Nysal reported that userspace backtraces are missing in offcputime bcc
tool. As an example:
$ sudo ./bcc/tools/offcputime.py -uU
Tracing off-CPU time (us) of user threads by user stack... Hit Ctrl-C to end.
^C
write
- python (9107)
8
write
- sudo (9105)
9
mmap
- python (9107)
16
clock_nanosleep
- multipathd (697)
3001604
The offcputime bcc tool attaches a bpf program to a kprobe on
finish_task_switch(), which is usually hit on a syscall from userspace.
With the switch to system call vectored, we started setting
pt_regs->link to zero. This is because system call vectored behaves like
a function call with LR pointing to the system call return address, and
with no modification to SRR0/SRR1. The LR value does indicate our next
instruction, so it is being saved as pt_regs->nip, and pt_regs->link is
being set to zero. This is not a problem by itself, but BPF uses perf
callchain infrastructure for capturing stack traces, and that stores LR
as the second entry in the stack trace. perf has code to cope with the
second entry being zero, and skips over it. However, generic userspace
unwinders assume that a zero entry indicates end of the stack trace,
resulting in a truncated userspace stack trace.
Rather than fixing all userspace unwinders to ignore/skip past the
second entry, store the real LR value in pt_regs->link so that there
continues to be a valid, though duplicate entry in the stack trace.
With this change:
$ sudo ./bcc/tools/offcputime.py -uU
Tracing off-CPU time (us) of user threads by user stack... Hit Ctrl-C to end.
^C
write
write
[unknown]
[unknown]
[unknown]
[unknown]
[unknown]
PyObject_VectorcallMethod
[unknown]
[unknown]
PyObject_CallOneArg
PyFile_WriteObject
PyFile_WriteString
[unknown]
[unknown]
PyObject_Vectorcall
_PyEval_EvalFrameDefault
PyEval_EvalCode
[unknown]
[unknown]
[unknown]
_PyRun_SimpleFileObject
_PyRun_AnyFileObject
Py_RunMain
[unknown]
Py_BytesMain
[unknown]
__libc_start_main
- python (1293)
7
write
write
[unknown]
sudo_ev_loop_v1
sudo_ev_dispatch_v1
[unknown]
[unknown]
[unknown]
[unknown]
__libc_start_main
- sudo (1291)
7
syscall
syscall
bpf_open_perf_buffer_opts
[unknown]
[unknown]
[unknown]
[unknown]
_PyObject_MakeTpCall
PyObject_Vectorcall
_PyEval_EvalFrameDefault
PyEval_EvalCode
[unknown]
[unknown]
[unknown]
_PyRun_SimpleFileObject
_PyRun_AnyFileObject
Py_RunMain
[unknown]
Py_BytesMain
[unknown]
__libc_start_main
- python (1293)
11
clock_nanosleep
clock_nanosleep
nanosleep
sleep
[unknown]
[unknown]
__clone
- multipathd (698)
3001661
Fixes: 7fa95f9ada ("powerpc/64s: system call support for scv/rfscv instructions")
Cc: stable@vger.kernel.org
Reported-by: "Nysal Jan K.A" <nysal@linux.ibm.com>
Signed-off-by: Naveen N Rao <naveen@kernel.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20240202154316.395276-1-naveen@kernel.org
This macro is to be used in assembly where C functions are called.
pcrel addressing mode requires branches to functions with a
localentry value of 1 to have either a trailing nop or @notoc.
This macro permits the latter without changing callers.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Add dummy definitions to fix selftests build]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230408021752.862660-5-npiggin@gmail.com
Kernel created user threads start similarly to kernel threads in that
they call a kernel function after first returning from _switch, so
they share ret_from_kernel_thread for this. Kernel threads never return
from that function though, whereas user threads often do (although some
don't, e.g., IO threads).
