kdump: Allow removal of page tables for crashkernel memory

From: Michael Holzheu <[email protected]>

This patch introduces a mechanism that allows architecture backends to remove page tables for the crashkernel memory. This can protect the loaded kdump kernel from being overwritten by broken kernel code.
A new function crash_map_pages() is added that can be implemented by architecture code. This function has the following syntax:

void crash_map_pages(int enable);

"enable" can be 0 for removing or 1 for adding page tables. The function is called before and after the crashkernel segments are loaded. It is also called in crash_shrink_memory() to create new page tables when the crashkernel memory size is reduced.

To support architectures that have large pages this patch also introduces a new define KEXEC_CRASH_MEM_ALIGN. The crashkernel start and size must always be aligned with KEXEC_CRASH_MEM_ALIGN.

Signed-off-by Michael Holzheu <[email protected]>

---
 include/linux/kexec.h |    5 +++++
 kernel/kexec.c        |   16 ++++++++++++++--
 2 files changed, 19 insertions(+), 2 deletions(-)

include/linux/kexec.h +5 -0

@@ -37,6 +37,10 @@
 #define KEXEC_CRASH_CONTROL_MEMORY_LIMIT KEXEC_CONTROL_MEMORY_LIMIT
 #endif
 
+#ifndef KEXEC_CRASH_MEM_ALIGN
+#define KEXEC_CRASH_MEM_ALIGN PAGE_SIZE
+#endif
+
 #define KEXEC_NOTE_HEAD_BYTES ALIGN(sizeof(struct elf_note), 4)
 #define KEXEC_CORE_NOTE_NAME "CORE"
 #define KEXEC_CORE_NOTE_NAME_BYTES ALIGN(sizeof(KEXEC_CORE_NOTE_NAME), 4)

@@ -133,6 +137,7 @@ extern void crash_kexec(struct pt_regs *
 int kexec_should_crash(struct task_struct *);
 void crash_save_cpu(struct pt_regs *regs, int cpu);
 void crash_save_vmcoreinfo(void);
+void crash_map_pages(int enable);
 void arch_crash_save_vmcoreinfo(void);
 void vmcoreinfo_append_str(const char *fmt, ...)
 	__attribute__ ((format (printf, 1, 2)));

kernel/kexec.c +14 -2

@@ -999,6 +999,7 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 			kimage_free(xchg(&kexec_crash_image, NULL));
 			result = kimage_crash_alloc(&image, entry,
 						     nr_segments, segments);
+			crash_map_pages(1);
 		}
 		if (result)
 			goto out;

@@ -1015,6 +1016,8 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 				goto out;
 		}
 		kimage_terminate(image);
+		if (flags & KEXEC_ON_CRASH)
+			crash_map_pages(0);
 	}
 	/* Install the new kernel, and  Uninstall the old */
 	image = xchg(dest_image, image);

@@ -1026,6 +1029,13 @@ out:
 	return result;
 }
 
+/*
+ * provide an empty default implementation here -- architecture
+ * code may override this
+ */
+void __weak crash_map_pages(int enable)
+{}
+
 #ifdef CONFIG_COMPAT
 asmlinkage long compat_sys_kexec_load(unsigned long entry,
 				unsigned long nr_segments,

@@ -1134,14 +1144,16 @@ int crash_shrink_memory(unsigned long ne
 		goto unlock;
 	}
 
-	start = roundup(start, PAGE_SIZE);
-	end = roundup(start + new_size, PAGE_SIZE);
+	start = roundup(start, KEXEC_CRASH_MEM_ALIGN);
+	end = roundup(start + new_size, KEXEC_CRASH_MEM_ALIGN);
 
+	crash_map_pages(1);
 	crash_free_reserved_phys_range(end, crashk_res.end);
 
 	if ((start == end) && (crashk_res.parent != NULL))
 		release_resource(&crashk_res);
 	crashk_res.end = end - 1;
+	crash_map_pages(0);
 
 unlock:
 	mutex_unlock(&kexec_mutex);

Quoted 7 lines · Michael Holzheu

From: Michael Holzheu <[email protected]>

This patch introduces a mechanism that allows architecture backends to remove page tables for the crashkernel memory. This can protect the loaded kdump kernel from being overwritten by broken kernel code.
A new function crash_map_pages() is added that can be implemented by architecture code. This function has the following syntax:

I guess having separate functions for mapping and unmapping pages might look cleaner. Because we are not passing a page range, so specifying what pages we are talking about in function name might make it more clear.

crash_map_reserved_pages()
crash_unmap_reserved_pages().

Secondly, what happens to the code which runs after crash (crash_kexec()). Current x86 code assumes that reserved region is mapped at the time of crash and does few things with control page there.

So this generic approach is not valid atleast for x86, because it does not tackle the scenario about how to map reserved range again once kernel crashes. It will only work if there is assumption that after a crash, we don't expect reserved range/pages to be mapped.

