linux.git/arch/arm64/kernel, branch v5.10

Merge tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux

2020-12-02T20:27:37+00:00

Pull arm64 fixes from Will Deacon:
 "I'm sad to say that we've got an unusually large arm64 fixes pull for
  rc7 which addresses numerous significant instrumentation issues with
  our entry code.

  Without these patches, lockdep is hopelessly unreliable in some
  configurations [1,2] and syzkaller is therefore not a lot of use
  because it's so noisy.

  Although much of this has always been broken, it appears to have been
  exposed more readily by other changes such as 044d0d6de9f5 ("lockdep:
  Only trace IRQ edges") and general lockdep improvements around IRQ
  tracing and NMIs.

  Fixing this properly required moving much of the instrumentation hooks
  from our entry assembly into C, which Mark has been working on for the
  last few weeks. We're not quite ready to move to the recently added
  generic functions yet, but the code here has been deliberately written
  to mimic that closely so we can look at cleaning things up once we
  have a bit more breathing room.

  Having said all that, the second version of these patches was posted
  last week and I pushed it into our CI (kernelci and cki) along with a
  commit which forced on PROVE_LOCKING, NOHZ_FULL and
  CONTEXT_TRACKING_FORCE. The result? We found a real bug in the
  md/raid10 code [3].

  Oh, and there's also a really silly typo patch that's unrelated.

  Summary:

   - Fix numerous issues with instrumentation and exception entry

   - Fix hideous typo in unused register field definition"

[1] https://lore.kernel.org/r/CACT4Y+aAzoJ48Mh1wNYD17pJqyEcDnrxGfApir=-j171TnQXhw@mail.gmail.com
[2] https://lore.kernel.org/r/20201119193819.GA2601289@elver.google.com
[3] https://lore.kernel.org/r/94c76d5e-466a-bc5f-e6c2-a11b65c39f83@redhat.com

* tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux:
  arm64: mte: Fix typo in macro definition
  arm64: entry: fix EL1 debug transitions
  arm64: entry: fix NMI {user, kernel}->kernel transitions
  arm64: entry: fix non-NMI kernel<->kernel transitions
  arm64: ptrace: prepare for EL1 irq/rcu tracking
  arm64: entry: fix non-NMI user<->kernel transitions
  arm64: entry: move el1 irq/nmi logic to C
  arm64: entry: prepare ret_to_user for function call
  arm64: entry: move enter_from_user_mode to entry-common.c
  arm64: entry: mark entry code as noinstr
  arm64: mark idle code as noinstr
  arm64: syscall: exit userspace before unmasking exceptions

arm64: entry: fix EL1 debug transitions

2020-11-30T12:11:38+00:00

In debug_exception_enter() and debug_exception_exit() we trace hardirqs
on/off while RCU isn't guaranteed to be watching, and we don't save and
restore the hardirq state, and so may return with this having changed.

Handle this appropriately with new entry/exit helpers which do the bare
minimum to ensure this is appropriately maintained, without marking
debug exceptions as NMIs. These are placed in entry-common.c with the
other entry/exit helpers.

In future we'll want to reconsider whether some debug exceptions should
be NMIs, but this will require a significant refactoring, and for now
this should prevent issues with lockdep and RCU.

Signed-off-by: Mark Rutland 
Cc: Catalin Marins 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-12-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: fix NMI {user, kernel}->kernel transitions

2020-11-30T12:11:38+00:00

Exceptions which can be taken at (almost) any time are consdiered to be
NMIs. On arm64 that includes:

* SDEI events
* GICv3 Pseudo-NMIs
* Kernel stack overflows
* Unexpected/unhandled exceptions

... but currently debug exceptions (BRKs, breakpoints, watchpoints,
single-step) are not considered NMIs.

As these can be taken at any time, kernel features (lockdep, RCU,
ftrace) may not be in a consistent kernel state. For example, we may
take an NMI from the idle code or partway through an entry/exit path.

While nmi_enter() and nmi_exit() handle most of this state, notably they
don't save/restore the lockdep state across an NMI being taken and
handled. When interrupts are enabled and an NMI is taken, lockdep may
see interrupts become disabled within the NMI code, but not see
interrupts become enabled when returning from the NMI, leaving lockdep
believing interrupts are disabled when they are actually disabled.

The x86 code handles this in idtentry_{enter,exit}_nmi(), which will
shortly be moved to the generic entry code. As we can't use either yet,
we copy the x86 approach in arm64-specific helpers. All the NMI
entrypoints are marked as noinstr to prevent any instrumentation
handling code being invoked before the state has been corrected.

