[3.12] gh-99108: Release the GIL around hashlib built-in computation (GH-104675) (#104776)

gh-99108: Release the GIL around hashlib built-in computation (GH-104675)

This matches the GIL releasing behavior of our existing `_hashopenssl`
module, extending it to the HACL* built-ins.

Includes adding comments to better describe the ENTER/LEAVE macros
purpose and explain the lock strategy in both existing and new code.
(cherry picked from commit 2e5d8a90aa)

Co-authored-by: Gregory P. Smith [Google] <greg@krypto.org>

Misc/NEWS.d/next/Library/2023-05-19-19-46-22.gh-issue-99108.wqCg0t.rst

@@ -0,0 +1,3 @@
+We now release the GIL around built-in :mod:`hashlib` computations of
+reasonable size for the SHA families and MD5 hash functions, matching
+what our OpenSSL backed hash computations already does.

Modules/_hashopenssl.c

@@ -227,12 +227,16 @@ get_hashlib_state(PyObject *module)
 typedef struct {
     PyObject_HEAD
     EVP_MD_CTX          *ctx;   /* OpenSSL message digest context */
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
     PyThread_type_lock   lock;  /* OpenSSL context lock */
 } EVPobject;
 
 typedef struct {
     PyObject_HEAD
     HMAC_CTX *ctx;            /* OpenSSL hmac context */
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
     PyThread_type_lock lock;  /* HMAC context lock */
 } HMACobject;
 
@@ -896,6 +900,8 @@ py_evp_fromname(PyObject *module, const char *digestname, PyObject *data_obj,
 
     if (view.buf && view.len) {
         if (view.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
             Py_BEGIN_ALLOW_THREADS
             result = EVP_hash(self, view.buf, view.len);
             Py_END_ALLOW_THREADS

Modules/hashlib.h

@@ -37,6 +37,13 @@
  * LEAVE_HASHLIB block or explicitly acquire and release the lock inside
  * a PY_BEGIN / END_ALLOW_THREADS block if they wish to release the GIL for
  * an operation.
+ *
+ * These only drop the GIL if the lock acquisition itself is likely to
+ * block. Thus the non-blocking acquire gating the GIL release for a
+ * blocking lock acquisition. The intent of these macros is to surround
+ * the assumed always "fast" operations that you aren't releasing the
+ * GIL around.  Otherwise use code similar to what you see in hash
+ * function update() methods.
  */
 
 #include "pythread.h"
@@ -53,7 +60,7 @@
         PyThread_release_lock((obj)->lock); \
     }
 
-/* TODO(gps): We should probably make this a module or EVPobject attribute
+/* TODO(gpshead): We should make this a module or class attribute
  * to allow the user to optimize based on the platform they're using. */
 #define HASHLIB_GIL_MINSIZE 2048
 

Modules/md5module.c

@@ -49,7 +49,9 @@ typedef long long MD5_INT64; /* 64-bit integer */
 
 typedef struct {
     PyObject_HEAD
-
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
+    PyThread_type_lock lock;
     Hacl_Streaming_MD5_state *hash_state;
 } MD5object;
 
@@ -72,6 +74,7 @@ static MD5object *
 newMD5object(MD5State * st)
 {
     MD5object *md5 = (MD5object *)PyObject_GC_New(MD5object, st->md5_type);
+    md5->lock = NULL;
     PyObject_GC_Track(md5);
     return md5;
 }
@@ -88,6 +91,9 @@ static void
 MD5_dealloc(MD5object *ptr)
 {
     Hacl_Streaming_MD5_legacy_free(ptr->hash_state);
+    if (ptr->lock != NULL) {
+        PyThread_free_lock(ptr->lock);
+    }
     PyTypeObject *tp = Py_TYPE(ptr);
     PyObject_GC_UnTrack(ptr);
     PyObject_GC_Del(ptr);
@@ -115,7 +121,9 @@ MD5Type_copy_impl(MD5object *self, PyTypeObject *cls)
     if ((newobj = newMD5object(st))==NULL)
         return NULL;
 
+    ENTER_HASHLIB(self);
     newobj->hash_state = Hacl_Streaming_MD5_legacy_copy(self->hash_state);
+    LEAVE_HASHLIB(self);
     return (PyObject *)newobj;
 }
 
