zrip 0.1.2 → 0.2.1

data/ext/zrip/src/lib.rs

@@ -1,44 +1,89 @@
-use magnus::{
-    exception::ExceptionClass, function, method, prelude::*, r_string::RString, value::Opaque,
-    Error, Ruby,
-};
-use std::cell::RefCell;
-use std::sync::{Mutex, OnceLock};
+mod rb;
 
+use std::ffi::c_void;
+use std::panic::{catch_unwind, AssertUnwindSafe};
+use std::sync::{Mutex, OnceLock, TryLockError};
+
+use rb_sys::{rb_data_type_struct__bindgen_ty_1, rb_data_type_t, size_t, VALUE};
 use zstd::dict::fastcover::FastCoverParams;
 use zstd::{CompressContext, DecompressContext, Dictionary};
 
-static DECOMPRESS_ERROR: OnceLock<Opaque<ExceptionClass>> = OnceLock::new();
-static COMPRESS_ERROR: OnceLock<Opaque<ExceptionClass>> = OnceLock::new();
-static MISSING_CONTENT_SIZE_ERROR: OnceLock<Opaque<ExceptionClass>> = OnceLock::new();
-static OUTPUT_SIZE_LIMIT_ERROR: OnceLock<Opaque<ExceptionClass>> = OnceLock::new();
+use crate::rb::{RbResult, RubyErr};
+
+static DECOMPRESS_ERROR: OnceLock<GlobalValue> = OnceLock::new();
+static COMPRESS_ERROR: OnceLock<GlobalValue> = OnceLock::new();
+static MISSING_CONTENT_SIZE_ERROR: OnceLock<GlobalValue> = OnceLock::new();
+static OUTPUT_SIZE_LIMIT_ERROR: OnceLock<GlobalValue> = OnceLock::new();
+
+const GVL_COMPRESS_THRESHOLD: usize = 64 * 1024;
+const GVL_FRAME_DECOMPRESS_THRESHOLD: usize = 64 * 1024;
 
-fn decompress_error(ruby: &Ruby) -> ExceptionClass {
-    ruby.get_inner(
-        *DECOMPRESS_ERROR
-            .get()
-            .expect("DecompressError not initialized"),
-    )
+#[derive(Copy, Clone)]
+struct GlobalValue(VALUE);
+
+unsafe impl Send for GlobalValue {}
+unsafe impl Sync for GlobalValue {}
+
+fn stored_value(lock: &OnceLock<GlobalValue>, name: &str) -> VALUE {
+    lock.get()
+        .unwrap_or_else(|| panic!("{name} not initialized"))
+        .0
 }
 
-fn compress_error(ruby: &Ruby) -> ExceptionClass {
-    ruby.get_inner(*COMPRESS_ERROR.get().expect("CompressError not initialized"))
+fn decompress_error() -> VALUE {
+    stored_value(&DECOMPRESS_ERROR, "DecompressError")
 }
 
-fn missing_content_size_error(ruby: &Ruby) -> ExceptionClass {
-    ruby.get_inner(
-        *MISSING_CONTENT_SIZE_ERROR
-            .get()
-            .expect("MissingContentSizeError not initialized"),
-    )
+fn compress_error() -> VALUE {
+    stored_value(&COMPRESS_ERROR, "CompressError")
 }
 
-fn output_size_limit_error(ruby: &Ruby) -> ExceptionClass {
-    ruby.get_inner(
-        *OUTPUT_SIZE_LIMIT_ERROR
-            .get()
-            .expect("OutputSizeLimitError not initialized"),
-    )
+fn output_size_limit_error() -> VALUE {
+    stored_value(&OUTPUT_SIZE_LIMIT_ERROR, "OutputSizeLimitError")
+}
+
+fn should_release_compress_gvl(input_len: usize) -> bool {
+    input_len >= GVL_COMPRESS_THRESHOLD
+}
+
+fn should_release_frame_decompress_gvl(output_len: usize) -> bool {
+    output_len >= GVL_FRAME_DECOMPRESS_THRESHOLD
+}
+
+fn with_mutex<T, R, F>(
+    mutex: &Mutex<T>,
+    release_gvl: bool,
+    name: &str,
+    input: &mut rb::InputBytes,
+    func: F,
+) -> RbResult<R>
+where
+    F: FnOnce(&mut T, &[u8]) -> RbResult<R>,
+{
+    if release_gvl {
+        input.lock_for_without_gvl(true)?;
+        return rb::maybe_without_gvl(true, || {
+            let mut guard = mutex
+                .lock()
+                .map_err(|_| RubyErr::runtime(format!("{name} mutex poisoned")))?;
+            func(&mut guard, input.as_slice())
+        })?;
+    }
+
+    match mutex.try_lock() {
+        Ok(mut guard) => func(&mut guard, input.as_slice()),
+        Err(TryLockError::WouldBlock) => {
+            // Even a small input needs protection when contention releases GVL.
+            input.lock_for_without_gvl(true)?;
+            rb::maybe_without_gvl(true, || {
+                let mut guard = mutex
+                    .lock()
+                    .map_err(|_| RubyErr::runtime(format!("{name} mutex poisoned")))?;
+                func(&mut guard, input.as_slice())
+            })?
+        }
+        Err(TryLockError::Poisoned(_)) => Err(RubyErr::runtime(format!("{name} mutex poisoned"))),
+    }
 }
 
