json 2.19.8 → 2.21.1

data/ext/json/ext/parser/parser.c

@@ -1,15 +1,18 @@
 #include "../json.h"
-#include "../vendor/ryu.h"
+#include "../vendor/fast_float_parser.h"
 #include "../simd/simd.h"
 
-static VALUE mJSON, eNestingError, Encoding_UTF_8;
-static VALUE CNaN, CInfinity, CMinusInfinity;
+static VALUE mJSON, eNestingError, eParserError, Encoding_UTF_8;
+static VALUE CNaN, CInfinity, CMinusInfinity, JSON_empty_string;
 
-static ID i_new, i_try_convert, i_uminus, i_encode;
+static ID i_new, i_try_convert, i_encode, i_at_line, i_at_column;
+#ifndef HAVE_RB_STR_TO_INTERNED_STR
+static ID i_uminus;
+#endif
 
-static VALUE sym_max_nesting, sym_allow_nan, sym_allow_trailing_comma, sym_allow_control_characters,
-             sym_allow_invalid_escape, sym_symbolize_names, sym_freeze, sym_decimal_class, sym_on_load,
-             sym_allow_duplicate_key;
+static VALUE sym_max_nesting, sym_allow_nan, sym_allow_trailing_comma, sym_allow_comments,
+             sym_allow_control_characters, sym_allow_invalid_escape, sym_symbolize_names,
+             sym_freeze, sym_decimal_class, sym_on_load, sym_allow_duplicate_key;
 
 static int binary_encindex;
 static int utf8_encindex;
@@ -58,6 +61,20 @@ typedef struct rvalue_cache_struct {
     VALUE entries[JSON_RVALUE_CACHE_CAPA];
 } rvalue_cache;
 
+static void rvalue_cache_mark(rvalue_cache *cache)
+{
+    for (int index = 0; index < cache->length; index++) {
+        rb_gc_mark_movable(cache->entries[index]);
+    }
+}
+
+static void rvalue_cache_compact(rvalue_cache *cache)
+{
+    for (int index = 0; index < cache->length; index++) {
+        cache->entries[index] = rb_gc_location(cache->entries[index]);
+    }
+}
+
 static rb_encoding *enc_utf8;
 
 #define JSON_RVALUE_CACHE_MAX_ENTRY_LENGTH 55
@@ -206,12 +223,12 @@ static rvalue_stack *rvalue_stack_spill(rvalue_stack *old_stack, VALUE *handle,
 
 static rvalue_stack *rvalue_stack_grow(rvalue_stack *stack, VALUE *handle, rvalue_stack **stack_ref)
 {
-    long required = stack->capa * 2;
+    long required = stack->capa ? stack->capa * 2 : RVALUE_STACK_INITIAL_CAPA;
 
     if (stack->type == RVALUE_STACK_STACK_ALLOCATED) {
         stack = rvalue_stack_spill(stack, handle, stack_ref);
     } else {
-        REALLOC_N(stack->ptr, VALUE, required);
+        JSON_SIZED_REALLOC_N(stack->ptr, VALUE, required, stack->capa);
         stack->capa = required;
     }
     return stack;
@@ -219,11 +236,15 @@ static rvalue_stack *rvalue_stack_grow(rvalue_stack *stack, VALUE *handle, rvalu
 
 static VALUE rvalue_stack_push(rvalue_stack *stack, VALUE value, VALUE *handle, rvalue_stack **stack_ref)
 {
+    JSON_ASSERT(stack->type != RVALUE_STACK_STACK_ALLOCATED || handle);
+
     if (RB_UNLIKELY(stack->head >= stack->capa)) {
         stack = rvalue_stack_grow(stack, handle, stack_ref);
     }
+
     stack->ptr[stack->head] = value;
     stack->head++;
+
     return value;
 }
 
@@ -243,14 +264,14 @@ static void rvalue_stack_mark(void *ptr)
     long index;
     if (stack && stack->ptr) {
         for (index = 0; index < stack->head; index++) {
-            rb_gc_mark(stack->ptr[index]);
+            rb_gc_mark_movable(stack->ptr[index]);
         }
     }
 }
 
 static void rvalue_stack_free_buffer(rvalue_stack *stack)
 {
-    ruby_xfree(stack->ptr);
+    JSON_SIZED_FREE_N(stack->ptr, stack->capa);
     stack->ptr = NULL;
 }
 
@@ -260,7 +281,7 @@ static void rvalue_stack_free(void *ptr)
     if (stack) {
         rvalue_stack_free_buffer(stack);
 #ifndef HAVE_RUBY_TYPED_EMBEDDABLE
-        ruby_xfree(stack);
+        JSON_SIZED_FREE(stack);
 #endif
     }
 }
@@ -268,7 +289,22 @@ static void rvalue_stack_free(void *ptr)
 static size_t rvalue_stack_memsize(const void *ptr)
 {
     const rvalue_stack *stack = (const rvalue_stack *)ptr;
-    return sizeof(rvalue_stack) + sizeof(VALUE) * stack->capa;
+    size_t memsize = sizeof(VALUE) * stack->capa;
+#ifndef HAVE_RUBY_TYPED_EMBEDDABLE
+    memsize += sizeof(rvalue_stack);
+#endif
+    return memsize;
+}
+
+static void rvalue_stack_compact(void *ptr)
+{
+    rvalue_stack *stack = (rvalue_stack *)ptr;
+    long index;
+    if (stack && stack->ptr) {
+        for (index = 0; index < stack->head; index++) {
+            stack->ptr[index] = rb_gc_location(stack->ptr[index]);
+        }
+    }
 }
 
 static const rb_data_type_t JSON_Parser_rvalue_stack_type = {
@@ -277,8 +313,11 @@ static const rb_data_type_t JSON_Parser_rvalue_stack_type = {
         .dmark = rvalue_stack_mark,
         .dfree = rvalue_stack_free,
         .dsize = rvalue_stack_memsize,
+        .dcompact = rvalue_stack_compact,
     },
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_EMBEDDABLE,
+    // We deliberately don't declare rvalue_stack as RUBY_TYPED_WB_PROTECTED
+    // because it churns a lot of values so trigering write barriers every time is very costly.
+    .flags = RUBY_TYPED_THREAD_SAFE_FREE | RUBY_TYPED_EMBEDDABLE,
 };
 
 static rvalue_stack *rvalue_stack_spill(rvalue_stack *old_stack, VALUE *handle, rvalue_stack **stack_ref)
@@ -309,33 +348,62 @@ static void rvalue_stack_eagerly_release(VALUE handle)
     }
 }
 
-static int convert_UTF32_to_UTF8(char *buf, uint32_t ch)
-{
-    int len = 1;
-    if (ch <= 0x7F) {
-        buf[0] = (char) ch;
-    } else if (ch <= 0x07FF) {
-        buf[0] = (char) ((ch >> 6) | 0xC0);
-        buf[1] = (char) ((ch & 0x3F) | 0x80);
-        len++;
-    } else if (ch <= 0xFFFF) {
-        buf[0] = (char) ((ch >> 12) | 0xE0);
-        buf[1] = (char) (((ch >> 6) & 0x3F) | 0x80);
-        buf[2] = (char) ((ch & 0x3F) | 0x80);
-        len += 2;
-    } else if (ch <= 0x1fffff) {
-        buf[0] =(char) ((ch >> 18) | 0xF0);
-        buf[1] =(char) (((ch >> 12) & 0x3F) | 0x80);
-        buf[2] =(char) (((ch >> 6) & 0x3F) | 0x80);
-        buf[3] =(char) ((ch & 0x3F) | 0x80);
-        len += 3;
-    } else {
-        buf[0] = '?';
-    }
-    return len;
-}
+/* frame stack */
+
+// Iterative (non-recursive) parsing keeps an explicit stack of the containers
+// currently being built, instead of relying on the C call stack. Each frame
+// only needs enough bookkeeping to close its container: which kind it is, the
+// rvalue_stack position where its children start (so we know how many to pop),
+// and the cursor at its opening brace (used to rewind for duplicate key
+// errors). Frames hold no VALUEs, so this stack needs no GC marking; it reuses
+// the same stack-allocated-with-heap-spill strategy as the rvalue_stack so that
+// it's freed even if parsing raises.
+//
+// The lifecycle helpers below (grow/push/peek/pop/spill/free/eagerly_release
+// and the rb_data_type_t) deliberately mirror their rvalue_stack counterparts
+// -- the element type and the absence of a mark function are the only real
+// differences. Keep the two in sync: a fix to the spill/release or
+// HAVE_RUBY_TYPED_EMBEDDABLE handling in one almost certainly belongs in the
+// other.
+#define JSON_FRAME_STACK_INITIAL_CAPA 32
+
+enum json_frame_type {
+    JSON_FRAME_ROOT,   // == JSON_PHASE_DONE
+    JSON_FRAME_ARRAY,  // == JSON_PHASE_ARRAY_COMMA
+    JSON_FRAME_OBJECT, // = JSON_PHASE_OBJECT_COMMA
+};
+
+// Where a frame is within its container's grammar. This is the entirety of the
+// parser's "what to do next" state: json_parse_any dispatches on the top
+// frame's phase and holds no resume state in C locals, so a parse can stop at
+// any value boundary and be resumed purely from the (persistable) frame stack.
+//
+// The first three phases are deliberately equal to the corresponding json_frame_type
+// to simplify the transition of phase in json_value_completed.
+enum json_frame_phase {
+    JSON_PHASE_DONE         = JSON_FRAME_ROOT,   // root only: the document value has been parsed
+    JSON_PHASE_ARRAY_COMMA  = JSON_FRAME_ARRAY,  // after a value: expecting ',' or the closing ']'
+    JSON_PHASE_OBJECT_COMMA = JSON_FRAME_OBJECT,  // after a value: expecting ',' or the closing '}'
+    JSON_PHASE_VALUE,  // expecting a value (document root, array element, or object value after ':')
+    JSON_PHASE_OBJECT_KEY,    // expecting a '"' key (after '{' or ',')
+    JSON_PHASE_OBJECT_COLON,  // object only: after a key, expecting ':'
+};
+
+typedef struct json_frame_struct {
+    enum json_frame_type type;
+    enum json_frame_phase phase;
+    long value_stack_head;    // rvalue_stack->head when this container opened
+    size_t start_offset; // object frames only (the '{'); NULL otherwise
+} json_frame;
+
+typedef struct json_frame_stack_struct {
+    enum rvalue_stack_type type; // shared with rvalue_stack: is ptr stack- or heap-allocated
+    long capa;
+    long head;
+    json_frame *ptr;
+} json_frame_stack;
 
-enum duplicate_key_action {
+enum deprecatable_action {
     JSON_DEPRECATED = 0,
     JSON_IGNORE,
     JSON_RAISE,
@@ -345,7 +413,8 @@ typedef struct JSON_ParserStruct {
     VALUE on_load_proc;
     VALUE decimal_class;
     ID decimal_method_id;
-    enum duplicate_key_action on_duplicate_key;
+    enum deprecatable_action on_duplicate_key;
+    enum deprecatable_action on_comment;
     int max_nesting;
     bool allow_nan;
     bool allow_trailing_comma;
@@ -356,17 +425,152 @@ typedef struct JSON_ParserStruct {
 } JSON_ParserConfig;
 
 typedef struct JSON_ParserStateStruct {
-    VALUE *stack_handle;
+    VALUE *value_stack_handle;
+    VALUE *frame_stack_handle;
     const char *start;
     const char *cursor;
     const char *end;
-    rvalue_stack *stack;
+    rvalue_stack *value_stack;
+    json_frame_stack *frames;
     rvalue_cache name_cache;
     int in_array;
     int current_nesting;
     unsigned int emitted_deprecations;
+    VALUE parser;
 } JSON_ParserState;
 
