[PATCH v2 0/3] x86: Allow variable-sized event frame

[PATCH v2 0/3] x86: Allow variable-sized event frame

[Xin Li (Intel)]

This was initially posted as part of the FRED patch series and turned
down due to its unacceptable quality:
  https://lore.kernel.org/lkml/20230410081438.1750-31-xin3.li@intel.com/

Now try another attempt to meet the bar.


A FRED event frame could contain different amount of information for
different event types, e.g., #MCE could push extra bytes of information,
or perhaps even for different instances of the same event type. Thus
the size of an event frame pushed by a FRED CPU is not fixed and the
address of a pt_regs structure that is used to save the user level
context of current task is not at a fixed offset from top of current
task kernel stack.

This patch set adds a new field named 'user_pt_regs' in the thread_info
structure to save the address of user level context pt_regs structure,
thus to eliminate the need of any advance information of event frame
size and allow a FRED CPU to push variable-sized event frame.

With the above change, we can
1) Remove the padding space at top of the init stack because there is
   no user level context for init task.
2) Get rid of TOP_OF_KERNEL_STACK_PADDING on x86_64, which was defined
   to 0 for IDT to keep the code consistent with 32bit.

Link to v1: https://lore.kernel.org/lkml/20240617084516.1484390-1-xin@zytor.com/


Xin Li (Intel) (3):
  x86/fred: Allow variable-sized event frame
  x86: Remove the padding space at top of the init stack
  x86: Get rid of TOP_OF_KERNEL_STACK_PADDING on x86_64

 arch/x86/entry/entry_fred.c        | 22 +++++++++++++++++
 arch/x86/include/asm/processor.h   | 38 +++++++++++++++++++++++-------
 arch/x86/include/asm/thread_info.h | 19 +++++----------
 arch/x86/kernel/process.c          | 21 +++++++++++++++++
 arch/x86/kernel/vmlinux.lds.S      | 18 ++++++++++++--
 include/linux/thread_info.h        |  1 +
 kernel/fork.c                      |  6 +++++
 7 files changed, 101 insertions(+), 24 deletions(-)


base-commit: 6575d1b4a6ef3336608127c704b612bc5e7b0fdc
-- 
2.48.1

[PATCH v2 1/3] x86/fred: Allow variable-sized event frame

[Xin Li (Intel)]

A FRED event frame could contain different amount of information for
different event types, or perhaps even for different instances of the
same event type. Thus the size of an event frame pushed by a FRED CPU
is not fixed and the address of the pt_regs structure that is used to
save a user level context of current task is not at a fixed offset
from top of current task kernel stack.

Add a new field named 'user_pt_regs' in the thread_info structure to
save the address of user level context pt_regs structure, thus to
eliminate the need of any advance information of event frame size
and allow a FRED CPU to push variable-sized event frame.

For IDT user level event delivery, a pt_regs structure is pushed by
hardware and software _always_ at a fixed offset from top of current
task kernel stack, so simply initialize user_pt_regs to point to the
pt_regs structure no matter whether one is pushed or not.

While for FRED user level event delivery, user_pt_regs is updated with
a pt_regs structure pointer generated in asm_fred_entrypoint_user().

Suggested-by: H. Peter Anvin (Intel) <hpa@zytor.com>
Signed-off-by: Xin Li (Intel) <xin@zytor.com>
---
 arch/x86/entry/entry_fred.c        | 22 ++++++++++++++++++++++
 arch/x86/include/asm/processor.h   | 18 ++++++++++++------
 arch/x86/include/asm/thread_info.h |  9 ++++++---
 arch/x86/kernel/process.c          | 22 ++++++++++++++++++++++
 include/linux/thread_info.h        |  1 +
 kernel/fork.c                      |  6 ++++++
 6 files changed, 69 insertions(+), 9 deletions(-)

diff --git a/arch/x86/entry/entry_fred.c b/arch/x86/entry/entry_fred.c
index f004a4dc74c2..1d54d451acb6 100644
--- a/arch/x86/entry/entry_fred.c
+++ b/arch/x86/entry/entry_fred.c
@@ -228,6 +228,28 @@ __visible noinstr void fred_entry_from_user(struct pt_regs *regs)
 	/* Invalidate orig_ax so that syscall_get_nr() works correctly */
 	regs->orig_ax = -1;
 
