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, 127 insertions, 0 deletions

diff --git a/Documentation/ABI/testing/dev-kmsg b/Documentation/ABI/testing/dev-kmsg
new file mode 100644
index 000000000..a377b6c09
--- /dev/null
+++ b/Documentation/ABI/testing/dev-kmsg
@@ -0,0 +1,127 @@
+What:		/dev/kmsg
+Date:		Mai 2012
+KernelVersion:	3.5
+Contact:	Kay Sievers <kay@vrfy.org>
+Description:	The /dev/kmsg character device node provides userspace access
+		to the kernel's printk buffer.
+
+		Injecting messages:
+
+		Every write() to the opened device node places a log entry in
+		the kernel's printk buffer.
+
+		The logged line can be prefixed with a <N> syslog prefix, which
+		carries the syslog priority and facility. The single decimal
+		prefix number is composed of the 3 lowest bits being the syslog
+		priority and the next 8 bits the syslog facility number.
+
+		If no prefix is given, the priority number is the default kernel
+		log priority and the facility number is set to LOG_USER (1). It
+		is not possible to inject messages from userspace with the
+		facility number LOG_KERN (0), to make sure that the origin of
+		the messages can always be reliably determined.
+
+		Accessing the buffer:
+
+		Every read() from the opened device node receives one record
+		of the kernel's printk buffer.
+
+		The first read() directly following an open() always returns
+		first message in the buffer; there is no kernel-internal
+		persistent state; many readers can concurrently open the device
+		and read from it, without affecting other readers.
+
+		Every read() will receive the next available record. If no more
+		records are available read() will block, or if O_NONBLOCK is
+		used -EAGAIN returned.
+
+		Messages in the record ring buffer get overwritten as whole,
+		there are never partial messages received by read().
+
+		In case messages get overwritten in the circular buffer while
+		the device is kept open, the next read() will return -EPIPE,
+		and the seek position be updated to the next available record.
+		Subsequent reads() will return available records again.
+
+		Unlike the classic syslog() interface, the 64 bit record
+		sequence numbers allow to calculate the amount of lost
+		messages, in case the buffer gets overwritten. And they allow
+		to reconnect to the buffer and reconstruct the read position
+		if needed, without limiting the interface to a single reader.
+
+		The device supports seek with the following parameters:
+
+		SEEK_SET, 0
+		  seek to the first entry in the buffer
+		SEEK_END, 0
+		  seek after the last entry in the buffer
+		SEEK_DATA, 0
+		  seek after the last record available at the time
+		  the last SYSLOG_ACTION_CLEAR was issued.
+
+		Other seek operations or offsets are not supported because of
+		the special behavior this device has. The device allows to read
+		or write only whole variable length messages (records) that are
+		stored in a ring buffer.
+
+		Because of the non-standard behavior also the error values are
+		non-standard. -ESPIPE is returned for non-zero offset. -EINVAL
+		is returned for other operations, e.g. SEEK_CUR. This behavior
+		and values are historical and could not be modified without the
+		risk of breaking userspace.
+
+		The output format consists of a prefix carrying the syslog
+		prefix including priority and facility, the 64 bit message
+		sequence number and the monotonic timestamp in microseconds,
+		and a flag field. All fields are separated by a ','.
+
+		Future extensions might add more comma separated values before
+		the terminating ';'. Unknown fields and values should be
+		gracefully ignored.
+
+		The human readable text string starts directly after the ';'
+		and is terminated by a '\n'. Untrusted values derived from
+		hardware or other facilities are printed, therefore
+		all non-printable characters and '\' itself in the log message
+		are escaped by "\x00" C-style hex encoding.
+
+		A line starting with ' ', is a continuation line, adding
+		key/value pairs to the log message, which provide the machine
+		readable context of the message, for reliable processing in
+		userspace.
+
+		Example::
+
+		  7,160,424069,-;pci_root PNP0A03:00: host bridge window [io  0x0000-0x0cf7] (ignored)
+		   SUBSYSTEM=acpi
+		   DEVICE=+acpi:PNP0A03:00
+		  6,339,5140900,-;NET: Registered protocol family 10
+		  30,340,5690716,-;udevd[80]: starting version 181
+
+		The DEVICE= key uniquely identifies devices the following way:
+
+		  ============  =================
+		  b12:8         block dev_t
+		  c127:3        char dev_t
+		  n8            netdev ifindex
+		  +sound:card0  subsystem:devname
+		  ============  =================
+
+		The flags field carries '-' by default. A 'c' indicates a
+		fragment of a line. Note, that these hints about continuation
+		lines are not necessarily correct, and the stream could be
+		interleaved with unrelated messages, but merging the lines in
+		the output usually produces better human readable results. A
+		similar logic is used internally when messages are printed to
+		the console, /proc/kmsg or the syslog() syscall.
+
+		By default, kernel tries to avoid fragments by concatenating
+		when it can and fragments are rare; however, when extended
+		console support is enabled, the in-kernel concatenation is
+		disabled and /dev/kmsg output will contain more fragments. If
+		the log consumer performs concatenation, the end result
+		should be the same. In the future, the in-kernel concatenation
+		may be removed entirely and /dev/kmsg users are recommended to
+		implement fragment handling.
+
+Users:		dmesg(1), userspace kernel log consumers