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-11-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: fix non-NMI kernel<->kernel transitions

2020-11-30T12:11:38+00:00

There are periods in kernel mode when RCU is not watching and/or the
scheduler tick is disabled, but we can still take exceptions such as
interrupts. The arm64 exception handlers do not account for this, and
it's possible that RCU is not watching while an exception handler runs.

The x86/generic entry code handles this by ensuring that all (non-NMI)
kernel exception handlers call irqentry_enter() and irqentry_exit(),
which handle RCU, lockdep, and IRQ flag tracing. We can't yet move to
the generic entry code, and already hadnle the user<->kernel transitions
elsewhere, so we add new kernel<->kernel transition helpers alog the
lines of the generic entry code.

Since we now track interrupts becoming masked when an exception is
taken, local_daif_inherit() is modified to track interrupts becoming
re-enabled when the original context is inherited. To balance the
entry/exit paths, each handler masks all DAIF exceptions before
exit_to_kernel_mode().

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-10-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: fix non-NMI user<->kernel transitions

2020-11-30T12:11:38+00:00

When built with PROVE_LOCKING, NO_HZ_FULL, and CONTEXT_TRACKING_FORCE
will WARN() at boot time that interrupts are enabled when we call
context_tracking_user_enter(), despite the DAIF flags indicating that
IRQs are masked.

The problem is that we're not tracking IRQ flag changes accurately, and
so lockdep believes interrupts are enabled when they are not (and
vice-versa). We can shuffle things so to make this more accurate. For
kernel->user transitions there are a number of constraints we need to
consider:

1) When we call __context_tracking_user_enter() HW IRQs must be disabled
   and lockdep must be up-to-date with this.

2) Userspace should be treated as having IRQs enabled from the PoV of
   both lockdep and tracing.

3) As context_tracking_user_enter() stops RCU from watching, we cannot
   use RCU after calling it.

4) IRQ flag tracing and lockdep have state that must be manipulated
   before RCU is disabled.

... with similar constraints applying for user->kernel transitions, with
the ordering reversed.

The generic entry code has enter_from_user_mode() and
exit_to_user_mode() helpers to handle this. We can't use those directly,
so we add arm64 copies for now (without the instrumentation markers
which aren't used on arm64). These replace the existing user_exit() and
user_exit_irqoff() calls spread throughout handlers, and the exception
unmasking is left as-is.

Note that:

* The accounting for debug exceptions from userspace now happens in
  el0_dbg() and ret_to_user(), so this is removed from
  debug_exception_enter() and debug_exception_exit(). As
  user_exit_irqoff() wakes RCU, the userspace-specific check is removed.

* The accounting for syscalls now happens in el0_svc(),
  el0_svc_compat(), and ret_to_user(), so this is removed from
  el0_svc_common(). This does not adversely affect the workaround for
  erratum 1463225, as this does not depend on any of the state tracking.

* In ret_to_user() we mask interrupts with local_daif_mask(), and so we
  need to inform lockdep and tracing. Here a trace_hardirqs_off() is
  sufficient and safe as we have not yet exited kernel context and RCU
  is usable.

* As PROVE_LOCKING selects TRACE_IRQFLAGS, the ifdeferry in entry.S only
  needs to check for the latter.

* EL0 SError handling will be dealt with in a subsequent patch, as this
  needs to be treated as an NMI.

Prior to this patch, booting an appropriately-configured kernel would
result in spats as below:

| DEBUG_LOCKS_WARN_ON(lockdep_hardirqs_enabled())
| WARNING: CPU: 2 PID: 1 at kernel/locking/lockdep.c:5280 check_flags.part.54+0x1dc/0x1f0
| Modules linked in:
| CPU: 2 PID: 1 Comm: init Not tainted 5.10.0-rc3 #3
| Hardware name: linux,dummy-virt (DT)
| pstate: 804003c5 (Nzcv DAIF +PAN -UAO -TCO BTYPE=--)
| pc : check_flags.part.54+0x1dc/0x1f0
| lr : check_flags.part.54+0x1dc/0x1f0
| sp : ffff80001003bd80
| x29: ffff80001003bd80 x28: ffff66ce801e0000
| x27: 00000000ffffffff x26: 00000000000003c0
| x25: 0000000000000000 x24: ffffc31842527258
| x23: ffffc31842491368 x22: ffffc3184282d000
| x21: 0000000000000000 x20: 0000000000000001
| x19: ffffc318432ce000 x18: 0080000000000000
| x17: 0000000000000000 x16: ffffc31840f18a78
| x15: 0000000000000001 x14: ffffc3184285c810
| x13: 0000000000000001 x12: 0000000000000000
| x11: ffffc318415857a0 x10: ffffc318406614c0
| x9 : ffffc318415857a0 x8 : ffffc31841f1d000
| x7 : 647261685f706564 x6 : ffffc3183ff7c66c
| x5 : ffff66ce801e0000 x4 : 0000000000000000
| x3 : ffffc3183fe00000 x2 : ffffc31841500000
| x1 : e956dc24146b3500 x0 : 0000000000000000
| Call trace:
|  check_flags.part.54+0x1dc/0x1f0
|  lock_is_held_type+0x10c/0x188
|  rcu_read_lock_sched_held+0x70/0x98
|  __context_tracking_enter+0x310/0x350
|  context_tracking_enter.part.3+0x5c/0xc8
|  context_tracking_user_enter+0x6c/0x80
|  finish_ret_to_user+0x2c/0x13cr