+static json_frame_stack *json_frame_stack_spill(json_frame_stack *old_stack, VALUE *handle, json_frame_stack **stack_ref);
+
+static json_frame_stack *json_frame_stack_grow(json_frame_stack *stack, VALUE *handle, json_frame_stack **stack_ref)
+{
+    long required = stack->capa ? stack->capa * 2 : JSON_FRAME_STACK_INITIAL_CAPA;
+
+    if (stack->type == RVALUE_STACK_STACK_ALLOCATED) {
+        stack = json_frame_stack_spill(stack, handle, stack_ref);
+    } else {
+        JSON_SIZED_REALLOC_N(stack->ptr, json_frame, required, stack->capa);
+        stack->capa = required;
+    }
+    return stack;
+}
+
+static json_frame *json_frame_stack_push(JSON_ParserState *state, json_frame frame)
+{
+    json_frame_stack *stack = state->frames;
+
+    JSON_ASSERT(stack->type != RVALUE_STACK_STACK_ALLOCATED || state->frame_stack_handle);
+
+    if (RB_UNLIKELY(stack->head >= stack->capa)) {
+        stack = json_frame_stack_grow(stack, state->frame_stack_handle, &state->frames);
+    }
+
+    json_frame *frame_ptr = &stack->ptr[stack->head++];
+    *frame_ptr = frame;
+    return frame_ptr;
+}
+
+static inline json_frame *json_frame_stack_peek(json_frame_stack *stack)
+{
+    return &stack->ptr[stack->head - 1];
+}
+
+static inline void json_frame_stack_pop(json_frame_stack *stack)
+{
+    stack->head--;
+}
+
+static void json_frame_stack_free_buffer(json_frame_stack *stack)
+{
+    JSON_SIZED_FREE_N(stack->ptr, stack->capa);
+    stack->ptr = NULL;
+}
+
+static void json_frame_stack_free(void *ptr)
+{
+    json_frame_stack *stack = (json_frame_stack *)ptr;
+    if (stack) {
+        json_frame_stack_free_buffer(stack);
+#ifndef HAVE_RUBY_TYPED_EMBEDDABLE
+        JSON_SIZED_FREE(stack);
+#endif
+    }
+}
+
+static size_t json_frame_stack_memsize(const void *ptr)
+{
+    const json_frame_stack *stack = (const json_frame_stack *)ptr;
+
+    size_t memsize = sizeof(json_frame) * stack->capa;
+#ifndef HAVE_RUBY_TYPED_EMBEDDABLE
+    memsize += sizeof(json_frame_stack);
+#endif
+    return memsize;
+}
+
+static const rb_data_type_t JSON_Parser_frame_stack_type = {
+    .wrap_struct_name = "JSON::Ext::Parser/frame_stack",
+    .function = {
+        .dmark = NULL,
+        .dfree = json_frame_stack_free,
+        .dsize = json_frame_stack_memsize,
+    },
+    .flags = RUBY_TYPED_THREAD_SAFE_FREE | RUBY_TYPED_WB_PROTECTED | RUBY_TYPED_EMBEDDABLE,
+};
+
+static json_frame_stack *json_frame_stack_spill(json_frame_stack *old_stack, VALUE *handle, json_frame_stack **stack_ref)
+{
+    json_frame_stack *stack;
+    *handle = TypedData_Make_Struct(0, json_frame_stack, &JSON_Parser_frame_stack_type, stack);
+    *stack_ref = stack;
+    MEMCPY(stack, old_stack, json_frame_stack, 1);
+
+    stack->capa = old_stack->capa << 1;
+    stack->ptr = ALLOC_N(json_frame, stack->capa);
+    stack->type = RVALUE_STACK_HEAP_ALLOCATED;
+    MEMCPY(stack->ptr, old_stack->ptr, json_frame, old_stack->head);
+    return stack;
+}
+
+static void json_frame_stack_eagerly_release(VALUE handle)
+{
+    if (handle) {
+        json_frame_stack *stack;
+        TypedData_Get_Struct(handle, json_frame_stack, &JSON_Parser_frame_stack_type, stack);
+#ifdef HAVE_RUBY_TYPED_EMBEDDABLE
+        json_frame_stack_free_buffer(stack);
+#else
+        json_frame_stack_free(stack);
+        RTYPEDDATA_DATA(handle) = NULL;
+#endif
+    }
+}
+
+static int convert_UTF32_to_UTF8(char *buf, uint32_t ch)
+{
+    int len = 1;
+    if (ch <= 0x7F) {
+        buf[0] = (char) ch;
+    } else if (ch <= 0x07FF) {
+        buf[0] = (char) ((ch >> 6) | 0xC0);
+        buf[1] = (char) ((ch & 0x3F) | 0x80);
+        len++;
+    } else if (ch <= 0xFFFF) {
+        buf[0] = (char) ((ch >> 12) | 0xE0);
+        buf[1] = (char) (((ch >> 6) & 0x3F) | 0x80);
+        buf[2] = (char) ((ch & 0x3F) | 0x80);
+        len += 2;
+    } else if (ch <= 0x1fffff) {
+        buf[0] =(char) ((ch >> 18) | 0xF0);
+        buf[1] =(char) (((ch >> 12) & 0x3F) | 0x80);
+        buf[2] =(char) (((ch >> 6) & 0x3F) | 0x80);
+        buf[3] =(char) ((ch & 0x3F) | 0x80);
+        len += 3;
+    } else {
+        buf[0] = '?';
+    }
+    return len;
+}
+
 static inline size_t rest(JSON_ParserState *state) {
     return state->end - state->cursor;
 }
@@ -398,6 +602,7 @@ static void cursor_position(JSON_ParserState *state, long *line_out, long *colum
 
     while (cursor >= state->start) {
         if (*cursor-- == '\n') {
+            line++;
             break;
         }
         column++;
@@ -412,6 +617,8 @@ static void cursor_position(JSON_ParserState *state, long *line_out, long *colum
     *column_out = column;
 }
 
+static const unsigned int MAX_DEPRECATIONS = 5;
+
 static void emit_parse_warning(const char *message, JSON_ParserState *state)
 {
     long line, column;
@@ -423,7 +630,7 @@ static void emit_parse_warning(const char *message, JSON_ParserState *state)
 
 #define PARSE_ERROR_FRAGMENT_LEN 32
 
-static VALUE build_parse_error_message(const char *format, JSON_ParserState *state, long line, long column)
+static VALUE build_parse_error_message(const char *format, JSON_ParserState *state)
 {
     unsigned char buffer[PARSE_ERROR_FRAGMENT_LEN + 3];
 
@@ -457,31 +664,61 @@ static VALUE build_parse_error_message(const char *format, JSON_ParserState *sta
         }
     }
 
-    VALUE message = rb_enc_sprintf(enc_utf8, format, ptr);
-    rb_str_catf(message, " at line %ld column %ld", line, column);
-    return message;
+    return rb_enc_sprintf(enc_utf8, format, ptr);
 }
 
-static VALUE parse_error_new(VALUE message, long line, long column)
+static VALUE parse_error_new(JSON_ParserState *state, VALUE message, long line, long column, bool eos)
 {
-    VALUE exc = rb_exc_new_str(rb_path2class("JSON::ParserError"), message);
-    rb_ivar_set(exc, rb_intern("@line"), LONG2NUM(line));
-    rb_ivar_set(exc, rb_intern("@column"), LONG2NUM(column));
+    VALUE exc = rb_exc_new_str(eParserError, message);
+    rb_ivar_set(exc, i_at_line, LONG2NUM(line));
+    rb_ivar_set(exc, i_at_column, LONG2NUM(column));
     return exc;
 }
 
-NORETURN(static) void raise_parse_error(const char *format, JSON_ParserState *state)
+NORETURN(static) void raise_parse_error(const char *format, JSON_ParserState *state, bool eos)
 {
-    long line, column;
-    cursor_position(state, &line, &column);
-    VALUE message = build_parse_error_message(format, state, line, column);
-    rb_exc_raise(parse_error_new(message, line, column));
+    if (state->parser) { 
+        if (eos) {
+            // the error will be swallowed by ResumableParser#parse, so no
+            // point building a message or backtrace.
+            rb_throw_obj(state->parser, state->parser);
+        } else {
+            // line and columns can't be accurate in resumable
+            rb_exc_raise(parse_error_new(state, build_parse_error_message(format, state), 0, 0, eos));
+        }
+    } else {
+        VALUE message = build_parse_error_message(format, state);
+        long line, column;
+        cursor_position(state, &line, &column);
+        rb_str_catf(message, " at line %ld column %ld", line, column);
+        rb_exc_raise(parse_error_new(state, message, line, column, eos));
+    }
+}
+
+NORETURN(static) void raise_eos_error(const char *format, JSON_ParserState *state)
+{
+    raise_parse_error(format, state, true);
 }
 
-NORETURN(static) void raise_parse_error_at(const char *format, JSON_ParserState *state, const char *at)
+NORETURN(static) void raise_syntax_error(const char *format, JSON_ParserState *state)
+{
+    raise_parse_error(format, state, false);
+}
+
+NORETURN(static) void raise_parse_error_at(const char *format, JSON_ParserState *state, const char *at, bool eos)
 {
     state->cursor = at;
-    raise_parse_error(format, state);
+    raise_parse_error(format, state, eos);
+}
+
+NORETURN(static) void raise_eos_error_at(const char *format, JSON_ParserState *state, const char *at)
+{
+    raise_parse_error_at(format, state, at, true);
+}
+
+NORETURN(static) void raise_syntax_error_at(const char *format, JSON_ParserState *state, const char *at)
+{
+    raise_parse_error_at(format, state, at, false);
 }
 
 /* unicode */
@@ -506,7 +743,7 @@ static const signed char digit_values[256] = {
 static uint32_t unescape_unicode(JSON_ParserState *state, const char *sp, const char *spe)
 {
     if (RB_UNLIKELY(sp > spe - 4)) {
-        raise_parse_error_at("incomplete unicode character escape sequence at %s", state, sp - 2);
+        raise_eos_error_at("incomplete unicode character escape sequence at %s", state, sp - 2);
     }
 
     const unsigned char *p = (const unsigned char *)sp;
@@ -517,7 +754,7 @@ static uint32_t unescape_unicode(JSON_ParserState *state, const char *sp, const
     const signed char b3 = digit_values[p[3]];
 
     if (RB_UNLIKELY((signed char)(b0 | b1 | b2 | b3) < 0)) {
-        raise_parse_error_at("incomplete unicode character escape sequence at %s", state, sp - 2);
+        raise_syntax_error_at("incomplete unicode character escape sequence at %s", state, sp - 2);
     }
 
     return ((uint32_t)b0 << 12) | ((uint32_t)b1 << 8) | ((uint32_t)b2 << 4) | (uint32_t)b3;
@@ -529,19 +766,36 @@ static uint32_t unescape_unicode(JSON_ParserState *state, const char *sp, const
 
 static const rb_data_type_t JSON_ParserConfig_type;
 
-static void
-json_eat_comments(JSON_ParserState *state)
+const char *COMMENT_DEPRECATION_MESSAGE = "Encountered comment in JSON. This will raise an error in json 3.0 unless enabled via `allow_comments: true`";
+NOINLINE(static) void
+json_eat_comments(JSON_ParserState *state, JSON_ParserConfig *config, const char *resume_pos)
 {
+    if (config->on_comment == JSON_RAISE) {
+        raise_syntax_error("unexpected token %s", state);
+    }
+
     const char *start = state->cursor;
+    // An incomplete comment suspends a resumable parse by rewinding the cursor
+    // and throwing. Callers that already consumed a token not yet committed to
+    // the frame stack pass resume_pos so the rewind re-reads that token too.
+    // Non-resumable error positions keep pointing at the comment either way.
+    const char *rewind_pos = (state->parser && resume_pos) ? resume_pos : start;
     state->cursor++;
 
     switch (peek(state)) {
         case '/': {
-            state->cursor = memchr(state->cursor, '\n', state->end - state->cursor);
-            if (!state->cursor) {
+            const char *newline = memchr(state->cursor, '\n', state->end - state->cursor);
+            if (!newline) {
+                // state->parser marks resumable mode, where the buffer end is only a
+                // chunk boundary: the terminating newline may still arrive, so leave
+                // the comment unterminated instead of consuming to end as a one-shot
+                // parse would.
+                if (state->parser) {
+                    raise_eos_error_at("unterminated comment, expected end of line", state, rewind_pos);
+                }
                 state->cursor = state->end;
             } else {
-                state->cursor++;
+                state->cursor = newline + 1;
             }
             break;
         }
@@ -551,7 +805,7 @@ json_eat_comments(JSON_ParserState *state)
             while (true) {
                 const char *next_match = memchr(state->cursor, '*', state->end - state->cursor);
                 if (!next_match) {
-                    raise_parse_error_at("unterminated comment, expected closing '*/'", state, start);
+                    raise_eos_error_at("unterminated comment, expected closing '*/'", state, rewind_pos);
                 }
 
                 state->cursor = next_match + 1;
@@ -563,13 +817,18 @@ json_eat_comments(JSON_ParserState *state)
             break;
         }
         default:
-            raise_parse_error_at("unexpected token %s", state, start);
+            raise_parse_error_at("unexpected token %s", state, eos(state) ? rewind_pos : start, eos(state));
             break;
     }
+
+    if (config->on_comment == JSON_DEPRECATED && state->emitted_deprecations < MAX_DEPRECATIONS) {
+        state->emitted_deprecations++;
+        emit_parse_warning(COMMENT_DEPRECATION_MESSAGE, state);
+    }
 }
 