Split these startup functions in two, and catch kernel threads that
improperly return from their function. This is intended to make the
complicated code a little bit easier to understand.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230325122904.2375060-7-npiggin@gmail.com
The kernel thread path in copy_thread creates a user interrupt frame on
stack and stores the function and arg parameters there, and
ret_from_kernel_thread loads them. This is a slightly confusing way to
overload that frame. Non-volatile registers are loaded from the switch
frame, so the parameters can be stored there. The user interrupt frame
is now only used by user threads when they return to user.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230325122904.2375060-4-npiggin@gmail.com
Zeroise user state in gprs (assign to zero) to reduce the influence of user
registers on speculation within kernel syscall handlers. Clears occur
at the very beginning of the sc and scv 0 interrupt handlers, with
restores occurring following the execution of the syscall handler.
Zeroise GPRS r0, r2-r11, r14-r31, on entry into the kernel for all
other interrupt sources. The remaining gprs are overwritten by
entry macros to interrupt handlers, irrespective of whether or not a
given handler consumes these register values. If an interrupt does not
select the IMSR_R12 IOption, zeroise r12.
Prior to this commit, r14-r31 are restored on a per-interrupt basis at
exit, but now they are always restored on 64bit Book3S. Remove explicit
REST_NVGPRS invocations on 64-bit Book3S. 32-bit systems do not clear
user registers on interrupt, and continue to depend on the return value
of interrupt_exit_user_prepare to determine whether or not to restore
non-volatiles.
The mmap_bench benchmark in selftests should rapidly invoke pagefaults.
See ~0.8% performance regression with this mitigation, but this
indicates the worst-case performance due to heavier-weight interrupt
handlers. This mitigation is able to be enabled/disabled through
CONFIG_INTERRUPT_SANITIZE_REGISTERS.
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221201071019.1953023-5-rmclure@linux.ibm.com
Interrupt code is shared between Book3E/S 64-bit systems for interrupt
handlers. Ensure that exit code correctly restores non-volatile gprs on
each system when CONFIG_INTERRUPT_SANITIZE_REGISTERS is enabled.
Also introduce macros for clearing/restoring registers on interrupt
entry for when this configuration option is either disabled or enabled.
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221201071019.1953023-3-rmclure@linux.ibm.com
Commit a4cb3651a1 ("powerpc/64s/interrupt: Fix lost interrupts when
returning to soft-masked context") fixed the problem of pending irqs
being cleared when clearing the HARD_DIS bit, but then it didn't clear
the bit at all. This change clears HARD_DIS without affecting other bits
in the mask.
When an interrupt hits in a soft-masked section that has MSR[EE]=1, it
can hard disable and set PACA_IRQS_HARD_DIS, which must be cleared when
returning to the EE=1 caller (unless it was set due to a MUST_HARD_MASK
interrupt becoming pending). Failure to clear this leaves the
returned-to context running with MSR[EE]=1 and PACA_IRQS_HARD_DIS, which
confuses irq assertions and could be dangerous for code that might test
the flag.
This was observed in a hash MMU kernel where a kernel hash fault hits in
a local_irqs_disabled region that has EE=1. The hash fault also runs
with EE=1, then as it returns, a decrementer hits in the restart section
and the irq restart code hard-masks which sets the PACA_IRQ_HARD_DIS
flag, which is not clear when the original context is returned to.
Reported-by: Sachin Sant <sachinp@linux.ibm.com>
Fixes: a4cb3651a1 ("powerpc/64s/interrupt: Fix lost interrupts when returning to soft-masked context")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Sachin Sant <sachinp@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221022052207.471328-1-npiggin@gmail.com
It's possible for an interrupt returning to an irqs-disabled context to
lose a pending soft-masked irq because it branches to part of the exit
code for irqs-enabled contexts, which is meant to clear only the
PACA_IRQS_HARD_DIS flag from PACAIRQHAPPENED by zeroing the byte. This
just looks like a simple thinko from a recent commit (if there was no
hard mask pending, there would be no reason to clear it anyway).
This also adds comment to the code that actually does need to clear the
flag.
Fixes: e485f6c751 ("powerpc/64/interrupt: Fix return to masked context after hard-mask irq becomes pending")
Reported-by: Sachin Sant <sachinp@linux.ibm.com>
Reported-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221013064418.1311104-1-npiggin@gmail.com
Using a 32-bit constant for this marker allows it to be loaded with
two ALU instructions, like 32-bit. This avoids a TOC entry and a
TOC load that depends on the r2 value that has just been loaded from
the PACA.
