On x86-32, with CONFIG_FIRMWARE and multiple CPUs, if you enable function
graph tracing and then suspend to RAM, it will triple fault and reboot when
it resumes.
The first fault happens when booting a secondary CPU:
The early head_32.S code calls into load_ucode_ap(), which has an an
ftrace hook, so it calls prepare_ftrace_return(), which calls
ftrace_graph_is_dead(), which tries to access the global
'kill_ftrace_graph' variable with a virtual address, causing a fault
because the CPU is still in real mode.
The fix is to add a check in prepare_ftrace_return() to make sure it's
running in protected mode before continuing. The check makes sure the
stack pointer is a virtual kernel address. It's a bit of a hack, but
it's not very intrusive and it works well enough.
For reference, here are a few other (more difficult) ways this could
have potentially been fixed:
- Move startup_32_smp()'s call to load_ucode_ap() down to *after* paging
is enabled. (No idea what that would break.)
- Track down load_ucode_ap()'s entire callee tree and mark all the
functions 'notrace'. (Probably not realistic.)
- Pause graph tracing in ftrace_suspend_notifier_call() or bringup_cpu()
or __cpu_up(), and ensure that the pause facility can be queried from
real mode.
A client can append random data to the end of an NFSv2 or NFSv3 RPC call
without our complaining; we'll just stop parsing at the end of the
expected data and ignore the rest.
Encoded arguments and replies are stored together in an array of pages,
and if a call is too large it could leave inadequate space for the
reply. This is normally OK because NFS RPC's typically have either
short arguments and long replies (like READ) or long arguments and short
replies (like WRITE). But a client that sends an incorrectly long reply
can violate those assumptions. This was observed to cause crashes.
Also, several operations increment rq_next_page in the decode routine
before checking the argument size, which can leave rq_next_page pointing
well past the end of the page array, causing trouble later in
svc_free_pages.
So, following a suggestion from Neil Brown, add a central check to
enforce our expectation that no NFSv2/v3 call has both a large call and
a large reply.
As followup we may also want to rewrite the encoding routines to check
more carefully that they aren't running off the end of the page array.
We may also consider rejecting calls that have any extra garbage
appended. That would be safer, and within our rights by spec, but given
the age of our server and the NFS protocol, and the fact that we've
never enforced this before, we may need to balance that against the
possibility of breaking some oddball client.
Reported-by: Tuomas Haanpää <email address hidden>
Reported-by: Ari Kauppi <email address hidden>
Reviewed-by: NeilBrown <email address hidden>
Signed-off-by: J. Bruce Fields <email address hidden>
Signed-off-by: Jiri Slaby <email address hidden>
KGDB is a kernel debug stub and it can't be used to debug userland as it
can only safely access kernel memory.
On MIPS however KGDB has always got the register state of sleeping
processes from the userland register context at the beginning of the
kernel stack. This is meaningless for kernel threads (which never enter
userland), and for user threads it prevents the user seeing what it is
doing while in the kernel:
(gdb) info threads
Id Target Id Frame
...
3 Thread 2 (kthreadd) 0x0000000000000000 in ?? ()
2 Thread 1 (init) 0x000000007705c4b4 in ?? ()
1 Thread -2 (shadowCPU0) 0xffffffff8012524c in arch_kgdb_breakpoint () at arch/mips/kernel/kgdb.c:201
Get the register state instead from the (partial) kernel register
context stored in the task's thread_struct for resume() to restore. All
threads now correctly appear to be in context_switch():
(gdb) info threads
Id Target Id Frame
...
3 Thread 2 (kthreadd) context_switch (rq=<optimized out>, cookie=..., next=<optimized out>, prev=0x0) at kernel/sched/core.c:2903
2 Thread 1 (init) context_switch (rq=<optimized out>, cookie=..., next=<optimized out>, prev=0x0) at kernel/sched/core.c:2903
1 Thread -2 (shadowCPU0) 0xffffffff8012524c in arch_kgdb_breakpoint () at arch/mips/kernel/kgdb.c:201
Call clobbered registers which aren't saved and exception registers
(BadVAddr & Cause) which can't be easily determined without stack
unwinding are reported as 0. The PC is taken from the return address,
such that the state presented matches that found immediately after
returning from resume().
Fixes: 8854700115ec ("[MIPS] kgdb: add arch support for the kernel's kgdb core")
Signed-off-by: James Hogan <email address hidden>
Cc: Jason Wessel <email address hidden>
Cc: <email address hidden>
Patchwork: https://patchwork.linux-mips.org/patch/15829/
Signed-off-by: Ralf Baechle <email address hidden>
Signed-off-by: Jiri Slaby <email address hidden>
The snd_use_lock_sync() (thus its implementation
snd_use_lock_sync_helper()) has the 5 seconds timeout to break out of
the sync loop. It was introduced from the beginning, just to be
"safer", in terms of avoiding the stupid bugs.
However, as Ben Hutchings suggested, this timeout rather introduces a
potential leak or use-after-free that was apparently fixed by the
commit 2d7d54002e39 ("ALSA: seq: Fix race during FIFO resize"):
for example, snd_seq_fifo_event_in() -> snd_seq_event_dup() ->
copy_from_user() could block for a long time, and snd_use_lock_sync()
goes timeout and still leaves the cell at releasing the pool.
For fixing such a problem, we remove the break by the timeout while
still keeping the warning.
On slow platforms with unreliable TSC, such as QEMU emulated machines,
it is possible for the kernel to request the next event in the past. In
that case, in the current implementation of xen_vcpuop_clockevent, we
simply return -ETIME. To be precise the Xen returns -ETIME and we pass
it on. However the result of this is a missed event, which simply causes
the kernel to hang.
Instead it is better to always ask the hypervisor for a timer event,
even if the timeout is in the past. That way there are no lost
interrupts and the kernel survives. To do that, remove the
VCPU_SSHOTTMR_future flag.
Two different threads with different rds sockets may be in
rds_recv_rcvbuf_delta() via receive path. If their ports
both map to the same word in the congestion map, then
using non-atomic ops to update it could cause the map to
be incorrect. Lets use atomics to avoid such an issue.
Full credit to Wengang <email address hidden> for
finding the issue, analysing it and also pointing out
to offending code with spin lock based fix.
Reviewed-by: Leon Romanovsky <email address hidden>
Signed-off-by: Wengang Wang <email address hidden>
Signed-off-by: Santosh Shilimkar <email address hidden>
Signed-off-by: David S. Miller <email address hidden>
Cc: Julia Lawall <email address hidden>
Signed-off-by: Jiri Slaby <email address hidden>
a2c54e1...
by
Corey Minyard <email address hidden>
As part of handling a crash on an SMP system, an IPI is send to
all other CPUs to save their current registers and stop. It was
using task_pt_regs(current) to get the registers, but that will
only be accurate if the CPU was interrupted running in userland.
Instead allow the architecture to pass in the registers (all
pass NULL now, but allow for the future) and then use get_irq_regs()
which should be accurate as we are in an interrupt. Fall back to
task_pt_regs(current) if nothing else is available.