 ALWAYS_INLINE(static) void
-json_eat_whitespace(JSON_ParserState *state)
+json_eat_whitespace_resume_at(JSON_ParserState *state, JSON_ParserConfig *config, bool include_comments, const char *resume_pos)
 {
     while (true) {
         switch (peek(state)) {
@@ -600,7 +859,11 @@ json_eat_whitespace(JSON_ParserState *state)
                 state->cursor++;
                 break;
             case '/':
-                json_eat_comments(state);
+                if (!include_comments) {
+                    return;
+                }
+
+                json_eat_comments(state, config, resume_pos);
                 break;
 
             default:
@@ -609,6 +872,12 @@ json_eat_whitespace(JSON_ParserState *state)
     }
 }
 
+ALWAYS_INLINE(static) void
+json_eat_whitespace(JSON_ParserState *state, JSON_ParserConfig *config, bool include_comments)
+{
+    json_eat_whitespace_resume_at(state, config, include_comments, NULL);
+}
+
 static inline VALUE build_string(const char *start, const char *end, bool intern, bool symbolize)
 {
     if (symbolize) {
@@ -754,13 +1023,13 @@ NOINLINE(static) VALUE json_string_unescape(JSON_ParserState *state, JSON_Parser
                         uint32_t sur = unescape_unicode(state, pe + 2, stringEnd);
 
                         if (RB_UNLIKELY((sur & 0xFC00) != 0xDC00)) {
-                            raise_parse_error_at("invalid surrogate pair at %s", state, p);
+                            raise_syntax_error_at("invalid surrogate pair at %s", state, p);
                         }
 
                         ch = (((ch & 0x3F) << 10) | ((((ch >> 6) & 0xF) + 1) << 16) | (sur & 0x3FF));
                         pe += 5;
                     } else {
-                        raise_parse_error_at("incomplete surrogate pair at %s", state, p);
+                        raise_syntax_error_at("incomplete surrogate pair at %s", state, p);
                         break;
                     }
                 }
@@ -770,20 +1039,22 @@ NOINLINE(static) VALUE json_string_unescape(JSON_ParserState *state, JSON_Parser
                 p = ++pe;
                 break;
             }
+            case 0:
+              return Qundef;
             default:
                 if ((unsigned char)*pe < 0x20) {
                     if (!config->allow_control_characters) {
                         if (*pe == '\n') {
-                            raise_parse_error_at("Invalid unescaped newline character (\\n) in string: %s", state, pe - 1);
+                            raise_syntax_error_at("Invalid unescaped newline character (\\n) in string: %s", state, pe - 1);
                         }
-                        raise_parse_error_at("invalid ASCII control character in string: %s", state, pe - 1);
+                        raise_syntax_error_at("invalid ASCII control character in string: %s", state, pe - 1);
                     }
                 }
 
                 if (config->allow_invalid_escape) {
                     APPEND_CHAR(*pe);
                 } else {
-                    raise_parse_error_at("invalid escape character in string: %s", state, pe - 1);
+                    raise_syntax_error_at("invalid escape character in string: %s", state, pe - 1);
                 }
                 break;
         }
@@ -879,19 +1150,24 @@ static inline VALUE json_decode_float(JSON_ParserConfig *config, uint64_t mantis
         return rb_float_new(negative ? -0.0 : 0.0);
     }
 
-    // Fall back to rb_cstr_to_dbl for potential subnormals (rare edge case)
-    // Ryu has rounding issues with subnormals around 1e-310 (< 2.225e-308)
-    if (RB_UNLIKELY(mantissa_digits > 17 || mantissa_digits + exponent < -307)) {
+    if (RB_UNLIKELY(mantissa_digits > 18 || mantissa_digits + exponent < -307)) {
+        // If the value is so small that it definitely underflows to 0.0, return early
+        // to avoid triggering a "Float out of range" warning from rb_cstr_to_dbl.
+        // When mantissa_digits + exponent < -324, value < 10^(-324) < DBL_TRUE_MIN/2,
+        // so it rounds to 0 in IEEE 754 round-to-nearest.
+        if (RB_UNLIKELY(mantissa_digits + exponent < -324)) {
+            return rb_float_new(negative ? -0.0 : 0.0);
+        }
         return json_decode_large_float(start, end - start);
     }
 
-    return DBL2NUM(ryu_s2d_from_parts(mantissa, mantissa_digits, (int32_t)exponent, negative));
+    return DBL2NUM(ffp_s2d(exponent, mantissa, negative));
 }
 
 static inline VALUE json_decode_array(JSON_ParserState *state, JSON_ParserConfig *config, long count)
 {
-    VALUE array = rb_ary_new_from_values(count, rvalue_stack_peek(state->stack, count));
-    rvalue_stack_pop(state->stack, count);
+    VALUE array = rb_ary_new_from_values(count, rvalue_stack_peek(state->value_stack, count));
+    rvalue_stack_pop(state->value_stack, count);
 
     if (config->freeze) {
         RB_OBJ_FREEZE(array);
@@ -935,38 +1211,50 @@ NORETURN(static) void raise_duplicate_key_error(JSON_ParserState *state, VALUE d
         rb_inspect(duplicate_key)
     );
 
-    long line, column;
-    cursor_position(state, &line, &column);
-    rb_str_concat(message, build_parse_error_message("", state, line, column)) ;
-    rb_exc_raise(parse_error_new(message, line, column));
+    rb_str_concat(message, build_parse_error_message("", state));
+    if (state->parser) { // line and columns can't be accurate in resumable
+        rb_exc_raise(parse_error_new(state, message, 0, 0, false));
+    } else {
+        long line, column;
+        cursor_position(state, &line, &column);
+        rb_str_catf(message, " at line %ld column %ld", line, column);
+        rb_exc_raise(parse_error_new(state, message, line, column, false));
+    }
+}
+
+NOINLINE(static) void json_on_duplicate_key(JSON_ParserState *state, JSON_ParserConfig *config, size_t count, const VALUE *pairs)
+{
+    switch (config->on_duplicate_key) {
+        case JSON_IGNORE:
+            return;
+
+        case JSON_DEPRECATED:
+            // Only emit the first few deprecations to avoid spamming.
+            if (state->emitted_deprecations < MAX_DEPRECATIONS) {
+                state->emitted_deprecations++;
+                emit_duplicate_key_warning(state, json_find_duplicated_key(count, pairs));
+            }
+            return;
+
+        case JSON_RAISE:
+            raise_duplicate_key_error(state, json_find_duplicated_key(count, pairs));
+            return;
+    }
+    UNREACHABLE;
 }
 
 static inline VALUE json_decode_object(JSON_ParserState *state, JSON_ParserConfig *config, size_t count)
 {
     size_t entries_count = count / 2;
     VALUE object = rb_hash_new_capa(entries_count);
-    const VALUE *pairs = rvalue_stack_peek(state->stack, count);
+    const VALUE *pairs = rvalue_stack_peek(state->value_stack, count);
     rb_hash_bulk_insert(count, pairs, object);
 
     if (RB_UNLIKELY(RHASH_SIZE(object) < entries_count)) {
-        switch (config->on_duplicate_key) {
-            case JSON_IGNORE:
-                break;
-            case JSON_DEPRECATED:
-                // Only emit the first few deprecations to avoid spamming.
-                if (state->emitted_deprecations < 5) {
-                    emit_duplicate_key_warning(state, json_find_duplicated_key(count, pairs));
-                    state->emitted_deprecations++;
-                }
-
-                break;
-            case JSON_RAISE:
-                raise_duplicate_key_error(state, json_find_duplicated_key(count, pairs));
-                break;
-        }
+        json_on_duplicate_key(state, config, count, pairs);
     }
 
-    rvalue_stack_pop(state->stack, count);
+    rvalue_stack_pop(state->value_stack, count);
 
     if (config->freeze) {
         RB_OBJ_FREEZE(object);
@@ -980,7 +1268,7 @@ static inline VALUE json_push_value(JSON_ParserState *state, JSON_ParserConfig *
     if (RB_UNLIKELY(config->on_load_proc)) {
         value = rb_proc_call_with_block(config->on_load_proc, 1, &value, Qnil);
     }
-    rvalue_stack_push(state->stack, value, state->stack_handle, &state->stack);
+    rvalue_stack_push(state->value_stack, value, state->value_stack_handle, &state->value_stack);
     return value;
 }
 
@@ -1053,7 +1341,7 @@ static VALUE json_parse_escaped_string(JSON_ParserState *state, JSON_ParserConfi
             case '"': {
                 VALUE string = json_string_unescape(state, config, start, state->cursor, is_name, &positions);
                 state->cursor++;
-                return json_push_value(state, config, string);
+                return string;
             }
             case '\\': {
                 if (RB_LIKELY(positions.size < JSON_MAX_UNESCAPE_POSITIONS)) {
@@ -1067,7 +1355,7 @@ static VALUE json_parse_escaped_string(JSON_ParserState *state, JSON_ParserConfi
             }
             default:
                 if (!config->allow_control_characters) {
-                    raise_parse_error("invalid ASCII control character in string: %s", state);
+                    raise_syntax_error("invalid ASCII control character in string: %s", state);
                 }
                 break;
         }
@@ -1075,8 +1363,7 @@ static VALUE json_parse_escaped_string(JSON_ParserState *state, JSON_ParserConfi
         state->cursor++;
     } while (string_scan(state));
 
-    raise_parse_error("unexpected end of input, expected closing \"", state);
-    return Qfalse;
+    return Qundef;
 }
 
 ALWAYS_INLINE(static) VALUE json_parse_string(JSON_ParserState *state, JSON_ParserConfig *config, bool is_name)
@@ -1085,15 +1372,19 @@ ALWAYS_INLINE(static) VALUE json_parse_string(JSON_ParserState *state, JSON_Pars
     const char *start = state->cursor;
 
     if (RB_UNLIKELY(!string_scan(state))) {
-        raise_parse_error("unexpected end of input, expected closing \"", state);
+        return Qundef;
     }
 
+    VALUE string;
     if (RB_LIKELY(*state->cursor == '"')) {
-        VALUE string = json_string_fastpath(state, config, start, state->cursor, is_name);
+        string = json_string_fastpath(state, config, start, state->cursor, is_name);
         state->cursor++;
-        return json_push_value(state, config, string);
     }
-    return json_parse_escaped_string(state, config, is_name, start);
+    else {
+        string = json_parse_escaped_string(state, config, is_name, start);
+    }
+
+    return string;
 }
 
 #if JSON_CPU_LITTLE_ENDIAN_64BITS
@@ -1166,7 +1457,7 @@ static inline int json_parse_digits(JSON_ParserState *state, uint64_t *accumulat
     return (int)(state->cursor - start);
 }
 
-static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig *config, bool negative, const char *start)
+static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig *config, bool negative, const char *start, bool resumable)
 {
     bool integer = true;
     const char first_digit = *state->cursor;
@@ -1180,7 +1471,7 @@ static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig
     int mantissa_digits = json_parse_digits(state, &mantissa);
 
     if (RB_UNLIKELY((first_digit == '0' && mantissa_digits > 1) || (negative && mantissa_digits == 0))) {
-        raise_parse_error_at("invalid number: %s", state, start);
+        return Qundef;
     }
 