@@ -130,7 +138,9 @@ MD5Type_digest_impl(MD5object *self)
 /*[clinic end generated code: output=eb691dc4190a07ec input=bc0c4397c2994be6]*/
 {
     unsigned char digest[MD5_DIGESTSIZE];
+    ENTER_HASHLIB(self);
     Hacl_Streaming_MD5_legacy_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return PyBytes_FromStringAndSize((const char *)digest, MD5_DIGESTSIZE);
 }
 
@@ -145,7 +155,9 @@ MD5Type_hexdigest_impl(MD5object *self)
 /*[clinic end generated code: output=17badced1f3ac932 input=b60b19de644798dd]*/
 {
     unsigned char digest[MD5_DIGESTSIZE];
+    ENTER_HASHLIB(self);
     Hacl_Streaming_MD5_legacy_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return _Py_strhex((const char*)digest, MD5_DIGESTSIZE);
 }
 
@@ -177,7 +189,18 @@ MD5Type_update(MD5object *self, PyObject *obj)
 
     GET_BUFFER_VIEW_OR_ERROUT(obj, &buf);
 
+    if (self->lock == NULL && buf.len >= HASHLIB_GIL_MINSIZE) {
+        self->lock = PyThread_allocate_lock();
+    }
+    if (self->lock != NULL) {
+        Py_BEGIN_ALLOW_THREADS
+        PyThread_acquire_lock(self->lock, 1);
         update(self->hash_state, buf.buf, buf.len);
+        PyThread_release_lock(self->lock);
+        Py_END_ALLOW_THREADS
+    } else {
+        update(self->hash_state, buf.buf, buf.len);
+    }
 
     PyBuffer_Release(&buf);
     Py_RETURN_NONE;
@@ -279,7 +302,15 @@ _md5_md5_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update(new->hash_state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update(new->hash_state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 

Modules/sha1module.c

@@ -48,7 +48,9 @@ typedef long long SHA1_INT64;        /* 64-bit integer */
 
 typedef struct {
     PyObject_HEAD
-
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
+    PyThread_type_lock lock;
     Hacl_Streaming_SHA1_state *hash_state;
 } SHA1object;
 
@@ -71,6 +73,7 @@ static SHA1object *
 newSHA1object(SHA1State *st)
 {
     SHA1object *sha = (SHA1object *)PyObject_GC_New(SHA1object, st->sha1_type);
+    sha->lock = NULL;
     PyObject_GC_Track(sha);
     return sha;
 }
@@ -88,6 +91,9 @@ static void
 SHA1_dealloc(SHA1object *ptr)
 {
     Hacl_Streaming_SHA1_legacy_free(ptr->hash_state);
+    if (ptr->lock != NULL) {
+        PyThread_free_lock(ptr->lock);
+    }
     PyTypeObject *tp = Py_TYPE(ptr);
     PyObject_GC_UnTrack(ptr);
     PyObject_GC_Del(ptr);
@@ -115,7 +121,9 @@ SHA1Type_copy_impl(SHA1object *self, PyTypeObject *cls)
     if ((newobj = newSHA1object(st)) == NULL)
         return NULL;
 
+    ENTER_HASHLIB(self);
     newobj->hash_state = Hacl_Streaming_SHA1_legacy_copy(self->hash_state);
+    LEAVE_HASHLIB(self);
     return (PyObject *)newobj;
 }
 
@@ -130,7 +138,9 @@ SHA1Type_digest_impl(SHA1object *self)
 /*[clinic end generated code: output=2f05302a7aa2b5cb input=13824b35407444bd]*/
 {
     unsigned char digest[SHA1_DIGESTSIZE];
+    ENTER_HASHLIB(self);
     Hacl_Streaming_SHA1_legacy_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return PyBytes_FromStringAndSize((const char *)digest, SHA1_DIGESTSIZE);
 }
 
@@ -145,7 +155,9 @@ SHA1Type_hexdigest_impl(SHA1object *self)
 /*[clinic end generated code: output=4161fd71e68c6659 input=97691055c0c74ab0]*/
 {
     unsigned char digest[SHA1_DIGESTSIZE];
+    ENTER_HASHLIB(self);
     Hacl_Streaming_SHA1_legacy_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return _Py_strhex((const char *)digest, SHA1_DIGESTSIZE);
 }
 
