*start*
02295 00024 US
Date: 28 Jan. 1981 4:35 pm PST (Wednesday)
From: MBrown.PA
Subject: file server meeting, Thursday 29 January, 1pm
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
We'll meet in the ISL conference room (to the left as you pass the ISL coffee
area).
I want to use this message to indicate some topics that it might be useful to
discuss at this meeting and its successors. Other suggestions are welcome.
1) A name for the system. The name "Sierra" was suggested, but not used, as
the name for the Cedar language. I think it would be a nice name for a file
server. It has the advantage that individual servers machines could be named
by Sierra peaks, lakes, rivers, or what-have-you.
2) The division of responsibility between the system and the Cedar universal file
system. For instance, one point of view says that a file server is not concerned
with problems of locating files; to access a file you must communicate with the
right machine and specify the right volume, as well as supply the right unique
ID. In this view, file location is handled via a higher-level mechanism, perhaps
some combination of location hints (Gifford's file references) and a database to
handle exceptions. I favor this view.
Providing a directory system for mapping string filenames to unique IDs leads to
a responsibility question with a similar flavor.
The division of responsibility issue will inevitably lead to a specification of the
file server's interface, which is a primary goal of these meetings. Offhand it
seems that the three capabilities the file server must offer to the UFS are (1)
given volume and file ID, transfer whole file, (2) given volume and file ID and
page number, transfer page, (3) provide transactions and a technique for
maintaining remote caches of files or pages located on the file server. There are
also other, less UFS-specific requirements such as backup, archiving, etc.
3) The internal organization of the system. Discussion of this should probably
be deferred until the external interface has been examined. A general design
goal that seems quite important is the ability to build a server with special
properties, e.g. one that also includes a directory system or a database system or
whatever.
--mark
*start*
00230 00024 US
Date: 29 Jan. 1981 3:07 pm PST (Thursday)
From: Kolling.PA
Subject: Butler's paper
To: MBrown
Reply-To: Kolling
cc: Kolling
Do you have a pointer to the latest copy of Butler's and Howard's paper?
Karen
*start*
00802 00024 US
Date: 29 Jan. 1981 3:54 pm PST (Thursday)
From: MBrown.PA
Subject: Re: Butler's paper
In-reply-to: Your message of 29 Jan. 1981 3:07 pm PST (Thursday)
To: Kolling
cc: MBrown
no, as far as I know there is no paper that explains the way the current
algorithm works. the root-leaf idea is explained in the may 29, 1979 version (i
think a copy of this is on butler's directory, or you can look at mine), but the
notion of "compatible actions" in that paper is completely wrong.
i have been working on mesa-like code for the algorithm; this is located on
[ivy]<MBrown>CRSSCode.bravo. If you like, you can have a go at this; I will be
glad to spend some time discussing it with you after you've looked at it. My
own ideas on the algorithm are not as sharp as I'd like!
--mark
*start*
00548 00024 US
Date: 11 Feb. 1981 4:08 pm PST (Wednesday)
From: MBrown.PA
Subject: sierra meeting, Thursday 12 February, 2:30pm, ISL conference room
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
(This meeting will actually be held whenever Popek's seminar is finished).
Proposed agenda: (1) discussion of questions raised in the minutes of our last
meeting, (2) (if time remains) I will try to present my current understanding of
the algorithm for coordinating distributed transaction commit.
--mark
*start*
00461 00024 US
Date: 16 Feb. 1981 5:27 pm PST (Monday)
From: MBrown.PA
Subject: memo on proposed BFS interface
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
[Ivy]<Sierra>Doc>BFSDesign.bravo, .press is a proposal for the BFS interface. The
design is incomplete in many ways, but the multiple file system transaction stuff
is worked out in considerable detail. This might be a focus for Thursday's
meeting.
--mark
*start*
00376 00024 US
Date: 17 Feb. 1981 9:14 am PST (Tuesday)
From: Levin.PA
Subject: A bas "BFS"
To: Boggs, Kolling, Taft, MBrown, Birrell, Schroeder
cc: Levin
Before things get much further advanced, can we find a different name than
BFS? I dislike acronyms, and this one is already attached to a distinct but
similar entity. That's too many black marks for me.
Roy
*start*
00361 00024 US
Date: 17 Feb. 1981 9:32 am PST (Tuesday)
From: MBrown.PA
Subject: Re: A bas "BFS"
In-reply-to: Levin's message of 17 Feb. 1981 9:14 am PST (Tuesday)
To: Levin
cc: Boggs, Kolling, Taft, MBrown, Birrell, Schroeder
I actually thought it was nice to reuse the same name. But I can appreciate the
danger. How about "FileStore"?
--mark
*start*
00363 00024 US
Date: 19 Feb. 1981 9:43 am PST (Thursday)
From: MBrown.PA
Subject: sierra meeting, today, 1:00pm, ISL conference room
To: Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
I will be prepared to discuss the transaction facilities of the FileStore (formerly
BFS) interface, if that turns out to be appropriate.
--mark
*start*
00961 00024 US
Date: 19 Feb. 1981 9:07 pm PST (Thursday)
From: MBrown.PA
Subject: Version 1 of FileStore
To: Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
[ivy]<Sierra>Doc>FileStoreDesign1.bravo is a revised interface. It should be
somewhat clearer: all of the handles are gone, all procedures take unique ids as
parameters. I have put copies of this memo in mailboxes.
I am working on an implementation of the transaction stuff; my current version
is [ivy]<Sierra>Doc>FileStoreTransCode1.bravo. It begins with a 3-page
introduction to logging, recoverable transaction states, and the recovery process.
Its quality is uneven, but the memo may be of some help in puzzling out the
interface.
We concluded at today's meeting that increased understanding of RPC will be
required sooner than we thought, since it may impact the design of FileStore.
Our next meeting is scheduled for 1:30 Monday at the usual place.
--mark
*start*
00502 00024 US
Date: 24 Feb. 1981 9:50 pm PST (Tuesday)
From: MBrown.PA
Subject: Lock manager "concepts" document
To: Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
[ivy]<Sierra>Doc>LockConcepts0.bravo is a document containing a distillation of a
number of papers that I have read on the topic of locking. I have put copies of
this memo in mailboxes. The memo promises a LockDesign0.bravo, containing a
proposed interface; this may not show up until Friday.
--mark
*start*
00409 00024 US
Date: 25 Feb. 1981 3:23 pm PST (Wednesday)
From: MBrown.PA
Subject: Alpine
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
Gloria Warner has approved the name change from Sierra to Alpine.
I will distribute hardcopies of the draft lock manager interface later today. As
soon as an <Alpine> directory is established it will be stored there.
--mark
*start*
00525 00024 US
Date: 12 March 1981 11:31 am PST (Thursday)
From: MBrown.PA
Subject: New version of FileStoreDesign
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
I have distributed hardcopies of FileStoreDesign2.bravo. This version makes a
few minor improvements in the Transaction section, but mainly is intended to
flesh out the Files section (though this is still far from complete.)
I am working on another memo, on buffering, that with luck will also appear
today.
--mark
*start*
01094 00024 US
Date: 16 March 1981 9:17 am PST (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30pm, CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
Last week we had a preliminary discussion of RPC issues outside of the
remote-binding area. We still need to answer questions like: (1) what is inside
the box labeled "RPCImpl"? Do we expect to use stubs for most parameter and
result munging in remote procedures? (2) how does a caller who expects a large,
unstructured result (say a page) get access to the PupBuffer that holds it? (3)
what are the restrictions on the size of argument and result records? must they
fit into a single gateway-sized packet? If so, how to handle primitives that
transfer runs of pages? (4) what is the relationship between RPC and a fast
file-transfer protocol?
We also discussed the topic of buffering, especially relating to fast file transfers.
Since that meeting, two memos have appeared, dealing with internal file server
buffering and with the "files" section of FileStore.
--mark
*start*
02662 00024 US
Date: 17 March 1981 8:45 am PST (Tuesday)
From: MBrown.PA
Subject: McJones on the performance of disk commands on Dolphin
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
cc: MBrown
---------------------------
Date: 16 March 1981 4:24 pm PST (Monday)
From: MBrown.PA
Subject: performance of disk commands on Dolphin
To: McJones
Reply-To: MBrown
cc: MBrown
We had a minor controversy in today's Alpine meeting about the motivation for
runs of pages at the disk face (the motivation at higher levels is pretty clear.) I
claimed that if transfers to consecutive sectors are expressed as separate disk
commands on the Dolphin, a revolution of the disk is lost between the two
transfers. Is this true? Or is the problem in the software overhead of setting up
the second disk command, the storage for the disk command, or whatever?
--mark
------------------------------------------------------------
Date: 16 March 1981 6:38 pm PST (Monday)
From: McJones
Subject: Re: performance of disk commands on Dolphin
In-reply-to: Your message of 16 March 1981 4:24 pm PST (Monday)
To: MBrown
cc: McJones
The problem is keeping up with the disk. There are something like 700
microseconds between sectors, and with the disk task taking up many of the
memory cycles we found that with double buffering even stripped down
programs couldn't compute the next IOCB fast enough to keep up with the disk.
We could have computed a whole track's worth ahead of time, but that would
have taken a lot of space, and by the way a fair amount of computing time. (Of
course this computing time would often occur during a seek, and we might or
might not have another user process to run at the same time...)
Since the necessary computing is simple and regular, it turns out to be fairly
straightforward to do in microcode.
/Paul
------------------------------------------------------------
I take this to say that the microcode CAN interpret a new command in the
inter-sector time, but the software CANNOT construct a new command in a sector
time. Thus lots of short commands are definitely a loss, but it is ok to issue two
medium-sized commands for consecutive runs and let the disk controller run
them one after the other. This eases the contiguity requirements for disk buffers
somewhat.
