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, 245 insertions, 0 deletions

diff --git a/fs/ntfs/logfile.h b/fs/ntfs/logfile.h
new file mode 100644
index 000000000..dadb56fd0
--- /dev/null
+++ b/fs/ntfs/logfile.h
@@ -0,0 +1,245 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+/*
+ * Defines for NTFS kernel journal (LogFile) handling.
+ *
+ * Copyright (c) 2000-2005 Anton Altaparmakov
+ */
+
+#ifndef _LINUX_NTFS_LOGFILE_H
+#define _LINUX_NTFS_LOGFILE_H
+
+#include "layout.h"
+
+/*
+ * Journal (LogFile) organization:
+ *
+ * Two restart areas present in the first two pages (restart pages, one restart
+ * area in each page).  When the volume is dismounted they should be identical,
+ * except for the update sequence array which usually has a different update
+ * sequence number.
+ *
+ * These are followed by log records organized in pages headed by a log record
+ * header going up to log file size.  Not all pages contain log records when a
+ * volume is first formatted, but as the volume ages, all records will be used.
+ * When the log file fills up, the records at the beginning are purged (by
+ * modifying the oldest_lsn to a higher value presumably) and writing begins
+ * at the beginning of the file.  Effectively, the log file is viewed as a
+ * circular entity.
+ *
+ * NOTE: Windows NT, 2000, and XP all use log file version 1.1 but they accept
+ * versions <= 1.x, including 0.-1.  (Yes, that is a minus one in there!)  We
+ * probably only want to support 1.1 as this seems to be the current version
+ * and we don't know how that differs from the older versions.  The only
+ * exception is if the journal is clean as marked by the two restart pages
+ * then it doesn't matter whether we are on an earlier version.  We can just
+ * reinitialize the logfile and start again with version 1.1.
+ */
+
+/* Some LogFile related constants. */
+#define MaxLogFileSize		0x100000000ULL
+#define DefaultLogPageSize	4096
+#define MinLogRecordPages	48
+
+/*
+ * Log file restart page header (begins the restart area).
+ *
+ * @magic: The magic is "RSTR".
+ * @usa_ofs: See ntfs_record struct definition in layout.h.  When creating,
+ *   set this to be immediately after this header structure (without any
+ *   alignment).
+ * @usa_count: See ntfs_record struct definition in layout.h.
+ * @chkdsk_lsn: The last log file sequence number found by chkdsk.  Only
+ *   used when the magic is changed to "CHKD".  Otherwise this is zero.
+ * @system_page_size: Byte size of system pages when the log file was
+ *   created, has to be >= 512 and a power of 2.  Use this to calculate
+ *   the required size of the usa (usa_count) and add it to usa_ofs. Then
+ *   verify that the result is less than the value of
+ *   the restart_area_offset.
+ * @log_page_size: Byte size of log file pages, has to be >= 512 and
+ *   a power of 2.  The default is 4096 and is used when the system page
+ *   size is between 4096 and 8192.  Otherwise this is set to the system
+ *   page size instead.
+ * @restart_area_offset: Byte offset from the start of this header to
+ *   the RESTART_AREA. Value has to be aligned to 8-byte boundary.  When
+ *   creating, set this to be after the usa.
+ * @minor_ver: Log file minor version.  Only check if major version is 1.
+ * @major_ver: Log file major version.  We only support version 1.1.
+ */
+struct restart_page_header {
+	__le32 magic;
+	__le16 usa_ofs;
+	__le16 usa_count;
+	__le64 chkdsk_lsn;
+	__le32 system_page_size;
+	__le32 log_page_size;
+	__le16 restart_area_offset;
+	__le16 minor_ver;
+	__le16 major_ver;
+} __packed;
+
+/*
+ * Constant for the log client indices meaning that there are no client records
+ * in this particular client array.  Also inside the client records themselves,
+ * this means that there are no client records preceding or following this one.
+ */
+#define LOGFILE_NO_CLIENT	cpu_to_le16(0xffff)
+#define LOGFILE_NO_CLIENT_CPU	0xffff
+
+/*
+ * These are the so far known RESTART_AREA_* flags (16-bit) which contain
+ * information about the log file in which they are present.
+ * gcc: Force enum bit width to 16.