This changes the value for 32-bit as well, so both have the same
value in the low 4 bytes and 64-bit has 0 in the top bytes.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220926034057.2360083-2-npiggin@gmail.com
If a synchronous interrupt (e.g., hash fault) is taken inside an
irqs-disabled region which has MSR[EE]=1, then an asynchronous interrupt
that is PACA_IRQ_MUST_HARD_MASK (e.g., PMI) is taken inside the
synchronous interrupt handler, then the synchronous interrupt will
return with MSR[EE]=1 and the asynchronous interrupt fires again.
If the asynchronous interrupt is a PMI and the original context does not
have PMIs disabled (only Linux IRQs), the asynchronous interrupt will
fire despite having the PMI marked soft pending. This can confuse the
perf code and cause warnings.
This patch changes the interrupt return so that irqs-disabled MSR[EE]=1
contexts will be returned to with MSR[EE]=0 if a PACA_IRQ_MUST_HARD_MASK
interrupt has become pending in the meantime.
The longer explanation for what happens:
1. local_irq_disable()
2. Hash fault interrupt fires, do_hash_fault handler runs
3. interrupt_enter_prepare() sets IRQS_ALL_DISABLED
4. interrupt_enter_prepare() sets MSR[EE]=1
5. PMU interrupt fires, masked handler runs
6. Masked handler marks PMI pending
7. Masked handler returns with PACA_IRQ_HARD_DIS set, MSR[EE]=0
8. do_hash_fault interrupt return handler runs
9. interrupt_exit_kernel_prepare() clears PACA_IRQ_HARD_DIS
10. interrupt returns with MSR[EE]=1
11. PMU interrupt fires, perf handler runs
Fixes: 4423eb5ae3 ("powerpc/64/interrupt: make normal synchronous interrupts enable MSR[EE] if possible")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220926054305.2671436-4-npiggin@gmail.com
Change system_call_exception arguments to pass a pointer to a stack
frame container caller state, as well as the original r0, which
determines the number of the syscall. This has been observed to yield
improved performance to passing them by registers, circumventing the
need to allocate a stack frame.
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Retain clearing of high bits of args for compat tasks]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220921065605.1051927-21-rmclure@linux.ibm.com
Use the convenience macros for saving/clearing/restoring gprs in keeping
with syscall calling conventions. The plural variants of these macros
can store a range of registers for concision.
This works well when the user gpr value we are hoping to save is still
live. In the syscall interrupt handlers, user register state is
sometimes juggled between registers. Hold-off from issuing the SAVE_GPR
macro for applicable neighbouring lines to highlight the delicate
register save logic.
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220921065605.1051927-5-rmclure@linux.ibm.com
This reverts commit 8875f47b76 ("powerpc/syscall: Save r3 in regs->orig_r3
").
Save caller's original r3 state to the kernel stackframe before entering
system_call_exception. This allows for user registers to be cleared by
the time system_call_exception is entered, reducing the influence of
user registers on speculation within the kernel.
Prior to this commit, orig_r3 was saved at the beginning of
system_call_exception. Instead, save orig_r3 while the user value is
still live in r3.
Also replicate this early save in 32-bit. A similar save was removed in
commit 6f76a01173 ("powerpc/syscall: implement system call entry/exit
logic in C for PPC32") when 32-bit adopted system_call_exception. Revert
its removal of orig_r3 saves.
Signed-off-by: Rohan McLure <rmclure@linux.ibm.com>
Reviewed-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220921065605.1051927-3-rmclure@linux.ibm.com
In interrupt_64.S, formerly entry_64.S, there are two toc entries
created for sys_call_table and compat_sys_call_table.
These are no longer used, since the system call entry was converted from
asm to C, so remove them.
Fixes: 68b34588e2 ("powerpc/64/sycall: Implement syscall entry/exit logic in C")
Acked-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220913124545.2817825-1-mpe@ellerman.id.au
emulate_step() instruction emulation including sc instruction emulation
initially appeared in xmon. It was then moved into sstep.c where kprobes
could use it too, and later hw_breakpoint and uprobes started to use it.
Until uprobes, the only instruction emulation users were for kernel
mode instructions.
- xmon only steps / breaks on kernel addresses.
- kprobes is kernel only.