 // ---------- frame header parsing ----------
@@ -48,54 +93,17 @@ const ZSTD_FRAME_MAGIC: [u8; 4] = [0x28, 0xB5, 0x2F, 0xFD];
 #[derive(Debug)]
 enum BoundedError {
     BadMagic,
-    MissingContentSize,
-    OutputSizeLimit { declared: u64, limit: u64 },
+    OutputSizeLimit { limit: u64 },
     DecoderFailed(String),
 }
 
 fn parse_frame_content_size(input: &[u8]) -> Result<Option<u64>, BoundedError> {
-    if input.len() < 5 {
-        return Err(BoundedError::BadMagic);
-    }
-    if input[..4] != ZSTD_FRAME_MAGIC {
-        return Err(BoundedError::BadMagic);
-    }
-    let fhd = input[4];
-    let fcs_flag = (fhd >> 6) & 3;
-    let single_segment = (fhd >> 5) & 1;
-    let dict_id_flag = fhd & 3;
-
-    let window_desc_size = if single_segment == 0 { 1usize } else { 0 };
-    let dict_id_size = [0usize, 1, 2, 4][dict_id_flag as usize];
-    let fcs_field_size = match fcs_flag {
-        0 => {
-            if single_segment == 1 {
-                1usize
-            } else {
-                return Ok(None);
-            }
-        }
-        1 => 2,
-        2 => 4,
-        3 => 8,
-        _ => unreachable!(),
-    };
-
-    let fcs_offset = 5 + window_desc_size + dict_id_size;
-    if input.len() < fcs_offset + fcs_field_size {
-        return Err(BoundedError::BadMagic);
-    }
-
-    let fcs_bytes = &input[fcs_offset..fcs_offset + fcs_field_size];
-    let value = match fcs_field_size {
-        1 => fcs_bytes[0] as u64,
-        2 => u16::from_le_bytes([fcs_bytes[0], fcs_bytes[1]]) as u64 + 256,
-        4 => u32::from_le_bytes([fcs_bytes[0], fcs_bytes[1], fcs_bytes[2], fcs_bytes[3]]) as u64,
-        8 => u64::from_le_bytes(fcs_bytes.try_into().unwrap()),
-        _ => unreachable!(),
-    };
-
-    Ok(Some(value))
+    zstd::frame::header::parse_frame_header(input)
+        .map(|header| header.frame_content_size)
+        .map_err(|err| match err {
+            zstd::DecompressError::BadMagic => BoundedError::BadMagic,
+            _ => BoundedError::DecoderFailed(err.to_string()),
+        })
 }
 
 fn decompress_bounded(
@@ -107,71 +115,186 @@ fn decompress_bounded(
         return Err(BoundedError::BadMagic);
     }
 
-    let upper = match parse_frame_content_size(compressed)? {
-        Some(n) => {
-            if max_output != 0 && n > max_output as u64 {
-                return Err(BoundedError::OutputSizeLimit {
-                    declared: n,
-                    limit: max_output as u64,
-                });
-            }
-            if n > u64::from(u32::MAX) {
-                return Err(BoundedError::OutputSizeLimit {
-                    declared: n,
-                    limit: u64::from(u32::MAX),
-                });
-            }
-            n as usize
-        }
-        None => {
-            if max_output != 0 {
-                return Err(BoundedError::MissingContentSize);
-            }
-            1024 * 1024
-        }
+    let limit = if max_output == 0 {
+        usize::MAX
+    } else {
+        max_output
     };
-
-    let result = dctx
-        .decompress_with_limit(compressed, upper)
-        .map_err(|e| BoundedError::DecoderFailed(format!("{e}")))?;
-
-    Ok(result.into_owned())
+    dctx.decompress_with_limit(compressed, limit)
+        .map(|out| out.into_owned())
+        .map_err(|err| match err {
+            zstd::DecompressError::OutputTooSmall => BoundedError::OutputSizeLimit {
+                limit: max_output as u64,
+            },
+            _ => BoundedError::DecoderFailed(err.to_string()),
+        })
+}
+
+fn decompress_work_size(compressed: &[u8], max_output: usize) -> usize {
+    match parse_frame_content_size(compressed) {
+        Ok(Some(n)) => usize::try_from(n).unwrap_or(usize::MAX),
+        Ok(None) if max_output != 0 => max_output,
+        _ => compressed.len(),
+    }
 }
 
-fn raise_bounded(ruby: &Ruby, err: BoundedError, prefix: &str) -> Error {
+fn bounded_err(err: BoundedError, prefix: &str) -> RubyErr {
     match err {
-        BoundedError::BadMagic => Error::new(
-            decompress_error(ruby),
+        BoundedError::BadMagic => RubyErr::new(
+            decompress_error(),
             format!("{prefix}: bad magic (input is not a Zstd frame)"),
         ),
-        BoundedError::MissingContentSize => Error::new(
-            missing_content_size_error(ruby),
-            format!("{prefix}: Frame_Content_Size absent from frame header"),
-        ),
-        BoundedError::OutputSizeLimit { declared, limit } => Error::new(
-            output_size_limit_error(ruby),
-            format!("{prefix}: declared content size {declared} exceeds limit {limit}"),
+        BoundedError::OutputSizeLimit { limit } => RubyErr::new(
+            output_size_limit_error(),
+            format!("{prefix}: decompressed output exceeds limit {limit}"),
         ),
         BoundedError::DecoderFailed(msg) => {
-            Error::new(decompress_error(ruby), format!("{prefix}: {msg}"))
+            RubyErr::new(decompress_error(), format!("{prefix}: {msg}"))
         }
     }
 }
 
 // ---------- dict helper ----------
 
-fn load_dict(ruby: &Ruby, bytes: &[u8]) -> Result<Dictionary, Error> {
+fn load_dict(bytes: &[u8]) -> RbResult<Dictionary> {
     Dictionary::from_bytes(bytes).map_err(|_| {
-        Error::new(
-            ruby.exception_runtime_error(),
-            "dictionary must be in ZDICT format (use DictTrainer to train one)",
-        )
+        RubyErr::runtime("dictionary must be in ZDICT format (use DictTrainer to train one)")
     })
 }
 