+ */
+enum {
+	RESTART_VOLUME_IS_CLEAN	= cpu_to_le16(0x0002),
+	RESTART_SPACE_FILLER	= cpu_to_le16(0xffff),
+} __packed;
+
+/*
+ * Log file restart area record.  The offset of this record is found by adding
+ * the offset of the RESTART_PAGE_HEADER to the restart_area_offset value found
+ * in it.  See notes at restart_area_offset above.
+ *
+ * @current_lsn: The current, i.e. last LSN inside the log when
+ *   the restart area was last written.  This happens often but what is
+ *   the interval?  Is it just fixed time or is it every time a check point
+ *   is written or somethine else?  On create set to 0.
+ * @log_clients: Number of log client records in the array of log client
+ *   records which follows this restart area.  Must be 1.
+ * @client_free_list: The index of the first free log client record in
+ *   the array of log client records.  LOGFILE_NO_CLIENT means that there
+ *   are no free log client records in the array.  If != LOGFILE_NO_CLIENT,
+ *   check that log_clients > client_free_list.  On Win2k and presumably
+ *   earlier, on a clean volume this is != LOGFILE_NO_CLIENT, and it should
+ *   be 0, i.e. the first (and only) client record is free and thus
+ *   the logfile is closed and hence clean.  A dirty volume would have left
+ *   the logfile open and hence this would be LOGFILE_NO_CLIENT.  On WinXP
+ *   and presumably later, the logfile is always open, even on clean
+ *   shutdown so this should always be LOGFILE_NO_CLIENT.
+ * @client_in_use_list: The index of the first in-use log client record in
+ *   the array of log client records.  LOGFILE_NO_CLIENT means that there
+ *   are no in-use log client records in the array.
+ *   If != LOGFILE_NO_CLIENT check that log_clients > client_in_use_list.
+ *   On Win2k and presumably earlier, on a clean volume this is
+ *   LOGFILE_NO_CLIENT, i.e. there are no client records in use and thus
+ *   the logfile is closed and hence clean.  A dirty volume would have left
+ *   the logfile open and hence this would be != LOGFILE_NO_CLIENT, and it
+ *   should be 0, i.e. the first (and only) client record is in use.  On
+ *   WinXP and presumably later, the logfile is always open, even on clean
+ *   shutdown so this should always be 0.
+ * @flags: Flags modifying LFS behaviour.  On Win2k and presumably earlier
+ *   this is always 0.  On WinXP and presumably later, if the logfile was
+ *   shutdown cleanly, the second bit, RESTART_VOLUME_IS_CLEAN, is set.
+ *   This bit is cleared when the volume is mounted by WinXP and set when
+ *   the volume is dismounted, thus if the logfile is dirty, this bit is
+ *   clear.  Thus we don't need to check the Windows version to determine
+ *   if the logfile is clean.  Instead if the logfile is closed, we know
+ *   it must be clean.  If it is open and this bit is set, we also know
+ *   it must be clean.  If on the other hand the logfile is open and this
+ *   bit is clear, we can be almost certain that the logfile is dirty.
+ * @seq_number_bits: How many bits to use for the sequence number.  This
+ *   is calculated as 67 - the number of bits required to store the logfile
+ *   size in bytes and this can be used in with the specified file_size as
+ *   a consistency check.
+ * @restart_area_length: Length of the restart area including the client
+ *   array.  Following checks required if version matches.  Otherwise,
+ *   skip them.  restart_area_offset + restart_area_length has to be
+ *   <= system_page_size.  Also, restart_area_length has to be >=
+ *   client_array_offset + (log_clients * sizeof(log client record)).
+ * @client_array_offset: Offset from the start of this record to the first
+ *   log client record if versions are matched.  When creating, set this
+ *   to be after this restart area structure, aligned to 8-bytes boundary.
+ *   If the versions do not match, this is ignored and the offset is
+ *   assumed to be (sizeof(RESTART_AREA) + 7) &  ~7, i.e. rounded up to
+ *   first 8-byte boundary.  Either way, client_array_offset has to be
+ *   aligned to an 8-byte boundary.  Also, restart_area_offset +
+ *   client_array_offset has to be <= 510.  Finally, client_array_offset +
+ *   (log_clients * sizeof(log client record)) has to be <= system_page_size.
