x86/kprobes: Fix crash when probing CS CALL instructionsgrafted

When using eBPF to probe CS CALL instructions within a function,
a crash can be triggered.

The eBPF tool probes offset 257 of the __hrtimer_run_queues()
function:

<__hrtimer_run_queues+249>:  nopl   0x0(%rax,%rax,1)
<__hrtimer_run_queues+254>:  mov    %r14,%rdi
<__hrtimer_run_queues+257>:  cs call <__x86_indirect_thunk_r12>
<__hrtimer_run_queues+263>:  mov    %eax,%r12d
<__hrtimer_run_queues+266>:  xchg   %ax,%ax
<__hrtimer_run_queues+268>:  mov    %r13,%rdi

Which triggers this crash:

  BUG: unable to handle page fault for address: 00000000000f41c9
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP NOPTI
  CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P
  RIP: 0010:__hrtimer_run_queues+0x106/0x230

Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is
at the 6th byte of the above CS CALL instruction. Since the CS CALL
instruction occupies 6 bytes, the exception occurred in the middle of that
call instruction.

The root cause is that when using eBPF tools to probe in the middle of a
function, a kprobe with INT3 is used as the underlying implementation.

During single-step emulation of the original CALL instruction,
int3_emulate_call() assumes that the probed CALL instruction is 5 bytes
long. However, the actual CS-prefixed CALL instruction occupies 6 bytes,
so it constructs an incorrect exception return address. When the CPU
returns from the kprobe handler, the next instruction to be executed is at
the address of the last byte of that CS CALL instruction. Coincidentally,
starting from that address, the CPU fetches and decodes a completely
different instruction, which ultimately triggers a kernel crash.

Fix the issue by using the actual instruction length obtained from
the instruction decoder when constructing the exception return
address, rather than relying on the hardcoded CALL_INSN_SIZE macro.

[ mingo: Refined the changelog ]

Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step")
Suggested-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Jinke Han <jinkehan@didiglobal.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Acked-by: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000

1 files changed, 166 insertions, 0 deletions

diff --git a/arch/alpha/lib/memcpy.c b/arch/alpha/lib/memcpy.c
new file mode 100644
index 000000000..78b6850a9
--- /dev/null
+++ b/arch/alpha/lib/memcpy.c
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ *  linux/arch/alpha/lib/memcpy.c
+ *
+ *  Copyright (C) 1995  Linus Torvalds
+ */
+
+/*
+ * This is a reasonably optimized memcpy() routine.
+ */
+
+/*
+ * Note that the C code is written to be optimized into good assembly. However,
+ * at this point gcc is unable to sanely compile "if (n >= 0)", resulting in a
+ * explicit compare against 0 (instead of just using the proper "blt reg, xx" or
+ * "bge reg, xx"). I hope alpha-gcc will be fixed to notice this eventually..
+ */
+
+#include <linux/types.h>
+#include <linux/export.h>
+#include <linux/string.h>
+
+/*
+ * This should be done in one go with ldq_u*2/mask/stq_u. Do it
+ * with a macro so that we can fix it up later..
+ */
+#define ALIGN_DEST_TO8_UP(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define ALIGN_DEST_TO8_DN(d,s,n) \
+	while (d & 7) { \
+		if (n <= 0) return; \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should similarly be done with ldq_u*2/mask/stq. The destination
+ * is aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_UP(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		*(char *) d = *(char *) s; \
+		d++; s++; \
+	}
+#define DO_REST_DN(d,s,n) \
+	while (n > 0) { \
+		n--; \
+		d--; s--; \
+		*(char *) d = *(char *) s; \
+	}
+
+/*
+ * This should be done with ldq/mask/stq. The source and destination are
+ * aligned, but we don't fill in a full quad-word
+ */
+#define DO_REST_ALIGNED_UP(d,s,n) DO_REST_UP(d,s,n)
+#define DO_REST_ALIGNED_DN(d,s,n) DO_REST_DN(d,s,n)
+
+/*
+ * This does unaligned memory copies. We want to avoid storing to
+ * an unaligned address, as that would do a read-modify-write cycle.
+ * We also want to avoid double-reading the unaligned reads.
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_unaligned_up (unsigned long d, unsigned long s,
+					  long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;			/* to avoid compare against 8 in the loop */
+	if (n >= 0) {
+		unsigned long low_word, high_word;
+		__asm__("ldq_u %0,%1":"=r" (low_word):"m" (*(unsigned long *) s));
+		do {
+			unsigned long tmp;
+			__asm__("ldq_u %0,%1":"=r" (high_word):"m" (*(unsigned long *)(s+8)));
+			n -= 8;
+			__asm__("extql %1,%2,%0"
+				:"=r" (low_word)
+				:"r" (low_word), "r" (s));
+			__asm__("extqh %1,%2,%0"
+				:"=r" (tmp)
+				:"r" (high_word), "r" (s));
+			s += 8;
+			*(unsigned long *) d = low_word | tmp;
+			d += 8;
+			low_word = high_word;
+		} while (n >= 0);
+	}
+	n += 8;
+	DO_REST_UP(d,s,n);
+}
+
+static inline void __memcpy_unaligned_dn (unsigned long d, unsigned long s,
+					  long n)
+{
+	/* I don't understand AXP assembler well enough for this. -Tim */
+	s += n;
+	d += n;
+	while (n--)
+		* (char *) --d = * (char *) --s;
+}
+
+/*
+ * Hmm.. Strange. The __asm__ here is there to make gcc use an integer register
+ * for the load-store. I don't know why, but it would seem that using a floating
+ * point register for the move seems to slow things down (very small difference,
+ * though).
+ *
+ * Note the ordering to try to avoid load (and address generation) latencies.
+ */
+static inline void __memcpy_aligned_up (unsigned long d, unsigned long s,
+					long n)
+{
+	ALIGN_DEST_TO8_UP(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		s += 8;
+		*(unsigned long *) d = tmp;
+		d += 8;
+	}
+	n += 8;
+	DO_REST_ALIGNED_UP(d,s,n);
+}
+static inline void __memcpy_aligned_dn (unsigned long d, unsigned long s,
+					long n)
+{
+	s += n;
+	d += n;
+	ALIGN_DEST_TO8_DN(d,s,n);
+	n -= 8;
+	while (n >= 0) {
+		unsigned long tmp;
+		s -= 8;
+		__asm__("ldq %0,%1":"=r" (tmp):"m" (*(unsigned long *) s));
+		n -= 8;
+		d -= 8;
+		*(unsigned long *) d = tmp;
+	}
+	n += 8;
+	DO_REST_ALIGNED_DN(d,s,n);
+}
+
+#undef memcpy
+
+void * memcpy(void * dest, const void *src, size_t n)
+{
+	if (!(((unsigned long) dest ^ (unsigned long) src) & 7)) {
+		__memcpy_aligned_up ((unsigned long) dest, (unsigned long) src,
+				     n);
+		return dest;
+	}
+	__memcpy_unaligned_up ((unsigned long) dest, (unsigned long) src, n);
+	return dest;
+}
+EXPORT_SYMBOL(memcpy);