[NativeAOT] enable background GC on Unix (#74735)

* enable for unix+arm64

* include softwarewritewatch.cpp

* make the unix crst recursive

* enable ww on Unix x64

* enabled card bundles

* comment

src/coreclr/nativeaot/Runtime/CMakeLists.txt

@@ -50,6 +50,7 @@ set(COMMON_RUNTIME_SOURCES
     ${GC_DIR}/handletablecore.cpp
     ${GC_DIR}/handletablescan.cpp
     ${GC_DIR}/objecthandle.cpp
+    ${GC_DIR}/softwarewritewatch.cpp
 )
 
 set(SERVER_GC_SOURCES
@@ -206,6 +207,12 @@ include_directories(${ARCH_SOURCES_DIR})
 
 add_definitions(-DFEATURE_BASICFREEZE)
 add_definitions(-DFEATURE_CONSERVATIVE_GC)
+
+if(CLR_CMAKE_TARGET_UNIX)
+  add_definitions(-DFEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP)
+  add_definitions(-DFEATURE_MANUALLY_MANAGED_CARD_BUNDLES)
+endif()
+
 add_definitions(-DFEATURE_CUSTOM_IMPORTS)
 add_definitions(-DFEATURE_DYNAMIC_CODE)
 add_compile_definitions($<$<OR:$<CONFIG:Debug>,$<CONFIG:Checked>>:FEATURE_GC_STRESS>)

src/coreclr/nativeaot/Runtime/amd64/WriteBarriers.S

@@ -84,6 +84,21 @@ LOCAL_LABEL(\BASENAME\()_UpdateShadowHeap_Done_\REFREG):
     // we're in a debug build and write barrier checking has been enabled).
     UPDATE_GC_SHADOW \BASENAME, \REFREG, rdi
 
+#ifdef FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP
+    mov     r11, [C_VAR(g_write_watch_table)]
+    cmp     r11, 0x0
+    je      LOCAL_LABEL(\BASENAME\()_CheckCardTable_\REFREG)
+
+    mov     r10, rdi
+    shr     r10, 0xC // SoftwareWriteWatch::AddressToTableByteIndexShift
+    add     r10, r11
+    cmp     byte ptr [r10], 0x0
+    jne     LOCAL_LABEL(\BASENAME\()_CheckCardTable_\REFREG)
+    mov     byte ptr [r10], 0xFF
+#endif
+
+LOCAL_LABEL(\BASENAME\()_CheckCardTable_\REFREG):
+
     // If the reference is to an object that's not in an ephemeral generation we have no need to track it
     // (since the object won't be collected or moved by an ephemeral collection).
     cmp     \REFREG, [C_VAR(g_ephemeral_low)]
@@ -95,17 +110,25 @@ LOCAL_LABEL(\BASENAME\()_UpdateShadowHeap_Done_\REFREG):
     // track this write. The location address is translated into an offset in the card table bitmap. We set
     // an entire byte in the card table since it's quicker than messing around with bitmasks and we only write
     // the byte if it hasn't already been done since writes are expensive and impact scaling.
-    shr     rdi, 11
-    add     rdi, [C_VAR(g_card_table)]
-    cmp     byte ptr [rdi], 0x0FF
-    jne     LOCAL_LABEL(\BASENAME\()_UpdateCardTable_\REFREG)
-
-LOCAL_LABEL(\BASENAME\()_NoBarrierRequired_\REFREG):
-    ret
+    shr     rdi, 0x0B
+    mov     r10, [C_VAR(g_card_table)]
+    cmp     byte ptr [rdi + r10], 0x0FF
+    je      LOCAL_LABEL(\BASENAME\()_NoBarrierRequired_\REFREG)
 
 // We get here if it's necessary to update the card table.
-LOCAL_LABEL(\BASENAME\()_UpdateCardTable_\REFREG):
-    mov     byte ptr [rdi], 0x0FF
+    mov     byte ptr [rdi + r10], 0xFF
+
+#ifdef FEATURE_MANUALLY_MANAGED_CARD_BUNDLES
+    // Shift rdi by 0x0A more to get the card bundle byte (we shifted by 0x0B already)
+    shr     rdi, 0x0A
+    add     rdi, [C_VAR(g_card_bundle_table)]
+    cmp     byte ptr [rdi], 0xFF
+    je      LOCAL_LABEL(\BASENAME\()_NoBarrierRequired_\REFREG)
+
+    mov     byte ptr [rdi], 0xFF
+#endif
+
+LOCAL_LABEL(\BASENAME\()_NoBarrierRequired_\REFREG):
     ret
 
 .endm
@@ -252,6 +275,21 @@ LEAF_ENTRY RhpByRefAssignRef, _TEXT
     // we're in a debug build and write barrier checking has been enabled).
     UPDATE_GC_SHADOW BASENAME, rcx, rdi
 