+// ---------- typed data ----------
+
+struct NativeDataType(rb_data_type_t);
+
+unsafe impl Send for NativeDataType {}
+unsafe impl Sync for NativeDataType {}
+
+static FRAME_CODEC_DATA_TYPE: OnceLock<NativeDataType> = OnceLock::new();
+static BLOCK_CODEC_DATA_TYPE: OnceLock<NativeDataType> = OnceLock::new();
+static DICT_TRAINER_DATA_TYPE: OnceLock<NativeDataType> = OnceLock::new();
+
+fn frame_codec_data_type() -> *const rb_data_type_t {
+    &FRAME_CODEC_DATA_TYPE
+        .get_or_init(|| NativeDataType(make_frame_codec_data_type()))
+        .0
+}
+
+fn block_codec_data_type() -> *const rb_data_type_t {
+    &BLOCK_CODEC_DATA_TYPE
+        .get_or_init(|| NativeDataType(make_block_codec_data_type()))
+        .0
+}
+
+fn dict_trainer_data_type() -> *const rb_data_type_t {
+    &DICT_TRAINER_DATA_TYPE
+        .get_or_init(|| NativeDataType(make_dict_trainer_data_type()))
+        .0
+}
+
+fn make_frame_codec_data_type() -> rb_data_type_t {
+    rb_data_type_t {
+        wrap_struct_name: c"zrip_frame_codec".as_ptr(),
+        function: rb_data_type_struct__bindgen_ty_1 {
+            dmark: None,
+            dfree: Some(frame_codec_free),
+            dsize: Some(frame_codec_native_size),
+            dcompact: None,
+            reserved: [std::ptr::null_mut(); 1],
+        },
+        parent: std::ptr::null(),
+        data: std::ptr::null_mut(),
+        flags: rb::frozen_shareable_flags(),
+    }
+}
+
+fn make_block_codec_data_type() -> rb_data_type_t {
+    rb_data_type_t {
+        wrap_struct_name: c"zrip_block_codec".as_ptr(),
+        function: rb_data_type_struct__bindgen_ty_1 {
+            dmark: None,
+            dfree: Some(block_codec_free),
+            dsize: Some(block_codec_native_size),
+            dcompact: None,
+            reserved: [std::ptr::null_mut(); 1],
+        },
+        parent: std::ptr::null(),
+        data: std::ptr::null_mut(),
+        flags: 1,
+    }
+}
+
+fn make_dict_trainer_data_type() -> rb_data_type_t {
+    rb_data_type_t {
+        wrap_struct_name: c"zrip_dict_trainer".as_ptr(),
+        function: rb_data_type_struct__bindgen_ty_1 {
+            dmark: None,
+            dfree: Some(dict_trainer_free),
+            dsize: Some(dict_trainer_native_size),
+            dcompact: None,
+            reserved: [std::ptr::null_mut(); 1],
+        },
+        parent: std::ptr::null(),
+        data: std::ptr::null_mut(),
+        flags: 1,
+    }
+}
+
+unsafe extern "C" fn frame_codec_free(ptr: *mut c_void) {
+    if ptr.is_null() {
+        return;
+    }
+
+    let _ = catch_unwind(AssertUnwindSafe(|| unsafe {
+        drop(Box::from_raw(ptr as *mut FrameCodec));
+    }));
+}
+
+unsafe extern "C" fn block_codec_free(ptr: *mut c_void) {
+    if ptr.is_null() {
+        return;
+    }
+
+    let _ = catch_unwind(AssertUnwindSafe(|| unsafe {
+        drop(Box::from_raw(ptr as *mut BlockCodec));
+    }));
+}
+
+unsafe extern "C" fn dict_trainer_free(ptr: *mut c_void) {
+    if ptr.is_null() {
+        return;
+    }
+
+    let _ = catch_unwind(AssertUnwindSafe(|| unsafe {
+        drop(Box::from_raw(ptr as *mut RbDictTrainer));
+    }));
+}
+
+unsafe extern "C" fn frame_codec_native_size(_ptr: *const c_void) -> size_t {
+    std::mem::size_of::<FrameCodec>() as size_t
+}
+
+unsafe extern "C" fn block_codec_native_size(_ptr: *const c_void) -> size_t {
+    std::mem::size_of::<BlockCodec>() as size_t
+}
+
+unsafe extern "C" fn dict_trainer_native_size(_ptr: *const c_void) -> size_t {
+    std::mem::size_of::<RbDictTrainer>() as size_t
+}
+
+unsafe fn frame_codec_ref(value: VALUE) -> RbResult<&'static FrameCodec> {
+    unsafe { rb::typed_data_ref(value, frame_codec_data_type(), "Zrip::FrameCodec") }
+}
+
+unsafe fn block_codec_ref(value: VALUE) -> RbResult<&'static BlockCodec> {
+    unsafe { rb::typed_data_ref(value, block_codec_data_type(), "Zrip::BlockCodec") }
+}
+
+unsafe fn dict_trainer_ref(value: VALUE) -> RbResult<&'static RbDictTrainer> {
+    unsafe { rb::typed_data_ref(value, dict_trainer_data_type(), "Zrip::DictTrainer") }
+}
+
 // ---------- FrameCodec ----------
 
-#[magnus::wrap(class = "Zrip::FrameCodec", free_immediately, size)]
 struct FrameCodec {
     dict_len: usize,
     dict_id: Option<u32>,
@@ -183,198 +306,311 @@ struct FrameCodec {
 unsafe impl Send for FrameCodec {}
 unsafe impl Sync for FrameCodec {}
 
-fn frame_codec_new(
-    ruby: &Ruby,
-    rb_dict: Option<RString>,
-    id: u32,
-    level: i32,
-) -> Result<FrameCodec, Error> {
+fn frame_codec_new_impl(class: VALUE, rb_dict: VALUE, id: VALUE, level: VALUE) -> RbResult<VALUE> {
+    let rb_dict = rb::value_to_option_bytes(rb_dict)?;
+    let id = rb::value_to_u32(id)?;
+    let level = rb::value_to_i32(level)?;
     let (dict_len, dict_id, cctx, dctx) = match rb_dict {
         None => {
             let cctx = CompressContext::new(level).map_err(|e| {
-                Error::new(
-                    compress_error(ruby),
+                RubyErr::new(
+                    compress_error(),
                     format!("CompressContext::new failed: {e}"),
                 )
             })?;
             (0, None, cctx, DecompressContext::new())
         }
-        Some(s) => {
-            let bytes: Vec<u8> = unsafe { s.as_slice().to_vec() };
-            s.freeze();
-            let dict = load_dict(ruby, &bytes)?;
+        Some(bytes) => {
+            let dict = load_dict(&bytes)?;
             let dl = bytes.len();
             let cctx = CompressContext::with_dict(level, dict.clone()).map_err(|e| {
-                Error::new(
-                    compress_error(ruby),
+                RubyErr::new(
+                    compress_error(),
                     format!("CompressContext::with_dict failed: {e}"),
                 )
             })?;
-            let dctx = DecompressContext::with_dict(dict);
+            let dctx = DecompressContext::with_dict(dict.clone());
             (dl, Some(id), cctx, dctx)
         }
     };
 
