Deferred String Passing

The Problem

A message contains some fixed-length data and capabilities, and a variable-length string of bytes. A server receiving messages from clients must allocate a buffer to receive the largest string it expects. This has a number of disadvantages:

The Proposal

This proposal allows the transfer of the string to be deferred until the server has determined the right time and place for the transfer. It introduces a new capability type called an Indirect Data capability. Let's consider first the case of a CALL invocation from a client to a server.

Assuming the server is Available, the fixed-length portion of the message is transferred. If the client's passed string is longer than the server's receive buffer, and the server's receive operation requested this new functionality, then as much of the string as will fit is copied, and the receiver receives an Indirect Data capability in the third capability parameter (as well as a Resume key in the fourth parameter). (This functionality cannot be used if there are requests that send both long strings and three keys.) The server is told the length of the string that the sender attempted to pass.

An Indirect Data capability refers to a process that is sending a message. The sending process is in a new process state called Sending. The address and length of the string being sent are stored in registers or pseudoregisters of the process. The Indirect Data capability grants the authority to read that region of the process's address space.

At any later time, the server calls the Indirect Data capability to obtain the rest of the string. The Indirect Data capability has the following methods:

The server can read any part of the string in any order and can read the same data more than once. (However, if reading more than once, beware of time-of-check-to-time-of-use errors.)

The server uses the Release method to signal that it is done receiving the string. The Release method changes the sender's state from Sending to Waiting (because the sender executed a CALL invocation), and voids all Indirect Data capabilities to the sending process.

The server completes the request and returns to the caller using the Resume key in the normal way. It is an error to invoke the Resume key without having first Released the send string.

Details

If the client's invocation was a SEND or RETURN rather than a CALL, then when the Indirect Data capability is Released, the sender's state is changed to Running or Available as appropriate. This ensures that the sender can safely overwrite his send buffer after the invocation completes.

There will be a way to specify on an invocation that the passed string should be taken from an existing Indirect Data capability that the sender holds. (The recipient may or may not get the Indirect Data capability, depending on whether he receives the whole string.) This may be useful for services that forward or proxy requests.

This mechanism allows us to remove the limit on the length of a string that may be passed. (The limit on the length of a string that may be received in a single operation (by receiving a message or calling an Indirect Data capability) remains, because kernel operations are unpreemptible.)

Invoking an Indirect Data capability to copy data may cause the sender's segment keeper to be invoked, unless the invoker specified that the invocation must be prompt. (A server with multiple independent clients generally would require promptness.) This behavior means that an Indirect Data capability cannot be proxied with a start key, only another Indirect Data capability. This means that the Indirect Data capability needs its own Discrim type.

The possibility that a segment keeper on a passed string won't be invoked is a change in behavior, but that shouldn't be serious, since most segment keepers don't take action on read access.

If the sender specified a prompt invocation, the Indirect Data capability is not sent. This prevents the situation where the sender is delayed until the recipient issues a Release. This generally occurs when a server returns to a client; the client must be prepared to accept the entire string.

Performance

This proposal requires more code to handle a gate invocation. On the sender's side, no new code is executed unless the sender's string is longer than the receiver's buffer, a case which was formerly an error. When returning to a Resume key, there was already a sanity check of the client's process state. Therefore I don't think there will be any performance penalty except in the case where you are using the new mechanism.

Using this mechanism presents a tradeoff between copying a string an extra time (or more), and making an extra system call. We estimate the crossover point to be about 256 bytes on ARM, which leaves ample opportunity to benefit.


SourceForge.net Logo Copyright 2007 by Strawberry Development Group. All rights reserved. For terms of redistribution, see the GNU General Public License This material is based upon work supported by the US Defense Advanced Research Projects Agency under Contract No. W31P4Q-07-C-0070. Approved for public release, distribution unlimited.