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, 414 insertions, 0 deletions

diff --git a/lib/decompress_unxz.c b/lib/decompress_unxz.c
new file mode 100644
index 000000000..05d5cb490
--- /dev/null
+++ b/lib/decompress_unxz.c
@@ -0,0 +1,414 @@
+// SPDX-License-Identifier: 0BSD
+
+/*
+ * Wrapper for decompressing XZ-compressed kernel, initramfs, and initrd
+ *
+ * Author: Lasse Collin <lasse.collin@tukaani.org>
+ */
+
+/*
+ * Important notes about in-place decompression
+ *
+ * At least on x86, the kernel is decompressed in place: the compressed data
+ * is placed to the end of the output buffer, and the decompressor overwrites
+ * most of the compressed data. There must be enough safety margin to
+ * guarantee that the write position is always behind the read position.
+ *
+ * The safety margin for XZ with LZMA2 or BCJ+LZMA2 is calculated below.
+ * Note that the margin with XZ is bigger than with Deflate (gzip)!
+ *
+ * The worst case for in-place decompression is that the beginning of
+ * the file is compressed extremely well, and the rest of the file is
+ * incompressible. Thus, we must look for worst-case expansion when the
+ * compressor is encoding incompressible data.
+ *
+ * The structure of the .xz file in case of a compressed kernel is as follows.
+ * Sizes (as bytes) of the fields are in parenthesis.
+ *
+ *    Stream Header (12)
+ *    Block Header:
+ *      Block Header (8-12)
+ *      Compressed Data (N)
+ *      Block Padding (0-3)
+ *      CRC32 (4)
+ *    Index (8-20)
+ *    Stream Footer (12)
+ *
+ * Normally there is exactly one Block, but let's assume that there are
+ * 2-4 Blocks just in case. Because Stream Header and also Block Header
+ * of the first Block don't make the decompressor produce any uncompressed
+ * data, we can ignore them from our calculations. Block Headers of possible
+ * additional Blocks have to be taken into account still. With these
+ * assumptions, it is safe to assume that the total header overhead is
+ * less than 128 bytes.
+ *
+ * Compressed Data contains LZMA2 or BCJ+LZMA2 encoded data. Since BCJ
+ * doesn't change the size of the data, it is enough to calculate the
+ * safety margin for LZMA2.
+ *
+ * LZMA2 stores the data in chunks. Each chunk has a header whose size is
+ * a maximum of 6 bytes, but to get round 2^n numbers, let's assume that
+ * the maximum chunk header size is 8 bytes. After the chunk header, there
+ * may be up to 64 KiB of actual payload in the chunk. Often the payload is
+ * quite a bit smaller though; to be safe, let's assume that an average
+ * chunk has only 32 KiB of payload.
+ *
+ * The maximum uncompressed size of the payload is 2 MiB. The minimum
+ * uncompressed size of the payload is in practice never less than the
+ * payload size itself. The LZMA2 format would allow uncompressed size
+ * to be less than the payload size, but no sane compressor creates such
+ * files. LZMA2 supports storing incompressible data in uncompressed form,
+ * so there's never a need to create payloads whose uncompressed size is
+ * smaller than the compressed size.
+ *
+ * The assumption, that the uncompressed size of the payload is never
+ * smaller than the payload itself, is valid only when talking about
+ * the payload as a whole. It is possible that the payload has parts where
+ * the decompressor consumes more input than it produces output. Calculating
+ * the worst case for this would be tricky. Instead of trying to do that,
+ * let's simply make sure that the decompressor never overwrites any bytes
+ * of the payload which it is currently reading.
+ *
+ * Now we have enough information to calculate the safety margin. We need
+ *   - 128 bytes for the .xz file format headers;
+ *   - 8 bytes per every 32 KiB of uncompressed size (one LZMA2 chunk header
+ *     per chunk, each chunk having average payload size of 32 KiB); and
+ *   - 64 KiB (biggest possible LZMA2 chunk payload size) to make sure that
+ *     the decompressor never overwrites anything from the LZMA2 chunk
+ *     payload it is currently reading.
+ *
+ * We get the following formula:
+ *
+ *    safety_margin = 128 + uncompressed_size * 8 / 32768 + 65536
+ *                  = 128 + (uncompressed_size >> 12) + 65536
+ *
+ * For comparison, according to arch/x86/boot/compressed/misc.c, the
+ * equivalent formula for Deflate is this:
+ *
+ *    safety_margin = 18 + (uncompressed_size >> 12) + 32768
+ *
+ * Thus, when updating Deflate-only in-place kernel decompressor to
+ * support XZ, the fixed overhead has to be increased from 18+32768 bytes
+ * to 128+65536 bytes.