We will eventually need more details on this point, in order to estimate whether
the standard run size should be two pages, four pages, or whatever. I am a little
surprised at the comment about the disk's interference with the emulator; I
thought that the Dolphin had plenty of memory bandwidth. I wonder what
impact the display has on this.
--mark
*start*
00653 00024 US
Date: 18 March 1981 5:13 pm PST (Wednesday)
From: MBrown.PA
Subject: Outline of a posible FileStore implementation
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
I have distributed hardcopies of a memo entitled "Outline of a posible FileStore
implementation." This may serve to clarify some of the constraints that I
(unconsciously) placed on the FSBuffer interface design. Roy suggested that it
may be helpful for me to give an oral presentation from this outline on Monday;
it might then be useful to attempt to work out alternative design proposals to a
similar degree of detail.
--mark
*start*
00416 00024 US
Date: 23 March 1981 10:35 am PST (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30pm, CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
I can talk about the "Outline of a possible FileStore implementation." We might
also revive the discussion of RPC issues left over from two weeks ago (as
outlined in my message of 16 March.)
--mark
*start*
00626 00024 US
Date: 30 March 1981 8:51 am PST (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30 (or after picnic), CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
Andrew and I were otherwise occupied last week, and no progress has been made
on the universal file system (cache manager) interface.
We might continue our discussion of RPC parameter functionality issues, such as
the interface from stub procedures to RPC pup buffers (for large unstructured
parameters or results). We might also discuss where to put in the provisions for
encryption.
--mark
*start*
00739 00024 US
Date: 31 March 1981 4:17 pm PST (Tuesday)
From: MBrown.PA
Subject: DBMS recovery memo
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
Reply-To: MBrown
cc: MBrown
I have distributed hardcopies of a draft memo entitled "Recovery in database
management systems." It explores the possible relationships between a file system
that provides transactions on files and a database system that provides
transactions on tuples (or whatever.) Naturally I have a particular interest in
seeing that Alpine is capable of participating as a useful partner in such a
relationship.
The memo is somewhat incomplete (I use < ... > to delimit unfinished sections)
but I think that the essential points come across ok.
--mark
*start*
00302 00024 US
Date: 6 April 1981 10:01 am PST (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30, CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
cc: MBrown
We can discuss the memo "Recovery in database management systems" that was
distributed last week.
--mark
*start*
00334 00024 US
Date: 20 April 1981 8:46 am PST (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30, CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
cc: MBrown
Maybe we'll discuss "file properties" such as creation date.
I would prefer to start on time since I must be at Stanford at 3pm.
--mark
*start*
00830 00024 US
Date: 27 April 1981 10:51 am PDT (Monday)
From: MBrown.PA
Subject: Alpine meeting today, 1:30, CSL alcove.
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
cc: MBrown
As an outgrowth of last week's discussion of file properties, Ed Taft agreed to
describe how archive and backup work on our present file systems. (These
processes seemed to be the only ones that cared about dates other than create.)
This might lead to a discussion of where archiving fits into the Alpine design.
Next Monday is in May, hence we should be starting to work on Alpine again
in earnest. It is time for those who want to make specific design proposals to
take responsibility for writing something down, and then do it. This meeting
seems like a good time to see who is going to do what, and roughly by when.
--mark
*start*
00976 00024 US
Date: 8 May 1981 6:44 pm PDT (Friday)
From: kolling.PA
Subject: question
To: MBrown
cc: Kolling
There's a paragraph in your dbms memo that says:
"If all updates to an object are undo-logged, and log record IDs are used as
update IDs, then the combination of object plus log has a very simple structure:
the object points to the most recent update, which points to the previous
(committed) update, etc. The entire history of the object can be found by
traversing the log in this way. Because of locking, only the most recent update
is possibly uncommitted and hence candidate for undoing. Undo pops the top
element from the stack of log records."
Actually, if the owner of the lock has been doing multiple writes to the same
object, the "previous" update of interest is not necessarily committed, yes? You
have to search back to find the most recently committed one, I think.....or do I
not understand when things get written into this log.....
*start*
00338 00024 US
Date: 9 May 1981 3:49 pm PDT (Saturday)
From: MBrown.PA
Subject: Re: question
In-reply-to: Your message of 8 May 1981 6:44 pm PDT (Friday)
To: kolling
cc: MBrown
yes, you are right. but it would not be too hard to do it the other way, by
having some (volatile) record of the previous committed update.
--mark
*start*
00385 00024 US
Date: 11 May 1981 9:11 am PDT (Monday)
From: MBrown.PA
Subject: No Alpine meeting today
To: Boggs, Kolling, Taft, Birrell, Levin, Schroeder
cc: MBrown
The subgroup that is to work on file system (as opposed to RPC) issues in depth
has not made any progress yet. If we make significant progress before Friday,
we'll try to call a meeting for Friday.
--mark
*start*
00386 00024 US
Date: 15 May 1981 11:52 am PDT (Friday)
From: Kolling.PA
Subject: name
To: mbrown
cc: Kolling
How about calling Get/SetLength Get/SetByteLength? Otherwise I will be
forever confusing them with the (different functioning) Get/SetLengths from
Juniper. Or maybe Get/SetSize should be called something that makes it easier to
not confuse with Get/SetLength.....
*start*
00977 00024 US
Date: 15 May 1981 12:02 pm PDT (Friday)
From: Kolling.PA
Subject: name
To: mbrown
cc: Kolling
I ran across your memo "Pilot development that would benefit alpine" when
browsing thru alpine>doc. About the error logging:
(1) I discussed this with Ed Taft et al at length when I upgraded Juniper's error
handling. The belief is that soft errors are due to transient glitches, dust, etc.,
and that they aren't an indication that the page is any more likely to go bad
than any other page.
(2) However, a high frequency of soft errors on many pages indicates is that the
drive and pack are not aligned; it is important to catch that and have
maintenance adjust the drive. Misalignment can even cause what look like hard
errors, but which aren't really.
(3) Generally speaking the number of hard bad pages on a T300 is no more than
3 or so ever, barring a major catastrophe of some sort. Pages do not in general
become hard bad during system use.
*start*
00646 00024 US
Date: 15 May 1981 12:10 pm PDT (Friday)
From: Kolling.PA
Subject: more info
To: mbrown
cc: Kolling
What Juniper does is log the hard and soft errors. This is most obvious to Ron
Weaver during the every other day Copydisk, as that sweeps each whole pack,
and he watches for it.
Juniper will replace a page that goes hard bad with a new page. However, it
deliberately takes the system down if it has seen more than ten new hard bad
pages since it came up, on the theory that maintenance should be called. This
count can be overidden if necessary. There is also a similar limit on the total
number of hard bad pages.
*start*
00363 00024 US
Date: 15 May 1981 12:21 pm PDT (Friday)
From: MBrown.PA
Subject: Re: name
In-reply-to: Your message of 15 May 1981 11:52 am PDT (Friday)
To: Kolling
cc: mbrown
well, size is standard usage from pilot, so i'd rather not change it. but saying
"byte" as part of "setlength" seems fine. watch for it in FileStoreDesign3.bravo ...
--mark
*start*
00501 00024 US
Date: 19 May 1981 6:04 pm PDT (Tuesday)
From: kolling.PA
Subject: ROUGH draft
To: taft
cc: mbrown, kolling
Ed, there is an EXTREMELY ROUGH draft of what we talked about on
[Juniper]<Kolling>Alpine>FilePageMgr.bravo. It is not fit to be taken seriously
yet, but since I will be on vacation until Friday, I thought I would give you a
pointer to it in case you wanted it. Feel free to either red pencil all over it or
just ignore it and I will work on it again Friday.
Karen
*start*
01608 00024 US
Date: 24 May 1981 7:39 pm PDT (Sunday)
From: MBrown.PA
Subject: FilePageMgr
To: Kolling, Taft
cc: MBrown
I just realized a basic problem with using mapped files to implement
FilePageMgr. The problem is that if files are mapped in "chunks" of 4 pages
(say), a 1 page write at the FilePageMgr interface must result in a 4 page write
to the underlying file. But this write is not guaranteed to be atomic: it may fail
part way through. If it fails while writing the page that the client actually
wrote, then no problem: the log contains the necessary information. But if it
fails while writing a page that is being written redundantly, then the only
recovery is from a backup copy of the file. This is not acceptable.
Of course, the contents of the entire chunk could be logged. This seems like a
very poor idea, both due to the volume of wasted logging and due to the lack of
separation between the log manager and the file page manager (log manager now
needs to be concerned with chunk alignment when writing, needs to predict
when the first page of a chunk is written whether the rest will be, etc.)
I conclude that we must restrict ourselves to single page swap units in the
implementation of FilePageMgr, if mapped files are used. Since reading/writing
runs is important for good performance, this is not a long-term solution, and
some notion of "chunk" is probably still necessary in the interface. In the long
run we'll have to find some way of getting Pilot to be more clever about doing
only the I/O that we request, and still optimizing runs when possible.
--mark
*start*
01349 00024 US
Date: 26 May 1981 7:21 pm PDT (Tuesday)
From: Taft.PA
Subject: Pilot implementation notes
To: MBrown
cc: Kolling, Taft
Your notes are consistent with everything I know (which isn't much). I have
the following amplifications and additions:
1. Yes, SimpleSpace.AllocateVM is used only during initialization, to set up the
permanent spaces (Space.virtualMemory, Space.mds, etc.) The "real" VM allocator
is buried inside Space.Crate (in SpaceImplA).