+ *   On Win2k and presumably earlier, this is 0x30, i.e. immediately
+ *   following this record.  On WinXP and presumably later, this is 0x40,
+ *   i.e. there are 16 extra bytes between this record and the client
+ *   array.  This probably means that the RESTART_AREA record is actually
+ *   bigger in WinXP and later.
+ * @file_size: Usable byte size of the log file.
+ *   If the restart_area_offset + the offset of the file_size are > 510
+ *   then corruption has occurred.  This is the very first check when
+ *   starting with the restart_area as if it fails it means that some of
+ *   the above values will be corrupted by the multi sector transfer
+ *   protection.  The file_size has to be rounded down to be a multiple
+ *   of the log_page_size in the RESTART_PAGE_HEADER and then it has to be
+ *   at least big enough to store the two restart pages and 48 (0x30) log
+ *   record pages.
+ * @last_lsn_data_length: Length of data of last LSN, not including the log
+ *   record header.  On create set to 0.
+ * @log_record_header_length: Byte size of the log record header.  If the
+ *   version matches then check that the value of log_record_header_length
+ *   is a multiple of 8, i.e. (log_record_header_length + 7) & ~7 ==
+ *   log_record_header_length.  When creating set it to
+ *   sizeof(LOG_RECORD_HEADER), aligned to 8 bytes.
+ * @log_page_data_offset: Offset to the start of data in a log record page.
+ *   Must be a multiple of 8.  On create set it to immediately after
+ *   the update sequence array of the log record page.
+ * @restart_log_open_count: A counter that gets incremented every time
+ *   the logfile is restarted which happens at mount time when the logfile
+ *   is opened. When creating set to a random value.  Win2k sets it to
+ *   the low 32 bits of the current system time in NTFS format (see time.h).
+ * @reserved: Reserved/alignment to 8-byte boundary.
+ */
+struct restart_area {
+	__le64 current_lsn;
+	__le16 log_clients;
+	__le16 client_free_list;
+	__le16 client_in_use_list;
+	__le16 flags;
+	__le32 seq_number_bits;
+	__le16 restart_area_length;
+	__le16 client_array_offset;
+	__le64 file_size;
+	__le32 last_lsn_data_length;
+	__le16 log_record_header_length;
+	__le16 log_page_data_offset;
+	__le32 restart_log_open_count;
+	__le32 reserved;
+} __packed;
+
+/*
+ * Log client record.  The offset of this record is found by adding the offset
+ * of the RESTART_AREA to the client_array_offset value found in it.
+ *
+ * @oldest_lsn: Oldest LSN needed by this client.  On create set to 0.
+ * @client_restart_lsn: LSN at which this client needs to restart
+ *   the volume, i.e. the current position within the log file.
+ *   At present, if clean this should = current_lsn in restart area but it
+ *   probably also = current_lsn when dirty most of the time.
+ *   At create set to 0.
+ * @prev_client: The offset to the previous log client record in the array
+ *   of log client records.  LOGFILE_NO_CLIENT means there is no previous
+ *   client record, i.e. this is the first one.  This is always
+ *   LOGFILE_NO_CLIENT.
+ * @next_client: The offset to the next log client record in the array of
+ *   log client records.  LOGFILE_NO_CLIENT means there are no next client
+ *   records, i.e. this is the last one.  This is always LOGFILE_NO_CLIENT.
+ * @seq_number: On Win2k and presumably earlier, this is set to zero every
+ *   time the logfile is restarted and it is incremented when the logfile
+ *   is closed at dismount time.  Thus it is 0 when dirty and 1 when clean.
+ *   On WinXP and presumably later, this is always 0.
+ * @reserved[6]: Reserved/alignment.
+ * @client_name_length: Length of client name in bytes.  Should always be 8.
+ * @client_name[64]: Name of the client in Unicode. Should always be "NTFS"
+ *   with the remaining bytes set to 0.
+ */
+struct log_client_record {
+	__le64 oldest_lsn;
+	__le64 client_restart_lsn;
+	__le16 prev_client;
+	__le16 next_client;
+	__le16 seq_number;
+	u8 reserved[6];
+	__le32 client_name_length;
+	__le16 client_name[64];
+} __packed;
+
+bool ntfs_check_logfile(struct inode *log_vi,
+		struct restart_page_header **rp);
+bool ntfs_empty_logfile(struct inode *log_vi);
+#endif /* _LINUX_NTFS_LOGFILE_H */