diff --git a/examples/client_test.cpp b/examples/client_test.cpp
index 3c51429d3..c7d512f13 100644
--- a/examples/client_test.cpp
+++ b/examples/client_test.cpp
@@ -389,7 +389,7 @@ int print_peer_info(std::string& out
 	if (print_ip) out += "IP                             ";
 	out += "progress        down     (total | peak   )  up      (total | peak   ) sent-req tmo bsy rcv flags         dn  up  source  ";
 	if (print_fails) out += "fail hshf ";
-	if (print_send_bufs) out += "rq sndb rcvb   q-bytes ";
+	if (print_send_bufs) out += "rq sndb (recvb |alloc | wmrk ) q-bytes ";
 	if (print_timers) out += "inactive wait timeout q-time ";
 	out += "  v disk ^    rtt  ";
 	if (print_block) out += "block-progress ";
@@ -471,9 +471,11 @@ int print_peer_info(std::string& out
 		}
 		if (print_send_bufs)
 		{
-			std::snprintf(str, sizeof(str), "%2d %6d %6d%5dkB "
+			std::snprintf(str, sizeof(str), "%2d %6d %6d|%6d|%6d%5dkB "
 				, i->requests_in_buffer, i->used_send_buffer
 				, i->used_receive_buffer
+				, i->receive_buffer_size
+				, i->receive_buffer_watermark
 				, i->queue_bytes / 1000);
 			out += str;
 		}
diff --git a/include/libtorrent/aux_/array_view.hpp b/include/libtorrent/aux_/array_view.hpp
index 2c9b438dc..f833860d7 100644
--- a/include/libtorrent/aux_/array_view.hpp
+++ b/include/libtorrent/aux_/array_view.hpp
@@ -34,6 +34,7 @@ POSSIBILITY OF SUCH DAMAGE.
 #define TORRENT_ARRAY_VIEW_HPP_INCLUDED
 
 #include <vector>
+#include <array>
 #include <type_traits> // for std::is_convertible
 #include "libtorrent/assert.hpp"
 
diff --git a/include/libtorrent/aux_/suggest_piece.hpp b/include/libtorrent/aux_/suggest_piece.hpp
index bfe4f4d63..6ff363cdf 100644
--- a/include/libtorrent/aux_/suggest_piece.hpp
+++ b/include/libtorrent/aux_/suggest_piece.hpp
@@ -107,10 +107,6 @@ struct suggest_piece
 		}
 
 		m_priority_pieces.push_back(index);
-
-		std::printf("SUGGEST: ");
-		for (int p : m_priority_pieces) std::printf(" %d", p);
-		std::printf("\n");
 	}
 
 private:
diff --git a/include/libtorrent/buffer.hpp b/include/libtorrent/buffer.hpp
index f1f26b727..192560849 100644
--- a/include/libtorrent/buffer.hpp
+++ b/include/libtorrent/buffer.hpp
@@ -40,12 +40,34 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include "libtorrent/invariant_check.hpp"
 #include "libtorrent/assert.hpp"
+#include "libtorrent/aux_/array_view.hpp"
+
+#if defined __GLIBC__
+#include <malloc.h>
+#elif defined _MSC_VER
+#include <malloc.h>
+#elif defined TORRENT_BSD
+#include <malloc/malloc.h>
+#endif
 
 namespace libtorrent {
 
+	namespace {
+		inline std::size_t round_up8(std::size_t const v)
+		{
+			return (v + 7) & (~std::size_t(0x7));
+		}
+	}
+
+// the buffer is allocated once and cannot be resized. The size() may be
+// larger than requested, in case the underlying allocator over allocated. In
+// order to "grow" an allocation, create a new buffer and initialize it by
+// the range of bytes from the existing, and move-assign the new over the
+// old.
 class buffer
 {
 public:
+
 	struct interval
 	{
 	interval()
@@ -110,31 +132,58 @@ public:
 		char const* end;
 	};
 
-	buffer(std::size_t n = 0)
-		: m_begin(0)
-		, m_size(0)
-		, m_capacity(0)
+	// allocate an uninitialized buffer of the specified size
+	buffer(std::size_t size = 0)
 	{
-		if (n) resize(n);
+		TORRENT_ASSERT(size < std::size_t(std::numeric_limits<std::int32_t>::max()));
+
+		if (size == 0) return;
+
+		size = round_up8(size);
+
+		m_begin = static_cast<char*>(std::malloc(size));
+		if (m_begin == nullptr)
+		{
+#ifndef BOOST_NO_EXCEPTIONS
+			throw std::bad_alloc();
+#else
+			std::terminate();
+#endif
+		}
+
+		// the actual allocation may be larger than we requested. If so, let the
+		// user take advantage of every single byte
+#if defined __GLIBC__
+		m_size = malloc_usable_size(m_begin);
+#elif defined _MSC_VER
+		m_size = _msize(m_begin);
+#elif defined TORRENT_BSD
+		m_size = malloc_size(m_begin);
+#else
+		m_size = size;
+#endif
 	}
 
-	buffer(buffer const& b)
-		: m_begin(0)
-		, m_size(0)
-		, m_capacity(0)
+	// allocate an uninitialized buffer of the specified size
+	// and copy the initialization range into the start of the buffer
+	buffer(std::size_t const size, aux::array_view<char const> initialize)
+		: buffer(size)
 	{
-		if (b.size() == 0) return;
-		resize(b.size());
-		std::memcpy(m_begin, b.begin(), b.size());
+		TORRENT_ASSERT(initialize.size() <= size);
+		if (initialize.size() > 0)
+		{
+			memcpy(m_begin, initialize.data(), std::min(initialize.size(), size));
+		}
 	}
 
+	buffer(buffer const& b) = delete;
+
 	buffer(buffer&& b)
 		: m_begin(b.m_begin)
 		, m_size(b.m_size)
-		, m_capacity(b.m_capacity)
 	{
 		b.m_begin = nullptr;
-		b.m_size = b.m_capacity = 0;
+		b.m_size = 0;
 	}
 
 	buffer& operator=(buffer&& b)
@@ -143,86 +192,30 @@ public:
 		std::free(m_begin);
 		m_begin = b.m_begin;
 		m_size = b.m_size;
-		m_capacity = b.m_capacity;
 		b.m_begin = nullptr;
-		b.m_size = b.m_capacity = 0;
+		b.m_size = 0;
 		return *this;
 	}
 
-	buffer& operator=(buffer const& b)
-	{
-		if (&b == this) return *this;
-		resize(b.size());
-		if (b.size() == 0) return *this;
-		std::memcpy(m_begin, b.begin(), b.size());
-		return *this;
-	}
+	buffer& operator=(buffer const& b) = delete;
 
-	~buffer()
-	{
-		std::free(m_begin);
-	}
+	~buffer() { std::free(m_begin); }
+
+	// TODO: 3 fix the naming convention here
+	char* ptr() { return m_begin; }
+	char const* ptr() const { return m_begin; }
 
 	buffer::interval data()
 	{ return interval(m_begin, m_begin + m_size); }
 	buffer::const_interval data() const
-	{ return const_interval(m_begin, m_begin + m_size); }
+	{ return interval(m_begin, m_begin + m_size); }
 
-	void resize(std::size_t n)
-	{
-		TORRENT_ASSERT(n < std::numeric_limits<std::uint32_t>::max());
-		reserve(n);
-		m_size = std::uint32_t(n);
-	}
+	operator aux::array_view<char>()
+	{ return aux::array_view<char>(m_begin, int(m_size)); }
+	operator aux::array_view<char const>() const
+	{ return aux::array_view<char const>(m_begin, int(m_size)); }
 
