Annotate anonymous mmap

Use PR_SET_VMA_ANON_NAME to set human-readable names for anonymous
virtual memory areas mapped by `mmap()` when compiled and run on Linux
5.17 or higher.  This makes it convenient for developers to debug mmap.

Author: wks

common.mk

@@ -8631,19 +8631,25 @@ io_buffer.$(OBJEXT): $(CCAN_DIR)/str/str.h
 io_buffer.$(OBJEXT): $(hdrdir)/ruby/ruby.h
 io_buffer.$(OBJEXT): $(hdrdir)/ruby/version.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/array.h
+io_buffer.$(OBJEXT): $(top_srcdir)/internal/basic_operators.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/bignum.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/bits.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/compilers.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/error.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/fixnum.h
+io_buffer.$(OBJEXT): $(top_srcdir)/internal/gc.h
+io_buffer.$(OBJEXT): $(top_srcdir)/internal/imemo.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/io.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/numeric.h
+io_buffer.$(OBJEXT): $(top_srcdir)/internal/sanitizers.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/serial.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/static_assert.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/string.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/thread.h
 io_buffer.$(OBJEXT): $(top_srcdir)/internal/vm.h
+io_buffer.$(OBJEXT): $(top_srcdir)/internal/warnings.h
 io_buffer.$(OBJEXT): {$(VPATH)}assert.h
+io_buffer.$(OBJEXT): {$(VPATH)}atomic.h
 io_buffer.$(OBJEXT): {$(VPATH)}backward/2/assume.h
 io_buffer.$(OBJEXT): {$(VPATH)}backward/2/attributes.h
 io_buffer.$(OBJEXT): {$(VPATH)}backward/2/bool.h
@@ -8657,6 +8663,7 @@ io_buffer.$(OBJEXT): {$(VPATH)}config.h
 io_buffer.$(OBJEXT): {$(VPATH)}defines.h
 io_buffer.$(OBJEXT): {$(VPATH)}encoding.h
 io_buffer.$(OBJEXT): {$(VPATH)}fiber/scheduler.h
+io_buffer.$(OBJEXT): {$(VPATH)}id.h
 io_buffer.$(OBJEXT): {$(VPATH)}intern.h
 io_buffer.$(OBJEXT): {$(VPATH)}internal.h
 io_buffer.$(OBJEXT): {$(VPATH)}internal/abi.h
@@ -8811,13 +8818,21 @@ io_buffer.$(OBJEXT): {$(VPATH)}internal/xmalloc.h
 io_buffer.$(OBJEXT): {$(VPATH)}io.h
 io_buffer.$(OBJEXT): {$(VPATH)}io/buffer.h
 io_buffer.$(OBJEXT): {$(VPATH)}io_buffer.c
+io_buffer.$(OBJEXT): {$(VPATH)}method.h
 io_buffer.$(OBJEXT): {$(VPATH)}missing.h
+io_buffer.$(OBJEXT): {$(VPATH)}node.h
 io_buffer.$(OBJEXT): {$(VPATH)}onigmo.h
 io_buffer.$(OBJEXT): {$(VPATH)}oniguruma.h
+io_buffer.$(OBJEXT): {$(VPATH)}ruby_assert.h
+io_buffer.$(OBJEXT): {$(VPATH)}ruby_atomic.h
+io_buffer.$(OBJEXT): {$(VPATH)}rubyparser.h
 io_buffer.$(OBJEXT): {$(VPATH)}st.h
 io_buffer.$(OBJEXT): {$(VPATH)}subst.h
 io_buffer.$(OBJEXT): {$(VPATH)}thread.h
+io_buffer.$(OBJEXT): {$(VPATH)}thread_$(THREAD_MODEL).h
 io_buffer.$(OBJEXT): {$(VPATH)}thread_native.h
+io_buffer.$(OBJEXT): {$(VPATH)}vm_core.h
+io_buffer.$(OBJEXT): {$(VPATH)}vm_opts.h
 iseq.$(OBJEXT): $(CCAN_DIR)/check_type/check_type.h
 iseq.$(OBJEXT): $(CCAN_DIR)/container_of/container_of.h
 iseq.$(OBJEXT): $(CCAN_DIR)/list/list.h

cont.c

@@ -475,18 +475,20 @@ fiber_pool_allocate_memory(size_t * count, size_t stride)
         }
 #else
         errno = 0;
-        void * base = mmap(NULL, (*count)*stride, PROT_READ | PROT_WRITE, FIBER_STACK_FLAGS, -1, 0);
+        size_t mmap_size = (*count)*stride;
+        void * base = mmap(NULL, mmap_size, PROT_READ | PROT_WRITE, FIBER_STACK_FLAGS, -1, 0);
 
