diff --git a/Lib/io.py b/Lib/io.py
index f75a1e1..f90a2f8 100644
--- a/Lib/io.py
+++ b/Lib/io.py
@@ -63,6 +63,7 @@
 import codecs
 import _fileio
 import warnings
+import threading
 
 # open() uses st_blksize whenever we can
 DEFAULT_BUFFER_SIZE = 8 * 1024  # bytes
@@ -908,6 +909,7 @@
         _BufferedIOMixin.__init__(self, raw)
         self.buffer_size = buffer_size
         self._reset_read_buf()
+        self._read_lock = threading.Lock()
 
     def _reset_read_buf(self):
         self._read_buf = b""
@@ -921,6 +923,10 @@
         mode. If n is negative, read until EOF or until read() would
         block.
         """
+        with self._read_lock:
+            return self._read_unlocked(n)
+
+    def _read_unlocked(self, n=None):
         nodata_val = b""
         empty_values = (b"", None)
         buf = self._read_buf
@@ -973,6 +979,10 @@
         do at most one raw read to satisfy it.  We never return more
         than self.buffer_size.
         """
+        with self._read_lock:
+            return self._peek_unlocked(n)
+
+    def _peek_unlocked(self, n=0):
         want = min(n, self.buffer_size)
         have = len(self._read_buf) - self._read_pos
         if have < want:
@@ -989,18 +999,21 @@
         # only return buffered bytes.  Otherwise, we do one raw read.
         if n <= 0:
             return b""
-        self.peek(1)
-        return self.read(min(n, len(self._read_buf) - self._read_pos))
+        with self._read_lock:
+            self._peek_unlocked(1)
+            return self._read_unlocked(
+                min(n, len(self._read_buf) - self._read_pos))
 
     def tell(self):
         return self.raw.tell() - len(self._read_buf) + self._read_pos
 
     def seek(self, pos, whence=0):
-        if whence == 1:
-            pos -= len(self._read_buf) - self._read_pos
-        pos = self.raw.seek(pos, whence)
-        self._reset_read_buf()
-        return pos
+        with self._read_lock:
+            if whence == 1:
+                pos -= len(self._read_buf) - self._read_pos
+            pos = self.raw.seek(pos, whence)
+            self._reset_read_buf()
+            return pos
 
 
 class BufferedWriter(_BufferedIOMixin):
@@ -1022,43 +1035,51 @@
                                 if max_buffer_size is None
                                 else max_buffer_size)
         self._write_buf = bytearray()
+        self._write_lock = threading.Lock()
 
     def write(self, b):
         if self.closed:
             raise ValueError("write to closed file")
         if isinstance(b, unicode):
             raise TypeError("can't write unicode to binary stream")
-        # XXX we can implement some more tricks to try and avoid partial writes
-        if len(self._write_buf) > self.buffer_size:
-            # We're full, so let's pre-flush the buffer
-            try:
-                self.flush()
-            except BlockingIOError as e:
-                # We can't accept anything else.
-                # XXX Why not just let the exception pass through?
-                raise BlockingIOError(e.errno, e.strerror, 0)
-        before = len(self._write_buf)
-        self._write_buf.extend(b)
-        written = len(self._write_buf) - before
-        if len(self._write_buf) > self.buffer_size:
-            try:
-                self.flush()
-            except BlockingIOError as e:
-                if (len(self._write_buf) > self.max_buffer_size):
-                    # We've hit max_buffer_size. We have to accept a partial
-                    # write and cut back our buffer.
-                    overage = len(self._write_buf) - self.max_buffer_size
-                    self._write_buf = self._write_buf[:self.max_buffer_size]
-                    raise BlockingIOError(e.errno, e.strerror, overage)
-        return written
+        with self._write_lock:
+            # XXX we can implement some more tricks to try and avoid
+            # partial writes
+            if len(self._write_buf) > self.buffer_size:
+                # We're full, so let's pre-flush the buffer
+                try:
+                    self._flush_unlocked()
+                except BlockingIOError as e:
+                    # We can't accept anything else.
+                    # XXX Why not just let the exception pass through?
+                    raise BlockingIOError(e.errno, e.strerror, 0)
+            before = len(self._write_buf)
+            self._write_buf.extend(b)
+            written = len(self._write_buf) - before
+            if len(self._write_buf) > self.buffer_size:
+                try:
+                    self._flush_unlocked()
+                except BlockingIOError as e:
+                    if len(self._write_buf) > self.max_buffer_size:
+                        # We've hit max_buffer_size. We have to accept a
+                        # partial write and cut back our buffer.
+                        overage = len(self._write_buf) - self.max_buffer_size
+                        self._write_buf = self._write_buf[:self.max_buffer_size]
+                        raise BlockingIOError(e.errno, e.strerror, overage)
+            return written
 
     def truncate(self, pos=None):
-        self.flush()
-        if pos is None:
-            pos = self.raw.tell()
-        return self.raw.truncate(pos)
+        with self._write_lock:
+            self._flush_unlocked()
+            if pos is None:
+                pos = self.raw.tell()
+            return self.raw.truncate(pos)
 