-	void insert(char* point, char const* first, char const* last)
-	{
-		std::size_t p = point - m_begin;
-		if (point == m_begin + m_size)
-		{
-			resize(size() + last - first);
-			std::memcpy(m_begin + p, first, last - first);
-			return;
-		}
-
-		resize(size() + last - first);
-		std::memmove(m_begin + p + (last - first), m_begin + p, last - first);
-		std::memcpy(m_begin + p, first, last - first);
-	}
-
-	void erase(char* b, char* e)
-	{
-		TORRENT_ASSERT(e <= m_begin + m_size);
-		TORRENT_ASSERT(b >= m_begin);
-		TORRENT_ASSERT(b <= e);
-		if (e == m_begin + m_size)
-		{
-			resize(b - m_begin);
-			return;
-		}
-		std::memmove(b, e, m_begin + m_size - e);
-		TORRENT_ASSERT(e >= b);
-		TORRENT_ASSERT(uintptr_t(e - b) <= std::numeric_limits<std::uint32_t>::max());
-		TORRENT_ASSERT(uintptr_t(e - b) <= m_size);
-		m_size -= std::uint32_t(e - b);
-	}
-
-	void clear() { m_size = 0; }
 	std::size_t size() const { return m_size; }
-	std::size_t capacity() const { return m_capacity; }
-	void reserve(std::size_t n)
-	{
-		if (n <= capacity()) return;
-		TORRENT_ASSERT(n > 0);
-		TORRENT_ASSERT(n < 0xffffffffu);
-
-		char* tmp = static_cast<char*>(std::realloc(m_begin, n));
-#ifndef BOOST_NO_EXCEPTIONS
-		if (tmp == nullptr) throw std::bad_alloc();
-#endif
-		m_begin = tmp;
-		m_capacity = std::uint32_t(n);
-	}
 
 	bool empty() const { return m_size == 0; }
 	char& operator[](std::size_t i) { TORRENT_ASSERT(i < size()); return m_begin[i]; }
@@ -238,15 +231,13 @@ public:
 		using std::swap;
 		swap(m_begin, b.m_begin);
 		swap(m_size, b.m_size);
-		swap(m_capacity, b.m_capacity);
 	}
 private:
-	char* m_begin;
-	std::uint32_t m_size;
-	std::uint32_t m_capacity;
+	char* m_begin = nullptr;
+	// m_begin points to an allocation of this size.
+	std::size_t m_size = 0;
 };
 
-
 }
 
 #endif // BTORRENT_BUFFER_HPP_INCLUDED
diff --git a/include/libtorrent/peer_connection.hpp b/include/libtorrent/peer_connection.hpp
index ffc28a047..2f87d7fe0 100644
--- a/include/libtorrent/peer_connection.hpp
+++ b/include/libtorrent/peer_connection.hpp
@@ -719,8 +719,6 @@ namespace libtorrent
 
 	protected:
 
-		size_t try_read(sync_t s, error_code& ec);
-
 		virtual void get_specific_peer_info(peer_info& p) const = 0;
 
 		virtual void write_choke() = 0;
@@ -742,6 +740,7 @@ namespace libtorrent
 		virtual void on_connected() = 0;
 		virtual void on_tick() {}
 
+		// implemented by concrete connection classes
 		virtual void on_receive(error_code const& error
 			, std::size_t bytes_transferred) = 0;
 		virtual void on_sent(error_code const& error
@@ -752,7 +751,10 @@ namespace libtorrent
 		virtual
 		std::tuple<int, aux::array_view<boost::asio::const_buffer>>
 		hit_send_barrier(aux::array_view<boost::asio::mutable_buffer> /* iovec */)
-		{ return std::make_tuple(INT_MAX, aux::array_view<boost::asio::const_buffer>()); }
+		{
+			return std::make_tuple(INT_MAX
+				, aux::array_view<boost::asio::const_buffer>());
+		}
 
 		void attach_to_torrent(sha1_hash const& ih);
 
@@ -760,21 +762,6 @@ namespace libtorrent
 
 		void update_desired_queue_size();
 
-		// called from the main loop when this connection has any
-		// work to do.
-		void on_send_data(error_code const& error
-			, std::size_t bytes_transferred);
-		void on_receive_data(error_code const& error
-			, std::size_t bytes_transferred);
-
-		// _nb means null_buffers. i.e. we just know the socket is
-		// readable at this point, we don't know how much has been received
-		void on_receive_data_nb(error_code const& error
-			, std::size_t bytes_transferred);
-
-		void receive_data_impl(error_code const& error
-			, std::size_t bytes_transferred, int read_loops);
-
 		void set_send_barrier(int bytes)
 		{
 			TORRENT_ASSERT(bytes == INT_MAX || bytes <= send_buffer_size());
@@ -788,6 +775,15 @@ namespace libtorrent
 		io_service& get_io_service() { return m_ios; }
 
 	private:
+
+		// callbacks for data being sent or received
+		void on_send_data(error_code const& error
+			, std::size_t bytes_transferred);
+		void on_receive_data(error_code const& error
+			, std::size_t bytes_transferred);
+
+		void account_received_bytes(int bytes_transferred);
+
 		// explicitly disallow assignment, to silence msvc warning
 		peer_connection& operator=(peer_connection const&);
 
diff --git a/include/libtorrent/peer_info.hpp b/include/libtorrent/peer_info.hpp
index e372b056a..5a5f371fd 100644
--- a/include/libtorrent/peer_info.hpp
+++ b/include/libtorrent/peer_info.hpp
@@ -246,6 +246,7 @@ namespace libtorrent
 		// allocated and used as receive buffer, respectively.
 		int receive_buffer_size;
 		int used_receive_buffer;
+		int receive_buffer_watermark;
 
 		// the number of pieces this peer has participated in sending us that
 		// turned out to fail the hash check.
diff --git a/include/libtorrent/receive_buffer.hpp b/include/libtorrent/receive_buffer.hpp
index a186171b3..b73ff84ce 100644
--- a/include/libtorrent/receive_buffer.hpp
+++ b/include/libtorrent/receive_buffer.hpp
@@ -35,6 +35,7 @@ POSSIBILITY OF SUCH DAMAGE.
 
 #include <libtorrent/buffer.hpp>
 #include <libtorrent/disk_buffer_holder.hpp>
+#include <libtorrent/sliding_average.hpp>
 #include <boost/asio/buffer.hpp>
 
 namespace libtorrent {
@@ -43,14 +44,6 @@ struct TORRENT_EXTRA_EXPORT receive_buffer
 {
 	friend struct crypto_receive_buffer;
 
-	receive_buffer()
-		: m_recv_start(0)
-		, m_recv_end(0)
-		, m_recv_pos(0)
-		, m_packet_size(0)
-		, m_soft_packet_size(0)
-	{}
-
 	int packet_size() const { return m_packet_size; }
 	int packet_bytes_remaining() const
 	{
@@ -59,19 +52,15 @@ struct TORRENT_EXTRA_EXPORT receive_buffer
 		return m_packet_size - m_recv_pos;
 	}
 
-	int max_receive();
+	int max_receive() const;
 
 	bool packet_finished() const { return m_packet_size <= m_recv_pos; }
 	int pos() const { return m_recv_pos; }
-	int capacity() const { return int(m_recv_buffer.capacity()); }
-
-	int regular_buffer_size() const
-	{
-		TORRENT_ASSERT(m_packet_size > 0);
-		return m_packet_size;
-	}
+	int capacity() const { return int(m_recv_buffer.size()); }
+	int watermark() const { return m_watermark.mean(); }
 
 	boost::asio::mutable_buffer reserve(int size);
+	void grow(int limit);
 
 	// tell the buffer we just received more bytes at the end of it. This will
 	// advance the end cursor
@@ -89,11 +78,6 @@ struct TORRENT_EXTRA_EXPORT receive_buffer
 	// has the read cursor reached the end cursor?
 	bool pos_at_end() { return m_recv_pos == m_recv_end; }
 
-	// make the buffer size divisible by 8 bytes (RC4 block size)
-	void clamp_size();
-
-	void set_soft_packet_size(int size) { m_soft_packet_size = size; }
-
 	// size = the packet size to remove from the receive buffer
 	// packet_size = the next packet size to receive in the buffer
 	// offset = the offset into the receive buffer where to remove `size` bytes
@@ -104,8 +88,7 @@ struct TORRENT_EXTRA_EXPORT receive_buffer
 	buffer::const_interval get() const;
 
 #if !defined(TORRENT_DISABLE_ENCRYPTION) && !defined(TORRENT_DISABLE_EXTENSIONS)
-	// returns the entire regular buffer
-	// should only be used during the handshake
+	// returns the entire buffer
 	buffer::interval mutable_buffer();
 
 	// returns the last 'bytes' from the receive buffer
@@ -114,7 +97,7 @@ struct TORRENT_EXTRA_EXPORT receive_buffer
 
 	// the purpose of this function is to free up and cut off all messages
 	// in the receive buffer that have been parsed and processed.
-	void normalize();
+	void normalize(int force_shrink = 0);
 	bool normalized() const { return m_recv_start == 0; }
 
 	void reset(int packet_size);
@@ -130,16 +113,15 @@ private:
 	// explicitly disallow assignment, to silence msvc warning
 	receive_buffer& operator=(receive_buffer const&);
 
