arm64: Implement HAVE_DYNAMIC_FTRACE_WITH_CALL_OPS

This patch enables support for DYNAMIC_FTRACE_WITH_CALL_OPS on arm64.
This allows each ftrace callsite to provide an ftrace_ops to the common
ftrace trampoline, allowing each callsite to invoke distinct tracer
functions without the need to fall back to list processing or to
allocate custom trampolines for each callsite. This significantly speeds
up cases where multiple distinct trace functions are used and callsites
are mostly traced by a single tracer.

The main idea is to place a pointer to the ftrace_ops as a literal at a
fixed offset from the function entry point, which can be recovered by
the common ftrace trampoline. Using a 64-bit literal avoids branch range
limitations, and permits the ops to be swapped atomically without
special considerations that apply to code-patching. In future this will
also allow for the implementation of DYNAMIC_FTRACE_WITH_DIRECT_CALLS
without branch range limitations by using additional fields in struct
ftrace_ops.

As noted in the core patch adding support for
DYNAMIC_FTRACE_WITH_CALL_OPS, this approach allows for directly invoking
ftrace_ops::func even for ftrace_ops which are dynamically-allocated (or
part of a module), without going via ftrace_ops_list_func.

Currently, this approach is not compatible with CLANG_CFI, as the
presence/absence of pre-function NOPs changes the offset of the
pre-function type hash, and there's no existing mechanism to ensure a
consistent offset for instrumented and uninstrumented functions. When
CLANG_CFI is enabled, the existing scheme with a global ops->func
pointer is used, and there should be no functional change. I am
currently working with others to allow the two to work together in
future (though this will liekly require updated compiler support).

I've benchamrked this with the ftrace_ops sample module [1], which is
not currently upstream, but available at:

  https://lore.kernel.org/lkml/20230103124912.2948963-1-mark.rutland@arm.com
  git://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git ftrace-ops-sample-20230109

Using that module I measured the total time taken for 100,000 calls to a
trivial instrumented function, with a number of tracers enabled with
relevant filters (which would apply to the instrumented function) and a
number of tracers enabled with irrelevant filters (which would not apply
to the instrumented function). I tested on an M1 MacBook Pro, running
under a HVF-accelerated QEMU VM (i.e. on real hardware).

Before this patch:

  Number of tracers     || Total time  | Per-call average time (ns)
  Relevant | Irrelevant || (ns)        | Total        | Overhead
  =========+============++=============+==============+============
         0 |          0 ||      94,583 |         0.95 |           -
         0 |          1 ||      93,709 |         0.94 |           -
         0 |          2 ||      93,666 |         0.94 |           -
         0 |         10 ||      93,709 |         0.94 |           -
         0 |        100 ||      93,792 |         0.94 |           -
  ---------+------------++-------------+--------------+------------
         1 |          1 ||   6,467,833 |        64.68 |       63.73
         1 |          2 ||   7,509,708 |        75.10 |       74.15
         1 |         10 ||  23,786,792 |       237.87 |      236.92
         1 |        100 || 106,432,500 |     1,064.43 |     1063.38
  ---------+------------++-------------+--------------+------------
         1 |          0 ||   1,431,875 |        14.32 |       13.37
         2 |          0 ||   6,456,334 |        64.56 |       63.62
        10 |          0 ||  22,717,000 |       227.17 |      226.22
       100 |          0 || 103,293,667 |      1032.94 |     1031.99
  ---------+------------++-------------+--------------+--------------

  Note: per-call overhead is estimated relative to the baseline case
  with 0 relevant tracers and 0 irrelevant tracers.

After this patch

  Number of tracers     || Total time  | Per-call average time (ns)
  Relevant | Irrelevant || (ns)        | Total        | Overhead
  =========+============++=============+==============+============
         0 |          0 ||      94,541 |         0.95 |           -
         0 |          1 ||      93,666 |         0.94 |           -
         0 |          2 ||      93,709 |         0.94 |           -
         0 |         10 ||      93,667 |         0.94 |           -
         0 |        100 ||      93,792 |         0.94 |           -
  ---------+------------++-------------+--------------+------------
         1 |          1 ||     281,000 |         2.81 |        1.86
         1 |          2 ||     281,042 |         2.81 |        1.87
         1 |         10 ||     280,958 |         2.81 |        1.86
         1 |        100 ||     281,250 |         2.81 |        1.87
  ---------+------------++-------------+--------------+------------
         1 |          0 ||     280,959 |         2.81 |        1.86
         2 |          0 ||   6,502,708 |        65.03 |       64.08
        10 |          0 ||  18,681,209 |       186.81 |      185.87
       100 |          0 || 103,550,458 |     1,035.50 |     1034.56
  ---------+------------++-------------+--------------+------------

  Note: per-call overhead is estimated relative to the baseline case
  with 0 relevant tracers and 0 irrelevant tracers.

As can be seen from the above:

a) Whenever there is a single relevant tracer function associated with a
   tracee, the overhead of invoking the tracer is constant, and does not
   scale with the number of tracers which are *not* associated with that
   tracee.

b) The overhead for a single relevant tracer has dropped to ~1/7 of the
   overhead prior to this series (from 13.37ns to 1.86ns). This is
   largely due to permitting calls to dynamically-allocated ftrace_ops
   without going through ftrace_ops_list_func.

I've run the ftrace selftests from v6.2-rc3, which reports:

| # of passed:  110
| # of failed:  0
| # of unresolved:  3
| # of untested:  0
| # of unsupported:  0
| # of xfailed:  1
| # of undefined(test bug):  0

... where the unresolved entries were the tests for DIRECT functions
(which are not supported), and the checkbashisms selftest (which is
irrelevant here):

| [8] Test ftrace direct functions against tracers        [UNRESOLVED]
| [9] Test ftrace direct functions against kprobes        [UNRESOLVED]
| [62] Meta-selftest: Checkbashisms       [UNRESOLVED]

... with all other tests passing (or failing as expected).

Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Cc: Florent Revest <revest@chromium.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20230123134603.1064407-9-mark.rutland@arm.com
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>

1 files changed, 4 insertions, 1 deletions

diff --git a/arch/arm64/Makefile b/arch/arm64/Makefile
index d62bd221828f..4c3be442fbb3 100644
--- a/arch/arm64/Makefile
+++ b/arch/arm64/Makefile
@@ -139,7 +139,10 @@ endif
 
 CHECKFLAGS	+= -D__aarch64__
 
-ifeq ($(CONFIG_DYNAMIC_FTRACE_WITH_ARGS),y)
+ifeq ($(CONFIG_DYNAMIC_FTRACE_WITH_CALL_OPS),y)
+  KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
+  CC_FLAGS_FTRACE := -fpatchable-function-entry=4,2
+else ifeq ($(CONFIG_DYNAMIC_FTRACE_WITH_ARGS),y)
   KBUILD_CPPFLAGS += -DCC_USING_PATCHABLE_FUNCTION_ENTRY
   CC_FLAGS_FTRACE := -fpatchable-function-entry=2
 endif