summaryrefslogtreecommitdiffstats
path: root/Documentation/fault-injection/provoke-crashes.rst
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/fault-injection/provoke-crashes.rst')
-rw-r--r--Documentation/fault-injection/provoke-crashes.rst57
1 files changed, 57 insertions, 0 deletions
diff --git a/Documentation/fault-injection/provoke-crashes.rst b/Documentation/fault-injection/provoke-crashes.rst
new file mode 100644
index 0000000000..1f087e502c
--- /dev/null
+++ b/Documentation/fault-injection/provoke-crashes.rst
@@ -0,0 +1,57 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+============================================================
+Provoking crashes with Linux Kernel Dump Test Module (LKDTM)
+============================================================
+
+The lkdtm module provides an interface to disrupt (and usually crash)
+the kernel at predefined code locations to evaluate the reliability of
+the kernel's exception handling and to test crash dumps obtained using
+different dumping solutions. The module uses KPROBEs to instrument the
+trigger location, but can also trigger the kernel directly without KPROBE
+support via debugfs.
+
+You can select the location of the trigger ("crash point name") and the
+type of action ("crash point type") either through module arguments when
+inserting the module, or through the debugfs interface.
+
+Usage::
+
+ insmod lkdtm.ko [recur_count={>0}] cpoint_name=<> cpoint_type=<>
+ [cpoint_count={>0}]
+
+recur_count
+ Recursion level for the stack overflow test. By default this is
+ dynamically calculated based on kernel configuration, with the
+ goal of being just large enough to exhaust the kernel stack. The
+ value can be seen at `/sys/module/lkdtm/parameters/recur_count`.
+
+cpoint_name
+ Where in the kernel to trigger the action. It can be
+ one of INT_HARDWARE_ENTRY, INT_HW_IRQ_EN, INT_TASKLET_ENTRY,
+ FS_SUBMIT_BH, MEM_SWAPOUT, TIMERADD, SCSI_QUEUE_RQ, or DIRECT.
+
+cpoint_type
+ Indicates the action to be taken on hitting the crash point.
+ These are numerous, and best queried directly from debugfs. Some
+ of the common ones are PANIC, BUG, EXCEPTION, LOOP, and OVERFLOW.
+ See the contents of `/sys/kernel/debug/provoke-crash/DIRECT` for
+ a complete list.
+
+cpoint_count
+ Indicates the number of times the crash point is to be hit
+ before triggering the action. The default is 10 (except for
+ DIRECT, which always fires immediately).
+
+You can also induce failures by mounting debugfs and writing the type to
+<debugfs>/provoke-crash/<crashpoint>. E.g.::
+
+ mount -t debugfs debugfs /sys/kernel/debug
+ echo EXCEPTION > /sys/kernel/debug/provoke-crash/INT_HARDWARE_ENTRY
+
+The special file `DIRECT` will induce the action directly without KPROBE
+instrumentation. This mode is the only one available when the module is
+built for a kernel without KPROBEs support::
+
+ # Instead of having a BUG kill your shell, have it kill "cat":
+ cat <(echo WRITE_RO) >/sys/kernel/debug/provoke-crash/DIRECT