-	// recv_buf.begin (start of actual receive buffer)
+	// m_recv_buffer.data() (start of actual receive buffer)
 	// |
-	// |      m_recv_start (logical start of current
-	// |      |  receive buffer, as perceived by upper layers)
+	// |      m_recv_start (tart of current packet)
 	// |      |
 	// |      |    m_recv_pos (number of bytes consumed
 	// |      |    |  by upper layer, from logical receive buffer)
 	// |      |    |
 	// |      x---------x
-	// |      |         |        recv_buf.end (end of actual receive buffer)
+	// |      |         |        m_recv_buffer.size() (end of actual receive buffer)
 	// |      |         |        |
 	// v      v         v        v
 	// *------==========---------
@@ -151,25 +133,23 @@ private:
 	// m_recv_buffer
 
 	// the start of the logical receive buffer
-	int m_recv_start;
+	int m_recv_start = 0;
 
 	// the number of valid, received bytes in m_recv_buffer
-	int m_recv_end;
+	int m_recv_end = 0;
 
 	// the byte offset in m_recv_buffer that we have
 	// are passing on to the upper layer. This is
 	// always <= m_recv_end
-	int m_recv_pos;
+	int m_recv_pos = 0;
 
 	// the size (in bytes) of the bittorrent message
 	// we're currently receiving
-	int m_packet_size;
+	int m_packet_size = 0;
 
-	// the number of bytes that the other
-	// end has to send us in order to respond
-	// to all outstanding piece requests we
-	// have sent to it
-	int m_soft_packet_size;
+	// keep track of how much of the receive buffer we use, if we're not using
+	// enuogh of it we shrink it
+	sliding_average<20> m_watermark;
 
 	buffer m_recv_buffer;
 };
@@ -183,10 +163,7 @@ private:
 struct crypto_receive_buffer
 {
 	crypto_receive_buffer(receive_buffer& next)
-		: m_recv_pos(INT_MAX)
-		, m_packet_size(0)
-		, m_soft_packet_size(0)
-		, m_connection_buffer(next)
+		: m_connection_buffer(next)
 	{}
 
 	buffer::interval mutable_buffer() { return m_connection_buffer.mutable_buffer(); }
@@ -219,8 +196,6 @@ struct crypto_receive_buffer
 	void reset(int packet_size);
 	void crypto_reset(int packet_size);
 
-	void set_soft_packet_size(int size);
-
 	int advance_pos(int bytes);
 
 	buffer::const_interval get() const;
@@ -231,9 +206,8 @@ private:
 	// explicitly disallow assignment, to silence msvc warning
 	crypto_receive_buffer& operator=(crypto_receive_buffer const&);
 
-	int m_recv_pos;
-	int m_packet_size;
-	int m_soft_packet_size;
+	int m_recv_pos = std::numeric_limits<int>::max();
+	int m_packet_size = 0;
 	receive_buffer& m_connection_buffer;
 };
 #endif // TORRENT_DISABLE_ENCRYPTION
@@ -241,3 +215,4 @@ private:
 } // namespace libtorrent
 
 #endif // #ifndef TORRENT_RECEIVE_BUFFER_HPP_INCLUDED
+
diff --git a/include/libtorrent/settings_pack.hpp b/include/libtorrent/settings_pack.hpp
index 0fb6751f3..159a641e7 100644
--- a/include/libtorrent/settings_pack.hpp
+++ b/include/libtorrent/settings_pack.hpp
@@ -1098,7 +1098,6 @@ namespace libtorrent
 			// and is disconnected.
 			max_rejects,
 
-			// ``recv_socket_buffer_size`` and ``send_socket_buffer_size``
 			// specifies the buffer sizes set on peer sockets. 0 (which is the
 			// default) means the OS default (i.e. don't change the buffer sizes).
 			// The socket buffer sizes are changed using setsockopt() with
@@ -1106,6 +1105,10 @@ namespace libtorrent
 			recv_socket_buffer_size,
 			send_socket_buffer_size,
 
+			// the max number of bytes a single peer connection's receive buffer is
+			// allowed to grow to.
+			max_peer_recv_buffer_size,
+
 			// ``file_checks_delay_per_block`` is the number of milliseconds to
 			// sleep in between disk read operations when checking torrents. This
 			// defaults to 0, but can be set to higher numbers to slow down the
diff --git a/include/libtorrent/sliding_average.hpp b/include/libtorrent/sliding_average.hpp
index ecd342e3d..b3b32518a 100644
--- a/include/libtorrent/sliding_average.hpp
+++ b/include/libtorrent/sliding_average.hpp
@@ -45,6 +45,8 @@ template <int inverted_gain>
 struct sliding_average
 {
 	sliding_average(): m_mean(0), m_average_deviation(0), m_num_samples(0) {}
+	sliding_average(sliding_average const&) = default;
+	sliding_average& operator=(sliding_average const&) = default;
 
 	void add_sample(int s)
 	{
@@ -75,19 +77,16 @@ struct sliding_average
 
 private:
 	// both of these are fixed point values (* 64)
-	int m_mean;
-	int m_average_deviation;
+	int m_mean = 0;
+	int m_average_deviation = 0;
 	// the number of samples we have received, but no more than inverted_gain
 	// this is the effective inverted_gain
-	int m_num_samples;
+	int m_num_samples = 0;
 };
 
 struct average_accumulator
 {
-	average_accumulator()
-		: m_num_samples(0)
-		, m_sample_sum(0)
-	{}
+	average_accumulator() {}
 
 	void add_sample(int s)
 	{
@@ -108,8 +107,10 @@ struct average_accumulator
 		return ret;
 	}
 
-	int m_num_samples;
-	std::uint64_t m_sample_sum;
+private:
+
+	int m_num_samples = 0;
+	std::uint64_t m_sample_sum = 0;
 };
 
 }
diff --git a/include/libtorrent/stack_allocator.hpp b/include/libtorrent/stack_allocator.hpp
index 692ec09c9..8b1468b71 100644
--- a/include/libtorrent/stack_allocator.hpp
+++ b/include/libtorrent/stack_allocator.hpp
@@ -42,6 +42,10 @@ namespace libtorrent { namespace aux
 	{
 		stack_allocator() {}
 
+		// non-copyable
+		stack_allocator(stack_allocator const&) = delete;
+		stack_allocator& operator=(stack_allocator const&) = delete;
+
 		int copy_string(std::string const& str)
 		{
 			int ret = int(m_storage.size());
@@ -100,11 +104,7 @@ namespace libtorrent { namespace aux
 
 	private:
 
-		// non-copyable
-		stack_allocator(stack_allocator const&);
-		stack_allocator& operator=(stack_allocator const&);
-
-		buffer m_storage;
+		std::vector<char> m_storage;
 	};
 
 } }
diff --git a/src/bt_peer_connection.cpp b/src/bt_peer_connection.cpp
index 6cbf6b926..117ef3d68 100644
--- a/src/bt_peer_connection.cpp
+++ b/src/bt_peer_connection.cpp
@@ -1105,12 +1105,11 @@ namespace libtorrent
 
 		boost::shared_ptr<torrent> t = associated_torrent().lock();
 		TORRENT_ASSERT(t);
-		bool merkle = static_cast<std::uint8_t>(recv_buffer.begin[0]) == 250;
+		bool const merkle = static_cast<std::uint8_t>(recv_buffer.begin[0]) == 250;
 		if (merkle)
 		{
 			if (recv_pos == 1)
 			{
-				m_recv_buffer.set_soft_packet_size(13);
 				received_bytes(0, received);
 				return;
 			}
@@ -1119,13 +1118,10 @@ namespace libtorrent
 				received_bytes(0, received);
 				return;
 			}
-			if (recv_pos == 13)
+			if (recv_pos >= 13)
 			{
-				const char* ptr = recv_buffer.begin + 9;
-				int list_size = detail::read_int32(ptr);
-				// now we know how long the bencoded hash list is
-				// and we can allocate the disk buffer and receive
-				// into it
+				char const* ptr = recv_buffer.begin + 9;
+				int const list_size = detail::read_int32(ptr);
 
 				if (list_size > m_recv_buffer.packet_size() - 13)
 				{
diff --git a/src/file_progress.cpp b/src/file_progress.cpp
index 9fa15aeca..c8c67cccc 100644
--- a/src/file_progress.cpp
+++ b/src/file_progress.cpp
@@ -91,12 +91,13 @@ namespace libtorrent { namespace aux
 			m_have_pieces.set_bit(piece);
 #endif
 
