RFC 2: MT safe fiber context switch on ARM32 (#15582)

We need a store-store barrier between pushing the registers to the
current stack and setting the resumable flag of the current fiber,
otherwise the CPU is allowed to reorder the instructions at runtime and
to store the resumable flag before or while storing the registers

thread 1: enqueues current fiber A
therad 1: swapcontext -> store resumable
thread 1: is preempted
thread 2: steals fiber A
thread 2: resumes fiber A
thread 2: loads registers => reads garbage => segfaults
thread 1: stores registers (too late)

The chosen assembly is compatible with armv6, but might not work on older architectures, and it doesn't take advantage of armv7 supporting the `dmb ish` instruction.

We might want to consider integrating #14524 (comment) into the compiler, so we could add flags based on features, for example `armv7` when any feature matches `/\+v7(ve|r|m|em|s|k|)/`.

This is the ARM32 equivalent of #15581.

Author: ysbaddaden

src/fiber/context/arm.cr

@@ -29,6 +29,17 @@ class Fiber
     #
     # Eventually reset LR to zero to avoid the ARM unwinder to mistake the
     # context switch as a regular call.
+    #
+    # NOTE: depending on the ARM architecture (v7, v6 or older) LLVM uses
+    # different strategies for atomics. By default it uses the "older"
+    # architecture that relies on the libgcc __sync_* symbols; when an armv6 CPU
+    # or +v6 feature is specified it uses the coprocessor instruction as used
+    # below, unless the +db (data barrier) feature is set, in which case it
+    # uses the `dmb ish` instruction. The +db feature is always enabled since
+    # armv7 / +v7.
+    #
+    # TODO: we should do the same, but we don't know the list of CPU features
+    # for the current target machine (and LLVM won't tell us).
 
     {% if compare_versions(Crystal::LLVM_VERSION, "9.0.0") >= 0 %}
       asm("
@@ -42,6 +53,10 @@ class Fiber
         vstmdb sp!, {d8-d15}          // push FPU registers
         {% end %}
         str    sp, [r0, #0]           // current_context.stack_top = sp
+        {% if flag?(:execution_context) %}
+        mov    r4, #0                 // barrier: ensure registers are stored
+        mcr    p15, #0, r4, c7, c10, #5
+        {% end %}
         mov    r4, #1                 // current_context.resumable = 1
         str    r4, [r0, #4]
 
@@ -72,6 +87,10 @@ class Fiber
         vstmdb sp!, {d8-d15}          // push FPU registers
         {% end %}
         str    sp, [$0, #0]           // current_context.stack_top = sp
+        {% if flag?(:execution_context) %}
+        mov    r4, #0                 // barrier: ensure registers are stored
+        mcr    p15, #0, r4, c7, c10, #5
+        {% end %}
         mov    r4, #1                 // current_context.resumable = 1
         str    r4, [$0, #4]