-    Ok(FrameCodec {
-        dict_len,
-        dict_id,
-        level,
-        cctx: Mutex::new(cctx),
-        dctx: Mutex::new(dctx),
-    })
+    let codec = unsafe {
+        rb::wrap_typed_data(
+            class,
+            Box::new(FrameCodec {
+                dict_len,
+                dict_id,
+                level,
+                cctx: Mutex::new(cctx),
+                dctx: Mutex::new(dctx),
+            }),
+            frame_codec_data_type(),
+        )?
+    };
+    rb::make_shareable(codec)?;
+    Ok(codec)
+}
+
+fn frame_codec_compress_impl(rb_self: VALUE, rb_input: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    let mut input = rb::input_bytes(rb_input)?;
+    let release_gvl = should_release_compress_gvl(input.len());
+    let out = with_mutex(
+        &rb_self.cctx,
+        release_gvl,
+        "FrameCodec CCtx",
+        &mut input,
+        |cctx, input| {
+            cctx.compress(input)
+                .map(|out| out.into_owned())
+                .map_err(|e| RubyErr::new(compress_error(), format!("zstd compress failed: {e}")))
+        },
+    )?;
+    rb::new_binary_string(&out)
+}
+
+fn frame_codec_decompress_impl(
+    rb_self: VALUE,
+    rb_input: VALUE,
+    max_output: VALUE,
+) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    let max_output = rb::value_to_usize(max_output)?;
+    let mut compressed = rb::input_bytes(rb_input)?;
+    let release_gvl = should_release_frame_decompress_gvl(decompress_work_size(
+        compressed.as_slice(),
+        max_output,
+    ));
+    let out = with_mutex(
+        &rb_self.dctx,
+        release_gvl,
+        "FrameCodec DCtx",
+        &mut compressed,
+        |dctx, compressed| {
+            decompress_bounded(compressed, max_output, dctx)
+                .map_err(|e| bounded_err(e, "zstd frame decode failed"))
+        },
+    )?;
+    rb::new_binary_string(&out)
 }
 
-fn frame_codec_compress(
-    ruby: &Ruby,
-    rb_self: &FrameCodec,
-    rb_input: RString,
-) -> Result<RString, Error> {
-    let input: &[u8] = unsafe { rb_input.as_slice() };
-    let mut cctx = rb_self.cctx.lock().expect("FrameCodec CCtx mutex poisoned");
-    let out = cctx
-        .compress(input)
-        .map_err(|e| Error::new(compress_error(ruby), format!("zstd compress failed: {e}")))?;
-    Ok(ruby.str_from_slice(&out))
+fn frame_codec_size_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    rb::usize_value(rb_self.dict_len)
 }
 
-fn frame_codec_decompress(
-    ruby: &Ruby,
-    rb_self: &FrameCodec,
-    rb_input: RString,
-    max_output: usize,
-) -> Result<RString, Error> {
-    let compressed: &[u8] = unsafe { rb_input.as_slice() };
-    let mut dctx = rb_self.dctx.lock().expect("FrameCodec DCtx mutex poisoned");
-    let out = decompress_bounded(compressed, max_output, &mut dctx)
-        .map_err(|e| raise_bounded(ruby, e, "zstd frame decode failed"))?;
-    Ok(ruby.str_from_slice(&out))
+fn frame_codec_has_dict_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    Ok(rb::bool_value(rb_self.dict_id.is_some()))
 }
 
-fn frame_codec_size(rb_self: &FrameCodec) -> usize {
-    rb_self.dict_len
+fn frame_codec_id_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    rb::u32_option_value(rb_self.dict_id)
 }
 
-fn frame_codec_has_dict(rb_self: &FrameCodec) -> bool {
-    rb_self.dict_id.is_some()
+fn frame_codec_level_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { frame_codec_ref(rb_self)? };
+    Ok(rb::i32_value(rb_self.level))
 }
 
-fn frame_codec_id(rb_self: &FrameCodec) -> Option<u32> {
-    rb_self.dict_id
+fn frame_codec_get_frame_content_size_impl(rb_input: VALUE) -> RbResult<VALUE> {
+    let bytes = rb::input_bytes(rb_input)?;
+    match parse_frame_content_size(bytes.as_slice()) {
+        Ok(v) => rb::u64_option_value(v),
+        Err(BoundedError::BadMagic) => Err(RubyErr::new(
+            decompress_error(),
+            "zstd frame header parse failed: bad magic (input is not a Zstd frame)",
+        )),
+        Err(e) => Err(bounded_err(e, "zstd frame header parse failed")),
+    }
 }
 
-fn frame_codec_level(rb_self: &FrameCodec) -> i32 {
-    rb_self.level
+unsafe extern "C" fn frame_codec_new(
+    class: VALUE,
+    rb_dict: VALUE,
+    id: VALUE,
+    level: VALUE,
+) -> VALUE {
+    rb::wrap(|| frame_codec_new_impl(class, rb_dict, id, level))
 }
 