-			int size = (std::min)(std::uint64_t(piece_size), total_size - off);
+			TORRENT_ASSERT(total_size >= off);
+			std::int64_t size = (std::min)(std::uint64_t(piece_size), total_size - off);
 			TORRENT_ASSERT(size >= 0);
 
 			while (size)
 			{
-				int add = (std::min)(std::int64_t(size), fs.file_size(file_index) - file_offset);
+				std::int64_t add = (std::min)(size, fs.file_size(file_index) - file_offset);
 				TORRENT_ASSERT(add >= 0);
 				m_file_progress[file_index] += add;
 
diff --git a/src/peer_connection.cpp b/src/peer_connection.cpp
index 8d1acdf50..9da90d46d 100644
--- a/src/peer_connection.cpp
+++ b/src/peer_connection.cpp
@@ -4485,6 +4485,7 @@ namespace libtorrent
 		p.used_send_buffer = m_send_buffer.size();
 		p.receive_buffer_size = m_recv_buffer.capacity();
 		p.used_receive_buffer = m_recv_buffer.pos();
+		p.receive_buffer_watermark = m_recv_buffer.watermark();
 		p.write_state = m_channel_state[upload_channel];
 		p.read_state = m_channel_state[download_channel];
 
@@ -5409,10 +5410,11 @@ namespace libtorrent
 		}
 	}
 
-	int peer_connection::request_bandwidth(int channel, int bytes)
+	int peer_connection::request_bandwidth(int const channel, int bytes)
 	{
 		TORRENT_ASSERT(is_single_thread());
 		INVARIANT_CHECK;
+
 		// we can only have one outstanding bandwidth request at a time
 		if (m_channel_state[channel] & peer_info::bw_limit) return 0;
 
@@ -5652,8 +5654,15 @@ namespace libtorrent
 
 		if (m_disconnecting) return;
 
+		if (m_recv_buffer.capacity() < 100
+			&& m_recv_buffer.max_receive() == 0)
+		{
+			m_recv_buffer.reserve(100);
+		}
+
 		// we may want to request more quota at this point
-		request_bandwidth(download_channel);
+		int const buffer_size = m_recv_buffer.max_receive();
+		request_bandwidth(download_channel, buffer_size);
 
 		if (m_channel_state[download_channel] & peer_info::bw_network) return;
 
@@ -5680,94 +5689,27 @@ namespace libtorrent
 			// from being at or exceeding the limit down to below the limit
 			return;
 		}
-		error_code ec;
-
-		try_read(read_async, ec);
-	}
-
-	size_t peer_connection::try_read(sync_t s, error_code& ec)
-	{
-		TORRENT_ASSERT(is_single_thread());
 		TORRENT_ASSERT(m_connected);
+		if (m_quota[download_channel] == 0) return;
 
-		if (m_quota[download_channel] == 0)
-		{
-			ec = boost::asio::error::would_block;
-			return 0;
-		}
+		int const quota_left = m_quota[download_channel];
+		int const max_receive = (std::min)(buffer_size, quota_left);
 
-		if (!can_read())
-		{
-			ec = boost::asio::error::would_block;
-			return 0;
-		}
-
-		int max_receive = m_recv_buffer.max_receive();
-
-		// only apply the contiguous receive buffer when we don't have any
-		// outstanding requests. When we're likely to receive pieces, we'll
-		// save more time from avoiding copying data from the socket
-		if (m_download_queue.empty())
-		{
-			if (s == read_sync)
-			{
-				ec = boost::asio::error::would_block;
-				return 0;
-			}
-
-			TORRENT_ASSERT((m_channel_state[download_channel] & peer_info::bw_network) == 0);
-			m_channel_state[download_channel] |= peer_info::bw_network;
-#ifndef TORRENT_DISABLE_LOGGING
-			peer_log(peer_log_alert::incoming, "ASYNC_READ");
-#endif
-
-			ADD_OUTSTANDING_ASYNC("peer_connection::on_receive_data_nb");
-			m_socket->async_read_some(null_buffers(), make_read_handler(
-				std::bind(&peer_connection::on_receive_data_nb, self(), _1, _2)));
-			return 0;
-		}
-
-		TORRENT_ASSERT(max_receive >= 0);
-
-		int quota_left = m_quota[download_channel];
-		if (max_receive > quota_left)
-			max_receive = quota_left;
-
-		if (max_receive == 0)
-		{
-			ec = boost::asio::error::would_block;
-			return 0;
-		}
+		if (max_receive == 0) return;
 
 		boost::asio::mutable_buffer const vec = m_recv_buffer.reserve(max_receive);
-
-		if (s == read_async)
-		{
-			TORRENT_ASSERT((m_channel_state[download_channel] & peer_info::bw_network) == 0);
-			m_channel_state[download_channel] |= peer_info::bw_network;
+		TORRENT_ASSERT((m_channel_state[download_channel] & peer_info::bw_network) == 0);
+		m_channel_state[download_channel] |= peer_info::bw_network;
 #ifndef TORRENT_DISABLE_LOGGING
-			peer_log(peer_log_alert::incoming, "ASYNC_READ"
-				, "max: %d bytes", max_receive);
+		peer_log(peer_log_alert::incoming, "ASYNC_READ"
+			, "max: %d bytes", max_receive);
 #endif
 
-			// utp sockets aren't thread safe...
-			ADD_OUTSTANDING_ASYNC("peer_connection::on_receive_data");
-			m_socket->async_read_some(
-				boost::asio::mutable_buffers_1(vec), make_read_handler(
-					std::bind(&peer_connection::on_receive_data, self(), _1, _2)));
-			return 0;
-		}
-
-		size_t const ret = m_socket->read_some(boost::asio::mutable_buffers_1(vec), ec);
-
-		// this is weird. You would imagine read_some() would do this
-		if (ret == 0 && !ec) ec = boost::asio::error::eof;
-
-#ifndef TORRENT_DISABLE_LOGGING
-		peer_log(peer_log_alert::incoming, "SYNC_READ", "max: %d ret: %d e: %s"
-			, max_receive, int(ret), ec ? ec.message().c_str() : "");
-#endif
-		return ret;
+		// utp sockets aren't thread safe...
+		ADD_OUTSTANDING_ASYNC("peer_connection::on_receive_data");
+		m_socket->async_read_some(
+			boost::asio::mutable_buffers_1(vec), make_read_handler(
+				std::bind(&peer_connection::on_receive_data, self(), _1, _2)));
 	}
 
 	void peer_connection::append_send_buffer(char* buffer, int size
@@ -5859,112 +5801,31 @@ namespace libtorrent
 		bool m_cond;
 	};
 
-	void peer_connection::on_receive_data_nb(const error_code& error
-		, std::size_t bytes_transferred)
-	{
-		TORRENT_ASSERT(is_single_thread());
-		COMPLETE_ASYNC("peer_connection::on_receive_data_nb");
-
-		// leave this bit set until we're done looping, reading from the socket.
-		// that way we don't trigger any async read calls until the end of this
-		// function.
-		TORRENT_ASSERT(m_channel_state[download_channel] & peer_info::bw_network);
-
-		// nb is short for null_buffers. In this mode we don't actually
-		// allocate a receive buffer up-front, but get notified when
-		// we can read from the socket, and then determine how much there
-		// is to read.
-
-		if (error)
-		{
-			TORRENT_ASSERT_VAL(error.value() != 0, error.value());
-#ifndef TORRENT_DISABLE_LOGGING
-			peer_log(peer_log_alert::info, "ERROR"
-				, "in peer_connection::on_receive_data_nb error: (%s:%d) %s"
-				, error.category().name(), error.value()
-				, error.message().c_str());
-#endif
-			on_receive(error, bytes_transferred);
-			disconnect(error, op_sock_read);
-			return;
-		}
-
-		error_code ec;
-		std::size_t buffer_size = m_socket->available(ec);
-		if (ec)
-		{
-			disconnect(ec, op_available);
-			return;
-		}
-
-#ifndef TORRENT_DISABLE_LOGGING
-		peer_log(peer_log_alert::incoming, "READ_AVAILABLE"
-			, "bytes: %d", int(buffer_size));
-#endif
-
-		// at this point the ioctl told us the socket doesn't have any
-		// pending bytes. This probably means some error happened.
-		// in order to find out though, we need to initiate a read
-		// operation
-		if (buffer_size == 0)
-		{
-			// try to read one byte. The socket is non-blocking anyway
-			// so worst case, we'll fail with EWOULDBLOCK
-			buffer_size = 1;
-		}
-		else
-		{
-			if (buffer_size > m_quota[download_channel])
-			{
-				request_bandwidth(download_channel, int(buffer_size));
-				buffer_size = m_quota[download_channel];
-			}
-			// we're already waiting to get some more
-			// quota from the bandwidth manager
-			if (buffer_size == 0)
-			{
-				// allow reading from the socket again
-				TORRENT_ASSERT(m_channel_state[download_channel] & peer_info::bw_network);
-				m_channel_state[download_channel] &= ~peer_info::bw_network;
-				return;
-			}
-		}
-
-		if (buffer_size > 2097152) buffer_size = 2097152;
-
-		boost::asio::mutable_buffer buffer = m_recv_buffer.reserve(int(buffer_size));
-		TORRENT_ASSERT(m_recv_buffer.normalized());
-
-		bytes_transferred = m_socket->read_some(boost::asio::mutable_buffers_1(buffer), ec);
-
-		if (ec)
-		{
-			if (ec == boost::asio::error::try_again
-				|| ec == boost::asio::error::would_block)
-			{
-				// allow reading from the socket again
-				TORRENT_ASSERT(m_channel_state[download_channel] & peer_info::bw_network);
-				m_channel_state[download_channel] &= ~peer_info::bw_network;
-				setup_receive();
-				return;
-			}
-			disconnect(ec, op_sock_read);
-			return;
-		}
-
-		receive_data_impl(error, bytes_transferred, 0);
-	}
-
 	// --------------------------
 	// RECEIVE DATA
 	// --------------------------
 