@@ -177,7 +189,18 @@ SHA1Type_update(SHA1object *self, PyObject *obj)
 
     GET_BUFFER_VIEW_OR_ERROUT(obj, &buf);
 
+    if (self->lock == NULL && buf.len >= HASHLIB_GIL_MINSIZE) {
+        self->lock = PyThread_allocate_lock();
+    }
+    if (self->lock != NULL) {
+        Py_BEGIN_ALLOW_THREADS
+        PyThread_acquire_lock(self->lock, 1);
         update(self->hash_state, buf.buf, buf.len);
+        PyThread_release_lock(self->lock);
+        Py_END_ALLOW_THREADS
+    } else {
+        update(self->hash_state, buf.buf, buf.len);
+    }
 
     PyBuffer_Release(&buf);
     Py_RETURN_NONE;
@@ -279,7 +302,15 @@ _sha1_sha1_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update(new->hash_state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update(new->hash_state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 

Modules/sha2module.c

@@ -52,12 +52,18 @@ class SHA512Type "SHA512object *" "&PyType_Type"
 typedef struct {
     PyObject_HEAD
     int digestsize;
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
+    PyThread_type_lock lock;
     Hacl_Streaming_SHA2_state_sha2_256 *state;
 } SHA256object;
 
 typedef struct {
     PyObject_HEAD
     int digestsize;
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
+    PyThread_type_lock lock;
     Hacl_Streaming_SHA2_state_sha2_512 *state;
 } SHA512object;
 
@@ -100,6 +106,7 @@ newSHA224object(sha2_state *state)
     if (!sha) {
         return NULL;
     }
+    sha->lock = NULL;
     PyObject_GC_Track(sha);
     return sha;
 }
@@ -112,6 +119,7 @@ newSHA256object(sha2_state *state)
     if (!sha) {
         return NULL;
     }
+    sha->lock = NULL;
     PyObject_GC_Track(sha);
     return sha;
 }
@@ -124,6 +132,7 @@ newSHA384object(sha2_state *state)
     if (!sha) {
         return NULL;
     }
+    sha->lock = NULL;
     PyObject_GC_Track(sha);
     return sha;
 }
@@ -136,6 +145,7 @@ newSHA512object(sha2_state *state)
     if (!sha) {
         return NULL;
     }
+    sha->lock = NULL;
     PyObject_GC_Track(sha);
     return sha;
 }
@@ -153,6 +163,9 @@ static void
 SHA256_dealloc(SHA256object *ptr)
 {
     Hacl_Streaming_SHA2_free_256(ptr->state);
+    if (ptr->lock != NULL) {
+        PyThread_free_lock(ptr->lock);
+    }
     PyTypeObject *tp = Py_TYPE(ptr);
     PyObject_GC_UnTrack(ptr);
     PyObject_GC_Del(ptr);
@@ -163,6 +176,9 @@ static void
 SHA512_dealloc(SHA512object *ptr)
 {
     Hacl_Streaming_SHA2_free_512(ptr->state);
+    if (ptr->lock != NULL) {
+        PyThread_free_lock(ptr->lock);
+    }
     PyTypeObject *tp = Py_TYPE(ptr);
     PyObject_GC_UnTrack(ptr);
     PyObject_GC_Del(ptr);
@@ -229,7 +245,9 @@ SHA256Type_copy_impl(SHA256object *self, PyTypeObject *cls)
         }
     }
 
+    ENTER_HASHLIB(self);
     SHA256copy(self, newobj);
+    LEAVE_HASHLIB(self);
     return (PyObject *)newobj;
 }
 
@@ -259,7 +277,9 @@ SHA512Type_copy_impl(SHA512object *self, PyTypeObject *cls)
         }
     }
 
+    ENTER_HASHLIB(self);
     SHA512copy(self, newobj);
+    LEAVE_HASHLIB(self);
     return (PyObject *)newobj;
 }
 