Hello Vivek,

Quoted 9 lines · Vivek Goyal

Quoted 7 lines · Michael Holzheu

From: Michael Holzheu <[email protected]>

This patch introduces a mechanism that allows architecture backends to remove page tables for the crashkernel memory. This can protect the loaded kdump kernel from being overwritten by broken kernel code.
A new function crash_map_pages() is added that can be implemented by architecture code. This function has the following syntax:

Quoted 7 lines

I guess having separate functions for mapping and unmapping pages might look cleaner. Because we are not passing a page range, so specifying what pages we are talking about in function name might make it more clear.

crash_map_reserved_pages()
crash_unmap_reserved_pages().

Ok fine, no problem.

Quoted 3 lines

Secondly, what happens to the code which runs after crash (crash_kexec()). Current x86 code assumes that reserved region is mapped at the time of crash and does few things with control page there.

For s390, purgatory code can run in real mode. No page tables are required.

Quoted 4 lines

So this generic approach is not valid atleast for x86, because it does not tackle the scenario about how to map reserved range again once kernel crashes. It will only work if there is assumption that after a crash, we don't expect reserved range/pages to be mapped.

All architectures that support unmapping of crashkernel memory have to deal with this problem somehow. Either remap the crashkernel memory in machine_kexec() again or be able to run in real mode.

I adjusted that patch regarding your comment above. Will the following patch be ok for you?

---
Subject: kdump: Add infrastructure for unmapping crashkernel memory

From: Michael Holzheu <[email protected]>

This patch introduces a mechanism that allows architecture backends to remove page tables for the crashkernel memory. This can protect the loaded kdump kernel from being overwritten by broken kernel code. Two new functions crash_map_reserved_pages() and crash_unmap_reserved_pages() are added that can be implemented by architecture code. The crash_map_reserved_pages() function is called before and crash_unmap_reserved_pages() after the crashkernel segments are loaded. The functions are also called in crash_shrink_memory() to create/remove page tables when the crashkernel memory size is reduced.

To support architectures that have large pages this patch also introduces a new define KEXEC_CRASH_MEM_ALIGN. The crashkernel start and size must always be aligned with KEXEC_CRASH_MEM_ALIGN.

Signed-off-by Michael Holzheu <[email protected]>

---
 include/linux/kexec.h |    6 ++++++
 kernel/kexec.c        |   21 +++++++++++++++++++--
 2 files changed, 25 insertions(+), 2 deletions(-)

include/linux/kexec.h +6 -0

@@ -37,6 +37,10 @@
 #define KEXEC_CRASH_CONTROL_MEMORY_LIMIT KEXEC_CONTROL_MEMORY_LIMIT
 #endif
 
+#ifndef KEXEC_CRASH_MEM_ALIGN
+#define KEXEC_CRASH_MEM_ALIGN PAGE_SIZE
+#endif
+
 #define KEXEC_NOTE_HEAD_BYTES ALIGN(sizeof(struct elf_note), 4)
 #define KEXEC_CORE_NOTE_NAME "CORE"
 #define KEXEC_CORE_NOTE_NAME_BYTES ALIGN(sizeof(KEXEC_CORE_NOTE_NAME), 4)

@@ -133,6 +137,8 @@ extern void crash_kexec(struct pt_regs *
 int kexec_should_crash(struct task_struct *);
 void crash_save_cpu(struct pt_regs *regs, int cpu);
 void crash_save_vmcoreinfo(void);
+void crash_map_reserved_pages(void);
+void crash_unmap_reserved_pages(void);
 void arch_crash_save_vmcoreinfo(void);
 void vmcoreinfo_append_str(const char *fmt, ...)
 	__attribute__ ((format (printf, 1, 2)));

kernel/kexec.c +19 -2

@@ -999,6 +999,7 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 			kimage_free(xchg(&kexec_crash_image, NULL));
 			result = kimage_crash_alloc(&image, entry,
 						     nr_segments, segments);
+			crash_map_reserved_pages();
 		}
 		if (result)
 			goto out;

@@ -1015,6 +1016,8 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 				goto out;
 		}
 		kimage_terminate(image);
+		if (flags & KEXEC_ON_CRASH)
+			crash_unmap_reserved_pages();
 	}
 	/* Install the new kernel, and  Uninstall the old */
 	image = xchg(dest_image, image);

@@ -1026,6 +1029,18 @@ out:
 	return result;
 }
 
+/*
+ * Add and remove page tables for crashkernel memory
+ *
+ * Provide an empty default implementation here -- architecture
+ * code may override this
+ */
+void __weak crash_map_reserved_pages(void)
+{}
+
+void __weak crash_unmap_reserved_pages(void)
+{}
+
 #ifdef CONFIG_COMPAT
 asmlinkage long compat_sys_kexec_load(unsigned long entry,
 				unsigned long nr_segments,