+	/*
+	 * A FRED event frame could contain different amount of information
+	 * for different event types, or perhaps even for different instances
+	 * of the same event type. Thus the size of an event frame pushed by
+	 * a FRED CPU is not fixed and the address of the pt_regs structure
+	 * that is used to save a user level context of current task is not
+	 * at a fixed offset from top of current task stack.
+	 *
+	 * Save the address of the pt_regs structure passed from and generated
+	 * in the caller function asm_fred_entrypoint_user() in thread_info so
+	 * that task_pt_regs() can be used to access the pt_regs structure
+	 * containing user level context after this point.
+	 *
+	 * What if another event happens before this point?
+	 *
+	 * Actually, another kernel event could happen earlier, even before the
+	 * pt_regs structure for saving user level context is completely saved.
+	 * It is guaranteed that the handler of the new event will NOT access
+	 * the pt_regs structure of the previous user level event.
+	 */
+	current->thread_info.user_pt_regs = regs;
+
 	switch (regs->fred_ss.type) {
 	case EVENT_TYPE_EXTINT:
 		return fred_extint(regs);
diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index 5d2f7e5aff26..79d8df28cd59 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -646,12 +646,18 @@ static __always_inline void prefetchw(const void *x)
 
 #define task_top_of_stack(task) ((unsigned long)(task_pt_regs(task) + 1))
 
-#define task_pt_regs(task) \
-({									\
-	unsigned long __ptr = (unsigned long)task_stack_page(task);	\
-	__ptr += THREAD_SIZE - TOP_OF_KERNEL_STACK_PADDING;		\
-	((struct pt_regs *)__ptr) - 1;					\
-})
+/*
+ * task_pt_regs() no longer converts a fixed offset from top of a task
+ * kernel stack to a pt_regs structure pointer, but rather returns
+ * whatever in the thread_info.user_pt_regs field, which contains the
+ * address of a pt_regs structure used to save a user level context of
+ * current task.
+ *
+ * Note, this can't be converted to an inline function as this header
+ * file defines 'struct thread_struct' which is used in the task_struct
+ * structure definition.
+ */
+#define task_pt_regs(task) ((task)->thread_info.user_pt_regs)
 
 #ifdef CONFIG_X86_32
 #define INIT_THREAD  {							  \
diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
index 9282465eea21..4372f171c65f 100644
--- a/arch/x86/include/asm/thread_info.h
+++ b/arch/x86/include/asm/thread_info.h
@@ -56,6 +56,7 @@
  */
 #ifndef __ASSEMBLER__
 struct task_struct;
+struct pt_regs;
 #include <asm/cpufeature.h>
 #include <linux/atomic.h>
 
@@ -66,11 +67,13 @@ struct thread_info {
 #ifdef CONFIG_SMP
 	u32			cpu;		/* current CPU */
 #endif
+	struct pt_regs		*user_pt_regs;
 };
 
-#define INIT_THREAD_INFO(tsk)			\
-{						\
-	.flags		= 0,			\
+#define INIT_THREAD_INFO(tsk)						\
+{									\
+	.flags		= 0,						\
+	.user_pt_regs	= (struct pt_regs *)TOP_OF_INIT_STACK - 1,	\
 }
 
 #else /* !__ASSEMBLER__ */
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 91f6ff618852..58c1cd4ca60a 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -108,6 +108,28 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 	return 0;
 }
 
+/*
+ * Initialize thread_info.user_pt_regs for IDT event delivery.
+ *
+ * For IDT user level event delivery, a pt_regs structure is pushed by both
+ * hardware and software and always resides at a fixed offset from top of
+ * current task kernel stack, thus thread_info.user_pt_regs is a per-task
+ * constant and NEVER changes after initialization.
+ *
+ * While for FRED user level event delivery, user_pt_regs is updated in
+ * fred_entry_from_user() immediately after user level event delivery.
+ *
+ * Note: thread_info.user_pt_regs of the init task is initialized at build
+ * time.
+ */
+void arch_init_user_pt_regs(struct task_struct *tsk)
+{
+	unsigned long top_of_stack = (unsigned long)task_stack_page(tsk) + THREAD_SIZE;
+
+	top_of_stack -= TOP_OF_KERNEL_STACK_PADDING;
+	tsk->thread_info.user_pt_regs = (struct pt_regs *)top_of_stack - 1;
+}
+
 #ifdef CONFIG_X86_64
 void arch_release_task_struct(struct task_struct *tsk)
 {
diff --git a/include/linux/thread_info.h b/include/linux/thread_info.h
index cf2446c9c30d..324667e0b8dd 100644
--- a/include/linux/thread_info.h
+++ b/include/linux/thread_info.h
@@ -273,6 +273,7 @@ void arch_task_cache_init(void); /* for CONFIG_SH */
 void arch_release_task_struct(struct task_struct *tsk);
 int arch_dup_task_struct(struct task_struct *dst,
 				struct task_struct *src);
+void arch_init_user_pt_regs(struct task_struct *tsk);
 