-fn frame_codec_get_frame_content_size(
-    ruby: &Ruby,
-    rb_input: RString,
-) -> Result<Option<u64>, Error> {
-    let bytes: &[u8] = unsafe { rb_input.as_slice() };
-    match parse_frame_content_size(bytes) {
-        Ok(v) => Ok(v),
-        Err(BoundedError::BadMagic) => Err(Error::new(
-            decompress_error(ruby),
-            "zstd frame header parse failed: bad magic (input is not a Zstd frame)",
-        )),
-        Err(e) => Err(raise_bounded(ruby, e, "zstd frame header parse failed")),
-    }
+unsafe extern "C" fn frame_codec_compress(rb_self: VALUE, rb_input: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_compress_impl(rb_self, rb_input))
+}
+
+unsafe extern "C" fn frame_codec_decompress(
+    rb_self: VALUE,
+    rb_input: VALUE,
+    max_output: VALUE,
+) -> VALUE {
+    rb::wrap(|| frame_codec_decompress_impl(rb_self, rb_input, max_output))
+}
+
+unsafe extern "C" fn frame_codec_size(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_size_impl(rb_self))
+}
+
+unsafe extern "C" fn frame_codec_has_dict(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_has_dict_impl(rb_self))
+}
+
+unsafe extern "C" fn frame_codec_id(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_id_impl(rb_self))
+}
+
+unsafe extern "C" fn frame_codec_level(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_level_impl(rb_self))
+}
+
+unsafe extern "C" fn frame_codec_get_frame_content_size(_class: VALUE, rb_input: VALUE) -> VALUE {
+    rb::wrap(|| frame_codec_get_frame_content_size_impl(rb_input))
 }
 
 // ---------- BlockCodec ----------
 
-#[magnus::wrap(class = "Zrip::BlockCodec", free_immediately, size)]
 struct BlockCodec {
     dict_len: usize,
     dict_id: Option<u32>,
     level: i32,
-    cctx: RefCell<CompressContext>,
-    dctx: RefCell<DecompressContext>,
+    cctx: Mutex<CompressContext>,
+    dctx: Mutex<DecompressContext>,
 }
 
-fn block_codec_new(
-    ruby: &Ruby,
-    rb_dict: Option<RString>,
-    id: u32,
-    level: i32,
-) -> Result<BlockCodec, Error> {
+fn block_codec_new_impl(class: VALUE, rb_dict: VALUE, id: VALUE, level: VALUE) -> RbResult<VALUE> {
+    let rb_dict = rb::value_to_option_bytes(rb_dict)?;
+    let id = rb::value_to_u32(id)?;
+    let level = rb::value_to_i32(level)?;
     let (dict_len, dict_id, cctx, dctx) = match rb_dict {
         None => {
             let cctx = CompressContext::new(level).map_err(|e| {
-                Error::new(
-                    compress_error(ruby),
+                RubyErr::new(
+                    compress_error(),
                     format!("CompressContext::new failed: {e}"),
                 )
             })?;
             (0, None, cctx, DecompressContext::new())
         }
-        Some(s) => {
-            let bytes: Vec<u8> = unsafe { s.as_slice().to_vec() };
-            let dict = load_dict(ruby, &bytes)?;
+        Some(bytes) => {
+            let dict = load_dict(&bytes)?;
             let dl = bytes.len();
             let cctx = CompressContext::with_dict(level, dict.clone()).map_err(|e| {
-                Error::new(
-                    compress_error(ruby),
+                RubyErr::new(
+                    compress_error(),
                     format!("CompressContext::with_dict failed: {e}"),
                 )
             })?;
-            let dctx = DecompressContext::with_dict(dict);
+            let dctx = DecompressContext::with_dict(dict.clone());
             (dl, Some(id), cctx, dctx)
         }
     };
 
-    Ok(BlockCodec {
-        dict_len,
-        dict_id,
-        level,
-        cctx: RefCell::new(cctx),
-        dctx: RefCell::new(dctx),
-    })
+    unsafe {
+        rb::wrap_typed_data(
+            class,
+            Box::new(BlockCodec {
+                dict_len,
+                dict_id,
+                level,
+                cctx: Mutex::new(cctx),
+                dctx: Mutex::new(dctx),
+            }),
+            block_codec_data_type(),
+        )
+    }
 }
 
-fn block_codec_compress(
-    ruby: &Ruby,
-    rb_self: &BlockCodec,
-    rb_input: RString,
-) -> Result<RString, Error> {
-    let input: &[u8] = unsafe { rb_input.as_slice() };
-    let mut cctx = rb_self.cctx.borrow_mut();
-    let out = cctx
-        .compress(input)
-        .map_err(|e| Error::new(compress_error(ruby), format!("zstd compress failed: {e}")))?;
-    Ok(ruby.str_from_slice(&out))
+fn block_codec_compress_impl(rb_self: VALUE, rb_input: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { block_codec_ref(rb_self)? };
+    let mut input = rb::input_bytes(rb_input)?;
+    let release_gvl = should_release_compress_gvl(input.len());
+    let out = with_mutex(
+        &rb_self.cctx,
+        release_gvl,
+        "BlockCodec CCtx",
+        &mut input,
+        |cctx, input| {
+            cctx.compress(input)
+                .map(|out| out.into_owned())
+                .map_err(|e| RubyErr::new(compress_error(), format!("zstd compress failed: {e}")))
+        },
+    )?;
+    rb::new_binary_string(&out)
+}
+
+fn block_codec_decompress_impl(
+    rb_self: VALUE,
+    rb_input: VALUE,
+    max_output: VALUE,
+) -> RbResult<VALUE> {
+    let rb_self = unsafe { block_codec_ref(rb_self)? };
+    let max_output = rb::value_to_usize(max_output)?;
+    let mut compressed = rb::input_bytes(rb_input)?;
+    let out = with_mutex(
+        &rb_self.dctx,
+        false,
+        "BlockCodec DCtx",
+        &mut compressed,
+        |dctx, compressed| {
+            decompress_bounded(compressed, max_output, dctx)
+                .map_err(|e| bounded_err(e, "zstd block decode failed"))
+        },
+    )?;
+    rb::new_binary_string(&out)
 }
 