+#ifdef FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP
+    mov     r11, [C_VAR(g_write_watch_table)]
+    cmp     r11, 0x0
+    je      LOCAL_LABEL(RhpByRefAssignRef_CheckCardTable)
+
+    mov     r10, rdi
+    shr     r10, 0xC // SoftwareWriteWatch::AddressToTableByteIndexShift
+    add     r10, r11
+    cmp     byte ptr [r10], 0x0
+    jne     LOCAL_LABEL(RhpByRefAssignRef_CheckCardTable)
+    mov     byte ptr [r10], 0xFF
+#endif
+
+LOCAL_LABEL(RhpByRefAssignRef_CheckCardTable):
+
     // If the reference is to an object that's not in an ephemeral generation we have no need to track it
     // (since the object won't be collected or moved by an ephemeral collection).
     cmp     rcx, [C_VAR(g_ephemeral_low)]
@@ -259,25 +297,30 @@ LEAF_ENTRY RhpByRefAssignRef, _TEXT
     cmp     rcx, [C_VAR(g_ephemeral_high)]
     jae     LOCAL_LABEL(RhpByRefAssignRef_NotInHeap)
 
-    // move current rdi value into rcx and then increment the pointers
+    // move current rdi value into rcx, we need to keep rdi and eventually increment by 8
     mov     rcx, rdi
-    add     rsi, 0x8
-    add     rdi, 0x8
 
     // We have a location on the GC heap being updated with a reference to an ephemeral object so we must
     // track this write. The location address is translated into an offset in the card table bitmap. We set
     // an entire byte in the card table since it's quicker than messing around with bitmasks and we only write
     // the byte if it hasn't already been done since writes are expensive and impact scaling.
-    shr     rcx, 11
-    add     rcx, [C_VAR(g_card_table)]
-    cmp     byte ptr [rcx], 0x0FF
-    jne     LOCAL_LABEL(RhpByRefAssignRef_UpdateCardTable)
-    ret
+    shr     rcx, 0x0B
+    mov     r10, [C_VAR(g_card_table)]
+    cmp     byte ptr [rcx + r10], 0x0FF
+    je      LOCAL_LABEL(RhpByRefAssignRef_NotInHeap)
 
 // We get here if it's necessary to update the card table.
-LOCAL_LABEL(RhpByRefAssignRef_UpdateCardTable):
-    mov     byte ptr [rcx], 0x0FF
-    ret
+    mov     byte ptr [rcx + r10], 0xFF
+
+#ifdef FEATURE_MANUALLY_MANAGED_CARD_BUNDLES
+    // Shift rcx by 0x0A more to get the card bundle byte (we shifted by 0x0B already)
+    shr     rcx, 0x0A
+    add     rcx, [C_VAR(g_card_bundle_table)]
+    cmp     byte ptr [rcx], 0xFF
+    je      LOCAL_LABEL(RhpByRefAssignRef_NotInHeap)
+
+    mov     byte ptr [rcx], 0xFF
+#endif
 
 LOCAL_LABEL(RhpByRefAssignRef_NotInHeap):
     // Increment the pointers before leaving

src/coreclr/nativeaot/Runtime/arm64/WriteBarriers.S

@@ -92,12 +92,11 @@
         //   destReg: location to be updated
         //   refReg:  objectref to be stored
         //   trash: register nr than can be trashed
-        //   trash2: register than can be trashed
         //
         // On exit:
         //   destReg:   trashed
         //
-        .macro INSERT_UNCHECKED_WRITE_BARRIER_CORE destReg, refReg, trash, trash2
+        .macro INSERT_UNCHECKED_WRITE_BARRIER_CORE destReg, refReg, trash
 
         // Update the shadow copy of the heap with the same value just written to the same heap. (A no-op unless
         // we are in a debug build and write barrier checking has been enabled).
@@ -129,27 +128,27 @@
         // Set this objects card, if it has not already been set.
 
         PREPARE_EXTERNAL_VAR_INDIRECT g_card_table, x\trash
-        add     \trash2, x\trash, \destReg, lsr #11
+        add     x17, x\trash, \destReg, lsr #11
 
         // Check that this card has not already been written. Avoiding useless writes is a big win on
         // multi-proc systems since it avoids cache thrashing.
-        ldrb    w\trash, [\trash2]
+        ldrb    w\trash, [x17]
         cmp     x\trash, 0xFF
         beq     0f
 
         mov     x\trash, 0xFF
-        strb    w\trash, [\trash2]
+        strb    w\trash, [x17]
 
 #ifdef FEATURE_MANUALLY_MANAGED_CARD_BUNDLES
         // Check if we need to update the card bundle table
         PREPARE_EXTERNAL_VAR_INDIRECT g_card_bundle_table, x\trash
-        add  \trash2, x\trash, \destReg, lsr #21
-        ldrb w\trash, [\trash2]
+        add  x17, x\trash, \destReg, lsr #21
+        ldrb w\trash, [x17]
         cmp  x\trash, 0xFF
         beq  0f
 
         mov  x\trash, 0xFF
-        strb w\trash, [\trash2]
+        strb w\trash, [x17]
 #endif
 