2. Pinning a region definitely prevents it from being written out. Note that the
unit of pinning is an entire region, not just a swap unit. (The semantics of this
are rather curious. Calling Apply with operation=pin swaps in and pins a single
specified swap unit, and also pins any other swap units in the region that
happen already to be in, but does nothing to swap units that are out. However,
SpecialSpace.MakeResident operates on an entire space and performs the pin
operation on all swap units in the space, so the extra precision available from
Apply is apparently inaccessible.)
3. If any page of a swap unit is dirty, the entire swap unit is written out.
4. A space that has swap units is never dead! I interpret this to mean that
Space.Kill is a no-op when applied to a space with swap units. (See the
comments at the beginning of CachedRegion.mesa.)
Ed
*start*
01411 00024 US
Date: 27 May 1981 10:09 am PDT (Wednesday)
From: Taft.PA
Subject: Pilot implementation notes
To: MBrown
cc: Kolling, Taft
The funny staggering of swap units you suggested for dealing with the
existence of leader pages is not necessary. It's perfectly ok to map a one-page
space into page zero of a file, and another (larger) space starting at page one.
Suppose the second space has uniform swap units of 4; then the swap units will
map to pages 1-4, 5-8, etc. (That is, there are no alignment restrictions.)
The restriction on changing the length of a mapped file is real: there is simply
no communication between the FileMgr and the VMMgr for maintaining
consistency in this case.
But this is not as onerous as it might seem, at least for extending a file. The
restriction is that there may not be a mapped space extending beyond the old or
new end of file. If we always create spaces of the right size, so as not to extend
beyond end of file, then we are not violating this restriction. (After extending a
file, we can of course create new space(s) to map onto the extension.)
Shortening a file is somewhat more problematical. But I think we can simply
outlaw the illegal case. That is, shortening a file causes all spaces mapped onto
(or overlapping) that part of the file to be deleted; and if there are any
outstanding references to those spaces, an error results.
Ed
*start*
01901 00024 US
Date: 27 May 1981 2:31 pm PDT (Wednesday)
From: Taft.PA
Subject: Pilot disk scheduler
To: MBrown, Kolling
---------------------------
Date: 27 May 1981 1:01 pm PDT (Wednesday)
From: Taft.PA
Subject: Disk scheduler removal
To: Luniewski
cc: Taft
I see that the disk scheduler was removed from DiskDriverSharedImpl and
replaced by a FIFO regime; your name is attached to this change. Was the
scheduler removed because it didn't work, or was it simply to save resident
memory on the Dandelion?
(I'm working on the design of a high-performance Pilot-based file server, and
I'm very concerned that absence of a disk scheduler will ruin performance.)
Ed
---------------------------
Date: 27 May 1981 1:49 pm PDT (Wednesday)
From: Luniewski.PA
Subject: Re: Disk scheduler removal
In-reply-to: Your message of 27 May 1981 1:01 pm PDT (Wednesday)
To: Taft
cc: Luniewski
The old elevator algorithm was removed for two reasons. First, removing it
saved a noticable amount of memory which was very important for meeting our
Rubicon Pilot working set goals. Second, Bob Ladner's measurements had
indicated that the algorithm probably was not buying us anything - the speed of
the disk and the compute time involved in handling a page fault were such that
there was little idle time while a page was transferring. That is, if a client
process ran while a disk transfer was in progress, Pilot would not complete the
computation involved in handling the page fault until after the first transfer was
complete. Thus, there tended to be only one disk request outstanding at any one
time. Whether this argument applies on the Dorado, which I presume your file
server will run on, is an open question as far as I know.
(To the best of my knowledge, the code that was there did work before I
commented it out.)
/Allen
------------------------------------------------------------
*start*
00757 00024 US
Date: 28 May 1981 12:15 pm PDT (Thursday)
From: Kolling.PA
Subject: comments on memo
To: Kolling
Theses are some notes I gave to mark:
Do we want to mention the high water mark in the implementation section?
About the pin/unpin calls, if the whole chunk is pinned/unpinned, is there any
guarantee that the writes of the pages in that chunk happen in the correct order
after unpinning?
What about the size of chunks (i.e., fixed or variable), is that decided or punted
for now? It should perhaps be mentioned as an issue.
About the one page USU causing demand paging to happen one page at a time:
are both ReadAhead and ReadPages going to do an Activate? So demand paging
will only happen if Pilot has swapped some pages out?
*start*
00731 00024 US
Date: 11 June 1981 5:41 pm PDT (Thursday)
From: kolling.PA
Subject: two things
To: MBrown
cc: Kolling
I've forgotten how locks work if the same transaction requests more than one lock
on the same object. Will a write lock block a read lock? Will it block other
write locks? Is there some other type of lock (exclusive?) that will block all
other requests, including those from the same transaction? Otherwise I have to
do a few special case checks to prevent a given transaction from making life
difficult (allocation tries after remove, etc.) This is not hard, I just want to
know if I should note it in the memo.
Don't forget: your section of the Activity Report is due in about 5 days.
Karen
*start*
01046 00024 US
Date: 12 June 1981 9:06 am PDT (Friday)
From: MBrown.PA
Subject: Re: two things
In-reply-to: Your message of 11 June 1981 5:41 pm PDT (Thursday)
To: kolling
cc: MBrown
If t requests a Read lock on e, is granted, then requests a Write lock on e, the
Read lock is converted to a Write lock. If t then requests a Write lock on e, it
will be granted immediately (since it already has it.) If this isn't what you
want, you'll have to special-case.
As you are discovering, the lock manager is not the solution to all
synchronization problems; it is not unusual for clients of the lock manager to
record the locks they hold in their own data structures, too, both to handle the
sort of situation you just mentioned and to avoid lots of re-requests. E.g. on
reading the first page of a file, we'll probably have to get some sort of lock on
the whole file; thereafter we can just check in some data structure associated
with the file to make sure we have a strong enough lock for the page access we
are about to do.
--mark
*start*
00481 00024 US
Date: 30 June 1981 2:58 pm PDT (Tuesday)
From: MBrown.PA
Subject: Alpine / RPC meeting
To: Birrell, Boggs, Levin, Kolling, Nelson, Schroeder, Taft
cc: MBrown
Lets have a meeting tomorrow, after dealer, to share status on RPC and Alpine. I
will sketch the Alpine design choices that Karen, Ed, and I have made since we
last held a large-group meeting. Andrew/David/Bruce will describe their
current thinking on RPC aspiration level and design.
--mark
*start*
05956 00024 US
Date: 30 June 1981 4:10 pm PDT (Tuesday)
From: Nelson.PA
Subject: RPC advance info
To: Taft, Kolling
cc: Birrell
Ed and Karen,
To speed up the meeting tomorrow, I suggest you read through the following
memo on how we propose to handle parameters. Mark has suggested that Ed, in
particular, may have some objections.
If he feels up to it, Andrew will also send you two a companion memo on how a
typical remote call is handled. This should save even more hand waving
tomorrow.
Bruce
-- Messages from file: [PARC-MAXC]<NELSON>RPC.;2
-- TUESDAY, JUNE 30, 1981 16:09:12-PDT --
Date: 29 June 1981 4:08 pm PDT (Monday)
Sender: Nelson.PA
From: Nelson, Birrell.pa
Subject: Parameter Passing in Cedar RPC
To: Satterthwaite, Morris, Horning, Rovner, Levin, Mitchell, MBrown
cc: Nelson, Birrell, Boggs
This memo describes a language change designed to allow Cedar RPC to support
reasonably flexible parameter passing. The details of the modification may, of
course, change when reviewed by the Cedar language committee. This memo is
circulated now because, lacking a better idea, the absence of the modification
will be a bottleneck in 4-6 weeks. Fairly prompt action on this proposal is
necessary.
We have been designing a remote procedure call mechanism for Cedar. Our
emphasis is not on parameter functionality, but even the bare minimum we find
acceptable requires a slight change to Mesa. The area that causes problems is
when a client wants to pass large amounts of data. The typical client we're
thinking about is Alpine, passing the contents of runs of disk pages. The two
procedures that are interesting to consider are "ReadPages" and "WritePages".
Both deal in runs of pages. Mark (reasonably) wants to have the data delivered
by ReadPages end up in storage that Alpine provided (so that he can arrange
contiguity, etc). That, and consideration of the efficiency of large
argument/result records, makes it unreasonable for his data to be passed by value
in the "obvious" way, such as having the declaration:
ReadPages: PROC[file: FileID, startPage: PageID]
RETURNS[buffer: DESCRIPTOR FOR ARRAY OF WORD]
If that was the only way he could pass the data, Mark would edit the
automatically generated stubs to do something different. The same is true if we
allowed him to return a pointer/reference (and we allocated collectable storage
int the user stub). We believe that it's not reasonable to build a large system if
you need to hand-edit the stubs, and in any case it seems unreasonable that our
first client should find the RPC facilities inadequate.
In a local interface, the "natural" way for ReadPages to be declared would be:
ReadPages: PROC [file: File, startPage: PageID,
buffer: DESCRIPTOR FOR ARRAY OF WORD]
In the RPC world, we could implement the effect wanted by allocating storage
for "buffer" in the server's RPC stub, calling the server's ReadPages procedure,
then copying the data back into the client machine's "buffer" storage.
Analogously in the WritePages case, we'd copy the data from the user's storage
into the server at call time. VAR parameters have some of the properties we'd
want, but they would force us to copy at call time and at return time, whether
or not it's needed. We can achieve the control we want by having qualifiers
"VALUE" and "RESULT" for address-containing argument types. (We might
need different keywords, though.) For example, in ReadPages, the buffer might
be declared
buffer: RESULT DESCRIPTOR FOR ARRAY OF WORD;
for WritePages, it would be
buffer: VALUE DESCRIPTOR FOR ARRAY OF WORD.