 #endif	/* __KERNEL__ */
 
diff --git a/kernel/fork.c b/kernel/fork.c
index 8695c345ca49..21af95645d84 100644
--- a/kernel/fork.c
+++ b/kernel/fork.c
@@ -1105,6 +1105,10 @@ int __weak arch_dup_task_struct(struct task_struct *dst,
 	return 0;
 }
 
+void __weak arch_init_user_pt_regs(struct task_struct *tsk)
+{
+}
+
 void set_task_stack_end_magic(struct task_struct *tsk)
 {
 	unsigned long *stackend;
@@ -1132,6 +1136,8 @@ static struct task_struct *dup_task_struct(struct task_struct *orig, int node)
 	if (err)
 		goto free_tsk;
 
+	arch_init_user_pt_regs(tsk);
+
 #ifdef CONFIG_THREAD_INFO_IN_TASK
 	refcount_set(&tsk->stack_refcount, 1);
 #endif
-- 
2.48.1

[PATCH v2 2/3] x86: Remove the padding space at top of the init stack

[Xin Li (Intel)]

Because the owner of the init stack, init task, doesn't have any user
level context, there will NEVER be an actual pt_regs structure pushed
at top of the init stack.

However a zeroed pt_regs structure is created at build time and kept
at top of the init stack for task_pt_regs() to function properly with
the init task in the same manner as a normal task with user level
context.

Besides, task_pt_regs() no longer converts a fixed offset from top of
a task kernel stack to a pt_regs structure pointer, but rather returns
whatever in the thread_info.user_pt_regs field, which is initialized
at build time to '(struct pt_regs *)TOP_OF_INIT_STACK - 1' for the
init task.

As a result, there is no point to reserve any padding space at top of
the init stack, so remove the padding space.

Signed-off-by: Xin Li (Intel) <xin@zytor.com>
---
 arch/x86/include/asm/processor.h | 16 ++++++++++++++--
 arch/x86/kernel/vmlinux.lds.S    | 18 ++++++++++++++++--
 2 files changed, 30 insertions(+), 4 deletions(-)

diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index 79d8df28cd59..cbb5d2158075 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -641,8 +641,20 @@ static __always_inline void prefetchw(const void *x)
 			  "m" (*(const char *)x));
 }
 
-#define TOP_OF_INIT_STACK ((unsigned long)&init_stack + sizeof(init_stack) - \
-			   TOP_OF_KERNEL_STACK_PADDING)
+extern unsigned long __end_init_stack[];
+
+/*
+ * No need to reserve extra padding space above the pt_regs structure
+ * at top of the init stack, because its owner init task doesn't have
+ * any user level context, thus there will NEVER be an actual pt_regs
+ * structure pushed at top of the init stack.
+ *
+ * However a zeroed pt_regs structure is created at build time and kept
+ * at top of the init stack for task_pt_regs() to function properly with
+ * the init task in the same manner as a normal task with user level
+ * context.
+ */
+#define TOP_OF_INIT_STACK ((unsigned long)&__end_init_stack)
 
 #define task_top_of_stack(task) ((unsigned long)(task_pt_regs(task) + 1))
 
diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S
index ccdc45e5b759..553b65e8bb8a 100644
--- a/arch/x86/kernel/vmlinux.lds.S
+++ b/arch/x86/kernel/vmlinux.lds.S
@@ -178,8 +178,22 @@ SECTIONS
 		/* init_task */
 		INIT_TASK_DATA(THREAD_SIZE)
 
-		/* equivalent to task_pt_regs(&init_task) */
-		__top_init_kernel_stack = __end_init_stack - TOP_OF_KERNEL_STACK_PADDING - PTREGS_SIZE;
+		/*
+		 * No need to reserve extra padding space above the pt_regs
+		 * structure at top of the init stack, because its owner
+		 * init task doesn't have any user level context, thus there
+		 * will NEVER be an actual pt_regs structure pushed at top
+		 * of the init stack.
+		 *
+		 * However a zeroed pt_regs structure is created at build
+		 * time and kept at top of the init stack for task_pt_regs()
+		 * to function properly with the init task in the same manner
+		 * as a normal task with user level context.
+		 *
+		 * task_pt_regs(&init_task) is now:
+		 *	'(struct pt_regs *)&__end_init_stack - 1'
+		 */
+		__top_init_kernel_stack = __end_init_stack - PTREGS_SIZE;
 