     def flush(self):
+        with self._write_lock:
+            self._flush_unlocked()
+
+    def _flush_unlocked(self):
         if self.closed:
             raise ValueError("flush of closed file")
         written = 0
@@ -1077,8 +1098,9 @@
         return self.raw.tell() + len(self._write_buf)
 
     def seek(self, pos, whence=0):
-        self.flush()
-        return self.raw.seek(pos, whence)
+        with self._write_lock:
+            self._flush_unlocked()
+            return self.raw.seek(pos, whence)
 
 
 class BufferedRWPair(BufferedIOBase):
@@ -1168,7 +1190,8 @@
         # First do the raw seek, then empty the read buffer, so that
         # if the raw seek fails, we don't lose buffered data forever.
         pos = self.raw.seek(pos, whence)
-        self._reset_read_buf()
+        with self._read_lock:
+            self._reset_read_buf()
         return pos
 
     def tell(self):
@@ -1205,8 +1228,9 @@
     def write(self, b):
         if self._read_buf:
             # Undo readahead
-            self.raw.seek(self._read_pos - len(self._read_buf), 1)
-            self._reset_read_buf()
+            with self._read_lock:
+                self.raw.seek(self._read_pos - len(self._read_buf), 1)
+                self._reset_read_buf()
         return BufferedWriter.write(self, b)
 
 
diff --git a/Lib/test/test_io.py b/Lib/test/test_io.py
index b93ce02..7fad7c7 100644
--- a/Lib/test/test_io.py
+++ b/Lib/test/test_io.py
@@ -6,8 +6,10 @@
 import sys
 import time
 import array
+import threading
+import random
 import unittest
-from itertools import chain
+from itertools import chain, cycle
 from test import test_support
 
 import codecs
@@ -390,6 +392,49 @@
         # this test. Else, write it.
         pass
 
+    def testThreads(self):
+        try:
+            # Write out many bytes with exactly the same number of 0's,
+            # 1's... 255's. This will help us check that concurrent reading
+            # doesn't duplicate or forget contents.
+            N = 1000
+            l = range(256) * N
+            random.shuffle(l)
+            s = bytes(bytearray(l))
+            with io.open(test_support.TESTFN, "wb") as f:
+                f.write(s)
+            with io.open(test_support.TESTFN, "rb", buffering=0) as raw:
+                bufio = io.BufferedReader(raw, 8)
+                errors = []
+                results = []
+                def f():
+                    try:
+                        # Intra-buffer read then buffer-flushing read
+                        for n in cycle([1, 19]):
+                            s = bufio.read(n)
+                            if not s:
+                                break
+                            # list.append() is atomic
+                            results.append(s)
+                    except Exception as e:
+                        errors.append(e)
+                        raise
+                threads = [threading.Thread(target=f) for x in range(20)]
+                for t in threads:
+                    t.start()
+                time.sleep(0.02) # yield
+                for t in threads:
+                    t.join()
+                self.assertFalse(errors,
+                    "the following exceptions were caught: %r" % errors)
+                s = b''.join(results)
+                for i in range(256):
+                    c = bytes(bytearray([i]))
+                    self.assertEqual(s.count(c), N)
+        finally:
+            test_support.unlink(test_support.TESTFN)
+
+
 
 class BufferedWriterTest(unittest.TestCase):
 
@@ -446,6 +491,38 @@
 
         self.assertEquals(b"abc", writer._write_stack[0])
 
+    def testThreads(self):
+        # BufferedWriter should not raise exceptions or crash
+        # when called from multiple threads.
+        try:
+            # We use a real file object because it allows us to
+            # exercise situations where the GIL is released before
+            # writing the buffer to the raw streams. This is in addition
+            # to concurrency issues due to switching threads in the middle
+            # of Python code.
+            with io.open(test_support.TESTFN, "wb", buffering=0) as raw:
+                bufio = io.BufferedWriter(raw, 8)
+                errors = []
+                def f():
+                    try:
+                        # Write enough bytes to flush the buffer
+                        s = b"a" * 19
+                        for i in range(50):
+                            bufio.write(s)
+                    except Exception as e:
+                        errors.append(e)
+                        raise
+                threads = [threading.Thread(target=f) for x in range(20)]
+                for t in threads:
+                    t.start()
+                time.sleep(0.02) # yield
+                for t in threads:
+                    t.join()
+                self.assertFalse(errors,
+                    "the following exceptions were caught: %r" % errors)
+        finally:
+            test_support.unlink(test_support.TESTFN)
+
 
 class BufferedRWPairTest(unittest.TestCase):
 
diff --git a/Misc/NEWS b/Misc/NEWS
index 5349c55..70a06ff 100644
--- a/Misc/NEWS
+++ b/Misc/NEWS
@@ -48,6 +48,9 @@
 Library
 -------
 
+- Issue #3476: binary buffered reading through the new "io" library is now
+  thread-safe.
+
 - Silence the DeprecationWarning of rfc822 when it is imported by mimetools
   since mimetools itself is deprecated. Because modules are cached, all
   subsequent imports of rfc822 will not raise a visible DeprecationWarning.