For ReadPages, the RESULT specification forces call-by-result (as in AlgolW).
Thus the client stub sends only the length of buffer as an argument, NOT the
array itself. The server stub allocates this much storage, calls the real ReadPages
to fill it, and sends back the entire buffer contents as a result. The client stub
copies this result buffer into the client caller's buffer. Thus RESULT means do
very little on the call, but copy back on the return.
For WritePages, the VALUE specification forces call-by-value of the array as
well as the descriptor. The client stub sends the entire buffer and the server
stub allocates space for it and receives the entire buffer. This buffer is passed
directy to the real WritePages. There are no results in this case.
If a parameter says "VALUE RESULT", we copy the data in both directions.
The fourth combination, when we copy the data in neither direction, corresponds
to the use of the address as a "handle" (which is outside the safe language, but
still useful). You can also achieve the effect of "handles" by using exported
types, which is perfectly safe and doesn't cause RPC any problems.
We need to consider the meaning when we have neither VALUE nor RESULT.
We propose that that case means "VALUE RESULT". The argument for this
default is mainly because that's in the safe language, whereas the "handle"
semantics are unsafe.
The exact details of syntax do not concern us. What we need, as RPC
implementors, is a symbol table indication, available through RTTypes, of
whether a parameter record component is <value>, <result>, or both, or neither.
Right now, we are quite willing to restrict this info to the top level, as in
AlgolW. We intend that VALUE and RESULT can be used to qualify any
address-containing argument, including REFs, (LONG) POINTERs,
(LONG) DESCRIPTORs, and (LONG) STRINGs.
We imagine that the compiler would ignore VALUE and RESULT after pass
three, at least for the time being. They would serve only as pragmats for RPC.
Your comments and suggestions are welcomed. Please bear in mind that Cedar
RPC is a modest project with emphasis on performance, rather than functionality,
goals. The very pragmatic decision to use stubs is our opening declaration to
this effect. No realistic comments will go unconsidered.
Bruce and Andrew
*start*
08146 00024 US
Date: 1 July 1981 11:19 am PDT (Wednesday)
From: Birrell.pa
Subject: Background reading
To: Levin, Kolling, Schroeder, Taft, MBrown
cc: Nelson, Boggs
This is a highly informal description of how we think simple RPC's go. It might
help if you read it before tomorrow.
Andrew
---------------------------
Date: 29 June 1981 2:03 pm PDT (Monday)
From: Birrell.pa
Subject: Simple RPC's
To: Boggs, Nelson
cc: , Birrell
This is intended to describe our present beliefs about how we should implement
the fast cases of RPC. I think I haven't introduced anything we haven't
discussed, but I might have. It's a long message.
The simplest and fastest special case of RPC is where the call and result will
each fit into a single packet, and where the user machine is on the same net as
the server machine. I believe that everyone is now convinced that the correct
way to build RPC is with "stubs". We believe the stubs should be automatically
generated from the DEF's file, producing Mesa source which is compiled into a
user-stub BCD and a server-stub BCD. The stubs are responsible primarily for
packing and unpacking arguments and results into packet buffers. Separate
software, the RPC-runtime package, takes these packets and implements the
network protocol. The network protocol provides once-only call semantics in the
absence of machine crashes.
The sequence of events in the normal case of the simple case is as follows.
The user calls a procedure imported from the appropriate interface. The interface
was exported by the user-stub module, so this calls the appropriate procedure in
the stub. The user-stub procedure allocates a packet buffer (in its local frame),
and marshals the arguments into it. At bind-time, the user-stub obtained the
network address and MDS base of the corresponding server-stub and an "index"
hint (used by the server to validate procedure descriptors); [network-address,
MDS, procedure descriptor, index] represents the remote procedure we want to
call, and the user-stub places that in the packet.
The user-stub then calls the RPC-User-runtime to transmit the packet. The
runtime allocates a new serial number and places it in the packet, and places the
PSB of the current process in the packet, and encapsulates the packet as an
ethernet PUP packet, with PUP checksum. This procedure of the runtime has a
condition variable in its local frame, which it registers in an array of condition
variables (indexed by PSB) known to the microcode (see the handling of
results). It then passes the packet to the microcode for transmission. This process
then sits in the RPC-User-runtime, WAITing on that condition variable,
expecting the result packet. If the result doesn't come, it retransmits the packet
(marked to say that an ACK is wanted); eventually the call can be timed out. If
an ACK comes (with correct serial number, of course), this process moves to a
state where it's waiting on the condition variable expecting the result, but not
retransmitting (still with possible timeout).
When an RPC-call packet comes in to a server machine, the microcode recognizes
it and handles it specially. In each MDS, there is a fixed conditions variable,
"rpcCond", known to the microcode. There is a stock of RPC-server processes,
whose idle state is waiting on rpcCond. The microcode places the packet in a
known place and notifies rpcCond (in the appropriate MDS). One process
waiting on rpcCond wakes up and takes the packet and checks the checksum. It
takes the source host address and serial number, and looks them in a table. This
table contains the last serial number received from that [host,PSB], so the
RPC-server process can eliminate duplicate packets. If the packet requests an
Ack, it sends one. (The packet may also ack an earlier result packet - see later).
The RPC-server process takes the [procedure descriptor, index] from the packet
and indexes a table to verify that this procedure descriptor is reasonable to be
called from other machines. (Otherwise we never have any hope of offering
secure servers.) Finally the RPC-server process calls the given procedure
descriptor, giving it the packet.
This call get us into the appropriate server-stub procedure. This is in the
server-stub module, and imports the appropriate server-procedure. The
server-stub unmarshals the parameters (into a state vector, probably, doing the
right things with long argument records), frees the packet, then calls the
server-procedure.
The server procedure does its thing, then returns to the server-stub.
The server-stub allocates a result packet buffer (in its local frame) and marshals
the results into it. It then transfers to its caller (the RPC-server idle-loop
procedure) without freeing its own frame (using TRANSFER-WITH-STATE).
The RPC-server takes the packet, puts the caller's serial number and PSB into it,
encapsulates it, and gets the microcode to transmit it. Before doing so, it registers
a local condition variable in a "Waiting-for-ack" array, together with the calling
[host,PSB]. After sending the result packet, the RPC-server process WAITs on its
condition variable, hoping for an acknowledgement. After a timeout, the
RPC-server retransmits the result (marked to say that an ACK is wanted). An
acknowledgemnt can happen in two ways: after retransmission of the result, the
caller may send an explicit ACK; alternatively, the caller may start a new call,
which counts as an implicit ack. To notice these, the microcode treats a
result-ACK the same as a call packet: if there is a process waiting for an ack
from the originating [host,PSB], the microcode notifies that process instead of the
idle ones on rpcCond. When the rpc-server process receives its
acknowledgement, it frees the server-stub's frame (and therefore the packet
buffer), de-registers its condition variable, then goes idle by waiting on rpcCond
(but if the ack was implicit by an incoming call packet, it takes that packet and
works on the new call).
When the result packet comes in to the caller's machine, the microcode looks at
the array of waiting caller PSB's. If one of them is waiting for this result (as
indicated by the [PSB] in the result packet, it notifies that process. If no one is
waiting, it notifies a fixed process (this happens if result packets get duplicated
or retransmitted, or if result acks get lost).
If the fixed process gets a packet, it sends an ack if requested, then frees the
packet.
When the caller process (which was WAITing in rpc-user-runtime) gets a
packet, it checks the checksum, sends an ACK if requested, and checks the serial
number (for old duplicate packets). When it gets the correct result packet, it
returns that packet to the user-stub.
The user-stub unmarshals the results, frees the packet, and returns to the user!
For the "normal" case, this uses four procedure calls and two process switches.
There is one packet in each direction. There is an additional packet and an
additional process switch if we need explicit acking. Buffers are allocated from
local frames. The data structures used are as follows.
user-stub: [host, PSB, proc.desc., index] for each procedure
user-runtime: last-serial-number
ARRAY PSB OF CONDITION (for outstanding calls)
process for handling stray packets
server-runtime: array of [host, PSB, serial-number] (= connections)
rpcCond, and queue of idle rpc-server processes
array of acceptable procedure descriptors
array of CONDITION, for un-acked results
server-stub: none.
The only extra complexity for the non-local network case is getting the routing
info into outgoing packets.
The multi-packet case needs more protocol design.
The above can be used for single-machine inter-MDS calls, by short-circuiting
at packet transmit/receive time. In that case, a packet can always be assumed to
have been acked.
Chuck Thacker has gently suggested that I give a Dealer on the progress of
RPC. Next week sounds like a good idea, and this week is possible if we agree
about the above.
Andrew
------------------------------------------------------------
*start*
00266 00024 US
Date: 7 July 1981 4:46 pm PDT (Tuesday)
From: kolling.PA
Subject: My kingdom for a JFFO.
To: Taft
cc: kolling
Do you have any idea what the maximum number of groups is that anyone
(rational) has specified for an access field on Ivy?
Karen
*start*
00198 00024 US
Date: 7 July 1981 5:06 pm PDT (Tuesday)
From: kolling.PA
Subject: Also,
To: Taft
cc: kolling
Are 62 groups being defined enough for Ivy, or do you really need more?