- hw_breakpoint only emulates kernel instructions, single steps user.
At one point, there was support for the kernel to execute sc
instructions, although that is long removed and it's not clear whether
there were any in-tree users. So system call emulation is not required
by the above users.
uprobes uses emulate_step and it appears possible to emulate sc
instruction in userspace. Userspace system call emulation is broken and
it's not clear it ever worked well.
The big complication is that userspace takes an interrupt to the kernel
to emulate the instruction. The user->kernel interrupt sets up registers
and interrupt stack frame expecting to return to userspace, then system
call instruction emulation re-directs that stack frame to the kernel,
early in the system call interrupt handler. This means the interrupt
return code takes the kernel->kernel restore path, which does not
restore everything as the system call interrupt handler would expect
coming from userspace. regs->iamr appears to get lost for example,
because the kernel->kernel return does not restore the user iamr.
Accounting such as irqflags tracing and CPU accounting does not get
flipped back to user mode as the system call handler expects, so those
appear to enter the kernel twice without returning to userspace.
These things may be individually fixable with various complication, but
it is a big complexity for unclear real benefit.
Furthermore, it is not possible to single step a system call instruction
since it causes an interrupt. As such, a separate patch disables probing
on system call instructions.
This patch removes system call emulation and disables stepping system
calls.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[minor commit log edit, and also get rid of '#ifdef CONFIG_PPC64']
Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/a412e3b3791ed83de18704c8d90f492e7a0049c0.1648648712.git.naveen.n.rao@linux.vnet.ibm.com
Commit 314f6c23dd ("powerpc/64s: Mask NIP before checking against
SRR0") masked off the low 2 bits of the NIP value in the interrupt
stack frame in case they are non-zero and mis-compare against a SRR0
register value of a CPU which always reads back 0 from the 2 low bits
which are reserved.
This now causes the opposite problem that an implementation which does
implement those bits in SRR0 will mis-compare against the masked NIP
value in which they have been cleared. QEMU is one such implementation,
and this is allowed by the architecture.
This can be triggered by sigfuz by setting low bits of PT_NIP in the
signal context.
Fix this for now by masking the SRR0 bits as well. Cleaner is probably
to sanitise these values before putting them in registers or stack, but
this is the quick and backportable fix.
Fixes: 314f6c23dd ("powerpc/64s: Mask NIP before checking against SRR0")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20220117134403.2995059-1-npiggin@gmail.com
When CONFIG_PPC_RFI_SRR_DEBUG=y we check the SRR values before returning
from interrupts. This is done in asm using EMIT_BUG_ENTRY, and passing
BUGFLAG_WARNING.
However that fails to create an exception table entry for the warning,
and so do_program_check() fails the exception table search and proceeds
to call _exception(), resulting in an oops like:
Oops: Exception in kernel mode, sig: 5 [#1]
LE PAGE_SIZE=64K MMU=Radix SMP NR_CPUS=2048 NUMA pSeries
Modules linked in:
CPU: 2 PID: 1204 Comm: sigreturn_unali Tainted: P 5.16.0-rc2-00194-g91ca3d4f77c5 #12
NIP: c00000000000c5b0 LR: 0000000000000000 CTR: 0000000000000000
...
NIP [c00000000000c5b0] system_call_common+0x150/0x268
LR [0000000000000000] 0x0
Call Trace:
[c00000000db73e10] [c00000000000c558] system_call_common+0xf8/0x268 (unreliable)
...
Instruction dump:
7cc803a6 888d0931 2c240000 4082001c 38800000 988d0931 e8810170 e8a10178
7c9a03a6 7cbb03a6 7d7a02a6 e9810170 <7f0b6088> 7d7b02a6 e9810178 7f0b6088
We should instead use EMIT_WARN_ENTRY, which creates an exception table
entry for the warning, allowing the warning to be correctly recognised,
and the code to resume after printing the warning.
Note however that because this warning is buried deep in the interrupt
return path, we are not able to recover from it (due to MSR_RI being
clear), so we still end up in die() with an unrecoverable exception.
Fixes: 59dc5bfca0 ("powerpc/64s: avoid reloading (H)SRR registers if they are still valid")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211221135101.2085547-2-mpe@ellerman.id.au