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-8-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: move el1 irq/nmi logic to C

2020-11-30T12:11:38+00:00

In preparation for reworking the EL1 irq/nmi entry code, move the
existing logic to C. We no longer need the asm_nmi_enter() and
asm_nmi_exit() wrappers, so these are removed. The new C functions are
marked noinstr, which prevents compiler instrumentation and runtime
probing.

In subsequent patches we'll want the new C helpers to be called in all
cases, so we don't bother wrapping the calls with ifdeferry. Even when
the new C functions are stubs the trivial calls are unlikely to have a
measurable impact on the IRQ or NMI paths anyway.

Prototypes are added to  as otherwise (in some
configurations) GCC will complain about the lack of a forward
declaration. We already do this for existing function, e.g.
enter_from_user_mode().

The new helpers are marked as noinstr (which prevents all
instrumentation, tracing, and kprobes). Otherwise, there should be no
functional change as a result of this patch.

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-7-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: prepare ret_to_user for function call

2020-11-30T12:11:37+00:00

In a subsequent patch ret_to_user will need to make a C function call
(in some configurations) which may clobber x0-x18 at the start of the
finish_ret_to_user block, before enable_step_tsk consumes the flags
loaded into x1.

In preparation for this, let's load the flags into x19, which is
preserved across C function calls. This avoids a redundant reload of the
flags and ensures we operate on a consistent shapshot regardless.

There should be no functional change as a result of this patch. At this
point of the entry/exit paths we only need to preserve x28 (tsk) and the
sp, and x19 is free for this use.

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-6-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: move enter_from_user_mode to entry-common.c

2020-11-30T12:11:37+00:00

In later patches we'll want to extend enter_from_user_mode() and add a
corresponding exit_to_user_mode(). As these will be common for all
entries/exits from userspace, it'd be better for these to live in
entry-common.c with the rest of the entry logic.

This patch moves enter_from_user_mode() into entry-common.c. As with
other functions in entry-common.c it is marked as noinstr (which
prevents all instrumentation, tracing, and kprobes) but there are no
other functional changes.

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-5-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: entry: mark entry code as noinstr

2020-11-30T12:11:37+00:00

Functions in entry-common.c are marked as notrace and NOKPROBE_SYMBOL(),
but they're still subject to other instrumentation which may rely on
lockdep/rcu/context-tracking being up-to-date, and may cause nested
exceptions (e.g. for WARN/BUG or KASAN's use of BRK) which will corrupt
exceptions registers which have not yet been read.

Prevent this by marking all functions in entry-common.c as noinstr to
prevent compiler instrumentation. This also blacklists the functions for
tracing and kprobes, so we don't need to handle that separately.
Functions elsewhere will be dealt with in subsequent patches.

Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-4-mark.rutland@arm.com
Signed-off-by: Will Deacon

arm64: mark idle code as noinstr

2020-11-30T12:11:37+00:00

Core code disables RCU when calling arch_cpu_idle(), so it's not safe
for arch_cpu_idle() or its calees to be instrumented, as the
instrumentation callbacks may attempt to use RCU or other features which
are unsafe to use in this context.

Mark them noinstr to prevent issues.

The use of local_irq_enable() in arch_cpu_idle() is similarly
problematic, and the "sched/idle: Fix arch_cpu_idle() vs tracing" patch
queued in the tip tree addresses that case.

Reported-by: Marco Elver 
Signed-off-by: Mark Rutland 
Cc: Catalin Marinas 
Cc: James Morse 
Cc: Will Deacon 
Link: https://lore.kernel.org/r/20201130115950.22492-3-mark.rutland@arm.com
Signed-off-by: Will Deacon