         if (base == MAP_FAILED) {
             // If the allocation fails, count = count / 2, and try again.
             *count = (*count) >> 1;
         }
         else {
+            ruby_annotate_mmap(base, mmap_size, "Ruby:fiber_pool_allocate_memory");
 #if defined(MADV_FREE_REUSE)
             // On Mac MADV_FREE_REUSE is necessary for the task_info api
             // to keep the accounting accurate as possible when a page is marked as reusable
             // it can possibly not occurring at first call thus re-iterating if necessary.
-            while (madvise(base, (*count)*stride, MADV_FREE_REUSE) == -1 && errno == EAGAIN);
+            while (madvise(base, mmap_size, MADV_FREE_REUSE) == -1 && errno == EAGAIN);
 #endif
             return base;
         }

gc.c

@@ -74,6 +74,12 @@
 #include <emscripten.h>
 #endif
 
+/* For ruby_annotate_mmap */
+#ifdef __linux__
+#include <linux/prctl.h>
+#include <sys/prctl.h>
+#endif
+
 #undef LIST_HEAD /* ccan/list conflicts with BSD-origin sys/queue.h. */
 
 #include "constant.h"
@@ -4494,3 +4500,36 @@ Init_GC(void)
 
     rb_gc_impl_init();
 }
+
+// Set a name for the anonymous virtual memory area. `addr` is the starting
+// address of the area and `size` is its length in bytes. `name` is a
+// NUL-terminated human-readable string.
+//
+// This function is usually called after calling `mmap()`.  The human-readable
+// annotation helps developers identify the call site of `mmap()` that created
+// the memory mapping.
+//
+// This function currently only works on Linux 5.17 or higher.  After calling
+// this function, we can see annotations in the form of "[anon:...]" in
+// `/proc/self/maps`, where `...` is the content of `name`.  This function has
+// no effect when called on other platforms.
+void
+ruby_annotate_mmap(const void *addr, unsigned long size, const char *name)
+{
+#if defined(__linux__) && defined(PR_SET_VMA) && defined(PR_SET_VMA_ANON_NAME)
+    // The name length cannot exceed 80 (including the '\0').
+    RUBY_ASSERT(strlen(name) < 80);
+    prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, (unsigned long)addr, size, name);
+    // We ignore errors in prctl. prctl may set errno to EINVAL for several
+    // reasons.
+    // 1. The attr (PR_SET_VMA_ANON_NAME) is not a valid attribute.
+    // 2. addr is an invalid address.
+    // 3. The string pointed by name is too long.
+    // The first error indicates PR_SET_VMA_ANON_NAME is not available, and may
+    // happen if we run the compiled binary on an old kernel.  In theory, all
+    // other errors should result in a failure.  But since EINVAL cannot tell
+    // the first error from others, and this function is mainly used for
+    // debugging, we silently ignore the error.
+    errno = 0;
+#endif
+}

gc/default.c

@@ -5,6 +5,10 @@
 #ifndef _WIN32
 # include <sys/mman.h>
 # include <unistd.h>
+# ifdef __linux__
+#  include <linux/prctl.h>
+#  include <sys/prctl.h>
+# endif
 #endif
 
 #if !defined(PAGE_SIZE) && defined(HAVE_SYS_USER_H)
@@ -1859,12 +1863,23 @@ heap_page_body_allocate(void)
 #ifdef HAVE_MMAP
         GC_ASSERT(HEAP_PAGE_ALIGN % sysconf(_SC_PAGE_SIZE) == 0);
 
-        char *ptr = mmap(NULL, HEAP_PAGE_ALIGN + HEAP_PAGE_SIZE,
+        size_t mmap_size = HEAP_PAGE_ALIGN + HEAP_PAGE_SIZE;
+        char *ptr = mmap(NULL, mmap_size,
                          PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
         if (ptr == MAP_FAILED) {
             return NULL;
         }
 
+        // If we are building `default.c` as part of the ruby executable, we
+        // may just call `ruby_annotate_mmap`.  But if we are building
+        // `default.c` as a shared library, we will not have access to private
+        // symbols, and we have to either call prctl directly or make our own
+        // wrapper.
+#if defined(__linux__) && defined(PR_SET_VMA) && defined(PR_SET_VMA_ANON_NAME)
+        prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, ptr, mmap_size, "Ruby:GC:default:heap_page_body_allocate");
+        errno = 0;
+#endif
+
         char *aligned = ptr + HEAP_PAGE_ALIGN;
         aligned -= ((VALUE)aligned & (HEAP_PAGE_ALIGN - 1));
         GC_ASSERT(aligned > ptr);

internal/gc.h

@@ -260,6 +260,7 @@ RUBY_SYMBOL_EXPORT_END
 int rb_ec_stack_check(struct rb_execution_context_struct *ec);
 void rb_gc_writebarrier_remember(VALUE obj);
 const char *rb_obj_info(VALUE obj);
+void ruby_annotate_mmap(const void *addr, unsigned long size, const char *name);
 