Karen
*start*
01133 00024 US
Date: 13 July 1981 4:45 pm PDT (Monday)
From: kolling.PA
Subject: defaults
To: MBrown, Taft
cc: Kolling
How important do you think it is that each Owner have associated with it a
(user-settable) set of defaults to use for the file access lists on file creation, as
opposed to just defaulting to:
file read: owner and Alpine wheels
file modify: owner
file read and set these file access lists: owner and Alpine wheels
Presumably this owner-customized default would only be of use to ftp/chat-type
clients, since the UFS and db managers would not use Alpine defaults (I guess)
but rather maintain their own. The reason for omitting it is that doing so
removes all connection between the owner database and the file properties
database, and if it is included then two places (the file properties and also the
owner data routines) have to know about stuff like what the irreducible
minimum is that a file access list can be set to (ex: for read and set access lists
always the owner and Alpine wheels, for modify no one, etc.). I am inclined to
omit the owner-defaults in the name of simplicity.
Karen
*start*
00264 00024 US
Date: 13 July 1981 5:45 pm PDT (Monday)
From: MBrown.PA
Subject: Re: defaults
In-reply-to: kolling's message of 13 July 1981 4:45 pm PDT (Monday)
To: kolling
cc: MBrown, Taft
Yes, but I would default the file read list to World.
--mark
*start*
00952 00024 US
Date: 13 July 1981 7:28 pm PDT (Monday)
From: Taft.PA
Subject: Re: defaults
In-reply-to: kolling's message of 13 July 1981 4:45 pm PDT (Monday)
To: kolling
cc: MBrown, Taft
I think it's essential that defaults be settable on a per-owner (or per-something)
basis. Using the "owner" = "directory" analogy, I should point out that the
ability to set defaults on a per-directory basis in IFS and Tenex is heavily used
and extremely important. (Indeed, proper setting of the defaults is the reason
that most users never need to directly manipulate protections at all -- a good
thing, since most users don't understand how protections work anyway.)
In any event, I don't see why there has to be any explicit connection between
the Owner and FileProperties data basis, other than the ability for the Owner
data base to be able to store uninterpreted entries whose contents are defined by
the code implementing FileProperties.
Ed
*start*
01544 00024 US
Date: 14 July 1981 9:18 am PDT (Tuesday)
From: MBrown.PA
Subject: Re: defaults
In-reply-to: Taft's message of 13 July 1981 7:28 pm PDT (Monday)
To: Taft
cc: kolling, MBrown
Your point about defaults seems correct, but may not imply that Alpine should
implement the client-settable defaults itself.
For me, a telling consideration is that an owner-based access control default is
not logically associated with a particular FileStore, but is logically associated with
the owner and hence updates to it should probably be coordinated across all
FileStores. It is only a matter of time until the distributed directory system, file
replication machinery, and other pieces of a universal file system come along
and want to implement their on access control policies, and Alpine files will all
have access control lists granting the "UFS", and no real individuals, access. I
would like to avoid putting much effort into something that we hope to make
obsolete so quickly.
I am in favor of low aspirations for Alpine's internal databases in the near term.
This means the simplest possible owner database (getting control of space
allocation is essential) plus storing essential file properties (including some small
number of RNames for access control) in the leader page(s). In the longer term,
we should have (1) a more viable Cedar database facility, and (2) support from
Pilot for a variable number of "file property" pages; either or both of these could
change the cost of doing something more ambitious.
--mark
*start*
00295 00024 US
Date: 15 July 1981 3:24 pm PDT (Wednesday)
From: kolling.PA
Subject: Access control
To: MBrown, Taft
cc: kolling
I'm about to leave a new version of the access control memo in your mailboxes.
This is what I will be coding from, unless I hear screeches of protest.
KK
*start*
04074 00024 US
Date: 16 July 1981 5:33 pm PDT (Thursday)
From: kolling.PA
Subject: comments please
To: taft, mbrown
cc: kolling
Two topics for your comments, please, or maybe we should have a short
discussion:
(1) My model of "FileStore", FileProperties, and AccessControl has been basically
as follows, where by "FileStore" I mean some upper layer of the FileStore
implementation. Two factors I considered here were isolating the FileProperties
handling so it can easily be changed when Pilot file properties really exist, and
isolating the knowledge of the structure of "access lists" in AccessControl itself:
FileProperties: this simple module's function in life is to read and write file
properties. I don't think it knows anything about the internal structure of
a file's access lists. That is, it knows xxxx is a file property access list,
but it doesn't know how to decode it into RNames or anything. It should
be written with an eye to replacing it probably completely when Pilot
file properties really exist.
AccessControl: (In addition to keeping track of disk space), this module expects
questions of the form: can client foo do bar to such and such a file?
Therefore it will be making read requests to the FileProperties module to
acquire access lists, which it will then peruse since it knows their
internals.
FileStore: this is a high level fellow. It must ask AccessControl questions about
permissions, as it shouldn't be munging around decoding access lists, etc.
itself. It will stash file permissions it has received in its open file table.
Upon client requests to read or write file access lists, File Store asks
AccessControl `okay for this client to set this file's access lists?' and then
calls FileProperties to do the actual read or write.
If I understand correctly, Mark, in contrast, thinks that the internal structure of
a file access list should be known to FileStore, and its calls to AccessControl
should be of the form, is client foo in RName bar (or is the client the owner,
etc.). AccessControl never needs to call FileProperties directly. Code that was in
AccessControl in my proposed implementation would live in FileStore in this
one. Note that, in either case, AccessControl will be decoding access lists for
owner data. I would like to reserve the right to have an owner access list be
different in form from a file access list, but they will certainly be very similar.
(2) Permissions: The permissions I suggested were as follows:
owner lists: permission to allocate and deallocate space for this owner name.
permission to read and set these access lists.
file lists: permission to read the file.
permission to modify (write over) the file.
permission to read and set these access lists.
Mark's suggested ones are:
owner lists: permission to create or delete files for this owner name.
permission to read and set these access lists.
file lists: permission to read the file.
permission to modify (including extend and truncate but not
delete) the file.
permission to read and set these access lists.
Under the old permissions, the following checks would take place:
open: file access of appropriate type.
create: owner allocate.
delete: file modify and owner allocate.
set size: (cached) owner create and check space available.
Under the new permissions, the following checks would take place:
open: file access of appropriate type.
create: owner create.
delete: file modify and owner create.
set size: just check space available.
The old permissions have the disadvantage of a sight amount of extra code to
keep a tiny cache of creators; the new permission means giving someone write
access to a file means they can use up your space. On the other hand, you may
want people to be able to append to a file but not create new files......It is unclear
to me which of these sets is better.
With either set of permissions, we think the default for "file modify" permission
would not be owner, as I suggested, but rather creator (or creator plus owner?).
Karen
*start*
02104 00024 US
Date: 17 July 1981 9:49 am PDT (Friday)
From: MBrown.PA
Subject: Re: comments please
In-reply-to: kolling's message of 16 July 1981 5:33 pm PDT (Thursday)
To: kolling
cc: taft, mbrown
(1) The "FileStore" box already deals with RName values (we have been calling
them ownerIDs and clientIDs.)
All that I was suggesting is that we define another type, perhaps called
AccessList, whose implementation might be a record containing three bits for the
three special groups "world", "owner", and "alpineWheels", plus a sequence of
RNames. Only AccessControl knows the internal structure of this list. Assume
that FileProperties is capable of associating such AccessLists with a file (with
restrictions on the length of the RName sequences), and retrieving the
AccessLists associated with a file. Assume that AccessControl is capable of
answering a query of the form "is this RName contained in this AccessList?"
Then in the FileStore implementation of OpenFile, say, we first call
FileProperties to get the proper AccessList, then call AccessControl to check
containment. In CreateFile, we first call the Owner database to get the
AccessList, then call AccessControl to make the check.
There is a small flaw in the above: the "owner" bit of a file access list can only
be interpreted in the context of a file. Maybe this argues for the other
organization, in which AccessControl calls FileProperties.
I don't see a strong motive for making an owner access list to differ from a file
access list, and in the absence of one I would incline to making them the same.
(2) I don't feel very strongly either way on this issue. My prejudices for our
environment generally favor greater freedom, hence I prefer a broader
interpretation of "file modify" (to include the ability to allocate space). I am not
sure whether file delete should require owner create, file modify, or both.
My suggested "file modify" default is the "owner create" list for the owner of the
file (at the time the file is created.) This would always include the creator, but
might include others.
--mark
*start*
02287 00024 US
Date: 17 July 1981 12:43 pm PDT (Friday)
From: Kolling.PA
Subject: Re: comments/ Ed, note esp. point 4.
In-reply-to: MBrown's message of 17 July 1981 9:49 am PDT (Friday)
To: MBrown
cc: kolling, taft
(1) Oh, I thought you were planning to have FileStore pull the access lists apart
into their group RNames, owner fields, etc. before presenting them to
AccessControl. The residual question is, should it be FileStore or AccessControl
who asks FileProperties for the access lists. Since I don't know what you plan
for the internals of FileStore I have no opinion on this, so let me know what you
decide, so I can start coding AccessControl. As you pointed out, if FileStore will
be doing the work, AccessControl does need a way to ask it "who is the owner
for this thing?"
(2) The problem with using the "owner create" list for the "file modify" default
is that I am trying to keep the owner stuff separate from the file stuff. That's
why I proposed not storing defaults for file access lists in the owner database. If
we have to go snarfing into the owner database for file defaults, we might as
well put in real defaults, which we can do by increasing the size of the owner
records to two pages. An alternative is to interpret the owner bit to mean both
the ownerName and then the ownerCreate list. This does mean that if you
change the owner create lists, files with owner bits set will now be accessible to
different people than they were before. This also means that AccessControl's
"who is the owner for this thing?" call to FileStore needs to be prepared to
supply all this data.
