * instance and a copy of the value of module_region, just reuse the
+      * allocation directly.
+      */
+     work = module_region;
+     INIT_WORK(work, module_free_wq_worker);
+     queue_work(module_free_wq, work);
+}

Whilst I can see that this works, it's a bit grotty because we're duplicating much of the hoop-jumping that vfree() also has to go through. Given that we allocate the module memory in the arch code, could we grab the pages at alloc time and stash them in the mod_arch_specific field?

Quoted 4 lines

Quoted 1 line

Yeah that should work

Great!

I may have spoken too soon: module_memfree() receives a void* pointer to the allocation to be freed, and we can't access the module struct from that.

But looking into this a bit more, I wonder why this doesn not affect other architectures as well, and it might make sense to solve this generically, or at least, provide the hooks in generic code.

Quoted 39 lines

Quoted 38 lines

Quoted 21 lines

Quoted 21 lines

arch/arm64/kernel/module.c +3 -0

+#endif
+
 void *module_alloc(unsigned long size)
 {
      gfp_t gfp_mask = GFP_KERNEL;

@@ -65,6 +106,7 @@ void *module_alloc(unsigned long size)
              return NULL;
      }
 
+     remap_linear_module_alias(p, true);
      return p;
 }
 

arch/arm64/mm/mmu.c +0 -0

index 493ff75670ff..5492b691aafd 100644
@@ -432,7 +432,7 @@ static void __init map_mem(pgd_t *pgdp)
      struct memblock_region *reg;
      int flags = 0;
 

Quoted 10 lines

arch/arm64/mm/mmu.c static void __init map_mem(pgd_t *pgdp) · line 436

Quoted 3 lines

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

-     if (debug_pagealloc_enabled())
+     if (rodata_enabled || debug_pagealloc_enabled())
              flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;

I suppose this is the most contentious part of the patch. Perhaps a better solution would be to tweak CONFIG_STRICT_MODULE_RWX so that it's structured more like CONFIG_STACKPROTECTOR. That way, your patch here could be the "strong" version, which comes with more of a performance cost.

I'd like to understand whether this performance hit is theoretical or not. It is obviously less efficient to map the linear region at page granularity, but which workloads will actually notice the difference?

Also, I think this should be a runtime setting, perhaps rodata=full? With a CONFIG_ option to set the default.

Quoted 6 lines

Yes, that's a better idea. We could even default to "full" if we don't have any performance data (if people start shouting then we'll know it's an issue!).

Do you plan to spin a v2?

Yes, but I'll do some digging first.