-fn block_codec_decompress(
-    ruby: &Ruby,
-    rb_self: &BlockCodec,
-    rb_input: RString,
-    max_output: usize,
-) -> Result<RString, Error> {
-    let compressed: &[u8] = unsafe { rb_input.as_slice() };
-    let mut dctx = rb_self.dctx.borrow_mut();
-    let out = decompress_bounded(compressed, max_output, &mut dctx)
-        .map_err(|e| raise_bounded(ruby, e, "zstd block decode failed"))?;
-    Ok(ruby.str_from_slice(&out))
+fn block_codec_size_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { block_codec_ref(rb_self)? };
+    rb::usize_value(rb_self.dict_len)
+}
+
+fn block_codec_has_dict_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { block_codec_ref(rb_self)? };
+    Ok(rb::bool_value(rb_self.dict_id.is_some()))
 }
 
-fn block_codec_size(rb_self: &BlockCodec) -> usize {
-    rb_self.dict_len
+fn block_codec_level_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { block_codec_ref(rb_self)? };
+    Ok(rb::i32_value(rb_self.level))
 }
 
-fn block_codec_has_dict(rb_self: &BlockCodec) -> bool {
-    rb_self.dict_id.is_some()
+unsafe extern "C" fn block_codec_new(
+    class: VALUE,
+    rb_dict: VALUE,
+    id: VALUE,
+    level: VALUE,
+) -> VALUE {
+    rb::wrap(|| block_codec_new_impl(class, rb_dict, id, level))
 }
 
-fn block_codec_level(rb_self: &BlockCodec) -> i32 {
-    rb_self.level
+unsafe extern "C" fn block_codec_compress(rb_self: VALUE, rb_input: VALUE) -> VALUE {
+    rb::wrap(|| block_codec_compress_impl(rb_self, rb_input))
+}
+
+unsafe extern "C" fn block_codec_decompress(
+    rb_self: VALUE,
+    rb_input: VALUE,
+    max_output: VALUE,
+) -> VALUE {
+    rb::wrap(|| block_codec_decompress_impl(rb_self, rb_input, max_output))
+}
+
+unsafe extern "C" fn block_codec_size(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| block_codec_size_impl(rb_self))
+}
+
+unsafe extern "C" fn block_codec_has_dict(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| block_codec_has_dict_impl(rb_self))
+}
+
+unsafe extern "C" fn block_codec_level(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| block_codec_level_impl(rb_self))
 }
 
 // ---------- DictTrainer ----------
 
-#[magnus::wrap(class = "Zrip::DictTrainer", free_immediately, size)]
 struct RbDictTrainer {
-    inner: RefCell<Option<TrainerState>>,
+    inner: Mutex<Option<TrainerState>>,
     max_dict_size: usize,
 }
 
@@ -383,67 +619,78 @@ struct TrainerState {
     total_bytes: usize,
 }
 
-fn dict_trainer_new(_ruby: &Ruby, max_dict_size: usize) -> RbDictTrainer {
-    RbDictTrainer {
-        max_dict_size,
-        inner: RefCell::new(Some(TrainerState {
-            samples: Vec::new(),
-            total_bytes: 0,
-        })),
+fn dict_trainer_new_impl(class: VALUE, max_dict_size: VALUE) -> RbResult<VALUE> {
+    let max_dict_size = rb::value_to_usize(max_dict_size)?;
+    unsafe {
+        rb::wrap_typed_data(
+            class,
+            Box::new(RbDictTrainer {
+                max_dict_size,
+                inner: Mutex::new(Some(TrainerState {
+                    samples: Vec::new(),
+                    total_bytes: 0,
+                })),
+            }),
+            dict_trainer_data_type(),
+        )
     }
 }
 
-fn dict_trainer_add_sample(
-    ruby: &Ruby,
-    rb_self: &RbDictTrainer,
-    rb_data: RString,
-) -> Result<(), Error> {
-    let mut borrow = rb_self.inner.borrow_mut();
-    let state = borrow.as_mut().ok_or_else(|| {
-        Error::new(
-            ruby.exception_runtime_error(),
-            "DictTrainer already consumed by #train",
-        )
-    })?;
-    let data: Vec<u8> = unsafe { rb_data.as_slice().to_vec() };
+fn dict_trainer_add_sample_impl(rb_self: VALUE, rb_data: VALUE) -> RbResult<VALUE> {
+    let data = rb::value_to_bytes(rb_data)?;
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    let mut borrow = rb_self
+        .inner
+        .lock()
+        .map_err(|_| RubyErr::runtime("DictTrainer mutex poisoned"))?;
+    let state = borrow
+        .as_mut()
+        .ok_or_else(|| RubyErr::runtime("DictTrainer already consumed by #train"))?;
     if data.len() < 4 {
-        return Ok(());
+        return Ok(rb::qnil());
     }
     state.total_bytes += data.len();
     state.samples.push(data);
-    Ok(())
-}
-
-fn dict_trainer_sample_count(ruby: &Ruby, rb_self: &RbDictTrainer) -> Result<usize, Error> {
-    let borrow = rb_self.inner.borrow();
-    borrow.as_ref().map(|s| s.samples.len()).ok_or_else(|| {
-        Error::new(
-            ruby.exception_runtime_error(),
-            "DictTrainer already consumed by #train",
-        )
-    })
-}
-
-fn dict_trainer_total_bytes(ruby: &Ruby, rb_self: &RbDictTrainer) -> Result<usize, Error> {
-    let borrow = rb_self.inner.borrow();
-    borrow.as_ref().map(|s| s.total_bytes).ok_or_else(|| {
-        Error::new(
-            ruby.exception_runtime_error(),
-            "DictTrainer already consumed by #train",
-        )
-    })
-}
-
-fn dict_trainer_train(ruby: &Ruby, rb_self: &RbDictTrainer) -> Result<RString, Error> {
-    let state = rb_self.inner.borrow_mut().take().ok_or_else(|| {
-        Error::new(
-            ruby.exception_runtime_error(),
-            "DictTrainer already consumed by #train",
-        )
-    })?;
+    Ok(rb::qnil())
+}
+
+fn dict_trainer_sample_count_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    let borrow = rb_self
+        .inner
+        .lock()
+        .map_err(|_| RubyErr::runtime("DictTrainer mutex poisoned"))?;
+    let value = borrow
+        .as_ref()
+        .map(|s| s.samples.len())
+        .ok_or_else(|| RubyErr::runtime("DictTrainer already consumed by #train"))?;
+    rb::usize_value(value)
+}
+
+fn dict_trainer_total_bytes_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    let borrow = rb_self
+        .inner
+        .lock()
+        .map_err(|_| RubyErr::runtime("DictTrainer mutex poisoned"))?;
+    let value = borrow
+        .as_ref()
+        .map(|s| s.total_bytes)
+        .ok_or_else(|| RubyErr::runtime("DictTrainer already consumed by #train"))?;
+    rb::usize_value(value)
+}
+
+fn dict_trainer_train_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    let state = rb_self
+        .inner
+        .lock()
+        .map_err(|_| RubyErr::runtime("DictTrainer mutex poisoned"))?
+        .take()
+        .ok_or_else(|| RubyErr::runtime("DictTrainer already consumed by #train"))?;
 
