mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
synced 2026-04-03 04:05:33 -04:00
117eab5c6e
Pull coredump updates from Christian Brauner:
"This contains an extension to the coredump socket and a proper rework
of the coredump code.
- This extends the coredump socket to allow the coredump server to
tell the kernel how to process individual coredumps. This allows
for fine-grained coredump management. Userspace can decide to just
let the kernel write out the coredump, or generate the coredump
itself, or just reject it.
* COREDUMP_KERNEL
The kernel will write the coredump data to the socket.
* COREDUMP_USERSPACE
The kernel will not write coredump data but will indicate to the
parent that a coredump has been generated. This is used when
userspace generates its own coredumps.
* COREDUMP_REJECT
The kernel will skip generating a coredump for this task.
* COREDUMP_WAIT
The kernel will prevent the task from exiting until the coredump
server has shutdown the socket connection.
The flexible coredump socket can be enabled by using the "@@"
prefix instead of the single "@" prefix for the regular coredump
socket:
@@/run/systemd/coredump.socket
- Cleanup the coredump code properly while we have to touch it
anyway.
Split out each coredump mode in a separate helper so it's easy to
grasp what is going on and make the code easier to follow. The core
coredump function should now be very trivial to follow"
* tag 'vfs-6.17-rc1.coredump' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs: (31 commits)
cleanup: add a scoped version of CLASS()
coredump: add coredump_skip() helper
coredump: avoid pointless variable
coredump: order auto cleanup variables at the top
coredump: add coredump_cleanup()
coredump: auto cleanup prepare_creds()
cred: add auto cleanup method
coredump: directly return
coredump: auto cleanup argv
coredump: add coredump_write()
coredump: use a single helper for the socket
coredump: move pipe specific file check into coredump_pipe()
coredump: split pipe coredumping into coredump_pipe()
coredump: move core_pipe_count to global variable
coredump: prepare to simplify exit paths
coredump: split file coredumping into coredump_file()
coredump: rename do_coredump() to vfs_coredump()
selftests/coredump: make sure invalid paths are rejected
coredump: validate socket path in coredump_parse()
coredump: don't allow ".." in coredump socket path
...
coredump selftest
=================
Background context
------------------
`coredump` is a feature which dumps a process's memory space when the process terminates
unexpectedly (e.g. due to segmentation fault), which can be useful for debugging. By default,
`coredump` dumps the memory to the file named `core`, but this behavior can be changed by writing a
different file name to `/proc/sys/kernel/core_pattern`. Furthermore, `coredump` can be piped to a
user-space program by writing the pipe symbol (`|`) followed by the command to be executed to
`/proc/sys/kernel/core_pattern`. For the full description, see `man 5 core`.
The piped user program may be interested in reading the stack pointers of the crashed process. The
crashed process's stack pointers can be read from `procfs`: it is the `kstkesp` field in
`/proc/$PID/stat`. See `man 5 proc` for all the details.
The problem
-----------
While a thread is active, the stack pointer is unsafe to read and therefore the `kstkesp` field
reads zero. But when the thread is dead (e.g. during a coredump), this field should have valid
value.
However, this was broken in the past and `kstkesp` was zero even during coredump:
* commit 0a1eb2d474ed ("fs/proc: Stop reporting eip and esp in /proc/PID/stat") changed kstkesp to
always be zero
* commit fd7d56270b52 ("fs/proc: Report eip/esp in /prod/PID/stat for coredumping") fixed it for the
coredumping thread. However, other threads in a coredumping process still had the problem.
* commit cb8f381f1613 ("fs/proc/array.c: allow reporting eip/esp for all coredumping threads") fixed
for all threads in a coredumping process.
* commit 92307383082d ("coredump: Don't perform any cleanups before dumping core") broke it again
for the other threads in a coredumping process.
The problem has been fixed now, but considering the history, it may appear again in the future.
The goal of this test
---------------------
This test detects problem with reading `kstkesp` during coredump by doing the following:
#. Tell the kernel to execute the "stackdump" script when a coredump happens. This script
reads the stack pointers of all threads of crashed processes.
#. Spawn a child process who creates some threads and then crashes.
#. Read the output from the "stackdump" script, and make sure all stack pointer values are
non-zero.