From mboxrd@z Thu Jan  1 00:00:00 1970
From: "Michael A. Olson" <mao@sleepycat.com>
To: elix@sourceware.cygnus.com
Subject: EL/IX API levels, services
Date: Wed, 12 Jan 2000 16:21:00 -0000
Message-id: <200001130020.QAA39244@triplerock.olsons.net>
X-SW-Source: 2000-q1/msg00000.html

Hi folks,

As everyone emerges from their Y2K bunkers to survey the wreckage,
I thought I would reopen the API discussion.

My point of view is that of a database software vendor (Sleepycat
distributes Berkeley DB) who would like to port once, to EL/IX,
and not have to do ports for every embedded OS vendor.

The EL/IX spec characterizes the four levels of interfaces as:

	Level 1: RTOS compatible layer.
	Level 2: Linux single process only.
	Level 3: Linux multiprocess for embedded applications.
	Level 4: Full POSIX or Linux compliance.

When I think about deploying Berkeley DB on an embedded OS, I
characterize levels of service in an analogous way.  These are
basically the levels of service that applications that embed
Berkeley DB are looking for:

	Single-user, no transactions: Basically a dbm-alike.

	Single-user with transactions: Multiple changes can be
	grouped into single operations, system and app crashes
	don't lose data.

	Multi-user, no transactions: Multi-thread, maybe multi-
	process.  Dbm, except that multiple threads can read
	and write the database concurrently without trouble.

	Multi-user with transactions: Combine the last two.

Under the current draft, level 1 of the EL/IX spec is all
we need to deliver threaded, transaction-protected database
management.  If the application wants to use multiple
processes, rather than multiple threads, to operate on
a single database, then we jump to level 4.

The only thing missing from level 3 of the spec that makes
it impossible to do multi-process database services is shared
memory.  The shm_ interfaces show up in EL/IX level four.
Why is that?  It seems to me that the level 3 goal (Linux
multi-process) would be furthered by the availability of
named shared memory.  The high-end embedded vendors, like Wind
River and QNX, have good named shared memory abstractions
already.  I don't know about the embedded Linux vendors.

Also, I'm a little curious about the interface names.
Berkeley DB uses shmat()/shmdt()/shmget() now.  The
EL/IX spec calls for shm_open() and shm_unlink().  I
can imagine how those map to the calls we use now, but
I am not familiar with them.

My vote would be to move the named shared memory interfaces
into level 3, because my code would run on more devices that
way.  I don't know how hard the implementation would be for
the OS vendors, though.

The good news is that the draft spec would permit us to
port to EL/IX quite easily.  We've got customers on embedded
OSes who would be quite happy with the services we could
offer at level 2 of the current draft.

Cheers,
					mike