     if state.samples.len() < 2 {
-        return Ok(ruby.str_from_slice(b""));
+        return rb::new_binary_string(b"");
     }
 
     let refs: Vec<&[u8]> = state.samples.iter().map(|s| s.as_slice()).collect();
@@ -456,81 +703,146 @@ fn dict_trainer_train(ruby: &Ruby, rb_self: &RbDictTrainer) -> Result<RString, E
     let dict_bytes =
         zstd::dict::finalize::finalize_dictionary(&content, &refs, rb_self.max_dict_size);
 
-    Ok(ruby.str_from_slice(&dict_bytes))
+    rb::new_binary_string(&dict_bytes)
+}
+
+fn dict_trainer_max_dict_size_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    rb::usize_value(rb_self.max_dict_size)
+}
+
+fn dict_trainer_trained_impl(rb_self: VALUE) -> RbResult<VALUE> {
+    let rb_self = unsafe { dict_trainer_ref(rb_self)? };
+    let borrow = rb_self
+        .inner
+        .lock()
+        .map_err(|_| RubyErr::runtime("DictTrainer mutex poisoned"))?;
+    Ok(rb::bool_value(borrow.is_none()))
+}
+
+unsafe extern "C" fn dict_trainer_new(class: VALUE, max_dict_size: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_new_impl(class, max_dict_size))
+}
+
+unsafe extern "C" fn dict_trainer_add_sample(rb_self: VALUE, rb_data: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_add_sample_impl(rb_self, rb_data))
 }
 
-fn dict_trainer_max_dict_size(rb_self: &RbDictTrainer) -> usize {
-    rb_self.max_dict_size
+unsafe extern "C" fn dict_trainer_sample_count(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_sample_count_impl(rb_self))
 }
 
-fn dict_trainer_trained(rb_self: &RbDictTrainer) -> bool {
-    rb_self.inner.borrow().is_none()
+unsafe extern "C" fn dict_trainer_total_bytes(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_total_bytes_impl(rb_self))
+}
+
+unsafe extern "C" fn dict_trainer_train(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_train_impl(rb_self))
+}
+
+unsafe extern "C" fn dict_trainer_max_dict_size(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_max_dict_size_impl(rb_self))
+}
+
+unsafe extern "C" fn dict_trainer_trained(rb_self: VALUE) -> VALUE {
+    rb::wrap(|| dict_trainer_trained_impl(rb_self))
 }
 
 // ---------- module init ----------
 
-#[magnus::init]
-fn init(ruby: &Ruby) -> Result<(), Error> {
-    unsafe { rb_sys::rb_ext_ractor_safe(true) };
+/// # Safety
+///
+/// Ruby calls this function while loading the native extension. The Ruby VM
+/// must be initialized, and the symbol must only be entered by Ruby's extension
+/// loader.
+#[no_mangle]
+pub unsafe extern "C" fn Init_zrip() {
+    rb::wrap_init(init);
+}
 
-    let module = ruby.define_module("Zrip")?;
+fn init() -> RbResult<()> {
+    #[cfg(ruby_engine = "mri")]
+    unsafe {
+        rb_sys::rb_ext_ractor_safe(true);
+    }
+
+    let module = unsafe { rb::define_module(c"Zrip")? };
 
     let decompress_error_class =
-        module.define_error("DecompressError", ruby.exception_standard_error())?;
+        unsafe { rb::define_error_under(module, c"DecompressError", rb_sys::rb_eStandardError)? };
     DECOMPRESS_ERROR
-        .set(Opaque::from(decompress_error_class))
+        .set(GlobalValue(decompress_error_class))
         .unwrap_or_else(|_| panic!("init called more than once"));
 
     let compress_error_class =
-        module.define_error("CompressError", ruby.exception_standard_error())?;
+        unsafe { rb::define_error_under(module, c"CompressError", rb_sys::rb_eStandardError)? };
     COMPRESS_ERROR
-        .set(Opaque::from(compress_error_class))
+        .set(GlobalValue(compress_error_class))
         .unwrap_or_else(|_| panic!("init called more than once"));
 
-    let missing_content_size_error_class =
-        module.define_error("MissingContentSizeError", decompress_error_class)?;
+    let missing_content_size_error_class = unsafe {
+        rb::define_error_under(module, c"MissingContentSizeError", decompress_error_class)?
+    };
     MISSING_CONTENT_SIZE_ERROR
-        .set(Opaque::from(missing_content_size_error_class))
+        .set(GlobalValue(missing_content_size_error_class))
         .unwrap_or_else(|_| panic!("init called more than once"));
 
     let output_size_limit_error_class =
-        module.define_error("OutputSizeLimitError", decompress_error_class)?;
+        unsafe { rb::define_error_under(module, c"OutputSizeLimitError", decompress_error_class)? };
     OUTPUT_SIZE_LIMIT_ERROR
-        .set(Opaque::from(output_size_limit_error_class))
+        .set(GlobalValue(output_size_limit_error_class))
         .unwrap_or_else(|_| panic!("init called more than once"));
 