+ */
+
+/*
+ * STATIC is defined to "static" if we are being built for kernel
+ * decompression (pre-boot code). <linux/decompress/mm.h> will define
+ * STATIC to empty if it wasn't already defined. Since we will need to
+ * know later if we are being used for kernel decompression, we define
+ * XZ_PREBOOT here.
+ */
+#ifdef STATIC
+#	define XZ_PREBOOT
+#else
+#	include <linux/decompress/unxz.h>
+#endif
+#ifdef __KERNEL__
+#	include <linux/decompress/mm.h>
+#endif
+
+#ifndef XZ_PREBOOT
+#	include <linux/slab.h>
+#	include <linux/xz.h>
+#else
+/*
+ * Use the internal CRC32 code instead of kernel's CRC32 module, which
+ * is not available in early phase of booting.
+ */
+#define XZ_INTERNAL_CRC32 1
+
+/*
+ * For boot time use, we enable only the BCJ filter of the current
+ * architecture or none if no BCJ filter is available for the architecture.
+ */
+#ifdef CONFIG_X86
+#	define XZ_DEC_X86
+#endif
+#if defined(CONFIG_PPC) && defined(CONFIG_CPU_BIG_ENDIAN)
+#	define XZ_DEC_POWERPC
+#endif
+#ifdef CONFIG_ARM
+#	ifdef CONFIG_THUMB2_KERNEL
+#		define XZ_DEC_ARMTHUMB
+#	else
+#		define XZ_DEC_ARM
+#	endif
+#endif
+#ifdef CONFIG_ARM64
+#	define XZ_DEC_ARM64
+#endif
+#ifdef CONFIG_RISCV
+#	define XZ_DEC_RISCV
+#endif
+#ifdef CONFIG_SPARC
+#	define XZ_DEC_SPARC
+#endif
+
+/*
+ * This will get the basic headers so that memeq() and others
+ * can be defined.
+ */
+#include "xz/xz_private.h"
+
+/*
+ * Replace the normal allocation functions with the versions from
+ * <linux/decompress/mm.h>. vfree() needs to support vfree(NULL)
+ * when XZ_DYNALLOC is used, but the pre-boot free() doesn't support it.
+ * Workaround it here because the other decompressors don't need it.
+ */
+#undef kmalloc_obj
+#undef kfree
+#undef vmalloc
+#undef vfree
+#define kmalloc_obj(type) malloc(sizeof(type))
+#define kfree(ptr) free(ptr)
+#define vmalloc(size) malloc(size)
+#define vfree(ptr) do { if (ptr != NULL) free(ptr); } while (0)
+
+/*
+ * FIXME: Not all basic memory functions are provided in architecture-specific
+ * files (yet). We define our own versions here for now, but this should be
+ * only a temporary solution.
+ *
+ * memeq and memzero are not used much and any remotely sane implementation
+ * is fast enough. memcpy/memmove speed matters in multi-call mode, but
+ * the kernel image is decompressed in single-call mode, in which only
+ * memmove speed can matter and only if there is a lot of incompressible data
+ * (LZMA2 stores incompressible chunks in uncompressed form). Thus, the
+ * functions below should just be kept small; it's probably not worth
+ * optimizing for speed.
+ */
+
+#ifndef memeq
+static bool memeq(const void *a, const void *b, size_t size)
+{
+	const uint8_t *x = a;
+	const uint8_t *y = b;
+	size_t i;
+
+	for (i = 0; i < size; ++i)
+		if (x[i] != y[i])
+			return false;
+
+	return true;
+}
+#endif
+
+#ifndef memzero
+static void memzero(void *buf, size_t size)
+{
+	uint8_t *b = buf;
+	uint8_t *e = b + size;
+
+	while (b != e)
+		*b++ = '\0';
+}
+#endif
+
+#ifndef memmove
+/* Not static to avoid a conflict with the prototype in the Linux headers. */
+void *memmove(void *dest, const void *src, size_t size)
+{
+	uint8_t *d = dest;
+	const uint8_t *s = src;
+	size_t i;
+
+	if (d < s) {
+		for (i = 0; i < size; ++i)
+			d[i] = s[i];
+	} else if (d > s) {
+		i = size;
+		while (i-- > 0)
+			d[i] = s[i];
+	}
+
+	return dest;
+}
+#endif
+
+/*
+ * Since we need memmove anyway, we could use it as memcpy too.
+ * Commented out for now to avoid breaking things.
+ */
+/*
+#ifndef memcpy
+#	define memcpy memmove
+#endif
+*/
+
+#include "xz/xz_crc32.c"
+#include "xz/xz_dec_stream.c"
+#include "xz/xz_dec_lzma2.c"
+#include "xz/xz_dec_bcj.c"
+
+#endif /* XZ_PREBOOT */
+
+/* Size of the input and output buffers in multi-call mode */
+#define XZ_IOBUF_SIZE 4096
+
+/*
+ * This function implements the API defined in <linux/decompress/generic.h>.