(3) My hesitation about having the access lists have the same form for files as
for owners, is that clients can call procedures to change these access lists. For
files, they can change owner, world, etc. but for the owner data they cannot.
World is never set for owner lists, and owner and Alpine wheels always are set.
So we give them an argument which they can ask us to change but then we tell
them they can never change it.....We could cart the same form of list around
internally, of course, in case we later want to change our minds about setting
world, etc.
(4) Ed, what's your choice about the two proposed protection schemes in
yesterday's message?
KK
*start*
00229 00024 US
Date: 20 July 1981 4:54 pm PDT (Monday)
From: Kolling.PA
Subject: directory names?
To: taft
cc: Kolling
Somewhere is there a list of the directory names on Ivy (other than by editing
an ftp log)?
Karen
*start*
00272 00024 US
Date: 20 July 1981 5:01 pm PDT (Monday)
From: Taft.PA
Subject: Re: directory names?
In-reply-to: Your message of 20 July 1981 4:54 pm PDT (Monday)
To: Kolling
"List <*>!1" will give you the closest approximation to a list of directory names.
Ed
*start*
00710 00024 US
Date: 22 July 1981 8:14 pm PDT (Wednesday)
From: Kolling.PA
Subject: Alpine legal archiving
To: Boggs, Taft, Birrell, Levin, Schroeder, MBrown
cc: kolling
Any files put in the [Ivy]<Alpine> directory will be automatically put into the
legal archives by elves.
If from time to time you have info about Alpine which should be archived but
which you would not normally preserve for posterity by storing it on <Alpine>
(such as a mail file, for example), either store it on <Alpine>JustForArchiving>
or send me a pointer to the file and I will cause it to be archived. Files on
<Alpine>JustForArchiving> may be deleted after archiving.
Any file format is acceptable.
Thanks,
Karen
*start*
00739 00024 US
Date: 22 July 1981 12:47 pm PDT (Wednesday)
From: kolling.PA
Subject: decision revisited
To: MBrown
cc: Kolling
We decided a day or so ago that since we needed to go into the owner data base
to get the owner create-access-list (to use to set the default
who-can-modify-this-file-list on file creation), we might as well keep a real set
of default access lists for each owner in the owner data base. How about instead
forcing the CreateFile call to supply the access lists for the file? This is
consistent with expecting a future distributed client to be keeping the defaults
and in line with keeping things simple. It is already sort of specified that way
in the most recent version of the FileStore memo.
KK
*start*
02672 00024 US
Date: 23 July 1981 2:35 pm PDT (Thursday)
From: Nelson.PA
Subject: FYI--RPC interfaces, version 2
To: MBrown, Kolling, Taft
cc: Birrell, Nelson, Boggs
Andrew and I have settled on this for the (short) time being. Most of the
private RPC interface is omitted; the client interface that you will use is given
second. The import, export, and encryption routines you have to invoke
explicitly, of course, with the values you want. But once this is done, calls
through the SOURCE interface are remote, and transparent (except for
RPCErrors). Comments welcome, if you have any.
Andrew has backed off considerably from his position that multiple interface
instances be supported by records of procedures (instead of interface records).
Instead, we now favor the original handle approach--there are a static number of
interface instances, multiplexed (if desired) by client-specified handles. See
Andrew for more info on this. The dust has not settled.
-- Components of the private the RPC runtime interface, probably to be called
"RPC".
-- TYPES.
RemoteInterface: TYPE = RECORD [
type: STRING ← NIL, --Default is interface name, e.g, "FTP".
instance: STRING ← NIL, -- Default is a UID; a client e.g. is "Ivy".
version: VersionRange ← AnyVersion ];
VersionRange: TYPE = RECORD [
-- A closed interval, client specified.
firstAcceptable, lastAcceptable: CARDINAL];
AnyVersion: VersionRange = [0,0]; -- Distinguished value that means use any.
-- These next items are all a part of the authentication package, not RPC.
ConversationID: TYPE = []; --On a per-principal-pair basis.
EncryptionKey: TYPE = []; --Not yet specified.
RPCError: ERROR [type: Problem];
Problem: TYPE = {[], spare1, spare2, spare3};
-- Components of the user-visible RPC interface. For each user interface
SOURCE, the stub generator creates three files:
SourceRPC The types and control procs are shown below;
SourcRPCClient Client stubs; exports both Source and SourceRPC;
SourceRPCServer Server stubs; exports part of SourceRPC.
-- This is a SourceRPC interface.
-- TYPES.
RemoteInterface: TYPE = RPC.RemoteInterface;
VersionRange: TYPE = RPC.VersionRange;
ConversationID: TYPE = RPC.ConversationID;
EncryptionKey: TYPE = RPC.EncryptionKey ← RPC.ClearKey;
-- PROCEDURES.
RPCError: ERROR = RPC.RPCError;
Problem: TYPE = RPC.Problem;
ExportInterface: PROC [interface: RemoteInterface ← [] ];
UnExportInterface: PROC;
ImportInterface: PROC [interface: RemoteInterface ← [] ];
UnImportInterface PROC;
SetEncryptionKey: PROC [conversation: ConversationID, key: EncryptionKey];
ClearEncryptionKey: PROC [conversation: ConversationID];
*start*
00656 00024 US
Date: 23 July 1981 7:04 pm PDT (Thursday)
From: Kolling.PA
Subject: hashing
To: mbrown
cc: kolling
I'm going to get the hash code by using your suggestion of packing the low 5
bits of each character three to a word and xoring them, with a file 4*the data
size. This gives an exceedingly low (1.15) ratio of names per occuring codes and
more importantly the pathological cases are only: maximum of names/codes = 3
and only a handful of them showed up in the 500 names. Making the file 2*
the data size increased the maximum to 4, but also the number of cases shot up
quite a lot. Disk space is cheap compared to disk reads.
KK
*start*
00368 00024 US
Date: 23 July 1981 12:37 pm PDT (Thursday)
From: Kolling.PA
Subject: legal archiving
To: Anderson
cc: Kolling
Be sure to read Martin's memo about legal archiving. Sometime in the next week
or so, we will be dumping on you a whole nest of memos from Alpine to be put
into the archives (there's no hurry about getting it done, though).
Karen
*start*
00410 00024 US
Date: 23 July 1981 2:54 pm PDT (Thursday)
From: Kolling.PA
Subject: CRSSCode.bravo
To: MBrown
cc: Kolling
There's a file called CRSSCode.bravo on Alpine, containing transaction-oriented
code, with your name on it and Butler's. Perhaps an implementation of the
Howard-Butler stuff? I couldn't tell from glancing briefly at it if it should be
archived for Alpine. Should it be?
KK
*start*
01232 00024 US
Date: 24 July 1981 2:25 pm PDT (Friday)
From: Nelson.PA
To: Kolling
Sub: Forgot you
Subject: Multiple RPC exporters
To: (Alpine:) MBrown, Taft
To: (Other:) Rovner, Satterthwaite, Schmidt, Levin
cc: Birrell, Nelson
Mesa permits the implementation of an interface to come from multiple exporters,
each of which supplies components of the interface. Andrew (and I) have
considered this situation in our proposed RPC world. Our conclusion is that
restricting interfaces to be exported entirely from one host is a significant
reduction in complexity, primarily for the RPC binder, but also for the stubs
themselves.
While we believe we can do multiple exporters at the cost of storage in stubs and
hairer binding, we don't want to and are hereby sounding you out on this
decision. There are basically three, not two, possible responses:
1. Single exporter is fine. No problem.
2. Single exporter OK for now, but multiple are probably needed down the road.
Do single now, but leave the hooks needed for easy impl. of multiple (say,
.5-1.0 man month of extra work).
3. Multiple exporters wanted from the start, even if it takes a lot longer.
Other comments welcome too.
Bruce and Andrew
*start*
00440 00024 US
Date: 24 July 1981 10:37 am PDT (Friday)
From: MBrown.PA
Subject: Re: hashing
In-reply-to: Your message of 23 July 1981 7:04 pm PDT (Thursday)
To: Kolling
cc: mbrown
sounds good. did you use a prime table size? did you compute the average
number of probes per entry? I have a few more suggestions about hashing, e.g.
how to do collision resolution; let's try to talk today before I leave (around 1pm).
--mark
*start*
00706 00024 US
Date: 24 July 1981 11:49 am PDT (Friday)
From: Kolling.PA
Subject: Re: hashing
In-reply-to: Your message of 24 July 1981 10:37 am PDT (Friday)
To: MBrown
cc: Kolling
prime table size: yes.
average number of probes per entry: I didn't quite do this. The average of the
names that hashed to the same code was 1.15. The max was 3, and this only
happened for (as I recall) something like 5 codes out of the possible 1999. Since
the table is 4*data and the cases didn't cluster these numbers seemed like good
handles on the average and worst cases.
collision resolution: I was just going to search linearly. I realize this will slowly
degrade the table. Is there a better way?
*start*
01232 00024 US
Date: 24 July 1981 2:25 pm PDT (Friday)
From: Nelson.PA
To: Kolling
Sub: Forgot you
Subject: Multiple RPC exporters
To: (Alpine:) MBrown, Taft
To: (Other:) Rovner, Satterthwaite, Schmidt, Levin
cc: Birrell, Nelson
Mesa permits the implementation of an interface to come from multiple exporters,
each of which supplies components of the interface. Andrew (and I) have
considered this situation in our proposed RPC world. Our conclusion is that
restricting interfaces to be exported entirely from one host is a significant
reduction in complexity, primarily for the RPC binder, but also for the stubs
themselves.