 #ifdef CONFIG_X86_32
 		/* 32 bit has nosave before _edata */
-- 
2.48.1

[PATCH v2 3/3] x86: Get rid of TOP_OF_KERNEL_STACK_PADDING on x86_64

[Xin Li (Intel)]

Because task_pt_regs() is now just an alias of thread_info.user_pt_regs,
and no matter whether FRED is enabled or not a user level event frame on
x86_64 is always pushed from top of current task kernel stack, i.e.,
'(unsigned long)task_stack_page(task) + THREAD_SIZE', there is no meaning
to keep TOP_OF_KERNEL_STACK_PADDING on x86_64, thus remove it.

Signed-off-by: Xin Li (Intel) <xin@zytor.com>
---
 arch/x86/include/asm/processor.h   |  6 ++++--
 arch/x86/include/asm/thread_info.h | 10 ----------
 arch/x86/kernel/process.c          |  3 +--
 3 files changed, 5 insertions(+), 14 deletions(-)

diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index cbb5d2158075..ab1f9cb19b71 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -656,8 +656,6 @@ extern unsigned long __end_init_stack[];
  */
 #define TOP_OF_INIT_STACK ((unsigned long)&__end_init_stack)
 
-#define task_top_of_stack(task) ((unsigned long)(task_pt_regs(task) + 1))
-
 /*
  * task_pt_regs() no longer converts a fixed offset from top of a task
  * kernel stack to a pt_regs structure pointer, but rather returns
@@ -672,6 +670,9 @@ extern unsigned long __end_init_stack[];
 #define task_pt_regs(task) ((task)->thread_info.user_pt_regs)
 
 #ifdef CONFIG_X86_32
+#define task_top_of_stack(task) ((unsigned long)task_stack_page(task) + THREAD_SIZE	\
+				 - TOP_OF_KERNEL_STACK_PADDING)
+
 #define INIT_THREAD  {							  \
 	.sp0			= TOP_OF_INIT_STACK,			  \
 	.sysenter_cs		= __KERNEL_CS,				  \
@@ -679,6 +680,7 @@ extern unsigned long __end_init_stack[];
 
 #else
 extern unsigned long __top_init_kernel_stack[];
+#define task_top_of_stack(task) ((unsigned long)task_stack_page(task) + THREAD_SIZE)
 
 #define INIT_THREAD {							\
 	.sp	= (unsigned long)&__top_init_kernel_stack,		\
diff --git a/arch/x86/include/asm/thread_info.h b/arch/x86/include/asm/thread_info.h
index 4372f171c65f..8e9badd610bc 100644
--- a/arch/x86/include/asm/thread_info.h
+++ b/arch/x86/include/asm/thread_info.h
@@ -30,10 +30,6 @@
  *
  * In vm86 mode, the hardware frame is much longer still, so add 16
  * bytes to make room for the real-mode segments.
- *
- * x86-64 has a fixed-length stack frame, but it depends on whether
- * or not FRED is enabled. Future versions of FRED might make this
- * dynamic, but for now it is always 2 words longer.
  */
 #ifdef CONFIG_X86_32
 # ifdef CONFIG_VM86
@@ -41,12 +37,6 @@
 # else
 #  define TOP_OF_KERNEL_STACK_PADDING 8
 # endif
-#else /* x86-64 */
-# ifdef CONFIG_X86_FRED
-#  define TOP_OF_KERNEL_STACK_PADDING (2 * 8)
-# else
-#  define TOP_OF_KERNEL_STACK_PADDING 0
-# endif
 #endif
 
 /*
diff --git a/arch/x86/kernel/process.c b/arch/x86/kernel/process.c
index 58c1cd4ca60a..51020caac332 100644
--- a/arch/x86/kernel/process.c
+++ b/arch/x86/kernel/process.c
@@ -124,9 +124,8 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
  */
 void arch_init_user_pt_regs(struct task_struct *tsk)
 {
-	unsigned long top_of_stack = (unsigned long)task_stack_page(tsk) + THREAD_SIZE;
+	unsigned long top_of_stack = task_top_of_stack(tsk);
 
-	top_of_stack -= TOP_OF_KERNEL_STACK_PADDING;
 	tsk->thread_info.user_pt_regs = (struct pt_regs *)top_of_stack - 1;
 }
 
-- 
2.48.1