+ *
+ * This wrapper will automatically choose single-call or multi-call mode
+ * of the native XZ decoder API. The single-call mode can be used only when
+ * both input and output buffers are available as a single chunk, i.e. when
+ * fill() and flush() won't be used.
+ */
+STATIC int INIT unxz(unsigned char *in, long in_size,
+		     long (*fill)(void *dest, unsigned long size),
+		     long (*flush)(void *src, unsigned long size),
+		     unsigned char *out, long *in_used,
+		     void (*error)(char *x))
+{
+	struct xz_buf b;
+	struct xz_dec *s;
+	enum xz_ret ret;
+	bool must_free_in = false;
+
+#if XZ_INTERNAL_CRC32
+	xz_crc32_init();
+#endif
+
+	if (in_used != NULL)
+		*in_used = 0;
+
+	if (fill == NULL && flush == NULL)
+		s = xz_dec_init(XZ_SINGLE, 0);
+	else
+		s = xz_dec_init(XZ_DYNALLOC, (uint32_t)-1);
+
+	if (s == NULL)
+		goto error_alloc_state;
+
+	if (flush == NULL) {
+		b.out = out;
+		b.out_size = (size_t)-1;
+	} else {
+		b.out_size = XZ_IOBUF_SIZE;
+		b.out = malloc(XZ_IOBUF_SIZE);
+		if (b.out == NULL)
+			goto error_alloc_out;
+	}
+
+	if (in == NULL) {
+		must_free_in = true;
+		in = malloc(XZ_IOBUF_SIZE);
+		if (in == NULL)
+			goto error_alloc_in;
+	}
+
+	b.in = in;
+	b.in_pos = 0;
+	b.in_size = in_size;
+	b.out_pos = 0;
+
+	if (fill == NULL && flush == NULL) {
+		ret = xz_dec_run(s, &b);
+	} else {
+		do {
+			if (b.in_pos == b.in_size && fill != NULL) {
+				if (in_used != NULL)
+					*in_used += b.in_pos;
+
+				b.in_pos = 0;
+
+				in_size = fill(in, XZ_IOBUF_SIZE);
+				if (in_size < 0) {
+					/*
+					 * This isn't an optimal error code
+					 * but it probably isn't worth making
+					 * a new one either.
+					 */
+					ret = XZ_BUF_ERROR;
+					break;
+				}
+
+				b.in_size = in_size;
+			}
+
+			ret = xz_dec_run(s, &b);
+
+			if (flush != NULL && (b.out_pos == b.out_size
+					|| (ret != XZ_OK && b.out_pos > 0))) {
+				/*
+				 * Setting ret here may hide an error
+				 * returned by xz_dec_run(), but probably
+				 * it's not too bad.
+				 */
+				if (flush(b.out, b.out_pos) != (long)b.out_pos)
+					ret = XZ_BUF_ERROR;
+
+				b.out_pos = 0;
+			}
+		} while (ret == XZ_OK);
+
+		if (must_free_in)
+			free(in);
+
+		if (flush != NULL)
+			free(b.out);
+	}
+
+	if (in_used != NULL)
+		*in_used += b.in_pos;
+
+	xz_dec_end(s);
+
+	switch (ret) {
+	case XZ_STREAM_END:
+		return 0;
+
+	case XZ_MEM_ERROR:
+		/* This can occur only in multi-call mode. */
+		error("XZ decompressor ran out of memory");
+		break;
+
+	case XZ_FORMAT_ERROR:
+		error("Input is not in the XZ format (wrong magic bytes)");
+		break;
+
+	case XZ_OPTIONS_ERROR:
+		error("Input was encoded with settings that are not "
+				"supported by this XZ decoder");
+		break;
+
+	case XZ_DATA_ERROR:
+	case XZ_BUF_ERROR:
+		error("XZ-compressed data is corrupt");
+		break;
+
+	default:
+		error("Bug in the XZ decompressor");
+		break;
+	}
+
+	return -1;
+
+error_alloc_in:
+	if (flush != NULL)
+		free(b.out);
+
+error_alloc_out:
+	xz_dec_end(s);
+
+error_alloc_state:
+	error("XZ decompressor ran out of memory");
+	return -1;
+}
+
+/*
+ * This function is used by architecture-specific files to decompress
+ * the kernel image.
+ */
+#ifdef XZ_PREBOOT
+STATIC int INIT __decompress(unsigned char *in, long in_size,
+			     long (*fill)(void *dest, unsigned long size),
+			     long (*flush)(void *src, unsigned long size),
+			     unsigned char *out, long out_size,
+			     long *in_used,
+			     void (*error)(char *x))
+{
+	return unxz(in, in_size, fill, flush, out, in_used, error);
+}
+#endif