While we believe we can do multiple exporters at the cost of storage in stubs and
hairer binding, we don't want to and are hereby sounding you out on this
decision. There are basically three, not two, possible responses:
1. Single exporter is fine. No problem.
2. Single exporter OK for now, but multiple are probably needed down the road.
Do single now, but leave the hooks needed for easy impl. of multiple (say,
.5-1.0 man month of extra work).
3. Multiple exporters wanted from the start, even if it takes a lot longer.
Other comments welcome too.
Bruce and Andrew
*start*
00686 00024 US
Date: 7 Aug. 1981 4:32 pm PDT (Friday)
From: kolling.PA
Subject: space manager transaction
To: mbrown
cc: kolling
When the system starts up, and FileStore calls either InitializeAccessControl or
RecoverAccessControl, who should request the creation of the Space Manager
transaction? Is it okay for AccessControl to do this, or would FileStore perhaps
want to hand the transID to AccessControl in these calls for some reasons of its
own? Does FileStore ever need to know the transID for SpaceManager, like
perhaps might it somewhere want to wait for all current trans to complete or
something? (It would have a long wait for the SpaceManager one....)
Karen
*start*
02381 00024 US
Date: 7 Aug. 1981 5:41 pm PDT (Friday)
From: MBrown.PA
Subject: Re: space manager transaction
In-reply-to: Your message of 7 Aug. 1981 4:32 pm PDT (Friday)
To: kolling
cc: mbrown
"Managers" who write log records will have to be specially registered with
FileStore somehow. The form of this registration is not yet designed, but I'm
getting to it.
One of the reasons for knowing the identity of managers is for handling "major"
checkpoints, such as warm shutdowns, in a clean way. When the system is shut
down, we'd like to give all of the managers a chance to clean out their buffer
pools and do a final commit, and notify the system that they have no more
interest in the contents of the log. This way the archive log does not have to
contain any of the client-written log records, only the basic FileStore log
records. Doing the shutdown can be delicate, since presumably we want to start
no more transactions, but might have to do so for the purposes of cleaning up
one of the managers ... Getting this right should be interesting.
A somewhat related question is whether or not managers' exclusive locks on the
files they manage have any special status. Having the lock manager choose one
of these locks to break in order to break a deadlock cycle is probably wrong. In
this case the lock manager must distinguish a manager's transaction from
garden-variety ones.
Offhand I see no reason why the manager can't create its own transaction.
I am glad that you are forcing me to get into issues like these. But you should
remember that in the interests of simplicity in the initial implementation, it may
turn out to be desirable to perform all of the owner-DB writes under the
transaction that requested the space change, not a shared one (even though this
may reduce the potential concurency somewhat.) I am not sure precisely how
much of your design would be impacted by this change (except that some parts,
like supplying recovery procedures and interfacing to the system as a "manager",
would just go away entirely.) My guess is that much of the complexity in your
implementation goes into the volatile data structures required to support the
deferred update, which should be nearly the same either way. Also, the file
organization is not a function of the style of transaction usage, as long as you
stick to one-page records.
--mark
*start*
00187 00024 US
Date: 10 Aug. 1981 2:15 pm PDT (Monday)
From: Kolling.PA
Subject: [Ivy]<Alpine>Doc>AccessControlDesign5.bravo.....
To: MBrown, Taft
cc: Kolling
.....now exists.
*start*
00280 00024 US
Date: 10 Aug. 1981 4:18 pm PDT (Monday)
From: kolling.PA
Subject: Re: [Ivy]<Alpine>Doc>AccessControlDesign5.bravo.....
In-reply-to: Kolling's message of 10 Aug. 1981 2:15 pm PDT (Monday)
To: Kolling
cc: MBrown, Taft
And also AccessControlReorganize.sil.
*start*
02252 00024 US
Date: 10 Aug. 1981 12:19 pm PDT (Monday)
From: kolling.PA
Subject: Re: space manager transaction
In-reply-to: Your message of 7 Aug. 1981 5:41 pm PDT (Friday)
To: kolling
We ought to talk when you have some time.....
(1) The space manager requests locks on "strange objects" sometimes in the name
of itself and sometimes in the name of the "real" transactions it sees (remember,
for example, that's how owner operations and space operations are mutually
exclusive). Although the sm goes to trouble to request locks in order, etc. it
would be possible for a deadlock to occur involving these locks, for example, as
follows:
client one does an allocation for owner foo (read lock on volatile owner foo)
client two does an allocation for owner bar (read lock on volatile owner bar)
client one tries to do a space operation for owner bar (requests write lock for
owner bar)
client two tries to do a space operation for owner foo (requests write lock for
owner foo)
Now, if the lock manager detects a deadlock and wants to break it, I don't see
how the sm can survive a broken lock, I think the only acceptable thing for the
lock manager to do is not just break the lock but abort the unlucky transaction.
Otherwise, what is the sm supposed to do? Does it even know the lock has been
broken, since the lock was gotten for the real transaction? If it doesn't know,
and someone else gets into the structure, the best it can do is detect a "this
should never happen" situation. If it does know, should it pretend the requested
operation never happened? Abort the transaction itself? etc. Perhaps the lock
manager needs a way to be told "this is a system internal lock" and breaking
those should be trans aborts.
(2) I have a highly concurrent design 98% fleshed out in gory detail. However,
it does have a lot of locks and monitors. Perhaps all this concurrency is not
worth the complexity. On the other hand, perhaps locks are cheap and it is.
The way it is designed, people will almost never have to wait for disk io, just
when they read in the owner record. When you have ten or 15 minutes, I'll run
it past your nose. How many simultaneous transactions do we expect to see
anyway?
Karen
*start*
00400 00024 US
Date: 10 Aug. 1981 5:05 pm PDT (Monday)
From: kolling.PA
Subject: tomorrow
To: mbrown
cc: kolling
I thought I'd spend tomorrow sketching out a "non-concurrent" algorithm. But,
if the idea is that each transaction updates the owner data file, doesn't that mean
that write access on that file has to be "world" ? (i.e., no access control on our
accesscontrol....).
Karen
*start*
02080 00024 US
Date: 10 Aug. 1981 6:47 pm PDT (Monday)
From: MBrown.PA
Subject: Re: tomorrow
In-reply-to: Your message of 10 Aug. 1981 5:05 pm PDT (Monday)
To: kolling
cc: mbrown
transactions do not have authentication associated with them; that is done at the
time a file is opened. hence there is no inherent conflict here. if there were,
we would be in trouble because it would be hard to write a directory system, or
any sort of database manager, without using the shared write transaction idea.
one reason for building transactions into the server is so that not all DBMSs have
to implement them themselves!
i haven't had a chance to go through your memo in detail yet, but for one
thing, I think that your aspiration level for the owner database reorganization
procedure is way too high. My idea was that this database might require
reorganization once every couple of years. Hence there is no problem in taking
the server down to do the reorg. Furthermore, I thought that reorg would simply
consist of printing out the current state of the database (in ascii, using
EnumerateOwners), setting the new hash table length (magic), formatting the
new hash table (magic), and inserting the database elements one by one (by
using CreateOwner.) Anything fancier than this seems like overkill to me. The
measurement stuff that you include in your code will tell us whether reorg is
required.
i am also willing to tolerate a very rough and ready design for updates to the
owner access lists, since these almost never happen; the updater is willing to
wait for quite a long time. Another possibility that I am willing to live with is
that the access lists are not maintained in a fully consistent way; e.g. client a
might create a file for owner 1 using transaction t1, then t2 might change the
owner access lists in such a way as to deny client a the right to create files. if
t2 commits before t1 commits then t1 should abort, but in this case i don't care
what happens.
i am sure that we can simplify this. i'll talk to you about it on wednesday.
--mark
*start*
02918 00024 US
Date: 11 Aug. 1981 12:24 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: tomorrow
In-reply-to: Your message of 10 Aug. 1981 6:47 pm PDT (Monday)
To: MBrown
cc: kolling
"transactions do not have authentication associated with them; that is done at the
time a file is opened." I thought the purpose of making "each transaction do the
updates on the owner database file" in the simple case was that these updates are
then tied to the commit for that transaction. The "openFileID" for a file maps to
[client, trans, fileID, accessrights], according to Ed's sil drawing. If I just hand
off the openFileID that some other transaction got for the owner file to trans t1,
how does the system know that those updates are associated with t1?
reorganizing the database:
(1) At one point I thought you told me that reorganizing the database should
not require taking the system down.....Since we need stops for the purposes of
warmshutdown, deferredupdate, etc. stopping for reorganizing "on-line" might be
"free", maybe or maybe not.
(2) "I thought that reorg would simply consist of printing out the current state of
the database (in ascii, using EnumerateOwners), setting the new hash table
length (magic), formatting the new hash table (magic), and inserting the
database elements one by one (by using CreateOwner.)": printing out the
current state? You don't mean that some poor operator has to then hand insert
500 owners, right? I could abandon my one-file reorg scheme and instead do
this as you suggest by using a second file as a buffer, and either replacing the
old id in Pilot's root directory (courtesy of FileStore perhaps) or doing a copy
over at the end. Fine by me.
"i am also willing to tolerate a very rough and ready design for updates to the
owner access lists": I'm not sure of the implications of this. I have to be careful
of updates to the owner space quotas anyway, so all(?) of the mechanism would
be there anyway. (As a side comment, I didn't plan to worry about FILE access
lists changing out from under a client who already had obtained a permission--
that's FileStore's problem, if it cares, which it probably shouldn't, or perhaps
locking prevents this from even happening.) As for OWNER access lists
changing, the trans t1 that had obtained create permission would automatically
block the trans t2 that wanted to change the owner access lists; otherwise I need
extra code to let change owner access list thru while create in progress but block
change space quota while create in progress.