@@ -275,9 +295,11 @@ SHA256Type_digest_impl(SHA256object *self)
 {
     uint8_t digest[SHA256_DIGESTSIZE];
     assert(self->digestsize <= SHA256_DIGESTSIZE);
+    ENTER_HASHLIB(self);
     // HACL* performs copies under the hood so that self->state remains valid
     // after this call.
     Hacl_Streaming_SHA2_finish_256(self->state, digest);
+    LEAVE_HASHLIB(self);
     return PyBytes_FromStringAndSize((const char *)digest, self->digestsize);
 }
 
@@ -293,9 +315,11 @@ SHA512Type_digest_impl(SHA512object *self)
 {
     uint8_t digest[SHA512_DIGESTSIZE];
     assert(self->digestsize <= SHA512_DIGESTSIZE);
+    ENTER_HASHLIB(self);
     // HACL* performs copies under the hood so that self->state remains valid
     // after this call.
     Hacl_Streaming_SHA2_finish_512(self->state, digest);
+    LEAVE_HASHLIB(self);
     return PyBytes_FromStringAndSize((const char *)digest, self->digestsize);
 }
 
@@ -311,7 +335,9 @@ SHA256Type_hexdigest_impl(SHA256object *self)
 {
     uint8_t digest[SHA256_DIGESTSIZE];
     assert(self->digestsize <= SHA256_DIGESTSIZE);
+    ENTER_HASHLIB(self);
     Hacl_Streaming_SHA2_finish_256(self->state, digest);
+    LEAVE_HASHLIB(self);
     return _Py_strhex((const char *)digest, self->digestsize);
 }
 
@@ -327,7 +353,9 @@ SHA512Type_hexdigest_impl(SHA512object *self)
 {
     uint8_t digest[SHA512_DIGESTSIZE];
     assert(self->digestsize <= SHA512_DIGESTSIZE);
+    ENTER_HASHLIB(self);
     Hacl_Streaming_SHA2_finish_512(self->state, digest);
+    LEAVE_HASHLIB(self);
     return _Py_strhex((const char *)digest, self->digestsize);
 }
 
@@ -348,7 +376,18 @@ SHA256Type_update(SHA256object *self, PyObject *obj)
 
     GET_BUFFER_VIEW_OR_ERROUT(obj, &buf);
 
+    if (self->lock == NULL && buf.len >= HASHLIB_GIL_MINSIZE) {
+        self->lock = PyThread_allocate_lock();
+    }
+    if (self->lock != NULL) {
+        Py_BEGIN_ALLOW_THREADS
+        PyThread_acquire_lock(self->lock, 1);
         update_256(self->state, buf.buf, buf.len);
+        PyThread_release_lock(self->lock);
+        Py_END_ALLOW_THREADS
+    } else {
+        update_256(self->state, buf.buf, buf.len);
+    }
 
     PyBuffer_Release(&buf);
     Py_RETURN_NONE;
@@ -371,7 +410,18 @@ SHA512Type_update(SHA512object *self, PyObject *obj)
 
     GET_BUFFER_VIEW_OR_ERROUT(obj, &buf);
 
+    if (self->lock == NULL && buf.len >= HASHLIB_GIL_MINSIZE) {
+        self->lock = PyThread_allocate_lock();
+    }
+    if (self->lock != NULL) {
+        Py_BEGIN_ALLOW_THREADS
+        PyThread_acquire_lock(self->lock, 1);
         update_512(self->state, buf.buf, buf.len);
+        PyThread_release_lock(self->lock);
+        Py_END_ALLOW_THREADS
+    } else {
+        update_512(self->state, buf.buf, buf.len);
+    }
 
     PyBuffer_Release(&buf);
     Py_RETURN_NONE;
@@ -560,7 +610,15 @@ _sha2_sha256_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update_256(new->state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update_256(new->state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 
@@ -606,7 +664,15 @@ _sha2_sha224_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update_256(new->state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update_256(new->state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 
@@ -651,7 +717,15 @@ _sha2_sha512_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update_512(new->state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update_512(new->state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 
@@ -696,7 +770,15 @@ _sha2_sha384_impl(PyObject *module, PyObject *string, int usedforsecurity)
         return NULL;
     }
     if (string) {
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             update_512(new->state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            update_512(new->state, buf.buf, buf.len);
+        }
         PyBuffer_Release(&buf);
     }
 