-	// nb is true if this callback is due to a null_buffers()
-	// invocation of async_read_some(). In that case, we need
-	// to disregard bytes_transferred.
-	// at all exit points of this function, one of the following MUST hold:
-	//  1. the socket is disconnecting
-	//  2. m_channel_state[download_channel] & peer_info::bw_network == 0
+	void peer_connection::account_received_bytes(int const bytes_transferred)
+	{
+		// tell the receive buffer we just fed it this many bytes of incoming data
+		TORRENT_ASSERT(bytes_transferred > 0);
+		m_recv_buffer.received(bytes_transferred);
+
+		// update the dl quota
+		TORRENT_ASSERT(bytes_transferred <= m_quota[download_channel]);
+		m_quota[download_channel] -= bytes_transferred;
+
+		// account receiver buffer size stats to the session
+		m_ses.received_buffer(bytes_transferred);
+
+		// estimage transport protocol overhead
+		trancieve_ip_packet(bytes_transferred, m_remote.address().is_v6());
+
+#ifndef TORRENT_DISABLE_LOGGING
+		peer_log(peer_log_alert::incoming, "READ"
+			, "%d bytes", int(bytes_transferred));
+#endif
+	}
 
 	void peer_connection::on_receive_data(const error_code& error
 		, std::size_t bytes_transferred)
@@ -5972,6 +5833,13 @@ namespace libtorrent
 		TORRENT_ASSERT(is_single_thread());
 		COMPLETE_ASYNC("peer_connection::on_receive_data");
 
+#ifndef TORRENT_DISABLE_LOGGING
+		peer_log(peer_log_alert::incoming, "ON_RECEIVE_DATA"
+			, "bytes: %d error: (%s:%d) %s"
+			, int(bytes_transferred), error.category().name(), error.value()
+			, error.message().c_str());
+#endif
+
 		// leave this bit set until we're done looping, reading from the socket.
 		// that way we don't trigger any async read calls until the end of this
 		// function.
@@ -5979,19 +5847,23 @@ namespace libtorrent
 
 		TORRENT_ASSERT(bytes_transferred > 0 || error);
 
-		receive_data_impl(error, bytes_transferred, 10);
-	}
+		m_counters.inc_stats_counter(counters::on_read_counter);
 
-	void peer_connection::receive_data_impl(const error_code& error
-		, std::size_t bytes_transferred, int read_loops)
-	{
-		TORRENT_ASSERT(is_single_thread());
+		INVARIANT_CHECK;
+
+		if (error)
+		{
 #ifndef TORRENT_DISABLE_LOGGING
-		peer_log(peer_log_alert::incoming, "ON_RECEIVE_DATA"
-			, "bytes: %d error: (%s:%d) %s"
-			, int(bytes_transferred), error.category().name(), error.value()
-			, error.message().c_str());
+			peer_log(peer_log_alert::info, "ERROR"
+				, "in peer_connection::on_receive_data_impl error: %s"
+				, error.message().c_str());
 #endif
+			on_receive(error, bytes_transferred);
+			disconnect(error, op_sock_read);
+			return;
+		}
+
+		m_last_receive = aux::time_now();
 
 		// submit all disk jobs later
 		m_ses.deferred_submit_jobs();
@@ -6005,27 +5877,14 @@ namespace libtorrent
 		// flush the send buffer at the end of this function
 		cork _c(*this);
 
-		INVARIANT_CHECK;
-
-		int bytes_in_loop = int(bytes_transferred);
-
-		if (error)
-		{
-#ifndef TORRENT_DISABLE_LOGGING
-			peer_log(peer_log_alert::info, "ERROR"
-				, "in peer_connection::on_receive_data_impl error: %s"
-				, error.message().c_str());
-#endif
-			trancieve_ip_packet(bytes_in_loop, m_remote.address().is_v6());
-			on_receive(error, bytes_transferred);
-			disconnect(error, op_sock_read);
-			return;
-		}
-
 		TORRENT_ASSERT(bytes_transferred > 0);
 
-		m_counters.inc_stats_counter(counters::on_read_counter);
-		m_ses.received_buffer(int(bytes_transferred));
+		// if we received exactly as many bytes as we provided a receive buffer
+		// for. There most likely are more bytes to read, and we should grow our
+		// receive buffer.
+		TORRENT_ASSERT(bytes_transferred <= m_recv_buffer.max_receive());
+		bool const grow_buffer = (bytes_transferred == m_recv_buffer.max_receive());
+		account_received_bytes(bytes_transferred);
 
 		if (m_extension_outstanding_bytes > 0)
 			m_extension_outstanding_bytes -= (std::min)(m_extension_outstanding_bytes, int(bytes_transferred));
@@ -6033,79 +5892,91 @@ namespace libtorrent
 		check_graceful_pause();
 		if (m_disconnecting) return;
 
-		int num_loops = 0;
-		do
+		// this is the case where we try to grow the receive buffer and try to
+		// drain the socket
+		if (grow_buffer)
 		{
-#ifndef TORRENT_DISABLE_LOGGING
-			peer_log(peer_log_alert::incoming, "READ"
-				, "%d bytes", int(bytes_transferred));
-#endif
-			// correct the dl quota usage, if not all of the buffer was actually read
-			TORRENT_ASSERT(int(bytes_transferred) <= m_quota[download_channel]);
-			m_quota[download_channel] -= int(bytes_transferred);
-
-			if (m_disconnecting)
-			{
-				trancieve_ip_packet(bytes_in_loop, m_remote.address().is_v6());
-				return;
-			}
-
-			TORRENT_ASSERT(bytes_transferred > 0);
-			m_recv_buffer.received(int(bytes_transferred));
-
-			int bytes = int(bytes_transferred);
-			int sub_transferred = 0;
-			do {
-// TODO: The stats checks can not be honored when authenticated encryption is in use
-// because we may have encrypted data which we cannot authenticate yet
-#if 0
-				std::int64_t cur_payload_dl = m_statistics.last_payload_downloaded();
-				std::int64_t cur_protocol_dl = m_statistics.last_protocol_downloaded();
-#endif
-				sub_transferred = m_recv_buffer.advance_pos(bytes);
-				on_receive(error, sub_transferred);
-				bytes -= sub_transferred;
-				TORRENT_ASSERT(sub_transferred > 0);
-
-#if 0
-				TORRENT_ASSERT(m_statistics.last_payload_downloaded() - cur_payload_dl >= 0);
-				TORRENT_ASSERT(m_statistics.last_protocol_downloaded() - cur_protocol_dl >= 0);
-				std::int64_t stats_diff = m_statistics.last_payload_downloaded() - cur_payload_dl +
-					m_statistics.last_protocol_downloaded() - cur_protocol_dl;
-				TORRENT_ASSERT(stats_diff == int(sub_transferred));
-#endif
-				if (m_disconnecting) return;
-
-			} while (bytes > 0 && sub_transferred > 0);
-
-			m_recv_buffer.normalize();
-
-			TORRENT_ASSERT(m_recv_buffer.pos_at_end());
-			TORRENT_ASSERT(m_recv_buffer.packet_size() > 0);
-
-			if (m_peer_choked)
-			{
-				m_recv_buffer.clamp_size();
-			}
-
-			if (num_loops > read_loops) break;
-
 			error_code ec;
-			bytes_transferred = try_read(read_sync, ec);
-			TORRENT_ASSERT(bytes_transferred > 0 || ec);
-			if (ec == boost::asio::error::would_block || ec == boost::asio::error::try_again) break;
+			std::size_t buffer_size = m_socket->available(ec);
 			if (ec)
 			{
-				trancieve_ip_packet(bytes_in_loop, m_remote.address().is_v6());
-				disconnect(ec, op_sock_read);
+				disconnect(ec, op_available);
 				return;
 			}
-			bytes_in_loop += int(bytes_transferred);
-			++num_loops;
-		}
-		while (bytes_transferred > 0);
 