     // Parse fractional part
@@ -1193,7 +1484,7 @@ static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig
         mantissa_digits += fractional_digits;
 
         if (RB_UNLIKELY(!fractional_digits)) {
-            raise_parse_error_at("invalid number: %s", state, start);
+            return Qundef;
         }
     }
 
@@ -1213,7 +1504,7 @@ static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig
         int exponent_digits = json_parse_digits(state, &abs_exponent);
 
         if (RB_UNLIKELY(!exponent_digits)) {
-            raise_parse_error_at("invalid number: %s", state, start);
+            return Qundef;
         }
 
         if (RB_UNLIKELY(exponent_digits >= 20 || abs_exponent > (uint64_t)INT64_MAX)) {
@@ -1223,6 +1514,16 @@ static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig
         }
     }
 
+    // A number touching the end of the buffer may still grow in a later chunk,
+    // so the caller will rewind and wait. Decoding it now would build a value
+    // -- for a long run of digits, an expensive bignum -- only to discard it,
+    // and repeating that on every resumed chunk is quadratic in the number's
+    // length. The digit scan above already advanced the cursor, which is all
+    // the caller needs to detect the incomplete number.
+    if (RB_UNLIKELY(resumable && eos(state))) {
+        return Qundef;
+    }
+
     if (integer) {
         return json_decode_integer(mantissa, mantissa_digits, negative, start, state->cursor);
     }
@@ -1235,229 +1536,420 @@ static inline VALUE json_parse_number(JSON_ParserState *state, JSON_ParserConfig
     return json_decode_float(config, mantissa, mantissa_digits, exponent, negative, start, state->cursor);
 }
 
-static inline VALUE json_parse_positive_number(JSON_ParserState *state, JSON_ParserConfig *config)
+// How many values (array elements, or interleaved object keys+values) have been
+// pushed onto the rvalue stack since this container opened. Used to size the
+// bulk decode on close, and to tell the first key/colon from later ones.
+static inline long json_frame_entry_count(const json_frame *frame, const rvalue_stack *value_stack)
 {
-    return json_parse_number(state, config, false, state->cursor);
+    return value_stack->head - frame->value_stack_head;
 }
 
-static inline VALUE json_parse_negative_number(JSON_ParserState *state, JSON_ParserConfig *config)
+// A complete value now sits on top of the rvalue stack. Advance the frame that
+// was waiting for it: the root document is done, or the enclosing container
+// moves on to expecting a ',' or its closing bracket. The caller passes the
+// frame it already has in hand -- the one that was expecting the value -- which
+// after a container close is the freshly re-exposed parent.
+static inline enum json_frame_phase json_value_completed(json_frame *frame)
 {
-    const char *start = state->cursor;
-    state->cursor++;
-    return json_parse_number(state, config, true, start);
+    JSON_ASSERT((int)JSON_PHASE_DONE         == (int)JSON_FRAME_ROOT);
+    JSON_ASSERT((int)JSON_PHASE_ARRAY_COMMA  == (int)JSON_FRAME_ARRAY);
+    JSON_ASSERT((int)JSON_PHASE_OBJECT_COMMA == (int)JSON_FRAME_OBJECT);
+
+    return frame->phase = (enum json_frame_phase) frame->type;
 }
 
-static VALUE json_parse_any(JSON_ParserState *state, JSON_ParserConfig *config)
+ALWAYS_INLINE(static) void json_match_keyword(JSON_ParserState *state, const char *keyword, size_t offset)
 {
-    json_eat_whitespace(state);
+    // It is assumed that since `keyword` is always a literal, the compiler is able to constantize this
+    // `strlen` and several other computations in that routine.
 
-    switch (peek(state)) {
-        case 'n':
-            if (rest(state) >= 4 && (memcmp(state->cursor, "null", 4) == 0)) {
-                state->cursor += 4;
-                return json_push_value(state, config, Qnil);
-            }
+    size_t len = strlen(keyword);
 
-            raise_parse_error("unexpected token %s", state);
-            break;
-        case 't':
-            if (rest(state) >= 4 && (memcmp(state->cursor, "true", 4) == 0)) {
-                state->cursor += 4;
-                return json_push_value(state, config, Qtrue);
-            }
+    // Note: memcmp with a small power of two and a literal string compile to an integer comparison /
+    // That's why we sometime compare starting from the first byte and sometimes from the second.
+    if (rest(state) >= len && (memcmp(state->cursor + offset, keyword + offset, len - offset) == 0)) {
+        state->cursor += len;
+        return;
+    }
 
-            raise_parse_error("unexpected token %s", state);
-            break;
-        case 'f':
-            // Note: memcmp with a small power of two compile to an integer comparison
-            if (rest(state) >= 5 && (memcmp(state->cursor + 1, "alse", 4) == 0)) {
-                state->cursor += 5;
-                return json_push_value(state, config, Qfalse);
-            }
+    bool eos = rest(state) < len && memcmp(state->cursor, keyword, rest(state)) == 0;
+    raise_parse_error("unexpected token %s", state, eos);
+}
 
-            raise_parse_error("unexpected token %s", state);
-            break;
-        case 'N':
-            // Note: memcmp with a small power of two compile to an integer comparison
-            if (config->allow_nan && rest(state) >= 3 && (memcmp(state->cursor + 1, "aN", 2) == 0)) {
-                state->cursor += 3;
-                return json_push_value(state, config, CNaN);
-            }
+// Parse an arbitrary JSON value iteratively. This is a state machine driven
+// entirely by the top frame's phase so it can stop at any value boundary and
+// resume purely from the frame stack. A JSON_FRAME_ROOT frame sits at the
+// bottom of the stack, so the stack is never empty mid-parse and the document
+// itself is just another frame whose value, once parsed, leaves its phase DONE.
+// When invoked in resumable mode, it returns true after parsing a complete document.
+// If reaching EOS without having parsed a complete document, either returns false
+// of raise a JSON::ParserError tagged with `@eos=true`.
+ALWAYS_INLINE(static) bool json_parse_any(JSON_ParserState *state, JSON_ParserConfig *config, bool resumable)
+{
+    json_frame *frame = json_frame_stack_peek(state->frames);
 
-            raise_parse_error("unexpected token %s", state);
-            break;
-        case 'I':
-            if (config->allow_nan && rest(state) >= 8 && (memcmp(state->cursor, "Infinity", 8) == 0)) {
-                state->cursor += 8;
-                return json_push_value(state, config, CInfinity);
-            }
+    switch (frame->phase) {
+        case JSON_PHASE_DONE:           JSON_UNREACHABLE_RETURN(false);
+        case JSON_PHASE_ARRAY_COMMA:    goto JSON_PHASE_ARRAY_COMMA;
+        case JSON_PHASE_OBJECT_COMMA:   goto JSON_PHASE_OBJECT_COMMA;
+        case JSON_PHASE_VALUE:          goto JSON_PHASE_VALUE;
+        case JSON_PHASE_OBJECT_KEY:     goto JSON_PHASE_OBJECT_KEY;
+        case JSON_PHASE_OBJECT_COLON:   goto JSON_PHASE_OBJECT_COLON;
+    }
+    JSON_UNREACHABLE_RETURN(false);
 
-            raise_parse_error("unexpected token %s", state);
-            break;
-        case '-': {
-            // Note: memcmp with a small power of two compile to an integer comparison
-            if (rest(state) >= 9 && (memcmp(state->cursor + 1, "Infinity", 8) == 0)) {
-                if (config->allow_nan) {
-                    state->cursor += 9;
-                    return json_push_value(state, config, CMinusInfinity);
-                } else {
-                    raise_parse_error("unexpected token %s", state);
-                }
-            }
-            return json_push_value(state, config, json_parse_negative_number(state, config));
-            break;
-        }
-        case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
-            return json_push_value(state, config, json_parse_positive_number(state, config));
-            break;
-        case '"': {
-            // %r{\A"[^"\\\t\n\x00]*(?:\\[bfnrtu\\/"][^"\\]*)*"}
-            return json_parse_string(state, config, false);
-            break;
+    JSON_PHASE_VALUE: {
+        json_eat_whitespace(state, config, true);
+
+        // A trailing comma lands us here expecting an element but finding the
+        // closing bracket; hand off to ARRAY_COMMA to close. An empty array
+        // closes inline at '[', so this position is only reached after a ','.
+        if (config->allow_trailing_comma && frame->type == JSON_FRAME_ARRAY && peek(state) == ']') {
+            goto JSON_PHASE_ARRAY_COMMA;
         }
-        case '[': {
-            state->cursor++;
-            json_eat_whitespace(state);
-            long stack_head = state->stack->head;
 
-            if (peek(state) == ']') {
-                state->cursor++;
-                return json_push_value(state, config, json_decode_array(state, config, 0));
-            } else {
-                state->current_nesting++;
-                if (RB_UNLIKELY(config->max_nesting && (config->max_nesting < state->current_nesting))) {
-                    rb_raise(eNestingError, "nesting of %d is too deep", state->current_nesting);
-                }
-                state->in_array++;
-                json_parse_any(state, config);
-            }
+        VALUE value;
+        const char *value_start = state->cursor;
 
-            while (true) {
-                json_eat_whitespace(state);
+        switch (peek(state)) {
+            case 'n':
+                json_match_keyword(state, "null", 0);
+                value = Qnil;
+                break;
 
-                const char next_char = peek(state);
+            case 't':
+                json_match_keyword(state, "true", 0);
+                value = Qtrue;
+                break;
 
-                if (RB_LIKELY(next_char == ',')) {
-                    state->cursor++;
-                    if (config->allow_trailing_comma) {
-                        json_eat_whitespace(state);
-                        if (peek(state) == ']') {
-                            continue;
-                        }
-                    }
-                    json_parse_any(state, config);
-                    continue;
-                }
+            case 'f':
+                json_match_keyword(state, "false", 1);
+                value = Qfalse;
+                break;
 
-                if (next_char == ']') {
-                    state->cursor++;
-                    long count = state->stack->head - stack_head;
-                    state->current_nesting--;
-                    state->in_array--;
-                    return json_push_value(state, config, json_decode_array(state, config, count));
+            case 'N':
+                if (!config->allow_nan) {
+                    raise_syntax_error("unexpected token %s", state);
                 }
 
-                raise_parse_error("expected ',' or ']' after array value", state);
-            }
-            break;
-        }
-        case '{': {
-            const char *object_start_cursor = state->cursor;
+                json_match_keyword(state, "NaN", 1);
+                value = CNaN;
+                break;
 
-            state->cursor++;
-            json_eat_whitespace(state);
-            long stack_head = state->stack->head;
+            case 'I':
+                if (!config->allow_nan) {
+                    raise_syntax_error("unexpected token %s", state);
+                }
 
-            if (peek(state) == '}') {
+                json_match_keyword(state, "Infinity", 0);
+                value = CInfinity;
+                break;
+
+            case '-': {
                 state->cursor++;
-                return json_push_value(state, config, json_decode_object(state, config, 0));
-            } else {
-                state->current_nesting++;
-                if (RB_UNLIKELY(config->max_nesting && (config->max_nesting < state->current_nesting))) {
-                    rb_raise(eNestingError, "nesting of %d is too deep", state->current_nesting);
+
+                value = json_parse_number(state, config, true, value_start, resumable);
+
+                if (RB_UNLIKELY(UNDEF_P(value) && config->allow_nan && peek(state) == 'I')) {
+                    state->cursor = value_start;
+                    json_match_keyword(state, "-Infinity", 1);
+                    value = CMinusInfinity;
+                    break;
+                }
+
+                // Top level numbers are ambiguous when parsing streams, we can't
+                // know if we parsed all the digits if we hit EOS.
+                if (RB_UNLIKELY(resumable && eos(state))) {
+                    state->cursor = value_start;
+                    return false;
                 }
 