Modules/sha3module.c

@@ -60,6 +60,9 @@ class _sha3.shake_256 "SHA3object *" "&SHAKE256type"
 
 typedef struct {
     PyObject_HEAD
+    // Prevents undefined behavior via multiple threads entering the C API.
+    // The lock will be NULL before threaded access has been enabled.
+    PyThread_type_lock lock;
     Hacl_Streaming_Keccak_state *hash_state;
 } SHA3object;
 
@@ -73,6 +76,7 @@ newSHA3object(PyTypeObject *type)
     if (newobj == NULL) {
         return NULL;
     }
+    newobj->lock = NULL;
     return newobj;
 }
 
@@ -133,7 +137,15 @@ py_sha3_new_impl(PyTypeObject *type, PyObject *data, int usedforsecurity)
 
     if (data) {
         GET_BUFFER_VIEW_OR_ERROR(data, &buf, goto error);
+        if (buf.len >= HASHLIB_GIL_MINSIZE) {
+            /* We do not initialize self->lock here as this is the constructor
+             * where it is not yet possible to have concurrent access. */
+            Py_BEGIN_ALLOW_THREADS
             sha3_update(self->hash_state, buf.buf, buf.len);
+            Py_END_ALLOW_THREADS
+        } else {
+            sha3_update(self->hash_state, buf.buf, buf.len);
+        }
     }
 
     PyBuffer_Release(&buf);
@@ -157,6 +169,9 @@ static void
 SHA3_dealloc(SHA3object *self)
 {
     Hacl_Streaming_Keccak_free(self->hash_state);
+    if (self->lock != NULL) {
+        PyThread_free_lock(self->lock);
+    }
     PyTypeObject *tp = Py_TYPE(self);
     PyObject_Free(self);
     Py_DECREF(tp);
@@ -181,7 +196,9 @@ _sha3_sha3_224_copy_impl(SHA3object *self)
     if ((newobj = newSHA3object(Py_TYPE(self))) == NULL) {
         return NULL;
     }
+    ENTER_HASHLIB(self);
     newobj->hash_state = Hacl_Streaming_Keccak_copy(self->hash_state);
+    LEAVE_HASHLIB(self);
     return (PyObject *)newobj;
 }
 
@@ -199,7 +216,9 @@ _sha3_sha3_224_digest_impl(SHA3object *self)
     unsigned char digest[SHA3_MAX_DIGESTSIZE];
     // This function errors out if the algorithm is Shake. Here, we know this
     // not to be the case, and therefore do not perform error checking.
+    ENTER_HASHLIB(self);
     Hacl_Streaming_Keccak_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return PyBytes_FromStringAndSize((const char *)digest,
         Hacl_Streaming_Keccak_hash_len(self->hash_state));
 }
@@ -216,7 +235,9 @@ _sha3_sha3_224_hexdigest_impl(SHA3object *self)
 /*[clinic end generated code: output=75ad03257906918d input=2d91bb6e0d114ee3]*/
 {
     unsigned char digest[SHA3_MAX_DIGESTSIZE];
+    ENTER_HASHLIB(self);
     Hacl_Streaming_Keccak_finish(self->hash_state, digest);
+    LEAVE_HASHLIB(self);
     return _Py_strhex((const char *)digest,
         Hacl_Streaming_Keccak_hash_len(self->hash_state));
 }
@@ -237,7 +258,18 @@ _sha3_sha3_224_update(SHA3object *self, PyObject *data)
 {
     Py_buffer buf;
     GET_BUFFER_VIEW_OR_ERROUT(data, &buf);
+    if (self->lock == NULL && buf.len >= HASHLIB_GIL_MINSIZE) {
+        self->lock = PyThread_allocate_lock();
+    }
+    if (self->lock != NULL) {
+        Py_BEGIN_ALLOW_THREADS
+        PyThread_acquire_lock(self->lock, 1);
         sha3_update(self->hash_state, buf.buf, buf.len);
+        PyThread_release_lock(self->lock);
+        Py_END_ALLOW_THREADS
+    } else {
+        sha3_update(self->hash_state, buf.buf, buf.len);
+    }
     PyBuffer_Release(&buf);
     Py_RETURN_NONE;
 }