-		m_last_receive = aux::time_now();
+#ifndef TORRENT_DISABLE_LOGGING
+			peer_log(peer_log_alert::incoming, "AVAILABLE"
+				, "%d bytes", int(buffer_size));
+#endif
+
+			request_bandwidth(download_channel, buffer_size);
+
+			int const quota_left = m_quota[download_channel];
+			if (buffer_size > quota_left) buffer_size = quota_left;
+			if (buffer_size > 0)
+			{
+				boost::asio::mutable_buffer const vec = m_recv_buffer.reserve(buffer_size);
+				size_t bytes = m_socket->read_some(boost::asio::mutable_buffers_1(vec), ec);
+
+				// this is weird. You would imagine read_some() would do this
+				if (bytes == 0 && !ec) ec = boost::asio::error::eof;
+
+#ifndef TORRENT_DISABLE_LOGGING
+				peer_log(peer_log_alert::incoming, "SYNC_READ", "max: %d ret: %d e: %s"
+					, int(buffer_size), int(bytes), ec ? ec.message().c_str() : "");
+#endif
+
+				TORRENT_ASSERT(bytes > 0 || ec);
+				if (ec == boost::asio::error::would_block || ec == boost::asio::error::try_again)
+				{
+					bytes = 0;
+				}
+				else if (ec)
+				{
+					disconnect(ec, op_sock_read);
+					return;
+				}
+
+				account_received_bytes(bytes);
+
+				bytes_transferred += bytes;
+			}
+		}
+
+		// feed bytes in receive buffer to upper layer by calling on_receive()
+
+		bool const prev_choked = m_peer_choked;
+		int bytes = int(bytes_transferred);
+		int sub_transferred = 0;
+		do {
+			sub_transferred = m_recv_buffer.advance_pos(bytes);
+			on_receive(error, sub_transferred);
+			bytes -= sub_transferred;
+			TORRENT_ASSERT(sub_transferred > 0);
+			if (m_disconnecting) return;
+		} while (bytes > 0 && sub_transferred > 0);
+
+		// if the peer went from unchoked to choked, suggest to the receive
+		// buffer that it shrinks to 100 bytes
+		int const force_shrink = (m_peer_choked && !prev_choked)
+			? 100 : 0;
+		m_recv_buffer.normalize(force_shrink);
+
+		if (m_recv_buffer.max_receive() == 0)
+		{
+			// the message we're receiving is larger than our receive
+			// buffer, we must grow.
+			int const buffer_size_limit
+				= m_settings.get_int(settings_pack::max_peer_recv_buffer_size);
+			m_recv_buffer.grow(buffer_size_limit);
+#ifndef TORRENT_DISABLE_LOGGING
+			peer_log(peer_log_alert::incoming, "GROW_BUFFER", "%d bytes"
+				, m_recv_buffer.capacity());
+#endif
+		}
+
+		TORRENT_ASSERT(m_recv_buffer.pos_at_end());
+		TORRENT_ASSERT(m_recv_buffer.packet_size() > 0);
 
 		if (is_seed())
 		{
@@ -6113,8 +5984,6 @@ namespace libtorrent
 			if (t) t->seen_complete();
 		}
 
-		trancieve_ip_packet(bytes_in_loop, m_remote.address().is_v6());
-
 		// allow reading from the socket again
 		TORRENT_ASSERT(m_channel_state[download_channel] & peer_info::bw_network);
 		m_channel_state[download_channel] &= ~peer_info::bw_network;
@@ -6345,16 +6214,14 @@ namespace libtorrent
 
 		time_point now = clock_type::now();
 
-		for (std::vector<pending_block>::iterator i = m_download_queue.begin()
-			, end(m_download_queue.end()); i != end; ++i)
+		for (auto& block : m_download_queue)
 		{
-			if (i->send_buffer_offset == pending_block::not_in_buffer) continue;
-			std::int32_t offset = i->send_buffer_offset;
-			offset -= int(bytes_transferred);
-			if (offset < 0)
-				i->send_buffer_offset = pending_block::not_in_buffer;
+			if (block.send_buffer_offset == pending_block::not_in_buffer)
+				continue;
+			if (block.send_buffer_offset < int(bytes_transferred))
+				block.send_buffer_offset = pending_block::not_in_buffer;
 			else
-				i->send_buffer_offset = offset;
+				block.send_buffer_offset -= int(bytes_transferred);
 		}
 
 		m_channel_state[upload_channel] &= ~peer_info::bw_network;
diff --git a/src/receive_buffer.cpp b/src/receive_buffer.cpp
index 3ff1cb8ad..05395590d 100644
--- a/src/receive_buffer.cpp
+++ b/src/receive_buffer.cpp
@@ -34,61 +34,65 @@ POSSIBILITY OF SUCH DAMAGE.
 
 namespace libtorrent {
 
-	namespace {
-		int round_up8(int v)
-		{
-			return ((v & 7) == 0) ? v : v + (8 - (v & 7));
-		}
-	}
-
-int receive_buffer::max_receive()
+int receive_buffer::max_receive() const
 {
-	int max = packet_bytes_remaining();
-	if (m_recv_pos >= m_soft_packet_size) m_soft_packet_size = 0;
-	if (m_soft_packet_size && max > m_soft_packet_size - m_recv_pos)
-		max = m_soft_packet_size - m_recv_pos;
-	return max;
+	return int(m_recv_buffer.size() - m_recv_end);
 }
 
 boost::asio::mutable_buffer receive_buffer::reserve(int size)
 {
 	TORRENT_ASSERT(size > 0);
 	TORRENT_ASSERT(m_recv_pos >= 0);
-	// this is unintuitive, but we used to use m_recv_pos in this function when
-	// we should have used m_recv_end. perhaps they always happen to be equal
-	TORRENT_ASSERT(m_recv_pos == m_recv_end);
 
 	// normalize() must be called before receiving more data
 	TORRENT_ASSERT(m_recv_start == 0);
 
-	m_recv_buffer.resize(m_recv_end + size);
+	if (m_recv_buffer.size() < m_recv_end + size)
+	{
+		int const new_size = std::max(m_recv_end + size, m_packet_size);
+		buffer new_buffer(new_size
+			, aux::array_view<char const>(m_recv_buffer.ptr(), m_recv_end));
+		m_recv_buffer = std::move(new_buffer);
+
+		// since we just increased the size of the buffer, reset the watermark to
+		// start at our new size (avoid flapping the buffer size)
+		m_watermark = sliding_average<20>();
+	}
+
 	return boost::asio::buffer(&m_recv_buffer[0] + m_recv_end, size);
 }
 
+void receive_buffer::grow(int const limit)
+{
+	int const current_size = int(m_recv_buffer.size());
+	TORRENT_ASSERT(current_size < std::numeric_limits<int>::max() / 3);
+
+	// first grow to one piece message, then grow by 50% each time
+	int const new_size = (current_size < m_packet_size)
+		? m_packet_size : std::min(current_size * 3 / 2, limit);
+
+	// re-allcoate the buffer and copy over the part of it that's used
+	buffer new_buffer(new_size
+		, aux::array_view<char const>(m_recv_buffer.ptr(), m_recv_end));
+	m_recv_buffer = std::move(new_buffer);
+
+	// since we just increased the size of the buffer, reset the watermark to
+	// start at our new size (avoid flapping the buffer size)
+	m_watermark = sliding_average<20>();
+}
+
 int receive_buffer::advance_pos(int bytes)
 {
-	int const packet_size = m_soft_packet_size ? m_soft_packet_size : m_packet_size;
-	int const limit = packet_size > m_recv_pos ? packet_size - m_recv_pos : packet_size;
+	int const limit = m_packet_size > m_recv_pos ? m_packet_size - m_recv_pos : m_packet_size;
 	int const sub_transferred = (std::min)(bytes, limit);
 	m_recv_pos += sub_transferred;
-	if (m_recv_pos >= m_soft_packet_size) m_soft_packet_size = 0;
 	return sub_transferred;
 }
 