-    // FrameCodec
-    let frame_codec_class = module.define_class("FrameCodec", ruby.class_object())?;
-    frame_codec_class.define_singleton_method("_native_new", function!(frame_codec_new, 3))?;
-    frame_codec_class.define_singleton_method(
-        "get_frame_content_size",
-        function!(frame_codec_get_frame_content_size, 1),
-    )?;
-    frame_codec_class.define_method("compress", method!(frame_codec_compress, 1))?;
-    frame_codec_class.define_method("_native_decompress", method!(frame_codec_decompress, 2))?;
-    frame_codec_class.define_method("size", method!(frame_codec_size, 0))?;
-    frame_codec_class.define_method("has_dict?", method!(frame_codec_has_dict, 0))?;
-    frame_codec_class.define_method("id", method!(frame_codec_id, 0))?;
-    frame_codec_class.define_method("level", method!(frame_codec_level, 0))?;
-
-    // BlockCodec
-    let block_codec_class = module.define_class("BlockCodec", ruby.class_object())?;
-    block_codec_class.define_singleton_method("_native_new", function!(block_codec_new, 3))?;
-    block_codec_class.define_method("compress", method!(block_codec_compress, 1))?;
-    block_codec_class.define_method("_native_decompress", method!(block_codec_decompress, 2))?;
-    block_codec_class.define_method("size", method!(block_codec_size, 0))?;
-    block_codec_class.define_method("has_dict?", method!(block_codec_has_dict, 0))?;
-    block_codec_class.define_method("level", method!(block_codec_level, 0))?;
-
-    // DictTrainer
-    let trainer_class = module.define_class("DictTrainer", ruby.class_object())?;
-    trainer_class.define_singleton_method("_native_new", function!(dict_trainer_new, 1))?;
-    trainer_class.define_method("add_sample", method!(dict_trainer_add_sample, 1))?;
-    trainer_class.define_method("sample_count", method!(dict_trainer_sample_count, 0))?;
-    trainer_class.define_method("total_bytes", method!(dict_trainer_total_bytes, 0))?;
-    trainer_class.define_method("train", method!(dict_trainer_train, 0))?;
-    trainer_class.define_method("max_dict_size", method!(dict_trainer_max_dict_size, 0))?;
-    trainer_class.define_method("trained?", method!(dict_trainer_trained, 0))?;
+    let frame_codec_class =
+        unsafe { rb::define_class_under(module, c"FrameCodec", rb_sys::rb_cObject)? };
+    unsafe {
+        rb::undef_alloc_func(frame_codec_class)?;
+        rb::define_singleton_method_3(frame_codec_class, c"_native_new", frame_codec_new)?;
+        rb::define_singleton_method_1(
+            frame_codec_class,
+            c"get_frame_content_size",
+            frame_codec_get_frame_content_size,
+        )?;
+        rb::define_method_1(frame_codec_class, c"compress", frame_codec_compress)?;
+        rb::define_method_2(
+            frame_codec_class,
+            c"_native_decompress",
+            frame_codec_decompress,
+        )?;
+        rb::define_method_0(frame_codec_class, c"size", frame_codec_size)?;
+        rb::define_method_0(frame_codec_class, c"has_dict?", frame_codec_has_dict)?;
+        rb::define_method_0(frame_codec_class, c"id", frame_codec_id)?;
+        rb::define_method_0(frame_codec_class, c"level", frame_codec_level)?;
+    }
+
+    let block_codec_class =
+        unsafe { rb::define_class_under(module, c"BlockCodec", rb_sys::rb_cObject)? };
+    unsafe {
+        rb::undef_alloc_func(block_codec_class)?;
+        rb::define_singleton_method_3(block_codec_class, c"_native_new", block_codec_new)?;
+        rb::define_method_1(block_codec_class, c"compress", block_codec_compress)?;
+        rb::define_method_2(
+            block_codec_class,
+            c"_native_decompress",
+            block_codec_decompress,
+        )?;
+        rb::define_method_0(block_codec_class, c"size", block_codec_size)?;
+        rb::define_method_0(block_codec_class, c"has_dict?", block_codec_has_dict)?;
+        rb::define_method_0(block_codec_class, c"level", block_codec_level)?;
+    }
+
+    let trainer_class =
+        unsafe { rb::define_class_under(module, c"DictTrainer", rb_sys::rb_cObject)? };
+    unsafe {
+        rb::undef_alloc_func(trainer_class)?;
+        rb::define_singleton_method_1(trainer_class, c"_native_new", dict_trainer_new)?;
+        rb::define_method_1(trainer_class, c"add_sample", dict_trainer_add_sample)?;
+        rb::define_method_0(trainer_class, c"sample_count", dict_trainer_sample_count)?;
+        rb::define_method_0(trainer_class, c"total_bytes", dict_trainer_total_bytes)?;
+        rb::define_method_0(trainer_class, c"train", dict_trainer_train)?;
+        rb::define_method_0(trainer_class, c"max_dict_size", dict_trainer_max_dict_size)?;
+        rb::define_method_0(trainer_class, c"trained?", dict_trainer_trained)?;
+    }
 
     Ok(())
 }
@@ -573,4 +885,22 @@ mod tests {
         let fcs = parse_frame_content_size(&compressed).unwrap();
         assert_eq!(fcs, Some(data.len() as u64));
     }
+
+    #[test]
+    fn bounded_decompress_uses_total_concatenated_limit() {
+        let mut dctx = DecompressContext::new();
+        let mut compressed = zstd::compress(&[b'a'; 100], 1).unwrap();
+        compressed.extend_from_slice(&zstd::compress(&[b'b'; 100], 1).unwrap());
+
+        assert!(matches!(
+            decompress_bounded(&compressed, 199, &mut dctx),
+            Err(BoundedError::OutputSizeLimit { .. })
+        ));
+        assert_eq!(
+            decompress_bounded(&compressed, 200, &mut dctx)
+                .unwrap()
+                .len(),
+            200
+        );
+    }
 }