 #if defined(HAVE_MALLOC_USABLE_SIZE) || defined(HAVE_MALLOC_SIZE) || defined(_WIN32)
 

io_buffer.c

@@ -16,6 +16,7 @@
 #include "internal/array.h"
 #include "internal/bits.h"
 #include "internal/error.h"
+#include "internal/gc.h"
 #include "internal/numeric.h"
 #include "internal/string.h"
 #include "internal/io.h"
@@ -83,6 +84,8 @@ io_buffer_map_memory(size_t size, int flags)
     if (base == MAP_FAILED) {
         rb_sys_fail("io_buffer_map_memory:mmap");
     }
+
+    ruby_annotate_mmap(base, size, "Ruby:io_buffer_map_memory");
 #endif
 
     return base;

rjit_c.c

@@ -85,6 +85,7 @@ rjit_reserve_addr_space(uint32_t mem_size)
 
             // If we succeeded, stop
             if (mem_block != MAP_FAILED) {
+                ruby_annotate_mmap(mem_block, mem_size, "Ruby:rjit_reserve_addr_space");
                 break;
             }
 
@@ -116,6 +117,10 @@ rjit_reserve_addr_space(uint32_t mem_size)
             -1,
             0
         );
+
+        if (mem_block != MAP_FAILED) {
+            ruby_annotate_mmap(mem_block, mem_size, "Ruby:rjit_reserve_addr_space:fallback");
+        }
     }
 
     // Check that the memory mapping was successful

shape.c

@@ -1232,11 +1232,15 @@ Init_default_shapes(void)
     rb_shape_tree_ptr = xcalloc(1, sizeof(rb_shape_tree_t));
 
 #ifdef HAVE_MMAP
-    rb_shape_tree_ptr->shape_list = (rb_shape_t *)mmap(NULL, rb_size_mul_or_raise(SHAPE_BUFFER_SIZE, sizeof(rb_shape_t), rb_eRuntimeError),
+    size_t shape_list_mmap_size = rb_size_mul_or_raise(SHAPE_BUFFER_SIZE, sizeof(rb_shape_t), rb_eRuntimeError);
+    rb_shape_tree_ptr->shape_list = (rb_shape_t *)mmap(NULL, shape_list_mmap_size,
                          PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
     if (GET_SHAPE_TREE()->shape_list == MAP_FAILED) {
         GET_SHAPE_TREE()->shape_list = 0;
     }
+    else {
+        ruby_annotate_mmap(rb_shape_tree_ptr->shape_list, shape_list_mmap_size, "Ruby:Init_default_shapes:shape_list");
+    }
 #else
     GET_SHAPE_TREE()->shape_list = xcalloc(SHAPE_BUFFER_SIZE, sizeof(rb_shape_t));
 #endif
@@ -1249,7 +1253,8 @@ Init_default_shapes(void)
     id_t_object = rb_make_internal_id();
 
 #ifdef HAVE_MMAP
-    rb_shape_tree_ptr->shape_cache = (redblack_node_t *)mmap(NULL, rb_size_mul_or_raise(REDBLACK_CACHE_SIZE, sizeof(redblack_node_t), rb_eRuntimeError),
+    size_t shape_cache_mmap_size = rb_size_mul_or_raise(REDBLACK_CACHE_SIZE, sizeof(redblack_node_t), rb_eRuntimeError);
+    rb_shape_tree_ptr->shape_cache = (redblack_node_t *)mmap(NULL, shape_cache_mmap_size,
                          PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
     rb_shape_tree_ptr->cache_size = 0;
 
@@ -1260,6 +1265,9 @@ Init_default_shapes(void)
         GET_SHAPE_TREE()->shape_cache = 0;
         GET_SHAPE_TREE()->cache_size = REDBLACK_CACHE_SIZE;
     }
+    else {
+        ruby_annotate_mmap(rb_shape_tree_ptr->shape_cache, shape_cache_mmap_size, "Ruby:Init_default_shapes:shape_cache");
+    }
 #endif
 
     // Root shape

thread_pthread_mn.c

@@ -194,6 +194,8 @@ nt_alloc_thread_stack_chunk(void)
         return NULL;
     }
 
+    ruby_annotate_mmap(m, MSTACK_CHUNK_SIZE, "Ruby:nt_alloc_thread_stack_chunk");
+
     size_t msz = nt_thread_stack_size();
     int header_page_cnt = 1;
     int stack_count = ((MSTACK_CHUNK_PAGE_NUM - header_page_cnt) * MSTACK_PAGE_SIZE) / msz;

yjit.c

@@ -291,6 +291,7 @@ rb_yjit_reserve_addr_space(uint32_t mem_size)
 
             // If we succeeded, stop
             if (mem_block != MAP_FAILED) {
+                ruby_annotate_mmap(mem_block, mem_size, "Ruby:rb_yjit_reserve_addr_space");
                 break;
             }
 
@@ -325,6 +326,10 @@ rb_yjit_reserve_addr_space(uint32_t mem_size)
             -1,
             0
         );
+
+        if (mem_block != MAP_FAILED) {
+            ruby_annotate_mmap(mem_block, mem_size, "Ruby:rb_yjit_reserve_addr_space:fallback");
+        }
     }
 
     // Check that the memory mapping was successful