-void receive_buffer::clamp_size()
-{
-	if (m_recv_pos == 0
-		&& (m_recv_buffer.capacity() - m_packet_size) > 128)
-	{
-		// round up to an even 8 bytes since that's the RC4 blocksize
-		buffer(round_up8(m_packet_size)).swap(m_recv_buffer);
-	}
-}
-
 // size = the packet size to remove from the receive buffer
 // packet_size = the next packet size to receive in the buffer
 // offset = the offset into the receive buffer where to remove `size` bytes
-void receive_buffer::cut(int size, int packet_size, int offset)
+void receive_buffer::cut(int const size, int const packet_size, int const offset)
 {
 	TORRENT_ASSERT(packet_size > 0);
 	TORRENT_ASSERT(int(m_recv_buffer.size()) >= size);
@@ -104,9 +108,11 @@ void receive_buffer::cut(int size, int packet_size, int offset)
 		TORRENT_ASSERT(m_recv_start - size <= m_recv_end);
 
 		if (size > 0)
+		{
 			std::memmove(&m_recv_buffer[0] + m_recv_start + offset
 				, &m_recv_buffer[0] + m_recv_start + offset + size
 				, m_recv_end - m_recv_start - size - offset);
+		}
 
 		m_recv_pos -= size;
 		m_recv_end -= size;
@@ -168,13 +174,41 @@ boost::asio::mutable_buffer receive_buffer::mutable_buffer(int const bytes)
 
 // the purpose of this function is to free up and cut off all messages
 // in the receive buffer that have been parsed and processed.
-void receive_buffer::normalize()
+// it may also shrink the size of the buffer allocation if we haven't been using
+// enough of it lately.
+void receive_buffer::normalize(int force_shrink)
 {
 	TORRENT_ASSERT(m_recv_end >= m_recv_start);
-	if (m_recv_start == 0) return;
 
-	if (m_recv_end > m_recv_start)
-		std::memmove(&m_recv_buffer[0], &m_recv_buffer[0] + m_recv_start, m_recv_end - m_recv_start);
+	m_watermark.add_sample(std::max(m_recv_end, m_packet_size));
+
+	// if the running average drops below half of the current buffer size,
+	// reallocate a smaller one.
+	bool const shrink_buffer = m_recv_buffer.size() / 2 > m_watermark.mean()
+		&& m_watermark.mean() > (m_recv_end - m_recv_start);
+
+	aux::array_view<char const> bytes_to_shift(
+		m_recv_buffer.ptr() + m_recv_start
+			, m_recv_end - m_recv_start);
+
+	if (force_shrink)
+	{
+		const int target_size = std::max(std::max(force_shrink
+			, int(bytes_to_shift.size())), m_packet_size);
+		buffer new_buffer(target_size, bytes_to_shift);
+		m_recv_buffer = std::move(new_buffer);
+	}
+	else if (shrink_buffer)
+	{
+		buffer new_buffer(m_watermark.mean(), bytes_to_shift);
+		m_recv_buffer = std::move(new_buffer);
+	}
+	else if (m_recv_end > m_recv_start
+		&& m_recv_start > 0)
+	{
+		std::memmove(m_recv_buffer.ptr(), bytes_to_shift.data()
+			, bytes_to_shift.size());
+	}
 
 	m_recv_end -= m_recv_start;
 	m_recv_start = 0;
@@ -276,24 +310,14 @@ void crypto_receive_buffer::crypto_reset(int packet_size)
 	}
 }
 
-void crypto_receive_buffer::set_soft_packet_size(int size)
-{
-	if (m_recv_pos == INT_MAX)
-		m_connection_buffer.set_soft_packet_size(size);
-	else
-		m_soft_packet_size = size;
-}
-
 int crypto_receive_buffer::advance_pos(int bytes)
 {
 	if (m_recv_pos == INT_MAX) return bytes;
 
-	int packet_size = m_soft_packet_size ? m_soft_packet_size : m_packet_size;
-	int limit = packet_size > m_recv_pos ? packet_size - m_recv_pos : packet_size;
-	int sub_transferred = (std::min)(bytes, limit);
+	int const limit = m_packet_size > m_recv_pos ? m_packet_size - m_recv_pos : m_packet_size;
+	int const sub_transferred = (std::min)(bytes, limit);
 	m_recv_pos += sub_transferred;
 	m_connection_buffer.cut(0, m_connection_buffer.packet_size() + sub_transferred);
-	if (m_recv_pos >= m_soft_packet_size) m_soft_packet_size = 0;
 	return sub_transferred;
 }
 
diff --git a/src/session.cpp b/src/session.cpp
index 7493c9805..f741df06c 100644
--- a/src/session.cpp
+++ b/src/session.cpp
@@ -84,6 +84,8 @@ namespace libtorrent
 		set.set_bool(settings_pack::contiguous_recv_buffer, false);
 #endif
 
+		set.set_int(settings_pack::max_peer_recv_buffer_size, 32 * 1024 + 200);
+
 		set.set_int(settings_pack::disk_io_write_mode, settings_pack::disable_os_cache);
 		set.set_int(settings_pack::disk_io_read_mode, settings_pack::disable_os_cache);
 
@@ -171,6 +173,8 @@ namespace libtorrent
 		set.set_int(settings_pack::max_out_request_queue, 1500);
 		set.set_int(settings_pack::max_allowed_in_request_queue, 2000);
 
+		set.set_int(settings_pack::max_peer_recv_buffer_size, 5 * 1024 * 1024);
+
 		// we will probably see a high rate of alerts, make it less
 		// likely to loose alerts
 		set.set_int(settings_pack::alert_queue_size, 10000);
diff --git a/src/settings_pack.cpp b/src/settings_pack.cpp
index 6529639ef..4a01ba1b7 100644
--- a/src/settings_pack.cpp
+++ b/src/settings_pack.cpp
@@ -279,6 +279,7 @@ namespace libtorrent
 		SET(max_rejects, 50, 0),
 		SET(recv_socket_buffer_size, 0, &session_impl::update_socket_buffer_size),
 		SET(send_socket_buffer_size, 0, &session_impl::update_socket_buffer_size),
+		SET_NOPREV(max_peer_recv_buffer_size, 2 * 1024 * 1024, 0),
 		SET(file_checks_delay_per_block, 0, 0),
 		SET(read_cache_line_size, 32, 0),
 		SET(write_cache_line_size, 16, 0),
diff --git a/test/test_buffer.cpp b/test/test_buffer.cpp
index 39252210f..20a54a0c5 100644
--- a/test/test_buffer.cpp
+++ b/test/test_buffer.cpp
@@ -43,125 +43,97 @@ POSSIBILITY OF SUCH DAMAGE.
 
 using namespace libtorrent;
 
-/*
-template<class T>
-T const& min_(T const& x, T const& y)
-{
-	return x < y ? x : y;
-}
-
-void test_speed()
-{
-	buffer b;
-
-	char data[32];
-
-	srand(0);
-
-	boost::timer t;
-
-	int const iterations = 5000000;
-	int const step = iterations / 20;
-
-	for (int i = 0; i < iterations; ++i)
-	{
-		int x = rand();
-
-		if (i % step == 0) std::cerr << ".";
-
-		std::size_t n = rand() % 32;
-		n = 32;
-
-		if (x % 2)
-		{
-			b.insert(data, data + n);
-		}
-		else
-		{
-			b.erase(min_(b.size(), n));
-		}
-	}
-
-	float t1 = t.elapsed();
-	std::cerr << "buffer elapsed: " << t.elapsed() << "\n";
-
-	std::vector<char> v;
-
-	srand(0);
-	t.restart();
-
-	for (int i = 0; i < iterations; ++i)
-	{
-		int x = rand();
-
-		if (i % step == 0) std::cerr << ".";
-
-		std::size_t n = rand() % 32;
-		n = 32;
-
-		if (x % 2)
-		{
-			v.insert(v.end(), data, data + n);
-		}
-		else
-		{
-			v.erase(v.begin(), v.begin() + min_(v.size(), n));
-		}
-	}
-
-	float t2 = t.elapsed();
-	std::cerr << "std::vector elapsed: " << t.elapsed() << "\n";
-
-	assert(t1 < t2);
-}
-*/
-
 // -- test buffer --
 
-TORRENT_TEST(buffer)
+static char const data[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
+
+TORRENT_TEST(buffer_constructor)
 {
-	char data[] = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};
 
-	buffer b;
+	{
+		buffer b;
+		TEST_CHECK(b.size() == 0);
+		TEST_CHECK(b.empty());
+	}
 