 0:
@@ -160,12 +159,11 @@
         //   destReg: location to be updated
         //   refReg:  objectref to be stored
         //   trash: register nr than can be trashed
-        //   trash2: register than can be trashed
         //
         // On exit:
         //   destReg:   trashed
         //
-        .macro INSERT_CHECKED_WRITE_BARRIER_CORE destReg, refReg, trash, trash2
+        .macro INSERT_CHECKED_WRITE_BARRIER_CORE destReg, refReg, trash
 
         // The "check" of this checked write barrier - is destReg
         // within the heap? if no, early out.
@@ -180,7 +178,7 @@
         ccmp    \destReg, x\trash, #0x2, hs
         bhs     0f
 
-        INSERT_UNCHECKED_WRITE_BARRIER_CORE \destReg, \refReg, \trash, \trash2
+        INSERT_UNCHECKED_WRITE_BARRIER_CORE \destReg, \refReg, \trash
 
 0:
         // Exit label
@@ -265,7 +263,7 @@ CmpXchgRetry:
         // The following barrier code takes the destination in x0 and the value in x1 so the arguments are
         // already correctly set up.
 
-        INSERT_CHECKED_WRITE_BARRIER_CORE x0, x1, 9, x0
+        INSERT_CHECKED_WRITE_BARRIER_CORE x0, x1, 9
 
 CmpXchgNoUpdate:
         // x10 still contains the original value.
@@ -307,7 +305,7 @@ ExchangeRetry:
         // The following barrier code takes the destination in x0 and the value in x1 so the arguments are
         // already correctly set up.
 
-        INSERT_CHECKED_WRITE_BARRIER_CORE x0, x1, 9, x0
+        INSERT_CHECKED_WRITE_BARRIER_CORE x0, x1, 9
 
         // x10 still contains the original value.
         mov     x0, x10
@@ -321,7 +319,7 @@ LEAF_ENTRY RhpAssignRefArm64, _TEXT
     ALTERNATE_ENTRY RhpAssignRefX1AVLocation
         stlr    x15, [x14]
 
-        INSERT_UNCHECKED_WRITE_BARRIER_CORE x14, x15, 12, X14
+        INSERT_UNCHECKED_WRITE_BARRIER_CORE x14, x15, 12
 
         ret
 LEAF_END RhpAssignRefArm64, _TEXT
@@ -343,9 +341,7 @@ LEAF_ENTRY RhpCheckedAssignRefArm64, _TEXT
 
         stlr    x15, [x14]
 
-        INSERT_CHECKED_WRITE_BARRIER_CORE x14, x15, 12, X15
-
-        add x14, x14, #8
+        INSERT_CHECKED_WRITE_BARRIER_CORE x14, x15, 12
 
         ret
 LEAF_END RhpCheckedAssignRefArm64, _TEXT
@@ -366,7 +362,7 @@ LEAF_ENTRY RhpByRefAssignRefArm64, _TEXT
     ldr     x15, [x13]
     stlr    x15, [x14]
 
-    INSERT_CHECKED_WRITE_BARRIER_CORE x14, x15, 12, X15
+    INSERT_CHECKED_WRITE_BARRIER_CORE x14, x15, 12
 
     add     X13, x13, #8
     add     x14, x14, #8

src/coreclr/nativeaot/Runtime/gcrhenv.cpp

@@ -1018,8 +1018,8 @@ void GCToEEInterface::DiagWalkBGCSurvivors(void* gcContext)
 #endif // FEATURE_EVENT_TRACE
 }
 
-#if defined(FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP) && (!defined(TARGET_ARM64) || !defined(TARGET_UNIX))
-#error FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP is only implemented for ARM64 and UNIX
+#if defined(FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP) && !defined(TARGET_UNIX)
+#error FEATURE_USE_SOFTWARE_WRITE_WATCH_FOR_GC_HEAP is only implemented for UNIX
 #endif
 
 void GCToEEInterface::StompWriteBarrier(WriteBarrierParameters* args)

src/coreclr/nativeaot/Runtime/unix/PalRedhawkUnix.cpp

@@ -869,7 +869,22 @@ extern "C" UInt32_BOOL DuplicateHandle(
 
 extern "C" UInt32_BOOL InitializeCriticalSection(CRITICAL_SECTION * lpCriticalSection)
 {
-    return pthread_mutex_init(&lpCriticalSection->mutex, NULL) == 0;
+    pthread_mutexattr_t mutexAttributes;
+    int st = pthread_mutexattr_init(&mutexAttributes);
+    if (st != 0)
+    {
+        return false;
+    }
+
+    st = pthread_mutexattr_settype(&mutexAttributes, PTHREAD_MUTEX_RECURSIVE);
+    if (st == 0)
+    {
+        st = pthread_mutex_init(&lpCriticalSection->mutex, &mutexAttributes);
+    }
+
+    pthread_mutexattr_destroy(&mutexAttributes);
+
+    return (st == 0);
 }
 
 extern "C" UInt32_BOOL InitializeCriticalSectionEx(CRITICAL_SECTION * lpCriticalSection, uint32_t arg2, uint32_t arg3)