-                if (peek(state) != '"') {
-                    raise_parse_error("expected object key, got %s", state);
+                if (RB_UNLIKELY(UNDEF_P(value))) {
+                    raise_syntax_error_at("invalid number: %s", state, value_start);
                 }
-                json_parse_string(state, config, true);
+                break;
+            }
+
+            case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9': {
+                value = json_parse_number(state, config, false, value_start, resumable);
 
-                json_eat_whitespace(state);
-                if (peek(state) != ':') {
-                    raise_parse_error("expected ':' after object key", state);
+                // Top level numbers are ambiguous when parsing streams, we can't
+                // know if we parsed all the digits if we hit EOS.
+                if (RB_UNLIKELY(resumable && eos(state))) {
+                    state->cursor = value_start;
+                    return false;
+                }
+
+                if (RB_UNLIKELY(UNDEF_P(value))) {
+                    raise_syntax_error_at("invalid number: %s", state, value_start);
+                }
+                break;
+            }
+
+            case '"': {
+                // %r{\A"[^"\\\t\n\x00]*(?:\\[bfnrtu\\/"][^"\\]*)*"}
+                value = json_parse_string(state, config, false);
+
+                if (RB_UNLIKELY(UNDEF_P(value))) {
+                    bool is_eos = eos(state);
+                    if (resumable && is_eos) {
+                        state->cursor = value_start;
+                        return false;
+                    }
+                    raise_parse_error("unexpected end of input, expected closing \"", state, is_eos);
                 }
+                break;
+            }
+
+            case '[': {
                 state->cursor++;
+                // The '[' is consumed but its frame is only pushed below, so a
+                // comment suspending here must resume from the bracket.
+                json_eat_whitespace_resume_at(state, config, true, value_start);
 
-                json_parse_any(state, config);
+                const char next = peek(state);
+                if (next == ']') {
+                    state->cursor++;
+                    value = json_decode_array(state, config, 0);
+                    break;
+                } else if (resumable && eos(state)) {
+                    state->cursor = value_start;
+                    return false;
+                }
+
+                state->current_nesting++;
+                if (RB_UNLIKELY(config->max_nesting && (config->max_nesting < state->current_nesting))) {
+                    rb_raise(eNestingError, "nesting of %d is too deep", state->current_nesting);
+                }
+                state->in_array++;
+
+                // Phase stays VALUE: the next iteration reads the first element.
+                frame = json_frame_stack_push(state, (json_frame){
+                    .type = JSON_FRAME_ARRAY,
+                    .phase = JSON_PHASE_VALUE,
+                    .value_stack_head = state->value_stack->head,
+                });
+                goto JSON_PHASE_VALUE;
             }
 
-            while (true) {
-                json_eat_whitespace(state);
+            case '{': {
+                state->cursor++;
+                // Same as '[': the frame is only pushed below.
+                json_eat_whitespace_resume_at(state, config, true, value_start);
 
-                const char next_char = peek(state);
-                if (next_char == '}') {
+                if (peek(state) == '}') {
                     state->cursor++;
-                    state->current_nesting--;
-                    size_t count = state->stack->head - stack_head;
+                    value = json_decode_object(state, config, 0);
+                    break;
+                } else if (resumable && eos(state)) {
+                    state->cursor = value_start;
+                    return false;
+                }
+
+                state->current_nesting++;
+                if (RB_UNLIKELY(config->max_nesting && (config->max_nesting < state->current_nesting))) {
+                    rb_raise(eNestingError, "nesting of %d is too deep", state->current_nesting);
+                }
 
-                    // Temporary rewind cursor in case an error is raised
-                    const char *final_cursor = state->cursor;
-                    state->cursor = object_start_cursor;
-                    VALUE object = json_decode_object(state, config, count);
-                    state->cursor = final_cursor;
+                // Phase KEY: the next iteration reads the first key.
+                frame = json_frame_stack_push(state, (json_frame){
+                    .type = JSON_FRAME_OBJECT,
+                    .phase = JSON_PHASE_OBJECT_KEY,
+                    .value_stack_head = state->value_stack->head,
+                    .start_offset = value_start - state->start,
+                });
+                goto JSON_PHASE_OBJECT_KEY;
+            }
 
-                    return json_push_value(state, config, object);
+            case 0:
+                // peek() returns 0 both at end-of-stream and for a literal NUL byte in the
+                // buffer. Only a genuine EOS means "feed me more"; a NUL byte that is not at
+                // EOS is just an invalid character.
+                if (eos(state)) {
+                    return false;
+                } else {
+                    raise_syntax_error("unexpected NULL byte: %s", state);
                 }
+            default:
+                raise_syntax_error("unexpected character: %s", state);
+        }
 
-                if (next_char == ',') {
-                    state->cursor++;
-                    json_eat_whitespace(state);
+        json_push_value(state, config, value);
+        json_value_completed(frame);
 
-                    if (config->allow_trailing_comma) {
-                        if (peek(state) == '}') {
-                            continue;
-                        }
-                    }
+        switch (frame->phase) {
+            case JSON_PHASE_DONE:           return true;
+            case JSON_PHASE_ARRAY_COMMA:    goto JSON_PHASE_ARRAY_COMMA;
+            case JSON_PHASE_OBJECT_COMMA:   goto JSON_PHASE_OBJECT_COMMA;
+            case JSON_PHASE_VALUE:          goto JSON_PHASE_VALUE;
+            case JSON_PHASE_OBJECT_KEY:     JSON_UNREACHABLE_RETURN(false);
+            case JSON_PHASE_OBJECT_COLON:   goto JSON_PHASE_OBJECT_COLON;
+        }
+        JSON_UNREACHABLE_RETURN(false);
+    }
 
-                    if (RB_UNLIKELY(peek(state) != '"')) {
-                        raise_parse_error("expected object key, got: %s", state);
-                    }
-                    json_parse_string(state, config, true);
+    JSON_PHASE_OBJECT_KEY: {
+        JSON_ASSERT(frame->type == JSON_FRAME_OBJECT);
 
-                    json_eat_whitespace(state);
-                    if (RB_UNLIKELY(peek(state) != ':')) {
-                        raise_parse_error("expected ':' after object key, got: %s", state);
-                    }
-                    state->cursor++;
+        json_eat_whitespace(state, config, true);
+
+        // A trailing comma lands us here expecting a key but finding the closing
+        // brace; hand off to OBJECT_COMMA to close. An empty object closes inline
+        // at '{', so this position is only reached after a ','.
+        if (config->allow_trailing_comma && peek(state) == '}') {
+            goto JSON_PHASE_OBJECT_COMMA;
+        }
 
-                    json_parse_any(state, config);
+        const char *start = state->cursor;
 
-                    continue;
+        if (RB_LIKELY(peek(state) == '"')) {
+            VALUE string = json_parse_string(state, config, true);
+            if (UNDEF_P(string)) {
+                if (resumable) {
+                    state->cursor = start;
+                    return false;
+                } else {
+                    raise_syntax_error("unexpected end of input, expected closing \"", state);
                 }
+            }
+            json_push_value(state, config, string);
+            frame->phase = JSON_PHASE_OBJECT_COLON;
+            goto JSON_PHASE_OBJECT_COLON;
+        } else if (resumable && eos(state)) {
+            return false;
+        } else {
+            // The message differs for the first key vs. a key after a
+            // ',': the first is the only one reached with nothing pushed
+            // for this object yet.
+            if (json_frame_entry_count(frame, state->value_stack) == 0) {
+                raise_syntax_error("expected object key, got %s", state);
+            } else {
+                raise_syntax_error("expected object key, got: %s", state);
+            }
+        }
+        JSON_UNREACHABLE_RETURN(false);
+    }
 
-                raise_parse_error("expected ',' or '}' after object value, got: %s", state);
+    JSON_PHASE_OBJECT_COLON: {
+        JSON_ASSERT(frame->type == JSON_FRAME_OBJECT);
+
+        json_eat_whitespace(state, config, true);
+
+        if (RB_LIKELY(peek(state) == ':')) {
+            state->cursor++;
+            frame->phase = JSON_PHASE_VALUE;
+            goto JSON_PHASE_VALUE;
+        } else if (resumable && eos(state)) {
+            return false;
+        } else {
+            // First colon (only the first pair's key is pushed, nothing
+            // else) vs. a later one.
+            if (json_frame_entry_count(frame, state->value_stack) == 1) {
+                raise_syntax_error("expected ':' after object key", state);
+            } else {
+                raise_syntax_error("expected ':' after object key, got: %s", state);
             }
-            break;
         }
+        JSON_UNREACHABLE_RETURN(false);
+    }
 
-        case 0:
-            raise_parse_error("unexpected end of input", state);
-            break;
+    JSON_PHASE_ARRAY_COMMA: {
+        JSON_ASSERT(frame->type == JSON_FRAME_ARRAY);
 
-        default:
-            raise_parse_error("unexpected character: %s", state);
-            break;
+        json_eat_whitespace(state, config, true);
+
+        const char next_char = peek(state);
+
+        if (RB_LIKELY(next_char == ',')) {
+            state->cursor++;
+            // Commit the phase before eating the whitespace that follows: an
+            // incomplete comment there would suspend the parse, and a phase not
+            // yet advanced past the ',' would drop it on resume. A trailing comma
+            // is recognized in JSON_PHASE_VALUE once the ']' is in the buffer.
+            frame->phase = JSON_PHASE_VALUE;
+            goto JSON_PHASE_VALUE;
+        } else if (next_char == ']') {
+            state->cursor++;
+            long count = json_frame_entry_count(frame, state->value_stack);
+            state->current_nesting--;
+            state->in_array--;
+
+            json_push_value(state, config, json_decode_array(state, config, count));
+            json_frame_stack_pop(state->frames);
+            frame = json_frame_stack_peek(state->frames);
+
+            json_value_completed(frame);
+
+            switch (frame->phase) {
+                case JSON_PHASE_DONE:           return true;
+                case JSON_PHASE_ARRAY_COMMA:    goto JSON_PHASE_ARRAY_COMMA;
+                case JSON_PHASE_OBJECT_COMMA:   goto JSON_PHASE_OBJECT_COMMA;
+                case JSON_PHASE_VALUE:          goto JSON_PHASE_VALUE;
+                case JSON_PHASE_OBJECT_KEY:     JSON_UNREACHABLE_RETURN(false);
+                case JSON_PHASE_OBJECT_COLON:   goto JSON_PHASE_OBJECT_COLON;
+            }
+        } else if (resumable && eos(state)) {
+            return false;
+        } else {
+            raise_syntax_error("expected ',' or ']' after array value", state);
+        }
+        JSON_UNREACHABLE_RETURN(false);
     }
 
-    raise_parse_error("unreachable: %s", state);
-    return Qundef;
+    JSON_PHASE_OBJECT_COMMA: {
+        JSON_ASSERT(frame->type == JSON_FRAME_OBJECT);
+
+        json_eat_whitespace(state, config, true);
+        const char next_char = peek(state);
+
+        if (RB_LIKELY(next_char == ',')) {
+            state->cursor++;
+            // Commit the phase before eating the whitespace that follows: an
+            // incomplete comment there would suspend the parse, and a phase not
+            // yet advanced past the ',' would drop it on resume. A trailing comma
+            // is recognized in JSON_PHASE_OBJECT_KEY once the '}' is in the buffer.
+            frame->phase = JSON_PHASE_OBJECT_KEY;
+            goto JSON_PHASE_OBJECT_KEY;
+        } else if (next_char == '}') {
+            state->cursor++;
+            state->current_nesting--;
+            size_t count = json_frame_entry_count(frame, state->value_stack);
+
+            // Temporary rewind cursor in case an error is raised
+            const char *final_cursor = state->cursor;
+            state->cursor = state->start + frame->start_offset;
+            VALUE object = json_decode_object(state, config, count);
+            state->cursor = final_cursor;
+
+            json_push_value(state, config, object);
+            json_frame_stack_pop(state->frames);
+            frame = json_frame_stack_peek(state->frames);
+            json_value_completed(frame);
+
+            switch (frame->phase) {
+                case JSON_PHASE_DONE:           return true;
+                case JSON_PHASE_ARRAY_COMMA:    goto JSON_PHASE_ARRAY_COMMA;
+                case JSON_PHASE_OBJECT_COMMA:   goto JSON_PHASE_OBJECT_COMMA;
+                case JSON_PHASE_VALUE:          goto JSON_PHASE_VALUE;
+                case JSON_PHASE_OBJECT_KEY:     JSON_UNREACHABLE_RETURN(false);
+                case JSON_PHASE_OBJECT_COLON:   goto JSON_PHASE_OBJECT_COLON;
+            }
+        } else if (resumable && eos(state)) {
+            return false;
+        } else {
+            raise_syntax_error("expected ',' or '}' after object value, got: %s", state);
+        }
+        JSON_UNREACHABLE_RETURN(false);
+    }
+
+    JSON_UNREACHABLE_RETURN(false);
 }
 