-	TEST_CHECK(b.size() == 0);
-	TEST_CHECK(b.capacity() == 0);
-	TEST_CHECK(b.empty());
+	{
+		buffer b(10);
+		TEST_CHECK(b.size() >= 10);
+	}
 
-	b.resize(10);
-	TEST_CHECK(b.size() == 10);
-	TEST_CHECK(b.capacity() == 10);
+	{
+		buffer b(50, data);
+		TEST_CHECK(std::memcmp(b.ptr(), data, 10) == 0);
+		TEST_CHECK(b.size() >= 50);
+	}
+}
 
-	std::memcpy(b.begin(), data, 10);
-	b.reserve(50);
-	TEST_CHECK(std::memcmp(b.begin(), data, 10) == 0);
-	TEST_CHECK(b.capacity() == 50);
+TORRENT_TEST(buffer_swap)
+{
+	buffer b1;
+	TEST_CHECK(b1.size() == 0);
+	buffer b2(10, data);
+	std::size_t const b2_size = b2.size();
+	TEST_CHECK(b2_size >= 10);
 
-	b.erase(b.begin() + 6, b.end());
-	TEST_CHECK(std::memcmp(b.begin(), data, 6) == 0);
-	TEST_CHECK(b.capacity() == 50);
-	TEST_CHECK(b.size() == 6);
+	b1.swap(b2);
 
-	b.insert(b.begin(), data + 5, data + 10);
-	TEST_CHECK(b.capacity() == 50);
-	TEST_CHECK(b.size() == 11);
-	TEST_CHECK(std::memcmp(b.begin(), data + 5, 5) == 0);
+	TEST_CHECK(b2.size() == 0);
+	TEST_CHECK(b1.size() == b2_size);
+	TEST_CHECK(std::memcmp(b1.ptr(), data, 10) == 0);
+}
 
-	b.clear();
-	TEST_CHECK(b.size() == 0);
-	TEST_CHECK(b.capacity() == 50);
+TORRENT_TEST(buffer_subscript)
+{
+	buffer b(50, data);
+	TEST_CHECK(std::memcmp(b.ptr(), data, 10) == 0);
+	TEST_CHECK(b.size() >= 50);
 
-	b.insert(b.end(), data, data + 10);
-	TEST_CHECK(b.size() == 10);
-	TEST_CHECK(std::memcmp(b.begin(), data, 10) == 0);
+	for (int i = 0; i < int(sizeof(data)/sizeof(data[0])); ++i)
+		TEST_CHECK(b[i] == data[i]);
+}
 
-	b.erase(b.begin(), b.end());
-	TEST_CHECK(b.capacity() == 50);
-	TEST_CHECK(b.size() == 0);
+TORRENT_TEST(buffer_subscript2)
+{
+	buffer b(1);
+	TEST_CHECK(b.size() >= 1);
 
-	buffer().swap(b);
-	TEST_CHECK(b.capacity() == 0);
+	for (int i = 0; i < int(b.size()); ++i)
+		b[i] = i & 0xff;
 
+	for (int i = 0; i < int(b.size()); ++i)
+		TEST_CHECK(b[i] == (i & 0xff));
+}
+
+TORRENT_TEST(buffer_move_construct)
+{
+	buffer b1(50, data);
+	TEST_CHECK(std::memcmp(b1.ptr(), data, 10) == 0);
+	TEST_CHECK(b1.size() >= 50);
+
+	buffer b2(std::move(b1));
+
+	TEST_CHECK(b1.size() == 0);
+
+	TEST_CHECK(std::memcmp(b2.ptr(), data, 10) == 0);
+	TEST_CHECK(b2.size() >= 50);
+}
+
+TORRENT_TEST(buffer_move_assign)
+{
+	buffer b1(50, data);
+	TEST_CHECK(std::memcmp(b1.ptr(), data, 10) == 0);
+	TEST_CHECK(b1.size() >= 50);
+
+	buffer b2;
+	TEST_CHECK(b2.size() == 0);
+
+	b2 = std::move(b1);
+
+	TEST_CHECK(b1.size() == 0);
+
+	TEST_CHECK(std::memcmp(b2.ptr(), data, 10) == 0);
+	TEST_CHECK(b2.size() >= 50);
 }
 
 // -- test chained buffer --
diff --git a/test/test_receive_buffer.cpp b/test/test_receive_buffer.cpp
index 5ec624122..781f1898c 100644
--- a/test/test_receive_buffer.cpp
+++ b/test/test_receive_buffer.cpp
@@ -92,6 +92,128 @@ TORRENT_TEST(recv_buffer_packet_finished)
 	TEST_EQUAL(b.packet_finished(), true);
 }
 
+TORRENT_TEST(recv_buffer_grow_floor)
+{
+	receive_buffer b;
+	b.reset(1337);
+	b.grow(100000);
+
+	// the exact size depends on the OS allocator. Technically there's no upper
+	// bound, but it's likely withint some reasonable size
+	TEST_CHECK(b.capacity() >= 1337);
+	TEST_CHECK(b.capacity() < 1337 + 1000);
+}
+
+TORRENT_TEST(recv_buffer_grow)
+{
+	receive_buffer b;
+	b.reserve(200);
+	b.grow(100000);
+	// grow by 50%
+	TEST_CHECK(b.capacity() >= 300);
+	TEST_CHECK(b.capacity() < 300 + 500);
+}
+
+TORRENT_TEST(recv_buffer_grow_limit)
+{
+	receive_buffer b;
+	b.reserve(2000);
+	b.grow(2100);
+	// grow by 50%, but capped by 2100 bytes
+	TEST_CHECK(b.capacity() >= 2100);
+	TEST_CHECK(b.capacity() < 2100 + 500);
+	printf("capacity: %d\n", b.capacity());
+}
+
+TORRENT_TEST(recv_buffer_reserve_minimum_grow)
+{
+	receive_buffer b;
+	b.reset(1337);
+	b.reserve(20);
+
+	// we only asked for 20 more bytes, but since the message size was set to
+	// 1337, that's the minimum size to grow to
+	TEST_CHECK(b.capacity() >= 1337);
+	TEST_CHECK(b.capacity() < 1337 + 1000);
+}
+
+TORRENT_TEST(recv_buffer_reserve_grow)
+{
+	receive_buffer b;
+	b.reserve(20);
+
+	TEST_CHECK(b.capacity() >= 20);
+	TEST_CHECK(b.capacity() < 20 + 500);
+}
+
+TORRENT_TEST(recv_buffer_reserve)
+{
+	receive_buffer b;
+	auto range1 = b.reserve(100);
+
+	int const capacity = b.capacity();
+
+	b.reset(20);
+	b.received(20);
+
+	TEST_EQUAL(b.capacity(), capacity);
+
+	auto range2 = b.reserve(50);
+
+	using namespace boost::asio;
+
+	TEST_EQUAL(b.capacity(), capacity);
+	TEST_EQUAL(buffer_cast<char*>(range1) + 20, buffer_cast<char*>(range2));
+	TEST_CHECK(buffer_size(range1) >= 20);
+	TEST_CHECK(buffer_size(range2) >= 50);
+}
+
+TORRENT_TEST(receive_buffer_normalize)
+{
+	receive_buffer b;
+	b.reset(16000);
+
+	// receive one large packet, to allocate a large receive buffer
+	for (int i = 0; i < 16; ++i)
+	{
+		b.reserve(1000);
+		b.received(1000);
+		b.normalize();
+	}
+
+	TEST_CHECK(b.capacity() >= 16000);
+	int const start_capacity = b.capacity();
+
+	// then receive lots of small packets. We should eventually re-allocate down
+	// to a smaller buffer
+	for (int i = 0; i < 15; ++i)
+	{
+		b.reset(160);
+		b.reserve(160);
+		b.received(160);
+		b.normalize();
+		printf("capacity: %d watermark: %d\n", b.capacity(), b.watermark());
+	}
+
+	TEST_CHECK(b.capacity() <= start_capacity / 2);
+	printf("capacity: %d\n", b.capacity());
+}
+
+TORRENT_TEST(receive_buffer_max_receive)
+{
+	receive_buffer b;
+	b.reset(2000);
+	b.reserve(2000);
+	b.received(2000);
+	b.normalize();
+
+	b.reset(20);
+	int const max_receive = b.max_receive();
+	TEST_CHECK(max_receive >= 2000);
+	b.received(20);
+	TEST_EQUAL(b.max_receive(), max_receive - 20);
+}
+
 #if !defined(TORRENT_DISABLE_ENCRYPTION) && !defined(TORRENT_DISABLE_EXTENSIONS)
 
 TORRENT_TEST(recv_buffer_mutable_buffers)