diff --git a/docs/manual.rst b/docs/manual.rst
index c18ffa80d..e2240bd42 100644
--- a/docs/manual.rst
+++ b/docs/manual.rst
@@ -3375,6 +3375,8 @@ that will be sent to the tracker. The user-agent is a good way to identify your
 
 		int recv_socket_buffer_size;
 		int send_socket_buffer_size;
+
+		bool optimize_hashing_for_speed;
 	};
 
 ``user_agent`` this is the client identification to the tracker.
@@ -3729,6 +3731,17 @@ the OS default (i.e. don't change the buffer sizes). The socket buffer
 sizes are changed using setsockopt() with SOL_SOCKET/SO_RCVBUF and
 SO_SNDBUFFER.
 
+``optimize_hashing_for_speed`` chooses between two ways of reading back
+piece data from disk when its complete and needs to be verified against
+the piece hash. This happens if some blocks were flushed to the disk
+out of order. Everything that is flushed in order is hashed as it goes
+along. Optimizing for speed will allocate space to fit all the the
+remaingin, unhashed, part of the piece, reads the data into it in a single
+call and hashes it. This is the default. If ``optimizing_hashing_for_speed``
+is false, a single block will be allocated (16 kB), and the unhashed parts
+of the piece are read, one at a time, and hashed in this single block. This
+is appropriate on systems that are memory constrained.
+
 
 pe_settings
 ===========
diff --git a/examples/client_test.cpp b/examples/client_test.cpp
index 2b42b8252..9888b2009 100644
--- a/examples/client_test.cpp
+++ b/examples/client_test.cpp
@@ -714,6 +714,7 @@ int main(int argc, char* argv[])
 	settings.user_agent = "client_test/" LIBTORRENT_VERSION;
 	settings.auto_upload_slots_rate_based = true;
 	settings.announce_to_all_trackers = true;
+	settings.optimize_hashing_for_speed = false;
 
 	std::deque<std::string> events;
 
diff --git a/include/libtorrent/session_settings.hpp b/include/libtorrent/session_settings.hpp
index 3fa90cf5f..c6544020e 100644
--- a/include/libtorrent/session_settings.hpp
+++ b/include/libtorrent/session_settings.hpp
@@ -163,6 +163,7 @@ namespace libtorrent
 			, max_rejects(50)
 			, recv_socket_buffer_size(0)
 			, send_socket_buffer_size(0)
+			, optimize_hashing_for_speed(true)
 		{}
 
 		// this is the user agent that will be sent to the tracker
@@ -541,6 +542,11 @@ namespace libtorrent
 		// 0 means OS default
 		int recv_socket_buffer_size;
 		int send_socket_buffer_size;
+
+		// if this is set to false, the hashing will be
+		// optimized for memory usage instead of the
+		// number of read operations
+		bool optimize_hashing_for_speed;
 	};
 
 #ifndef TORRENT_DISABLE_DHT
diff --git a/src/storage.cpp b/src/storage.cpp
index f958e23dc..e6ed9cbac 100644
--- a/src/storage.cpp
+++ b/src/storage.cpp
@@ -488,19 +488,42 @@ namespace libtorrent
 			if (disk_pool()) block_size = disk_pool()->block_size();
 			int size = slot_size;
 			int num_blocks = (size + block_size - 1) / block_size;
-			file::iovec_t* bufs = TORRENT_ALLOCA(file::iovec_t, num_blocks);
-			for (int i = 0; i < num_blocks; ++i)
-			{
-				bufs[i].iov_base = disk_pool()->allocate_buffer("hash temp");
-				bufs[i].iov_len = (std::min)(block_size, size);
-				size -= bufs[i].iov_len;
-			}
-			readv(bufs, slot, ph.offset, num_blocks);
 
-			for (int i = 0; i < num_blocks; ++i)
+			// when we optimize for speed we allocate all the buffers we
+			// need for the rest of the piece, and read it all in one call
+			// and then hash it. When optimizing for memory usage, we read
+			// one block at a time and hash it. This ends up only using a
+			// single buffer
+			if (settings().optimize_hashing_for_speed)
 			{
-				ph.h.update((char const*)bufs[i].iov_base, bufs[i].iov_len);
-				disk_pool()->free_buffer((char*)bufs[i].iov_base);
+				file::iovec_t* bufs = TORRENT_ALLOCA(file::iovec_t, num_blocks);
+				for (int i = 0; i < num_blocks; ++i)
+				{
+					bufs[i].iov_base = disk_pool()->allocate_buffer("hash temp");
+					bufs[i].iov_len = (std::min)(block_size, size);
+					size -= bufs[i].iov_len;
+				}
+				readv(bufs, slot, ph.offset, num_blocks);
+
+				for (int i = 0; i < num_blocks; ++i)
+				{
+					ph.h.update((char const*)bufs[i].iov_base, bufs[i].iov_len);
+					disk_pool()->free_buffer((char*)bufs[i].iov_base);
+				}
+			}
+			else
+			{
+				file::iovec_t buf;
+				disk_buffer_holder holder(*disk_pool(), disk_pool()->allocate_buffer("hash temp"));
+				buf.iov_base = holder.get();
+				for (int i = 0; i < num_blocks; ++i)
+				{
+					buf.iov_len = (std::min)(block_size, size);
+					readv(&buf, slot, ph.offset, 1);
+					ph.h.update((char const*)buf.iov_base, buf.iov_len);
+					ph.offset += buf.iov_len;
+					size -= buf.iov_len;
+				}
 			}
 			if (error()) return sha1_hash(0);
 		}