-static void json_ensure_eof(JSON_ParserState *state)
+static void json_ensure_eof(JSON_ParserState *state, JSON_ParserConfig *config)
 {
-    json_eat_whitespace(state);
+    json_eat_whitespace(state, config, true);
     if (!eos(state)) {
-        raise_parse_error("unexpected token at end of stream %s", state);
+        raise_syntax_error("unexpected token at end of stream %s", state);
     }
 }
 
@@ -1495,6 +1987,8 @@ static VALUE convert_encoding(VALUE source)
 struct parser_config_init_args {
     JSON_ParserConfig *config;
     VALUE self;
+    VALUE unknown_keywords;
+    bool strict;
 };
 
 static void parser_config_wb_write(VALUE self, VALUE *dest, VALUE val)
@@ -1512,6 +2006,7 @@ static int parser_config_init_i(VALUE key, VALUE val, VALUE data)
          if (key == sym_max_nesting)                { config->max_nesting = RTEST(val) ? FIX2INT(val) : 0; }
     else if (key == sym_allow_nan)                  { config->allow_nan = RTEST(val); }
     else if (key == sym_allow_trailing_comma)       { config->allow_trailing_comma = RTEST(val); }
+    else if (key == sym_allow_comments)             { config->on_comment = RTEST(val) ? JSON_IGNORE : JSON_RAISE; }
     else if (key == sym_allow_control_characters)   { config->allow_control_characters = RTEST(val); }
     else if (key == sym_allow_invalid_escape)       { config->allow_invalid_escape = RTEST(val); }
     else if (key == sym_symbolize_names)            { config->symbolize_names = RTEST(val); }
@@ -1547,27 +2042,42 @@ static int parser_config_init_i(VALUE key, VALUE val, VALUE data)
             }
         }
     }
+    else if (args->strict) {
+        if (!args->unknown_keywords) {
+            args->unknown_keywords = rb_obj_hide(rb_ary_new());
+        }
+        rb_ary_push(args->unknown_keywords, key);
+    }
 
     return ST_CONTINUE;
 }
 
-static void parser_config_init(JSON_ParserConfig *config, VALUE opts, VALUE self)
+static void parser_config_init(JSON_ParserConfig *config, VALUE opts, VALUE self, bool strict)
 {
     config->max_nesting = 100;
 
     struct parser_config_init_args args = {
         .config = config,
         .self = self,
+        .strict = strict,
     };
 
-    if (!NIL_P(opts)) {
-        Check_Type(opts, T_HASH);
-        if (RHASH_SIZE(opts) > 0) {
-            // We assume in most cases few keys are set so it's faster to go over
-            // the provided keys than to check all possible keys.
-            rb_hash_foreach(opts, parser_config_init_i, (VALUE)&args);
-        }
+    if (NIL_P(opts)) return;
+    Check_Type(opts, T_HASH);
+    if (RHASH_SIZE(opts) == 0) return;
 
+    // We assume in most cases few keys are set so it's faster to go over
+    // the provided keys than to check all possible keys.
+    rb_hash_foreach(opts, parser_config_init_i, (VALUE)&args);
+
+    if (RB_UNLIKELY(args.unknown_keywords)) {
+        if (RARRAY_LEN(args.unknown_keywords) == 1) {
+            rb_raise(rb_eArgError, "unknown keyword: %" PRIsVALUE, RARRAY_AREF(args.unknown_keywords, 0));
+        }
+        else {
+            VALUE keywords = rb_ary_join(args.unknown_keywords, rb_utf8_str_new_cstr(", "));
+            rb_raise(rb_eArgError, "unknown keywords: %" PRIsVALUE, keywords);
+        }
     }
 }
 
@@ -1576,30 +2086,16 @@ static void parser_config_init(JSON_ParserConfig *config, VALUE opts, VALUE self
  *
  * Creates a new JSON::Ext::ParserConfig instance.
  *
- * It will be configured by the _opts_ hash. _opts_ can have the following
- * keys:
+ * Argument +opts+, if given, contains a \Hash of options for the parsing.
+ * See {Parsing Options}[#module-JSON-label-Parsing+Options].
  *
- * _opts_ can have the following keys:
- * * *max_nesting*: The maximum depth of nesting allowed in the parsed data
- *   structures. Disable depth checking with :max_nesting => false|nil|0, it
- *   defaults to 100.
- * * *allow_nan*: If set to true, allow NaN, Infinity and -Infinity in
- *   defiance of RFC 4627 to be parsed by the Parser. This option defaults to
- *   false.
- * * *symbolize_names*: If set to true, returns symbols for the names
- *   (keys) in a JSON object. Otherwise strings are returned, which is
- *   also the default. It's not possible to use this option in
- *   conjunction with the *create_additions* option.
- * * *decimal_class*: Specifies which class to use instead of the default
- *    (Float) when parsing decimal numbers. This class must accept a single
- *    string argument in its constructor.
  */
 static VALUE cParserConfig_initialize(VALUE self, VALUE opts)
 {
     rb_check_frozen(self);
     GET_PARSER_CONFIG;
 
-    parser_config_init(config, opts, self);
+    parser_config_init(config, opts, self, false);
 
     return self;
 }
@@ -1616,35 +2112,64 @@ static VALUE cParser_parse(JSON_ParserConfig *config, VALUE src)
     }
 
     VALUE rvalue_stack_buffer[RVALUE_STACK_INITIAL_CAPA];
-    rvalue_stack stack = {
+    rvalue_stack value_stack = {
         .type = RVALUE_STACK_STACK_ALLOCATED,
         .ptr = rvalue_stack_buffer,
         .capa = RVALUE_STACK_INITIAL_CAPA,
     };
 
+    // Seed the frame stack with the root frame, establishing the invariant that
+    // json_parse_any always has a top frame to dispatch on (so the stack is never
+    // empty mid-parse).
+    json_frame frame_stack_buffer[JSON_FRAME_STACK_INITIAL_CAPA];
+    frame_stack_buffer[0] = (json_frame){
+        .type = JSON_FRAME_ROOT,
+        .phase = JSON_PHASE_VALUE,
+    };
+    json_frame_stack frames = {
+        .type = RVALUE_STACK_STACK_ALLOCATED,
+        .ptr = frame_stack_buffer,
+        .capa = JSON_FRAME_STACK_INITIAL_CAPA,
+        .head = 1,
+    };
+
     long len;
     const char *start;
 
     RSTRING_GETMEM(Vsource, start, len);
 
-    VALUE stack_handle = 0;
+    VALUE value_stack_handle = 0;
+    VALUE frame_stack_handle = 0;
     JSON_ParserState _state = {
         .start = start,
         .cursor = start,
         .end = start + len,
-        .stack = &stack,
-        .stack_handle = &stack_handle,
+        .value_stack = &value_stack,
+        .value_stack_handle = &value_stack_handle,
+        .frames = &frames,
+        .frame_stack_handle = &frame_stack_handle,
     };
     JSON_ParserState *state = &_state;
 
-    VALUE result = json_parse_any(state, config);
+    bool complete = json_parse_any(state, config, false);
+
+    // The root document value is parsed; it is the lone survivor on
+    // the rvalue stack.
+    VALUE result = complete ? *rvalue_stack_peek(state->value_stack, 1) : Qundef;
 
     // This may be skipped in case of exception, but
     // it won't cause a leak.
-    rvalue_stack_eagerly_release(stack_handle);
-    RB_GC_GUARD(stack_handle);
+    rvalue_stack_eagerly_release(value_stack_handle);
+    json_frame_stack_eagerly_release(frame_stack_handle);
+    RB_GC_GUARD(value_stack_handle);
+    RB_GC_GUARD(frame_stack_handle);
     RB_GC_GUARD(Vsource);
-    json_ensure_eof(state);
+
+    if (complete) {
+        json_ensure_eof(state, config);
+    } else {
+        raise_eos_error("unexpected end of input", state);
+    }
 
     return result;
 }
@@ -1666,7 +2191,7 @@ static VALUE cParser_m_parse(VALUE klass, VALUE Vsource, VALUE opts)
 {
     JSON_ParserConfig _config = {0};
     JSON_ParserConfig *config = &_config;
-    parser_config_init(config, opts, false);
+    parser_config_init(config, opts, Qfalse, false);
 
     return cParser_parse(config, Vsource);
 }
@@ -1674,23 +2199,35 @@ static VALUE cParser_m_parse(VALUE klass, VALUE Vsource, VALUE opts)
 static void JSON_ParserConfig_mark(void *ptr)
 {
     JSON_ParserConfig *config = ptr;
-    rb_gc_mark(config->on_load_proc);
-    rb_gc_mark(config->decimal_class);
+    rb_gc_mark_movable(config->on_load_proc);
+    rb_gc_mark_movable(config->decimal_class);
 }
 
 static size_t JSON_ParserConfig_memsize(const void *ptr)
 {
+#ifdef HAVE_RUBY_TYPED_EMBEDDABLE
+    return 0;
+#else
     return sizeof(JSON_ParserConfig);
+#endif
+}
+
+static void JSON_ParserConfig_compact(void *ptr)
+{
+    JSON_ParserConfig *config = ptr;
+    config->on_load_proc = rb_gc_location(config->on_load_proc);
+    config->decimal_class = rb_gc_location(config->decimal_class);
 }
 
 static const rb_data_type_t JSON_ParserConfig_type = {
     .wrap_struct_name = "JSON::Ext::Parser/ParserConfig",
     .function = {
-        JSON_ParserConfig_mark,
-        RUBY_DEFAULT_FREE,
-        JSON_ParserConfig_memsize,
+        .dmark = JSON_ParserConfig_mark,
+        .dfree = RUBY_DEFAULT_FREE,
+        .dsize = JSON_ParserConfig_memsize,
+        .dcompact = JSON_ParserConfig_compact,
     },
-    .flags = RUBY_TYPED_FREE_IMMEDIATELY | RUBY_TYPED_WB_PROTECTED | RUBY_TYPED_FROZEN_SHAREABLE | RUBY_TYPED_EMBEDDABLE,
+    .flags = RUBY_TYPED_THREAD_SAFE_FREE | RUBY_TYPED_WB_PROTECTED | RUBY_TYPED_FROZEN_SHAREABLE | RUBY_TYPED_EMBEDDABLE,
 };
 
 static VALUE cJSON_parser_s_allocate(VALUE klass)
@@ -1699,6 +2236,602 @@ static VALUE cJSON_parser_s_allocate(VALUE klass)
     return TypedData_Make_Struct(klass, JSON_ParserConfig, &JSON_ParserConfig_type, config);
 }
 