@@ -1134,14 +1149,16 @@ int crash_shrink_memory(unsigned long ne
 		goto unlock;
 	}
 
-	start = roundup(start, PAGE_SIZE);
-	end = roundup(start + new_size, PAGE_SIZE);
+	start = roundup(start, KEXEC_CRASH_MEM_ALIGN);
+	end = roundup(start + new_size, KEXEC_CRASH_MEM_ALIGN);
 
+	crash_map_reserved_pages();
 	crash_free_reserved_phys_range(end, crashk_res.end);
 
 	if ((start == end) && (crashk_res.parent != NULL))
 		release_resource(&crashk_res);
 	crashk_res.end = end - 1;
+	crash_unmap_reserved_pages();
 
 unlock:
 	mutex_unlock(&kexec_mutex);

From: Michael Holzheu <[email protected]>

This patch introduces a mechanism that allows architecture backends to remove page tables for the crashkernel memory. This can protect the loaded kdump kernel from being overwritten by broken kernel code.
A new function crash_map_pages() is added that can be implemented by architecture code. This function has the following syntax:

void crash_map_pages(int enable);

"enable" can be 0 for removing or 1 for adding page tables. The function is called before and after the crashkernel segments are loaded. It is also called in crash_shrink_memory() to create new page tables when the crashkernel memory size is reduced.

To support architectures that have large pages this patch also introduces a new define KEXEC_CRASH_MEM_ALIGN. The crashkernel start and size must always be aligned with KEXEC_CRASH_MEM_ALIGN.

Signed-off-by Michael Holzheu <[email protected]>

---
 include/linux/kexec.h |    5 +++++
 kernel/kexec.c        |   16 ++++++++++++++--
 2 files changed, 19 insertions(+), 2 deletions(-)

include/linux/kexec.h +5 -0

@@ -37,6 +37,10 @@
 #define KEXEC_CRASH_CONTROL_MEMORY_LIMIT KEXEC_CONTROL_MEMORY_LIMIT
 #endif
 
+#ifndef KEXEC_CRASH_MEM_ALIGN
+#define KEXEC_CRASH_MEM_ALIGN PAGE_SIZE
+#endif
+
 #define KEXEC_NOTE_HEAD_BYTES ALIGN(sizeof(struct elf_note), 4)
 #define KEXEC_CORE_NOTE_NAME "CORE"
 #define KEXEC_CORE_NOTE_NAME_BYTES ALIGN(sizeof(KEXEC_CORE_NOTE_NAME), 4)

@@ -133,6 +137,7 @@ extern void crash_kexec(struct pt_regs *
 int kexec_should_crash(struct task_struct *);
 void crash_save_cpu(struct pt_regs *regs, int cpu);
 void crash_save_vmcoreinfo(void);
+void crash_map_pages(int enable);
 void arch_crash_save_vmcoreinfo(void);
 void vmcoreinfo_append_str(const char *fmt, ...)
 	__attribute__ ((format (printf, 1, 2)));

kernel/kexec.c +14 -2

@@ -999,6 +999,7 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 			kimage_free(xchg(&kexec_crash_image, NULL));
 			result = kimage_crash_alloc(&image, entry,
 						     nr_segments, segments);
+			crash_map_pages(1);
 		}
 		if (result)
 			goto out;

@@ -1015,6 +1016,8 @@ SYSCALL_DEFINE4(kexec_load, unsigned lon
 				goto out;
 		}
 		kimage_terminate(image);
+		if (flags & KEXEC_ON_CRASH)
+			crash_map_pages(0);
 	}
 	/* Install the new kernel, and  Uninstall the old */
 	image = xchg(dest_image, image);

@@ -1026,6 +1029,13 @@ out:
 	return result;
 }
 
+/*
+ * provide an empty default implementation here -- architecture
+ * code may override this
+ */
+void __weak crash_map_pages(int enable)
+{}
+
 #ifdef CONFIG_COMPAT
 asmlinkage long compat_sys_kexec_load(unsigned long entry,
 				unsigned long nr_segments,

@@ -1134,14 +1144,16 @@ int crash_shrink_memory(unsigned long ne
 		goto unlock;
 	}
 
-	start = roundup(start, PAGE_SIZE);
-	end = roundup(start + new_size, PAGE_SIZE);
+	start = roundup(start, KEXEC_CRASH_MEM_ALIGN);
+	end = roundup(start + new_size, KEXEC_CRASH_MEM_ALIGN);
 
+	crash_map_pages(1);
 	crash_free_reserved_phys_range(end, crashk_res.end);
 
 	if ((start == end) && (crashk_res.parent != NULL))
 		release_resource(&crashk_res);
 	crashk_res.end = end - 1;
+	crash_map_pages(0);
 
 unlock:
 	mutex_unlock(&kexec_mutex);