+static void json_str_clear(VALUE str)
+{
+    if (RB_OBJ_FROZEN_RAW(str)) {
+        return;
+    }
+    rb_str_replace(str, JSON_empty_string);
+}
+
+typedef struct JSON_ResumableParserStruct {
+    JSON_ParserConfig config;
+    JSON_ParserState state;
+    rvalue_stack value_stack;
+    json_frame_stack frames;
+    VALUE buffer;
+    size_t parsed_bytes;
+    size_t incomplete_bytes;
+    bool complete;
+    bool in_use;
+} JSON_ResumableParser;
+
+static void JSON_ResumableParser_mark(void *ptr)
+{
+    JSON_ResumableParser *parser = (JSON_ResumableParser *)ptr;
+    JSON_ParserConfig_mark(&parser->config);
+    rvalue_stack_mark(&parser->value_stack);
+    rvalue_cache_mark(&parser->state.name_cache);
+    rb_gc_mark(parser->buffer); // pin the buffer
+    rb_gc_mark_movable(parser->state.parser);
+}
+
+static void JSON_ResumableParser_free(void *ptr)
+{
+    JSON_ResumableParser *parser = (JSON_ResumableParser *)ptr;
+    rvalue_stack_free_buffer(&parser->value_stack);
+    json_frame_stack_free_buffer(&parser->frames);
+}
+
+static size_t JSON_ResumableParser_memsize(const void *ptr)
+{
+    const JSON_ResumableParser *parser = (const JSON_ResumableParser *)ptr;
+    size_t memsize = JSON_ParserConfig_memsize(&parser->config);
+    memsize += rvalue_stack_memsize(&parser->value_stack);
+    memsize += json_frame_stack_memsize(&parser->frames);
+#ifndef HAVE_RUBY_TYPED_EMBEDDABLE
+    memsize += (
+        sizeof(JSON_ResumableParser)
+        - sizeof(JSON_ParserState)
+        - sizeof(JSON_ParserConfig)
+        - sizeof(rvalue_stack)
+        - sizeof(json_frame_stack)
+    );
+#endif
+    return memsize;
+}
+
+static void JSON_ResumableParser_compact(void *ptr)
+{
+    JSON_ResumableParser *parser = (JSON_ResumableParser *)ptr;
+    JSON_ParserConfig_compact(&parser->config);
+    rvalue_stack_compact(&parser->value_stack);
+    rvalue_cache_compact(&parser->state.name_cache);
+    parser->buffer = rb_gc_location(parser->buffer);
+    parser->state.parser = rb_gc_location(parser->state.parser);
+}
+
+static const rb_data_type_t JSON_ResumableParser_type = {
+    .wrap_struct_name = "JSON::Ext::ResumableParser",
+    .function = {
+        JSON_ResumableParser_mark,
+        JSON_ResumableParser_free,
+        JSON_ResumableParser_memsize,
+        JSON_ResumableParser_compact,
+    },
+    // RUBY_TYPED_WB_PROTECTED is deliberately not declared because
+    // this is a superset of JSON_Parser_rvalue_stack_type, so we'd need
+    // to trigger a lot of write barriers.
+    .flags = RUBY_TYPED_THREAD_SAFE_FREE | RUBY_TYPED_EMBEDDABLE,
+};
+
+static VALUE cResumableParser_allocate(VALUE klass)
+{
+    JSON_ResumableParser *parser;
+    VALUE obj = TypedData_Make_Struct(klass, JSON_ResumableParser, &JSON_ResumableParser_type, parser);
+    parser->state.in_array++;
+    parser->state.parser = obj;
+    return obj;
+}
+
+static inline JSON_ResumableParser *cResumableParser_get(VALUE self)
+{
+    JSON_ResumableParser *parser;
+    TypedData_Get_Struct(self, JSON_ResumableParser, &JSON_ResumableParser_type, parser);
+    return parser;
+}
+
+/*
+ * call-seq: new(opts => {})
+ *
+ * Creates a new JSON::ResumableParser instance.
+ *
+ * Argument +opts+, if given, contains a \Hash of options for the parsing.
+ * See {Parsing Options}[#module-JSON-label-Parsing+Options].
+ *
+ * A ResumableParser is able to parse partial documents and resume parsing later
+ * when more of the document is provided:
+ *
+ *   parser = JSON::ResumableParser.new
+ *   parser << '{"user": "george", "role": "ad'
+ *   parser.parse # => false
+ *   parser.eos? # => true
+ *   parser.partial_value # => { "user" => "george", "role" => nil }
+ *   parser.rest # => '"ad'
+ *
+ *   parser << 'min" }[1, 2, 3]'
+ *   parser.parse # => true
+ *   parser.value # => { "user" => "george", "role" => "admin" }
+ *
+ *   parser.parse # => true
+ *   parser.value # => [1, 2, 3]
+ *
+ * === Limitations
+ *
+ * While ResumableParser is able to parse streams of documents without any
+ * explicit separators between them, it is highly recommended to separate documents
+ * by either spaces or newlines, as otherwise the \JSON syntax for numbers may be ambiguous.
+ * When parsing a number, ResumableParser will not consider the number complete until something follows:
+ *
+ *   parser << '123'
+ *   parser.parse # => false
+ *   parser << ' '
+ *   parser.parse # => true
+ *   parser.value # => 123
+ *
+ * === Security
+ *
+ * An incomplete document is buffered in full and there is no size limit, so when reading
+ * from an untrusted source the caller is responsible for bounding how much data is fed.
+ * For example:
+ *
+ *  loop do
+ *    if parser.parsed_bytes > DOCUMENT_MAX_SIZE
+ *      raise "document too large"
+ *    end
+ *
+ *    parser << read_chunk
+ *    while parser.parse
+ *      process(parser.value)
+ *    end
+ *  end
+ */
+static VALUE cResumableParser_initialize(int argc, VALUE *argv, VALUE self)
+{
+    rb_check_frozen(self);
+
+    VALUE opts = Qfalse;
+    rb_scan_args_kw(RB_SCAN_ARGS_LAST_HASH_KEYWORDS, argc, argv, "0:", &opts);
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+
+    opts = argc > 0 ? argv[0] : Qnil;
+    parser_config_init(&parser->config, opts, self, true);
+
+    return self;
+}
+
+static JSON_ResumableParser *ResumableParser_acquire(VALUE self, bool lock);
+
+/*
+ * call-seq: self << string -> self
+ *
+ * Appends the given string to the parser's buffer.
+ */
+static VALUE cResumableParser_feed(VALUE self, VALUE str)
+{
+    rb_check_frozen(self);
+
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, false);
+
+    str = convert_encoding(str);
+    if (!RSTRING_LEN(str)) {
+        return self;
+    }
+
+    size_t offset = parser->state.cursor - parser->state.start;
+    const size_t remaining = parser->state.end - parser->state.cursor;
+
+    if (!remaining) {
+        if (parser->buffer) {
+            json_str_clear(parser->buffer);
+        }
+        parser->buffer = RB_OBJ_FROZEN_RAW(str) ? str : rb_obj_hide(rb_str_new_shared(str));
+        offset = 0;
+    } else {
+        JSON_ASSERT(parser->buffer);
+
+        const size_t size = parser->state.end - parser->state.start;
+        const size_t consumed = size - remaining;
+
+        if (RB_OBJ_FROZEN_RAW(parser->buffer)) {
+            VALUE new_buffer = rb_obj_hide(rb_str_buf_new(remaining + RSTRING_LEN(str)));
+            rb_enc_associate_index(new_buffer, utf8_encindex);
+
+            char *old_ptr = RSTRING_PTR(parser->buffer);
+            memcpy(RSTRING_PTR(new_buffer), old_ptr + consumed, remaining);
+            rb_str_set_len(new_buffer, remaining);
+            offset = 0;
+            parser->buffer = new_buffer;
+        } else if (consumed > (size / 2) && size >= 512) {
+            rb_str_modify(parser->buffer);
+            char *old_ptr = RSTRING_PTR(parser->buffer);
+            memmove(old_ptr, old_ptr + consumed, remaining);
+            rb_str_set_len(parser->buffer, remaining);
+            offset = 0;
+        }
+        rb_str_append(parser->buffer, str);
+    }
+
+    long len;
+    const char *start;
+    RSTRING_GETMEM(parser->buffer, start, len);
+    parser->state.start = start;
+    parser->state.end = start + len;
+    parser->state.cursor = parser->state.start + offset;
+
+    return self;
+}
+
+struct json_parse_any_args {
+    JSON_ParserState *state;
+    JSON_ParserConfig *config;
+    VALUE parser;
+};
+
+static VALUE json_parse_any_resumable_safe0(RB_BLOCK_CALL_FUNC_ARGLIST(yielded_arg, _args))
+{
+    struct json_parse_any_args *args = (struct json_parse_any_args *)_args;
+    return (VALUE)json_parse_any(args->state, args->config, true);
+}
+
+static VALUE json_parse_any_resumable_safe(VALUE _args)
+{
+    struct json_parse_any_args *args = (struct json_parse_any_args *)_args;
+    VALUE result = rb_catch_obj(args->parser, json_parse_any_resumable_safe0, _args);
+    return result == args->parser ? Qfalse : result;
+}
+
+static JSON_ResumableParser *ResumableParser_acquire(VALUE self, bool lock)
+{
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+
+    if (parser->in_use) {
+        rb_raise(rb_eArgError, "ResumableParser can't be used recursively");
+    }
+
+    if (lock) {
+        parser->in_use = true;
+    }
+
+    // self may have moved, so we need to update all pointers
+    // Investigate: We might be better off keeping JSON_ParserState on the stack
+    // and only persist what we need.
+    parser->state.value_stack = &parser->value_stack;
+    parser->state.frames = &parser->frames;
+
+    return parser;
+}
+
+/*
+ * call-seq: parse -> true or false
+ *
+ * Attemps to parse a JSON document from the internal buffer.
+ * Returns whether a complete document could be parsed.
+ *
+ * It does raise +JSON::ParserError+ when encountering invalid \JSON syntax.
+ *
+ * The parsed object can be retrieved by calling #value
+ */
+static VALUE cResumableParser_parse(VALUE self)
+{
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, true);
+
+    if (parser->complete) {
+        parser->parsed_bytes = 0;
+        parser->incomplete_bytes = 0;
+        parser->complete = false;
+    }
+
+    if (!parser->buffer) {
+        parser->in_use = false;
+        return Qfalse;
+    }
+
+    if (parser->frames.head == 0) {
+        json_frame_stack_push(&parser->state, (json_frame){
+            .type = JSON_FRAME_ROOT,
+            .phase = JSON_PHASE_VALUE,
+        });
+    }
+
+    VALUE Vsource = parser->buffer; // Prevent compaction
+
+    json_frame *frame = json_frame_stack_peek(&parser->frames);
+
+    if (frame->phase == JSON_PHASE_DONE) {
+        JSON_ASSERT(parser->value_stack.head == 1);
+        JSON_ASSERT(parser->frames.head == 1);
+
+        frame->phase = JSON_PHASE_VALUE;
+        rvalue_stack_pop(parser->state.value_stack, 1);
+    }
+
+    struct json_parse_any_args args = {
+        .state = &parser->state,
+        .config = &parser->config,
+        .parser = self,
+    };
+    int status;
+    const char *initial_cursor = parser->state.cursor;
+    parser->complete = rb_protect(json_parse_any_resumable_safe, (VALUE)&args, &status);
+
+    if (status) {
+        parser->complete = true; // a parse error is considered complete
+    }
+
+    parser->parsed_bytes += parser->state.cursor - initial_cursor;
+    parser->incomplete_bytes = parser->complete ? 0 : parser->state.end - parser->state.cursor;
+
+    json_eat_whitespace(&parser->state, &parser->config, false);
+    if (eos(&parser->state)) {
+        json_str_clear(parser->buffer);
+        parser->buffer = Qfalse;
+    }
+    parser->in_use = false;
+
+    if (status) {
+        rb_jump_tag(status); // reraise
+    }
+    RB_GC_GUARD(Vsource);
+    return parser->complete ? Qtrue : Qfalse;
+}
+
+/*
+ * call-seq: value? -> true or false
+ *
+ * Returns whether a parsed value is available.
+ */
+static VALUE cResumableParser_value_p(VALUE self)
+{
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, false);
+
+    if (parser->value_stack.head > 0) {
+        json_frame *frame = json_frame_stack_peek(&parser->frames);
+        if (frame->phase == JSON_PHASE_DONE) {
+            return Qtrue;
+        }
+    }
+    return Qfalse;
+}
+
+/*
+ * call-seq: value -> object
+ *
+ * Returns and consume the last parsed value.
+ * Raises ArgumentError if there is no parsed value or if it was already retrieved:
+ *   parser << '[1][2]'
+ *   parser.value # ArgumentError no ready value
+ *   parser.parse # => true
+ *   parser.value # => [1]
+ *   parser.value # ArgumentError no ready value
+ */
+static VALUE cResumableParser_value(VALUE self)
+{
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, false);
+
+    if (parser->frames.head > 0) {
+        json_frame *frame = json_frame_stack_peek(&parser->frames);
+
+        if (frame->phase == JSON_PHASE_DONE) {
+            VALUE result = *rvalue_stack_peek(parser->state.value_stack, 1);
+            rvalue_stack_pop(parser->state.value_stack, 1);
+            json_frame_stack_pop(parser->state.frames);
+            return result;
+        }
+    }
+    rb_raise(rb_eArgError, "no ready value");
+}
+
+/*
+ * call-seq: clear -> self
+ *
+ * Entirely reset the parser state and buffer.
+ */
+static VALUE cResumableParser_clear(VALUE self)
+{
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, false);
+    parser->buffer = 0;
+    parser->complete = true;
+    parser->parsed_bytes = 0;
+    parser->incomplete_bytes = 0;
+    parser->frames.head = 0;
+    parser->value_stack.head = 0;
+    parser->state.name_cache.length = 0;
+    parser->state.current_nesting = 0;
+    parser->state.in_array = 1;
+    parser->state.emitted_deprecations = 0;
+    parser->state.start = parser->state.cursor = parser->state.end = NULL;
+    return self;
+}
+
+static VALUE cResumableParser_partial_value_body(VALUE self)
+{
+    JSON_ResumableParser *original_parser = cResumableParser_get(self);
+    JSON_ResumableParser parser = *original_parser;
+
+    parser.state.frames = &parser.frames;
+    parser.state.value_stack = &parser.value_stack;
+
+    if (parser.value_stack.head == 0) {
+        return Qnil;
+    }
+
+    json_frame *frame = json_frame_stack_peek(parser.state.frames);
+    long missing_object_value = 0;
+    if (frame->type == JSON_FRAME_OBJECT && (frame->phase == JSON_PHASE_VALUE || frame->phase == JSON_PHASE_OBJECT_COLON)) {
+        missing_object_value = 1;
+    }
+
+    // Copy the value stack as we need to mutate it. The collapse loop folds each
+    // open container by popping its entries and pushing the single result, so a
+    // parent always reclaims its child's slot; head exceeds its live size by at
+    // most one, either for the missing-value placeholder pushed below or for the
+    // result of folding an empty innermost container. That one spare slot keeps
+    // rvalue_stack_push from growing (reallocating) this ALLOCV buffer.
+    long capa = parser.value_stack.head;
+    parser.value_stack.capa = capa + 1;
+    VALUE tmpbuf, *value_stack_buffer = ALLOCV_N(VALUE, tmpbuf, parser.value_stack.capa);
+    MEMCPY(value_stack_buffer, parser.value_stack.ptr, VALUE, capa);
+    parser.value_stack.ptr = value_stack_buffer;
+
+    JSON_ParserState *state = &parser.state;
+    JSON_ParserConfig *config = &parser.config;
+
+    if (missing_object_value) {
+        rvalue_stack_push(state->value_stack, Qnil, NULL, &state->value_stack);
+    }
+
+    VALUE partial_result = Qundef;
+
+    while (UNDEF_P(partial_result)) {
+        frame = json_frame_stack_peek(state->frames);
+
+        switch (frame->type) {
+            case JSON_FRAME_ROOT: {
+                partial_result = *rvalue_stack_peek(state->value_stack, 1);
+                break;
+            }
+
+            case JSON_FRAME_ARRAY: {
+                long count = json_frame_entry_count(frame, state->value_stack);
+                json_push_value(state, config, json_decode_array(state, config, count));
+                json_frame_stack_pop(state->frames);
+
+                break;
+            }
+
+            case JSON_FRAME_OBJECT: {
+                long count = json_frame_entry_count(frame, state->value_stack);
+                json_push_value(state, config, json_decode_object(state, config, count));
+                json_frame_stack_pop(state->frames);
+                break;
+            }
+
+            default: {
+                JSON_UNREACHABLE_RETURN(Qundef);
+                break;
+            }
+        }
+    }
+
+    ALLOCV_END(tmpbuf);
+    return partial_result;
+}
+
+/*
+ * call-seq: partial_value -> object
+ *
+ * Returns the Ruby objects parsed up to this point:
+ *   parser << '[1, [2, 3,'
+ *   parser.parse # => false
+ *   parser.value # ArgumentError no ready value
+ *   parser.partial_value # => [1, [2, 3]]
+ */
+static VALUE cResumableParser_partial_value(VALUE self)
+{
+    JSON_ResumableParser *parser = ResumableParser_acquire(self, true);
+
+    int status;
+    VALUE result = rb_protect(cResumableParser_partial_value_body, self, &status);
+    parser->in_use = false;
+    if (status) {
+        rb_jump_tag(status);
+    }
+    return result;
+}
+
+/*
+ * call-seq: rest -> string
+ *
+ * Returns a string containing what remains to be parsed in the buffer
+ *   parser << '{ "message": "unterminated message'
+ *   parser.parse # => false
+ *   parser.rest # => '"unterminated message"'
+ */
+static VALUE cResumableParser_rest(VALUE self)
+{
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+
+    if (!parser->buffer) {
+        return rb_utf8_str_new("", 0);
+    }
+
+    size_t offset = parser->state.cursor - parser->state.start;
+    const char *ptr;
+    long len;
+    RSTRING_GETMEM(parser->buffer, ptr, len);
+    return rb_utf8_str_new(ptr + offset, len - offset);
+}
+
+/*
+ * call-seq: eos? -> true or false
+ *
+ * Returns whether the internal buffer has been entirely consumed.
+ */
+static VALUE cResumableParser_eos_p(VALUE self)
+{
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+    return eos(&parser->state) ? Qtrue : Qfalse;
+}
+
+/*
+ * call-seq: partial_value? -> true or false
+ *
+ * Returns whether a document is currently under construction: an unclosed
+ * container, a key awaiting its value, etc.
+ *
+ * It answers the same question as <tt>!partial_value.nil?</tt>, but as a
+ * cheap predicate on the parser's internal state, without materializing the
+ * partially parsed Ruby objects:
+ *   parser << '{"a":1,'
+ *   parser.parse # => false
+ *   parser.partial_value? # => true
+ *
+ * A fully parsed document whose value hasn't been retrieved yet is not under
+ * construction: #value? returns true and #partial_value? returns false.
+ */
+static VALUE cResumableParser_partial_value_p(VALUE self)
+{
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+
+    // Mirror of #value?: values on the stack while the document isn't DONE
+    // belong to a partially built document. A container whose first key or
+    // element hasn't been parsed yet has no frame nor value registered (the
+    // tokenizer rewinds to the container start on EOS), so that state is
+    // observable through the buffer (#eos?/#rest) instead, keeping this
+    // predicate consistent with #partial_value returning nil.
+    if (parser->value_stack.head > 0) {
+        json_frame *frame = json_frame_stack_peek(&parser->frames);
+        if (frame->phase != JSON_PHASE_DONE) {
+            return Qtrue;
+        }
+    }
+    return Qfalse;
+}
+
+/*
+ * call-seq: parsed_bytes -> integer
+ *
+ * Returns the number of bytes parsed since the start of the current partial value.
+ * This is intended to be used for securing against untrusted input:
+ *
+ *  loop do
+ *    if parser.parsed_bytes > DOCUMENT_MAX_SIZE
+ *      raise "document too large"
+ *    end
+ *
+ *    parser << read_chunk
+ *    while parser.parse
+ *      process(parser.value)
+ *    end
+ *  end
+ */
+static VALUE cResumableParser_parsed_bytes(VALUE self)
+{
+    JSON_ResumableParser *parser = cResumableParser_get(self);
+    return ULL2NUM(parser->parsed_bytes + parser->incomplete_bytes);
+}
+
 void Init_parser(void)
 {
 #ifdef HAVE_RB_EXT_RACTOR_SAFE
@@ -1710,30 +2843,53 @@ void Init_parser(void)
     mJSON = rb_define_module("JSON");
     VALUE mExt = rb_define_module_under(mJSON, "Ext");
     VALUE cParserConfig = rb_define_class_under(mExt, "ParserConfig", rb_cObject);
+
+    rb_global_variable(&eParserError);
+    eParserError = rb_path2class("JSON::ParserError");
+
+    rb_global_variable(&eNestingError);
     eNestingError = rb_path2class("JSON::NestingError");
-    rb_gc_register_mark_object(eNestingError);
+
     rb_define_alloc_func(cParserConfig, cJSON_parser_s_allocate);
-    rb_define_method(cParserConfig, "initialize", cParserConfig_initialize, 1);
+    rb_define_private_method(cParserConfig, "initialize", cParserConfig_initialize, 1);
     rb_define_method(cParserConfig, "parse", cParserConfig_parse, 1);
 
     VALUE cParser = rb_define_class_under(mExt, "Parser", rb_cObject);
     rb_define_singleton_method(cParser, "parse", cParser_m_parse, 2);
 
+    VALUE cResumableParser = rb_define_class_under(mJSON, "ResumableParser", rb_cObject);
+    rb_define_alloc_func(cResumableParser, cResumableParser_allocate);
+    rb_define_private_method(cResumableParser, "initialize", cResumableParser_initialize, -1);
+    rb_define_method(cResumableParser, "<<", cResumableParser_feed, 1);
+    rb_define_method(cResumableParser, "parse", cResumableParser_parse, 0);
+    rb_define_method(cResumableParser, "value", cResumableParser_value, 0);
+    rb_define_method(cResumableParser, "value?", cResumableParser_value_p, 0);
+    rb_define_method(cResumableParser, "partial_value", cResumableParser_partial_value, 0);
+    rb_define_method(cResumableParser, "partial_value?", cResumableParser_partial_value_p, 0);
+    rb_define_method(cResumableParser, "clear", cResumableParser_clear, 0);
+    rb_define_method(cResumableParser, "rest", cResumableParser_rest, 0);
+    rb_define_method(cResumableParser, "eos?", cResumableParser_eos_p, 0);
+    rb_define_method(cResumableParser, "parsed_bytes", cResumableParser_parsed_bytes, 0);
+
+    rb_global_variable(&CNaN);
     CNaN = rb_const_get(mJSON, rb_intern("NaN"));
-    rb_gc_register_mark_object(CNaN);
 
+    rb_global_variable(&CInfinity);
     CInfinity = rb_const_get(mJSON, rb_intern("Infinity"));
-    rb_gc_register_mark_object(CInfinity);
 
+    rb_global_variable(&CMinusInfinity);
     CMinusInfinity = rb_const_get(mJSON, rb_intern("MinusInfinity"));
-    rb_gc_register_mark_object(CMinusInfinity);
 
     rb_global_variable(&Encoding_UTF_8);
     Encoding_UTF_8 = rb_const_get(rb_path2class("Encoding"), rb_intern("UTF_8"));
 
+    rb_global_variable(&JSON_empty_string);
+    JSON_empty_string = rb_obj_hide(rb_utf8_str_new("", 0));
+
     sym_max_nesting = ID2SYM(rb_intern("max_nesting"));
     sym_allow_nan = ID2SYM(rb_intern("allow_nan"));
     sym_allow_trailing_comma = ID2SYM(rb_intern("allow_trailing_comma"));
+    sym_allow_comments = ID2SYM(rb_intern("allow_comments"));
     sym_allow_control_characters = ID2SYM(rb_intern("allow_control_characters"));
     sym_allow_invalid_escape = ID2SYM(rb_intern("allow_invalid_escape"));
     sym_symbolize_names = ID2SYM(rb_intern("symbolize_names"));
@@ -1744,8 +2900,12 @@ void Init_parser(void)
 
     i_new = rb_intern("new");
     i_try_convert = rb_intern("try_convert");
+#ifndef HAVE_RB_STR_TO_INTERNED_STR
     i_uminus = rb_intern("-@");
+#endif
     i_encode = rb_intern("encode");
+    i_at_line = rb_intern("@line");
+    i_at_column = rb_intern("@column");
 
     binary_encindex = rb_ascii8bit_encindex();
     utf8_encindex = rb_utf8_encindex();