I think we need to settle in for a relatively long talk. It may be that a simple
solution is best if it can provide enough features. What worries me somewhat
about the scheme I propose in the memo is not the amount of code, but the
chance of dropping a lock on the floor. Let's talk (after you've had a chance to
go thru the memo?). Maybe after Dealer?
Karen
*start*
00817 00024 US
Date: 12 Aug. 1981 10:34 am PDT (Wednesday)
From: MBrown.PA
Subject: Re: tomorrow
In-reply-to: Your message of 11 Aug. 1981 12:24 pm PDT (Tuesday)
To: kolling
cc: MBrown
you would have to open the owner database file with each transaction that needs
to update that file.
i assumed that we would have a program read the text file produced by dumping
the owner database. i favor a text dump because you need a way of dumping it
in text form for debugging anyway. every data structure should have an
associated print routine.
ignoring the online reorganization part of your design, i don't think that the rest
is inherently too compilcated, but i think that we need to think about it some
more to develop the right abstractions. yes, maybe i'll have some good ideas
after dealer.
--mark
*start*
01985 00024 US
Date: 12 Aug. 1981 12:37 pm PDT (Wednesday)
From: Kolling.PA
Subject: volume groups
To: kolling
About one FileStore per volume group and about volumes going offline:
1. Who handles attempts to read and write files not on an online volume, both
for the case when "it is known" beforehand that the volume is offline and also
when ooops it disappears out from under us as we are doing io? Should
AccessControl be prepared to have its calls on FileProperties to read a file access
list get volumehasgoneoffline or something?
2. Do we want a way for FileStore to tell AccessControl that a volume has gone
offline so AccessControl will reject space change requests for that volume, and
that it has come back online? (VolumeStateChange: PROCEDURE[volumeID:
VolumeID, state: {online, offline}])? Or does FileStore never ask AccessControl
to do stuff for offline volumes?
3. Some proposed changes to Access Control's interface for volume groups: (Some
of this depends on who knows a volume is offline; also, who does various types
of error handling, such as volumeNotInGroup(horrors).)
(a) CheckAccessOwnerCreate: the volume is specified as a parameter (some
default that means any in the group), if default, CheckAccessOwnerCreate will
now determine the volume (the online one with the most free space as far as it
knows). new error returns: volumeNotInGroup.
(b) ChangeSpaceViaOwner and ChangeSpaceViaOpenFileID now expect a
VolumeID. new error returns: notEnoughSpaceInOwnerQuota (was
notEnoughSpace), notEnoughSpaceOnVolume, volumeOffline (assuming FileStore
told us this?), volumeNotInGroup.
(c) InitializeAccessControlDataBase: new parameters -- ownerFileVolumeID and
volumeGroup: LIST OF VolumeIDs.
(d) BeginRecoveryAccessControl new parameters -- ownerFileVolumeID and
old parameter openRootFile has a new parameter ownerFileVolumeID.
4. I assume we don't add or remove volumes for a group. Let me know if this is
incorrect.
Karen
*start*
00650 00024 US
Date: 13 Aug. 1981 2:30 pm PDT (Thursday)
From: kolling.PA
Subject: notes on locks
To: mbrown
cc: kolling
(1) In both AC designs, I ran into cases where I would like to have an exclusive
lock which I could release. In the Simple Design this would be nice, for
example, if someone is going after an owner who turns out not to be in the
"correct" page, it could then harmlessly release that exclusive lock so as not to
block someone who might want that page.
(2) In the complex design, I also kept running into cases where I wanted to lock
out my own transaction, although I got around those by redesign or hand
checking.
*start*
01777 00024 US
Date: 23 Aug. 1981 3:09 pm PDT (Sunday)
From: Taft.PA
Subject: Multi-pack Alpine volumes
To: MBrown
cc: Kolling, Taft
I finally got around to reading your memo. I agree with most of what you say.
An advantage of multiple-pack volumes that you left out is that they allow you
to store a file that is larger than the remaining free space on any single pack. In
a system with many packs that is close to being full (example: Ivy), this is an
important advantage. IFS requires each file to be entirely contained on one pack,
and that is a real problem: the largest file you can store right now is about 4000
pages. This is a large file, but not unreasonably so.
Of course, this problem can be dealt with by some sort of disk permuting
program, but that's a lot of work.
Ultimately, I think the right approach is a flexible one. A file server's "primary"
Alpine volume should encompass as many packs as is appropriate, since it doesn't
make much sense to operate the server with only part of the file store on-line.
On the other hand, dynamically mountable Alpine volumes should ordinarily
consist of a single pack so as to make them easily portable. (After all, not many
file servers are likely to have more than one extra drive.)
In the short term, I see little point in trying to support multiple-pack logical
volumes at the Alpine level when Pilot doesn't provide that capability
underneath. And I'm certainly not in favor of screwing around with Pilot in
order to finish the implementation.
So I think the simplest statement of your proposal is that Alpine volumes will be
one-to-one with Pilot logical volumes; and in the short term this means that an
Alpine volume is no larger than a disk pack, due to temporary Pilot limitations.
Ed
*start*
02532 00024 US
Date: 25 Aug. 1981 4:56 pm PDT (Tuesday)
From: MBrown.PA
Subject: File Store document, version 5
To: Kolling, Taft
cc: MBrown
I've placed a hardcopy of this memo in your mailboxes. The changes are pretty
major; see the change log in the memo. The document is getting rather large
and it may be time to create an "Alpine concepts and facilities" document
separate from the interface description. Such a document would be suitable for
outside review, which we should get (I was thinking of having Mitchell,
Schroeder, and Lampson read it) at some point before the implementation really
gets rolling.
I am now trying to flesh out the treatment of errors in the interface. I am in
general agreement with the stylistic conventions suggested in Roy's memo
"Guidelines for Signalling in Cedar"
(ivy]Levin>Cedar>SignallingGuidelines.bravo.) But we need to figure out how
to apply these ideas to our system.
One general principle that seems useful: if an error detected by a procedure is
likely to the the result of a sequence of calls outside the immediate scope of the
caller, the error should be raised as an ERROR and not by a return code. Failure
of credentials (except on Login itself) and transaction aborts are examples of this
type of error.
Abstraction failures represent a fairly well-delimited class. For instance, a create
owner failure due to the owner database being full is an abstraction failure, as is
a hard disk error. Is a SetLength or CreateFile failure due to out of disk quota an
abstraction failure? I guess so.
I am assuming that we shall not require SIGNALs in our interface, only
ERRORs. Remote RESUME is too much to ask for right now. Andrew seems
willing to provide remote ERROR in a timely fashion.
A remote ERROR will unwind the remote call before we see the error locally.
This means that what Roy calls "Client programming errors" (e.g. RName too long
or containing an invalid character) must return enough information to diagnose
the problem that the remote machine found.
The type of error for which the correct interface style is unclear is "read page
off end of file", "owner already exists in call to create owner", etc. Offhand I am
inclined to report these as ERRORs although they do not fit neatly into any of
Roy's categories (they come closest to being calling errors, but they clearly
should be documented in the interface, which calling errors are not.) Maybe
this means that I am wrong and these should really be reported by return codes.
--mark
*start*
00295 00024 US
Date: 27 Aug. 1981 1:18 pm PDT (Thursday)
From: kolling.PA
Subject: AccessControl
To: mbrown, taft
cc: kolling
The latest version of AccessControl, AccessControlDesign6.bravo, is in your
offices. Its internals are drastically different from the previous versions.
KK
*start*
00173 00024 US
Date: 28 Aug. 1981 12:42 pm PDT (Friday)
From: kolling.PA
Subject: [Ivy]<Walnut>Cedar>Grapevine>
To: kolling
is where the new rname defs etc. are.
*start*
00487 00024 US
Date: 28 Aug. 1981 1:29 pm PDT (Friday)
From: kolling.PA
Subject: owner file access
To: mbrown
cc: kolling
As I recall, the owner database file is supposed to be writeable by AlpineWheels
only for obvious reasons, and I was supposed to put an "AlpineWheels"
clientname into the OpenFile call on it for the client transactions I want to have
read and write pages of it. I don't see a way to do this in the FileStore interface,
as Open just takes a transID.....
*start*
00771 00024 US
Date: 29 Aug. 1981 12:16 am PDT (Saturday)
From: MBrown.PA
Subject: Alpine overview
To: Kolling, Taft
cc: MBrown
The time is approaching that we shall have to tell the rest of CSL what we think
Alpine is. For most of the lab I think this can be accomplished with a short
overview document and a Dealer. We should also have a few reviewers look at
what we are doing in more detail; the overview document can be their
introduction to the subject.
I have drafted an overview document and put hardcopies in your mailboxes. I
expect that I have some blind spots and that I have drawn some conclusions that
may not represent the group's feeling, so I will refrain from circulating the
document until each of you has suggested amendments.
--mark
*start*
03697 00024 US
Date: 31 Aug. 1981 4:05 pm PDT (Monday)
From: kolling.PA
Subject: Re: Alpine overview
In-reply-to: MBrown's message of 29 Aug. 1981 12:16 am PDT (Saturday)
To: MBrown
cc: Kolling
1. page 2 typo: "These access control facilites"
2. page 2: I'm not clear about what you're trying to say in the area starting
"Alpine must support server configurations that survive any single failure of the
disk storage medium .... This degree of redundancy should not be a requirement
for all servers, however." I thought we could survive a single failure by
playing back the log if we went back far enough. Or are you talking about
some other mechanism? Or about having it live on a workstation so there is
only one copy of the log?
3. page 3: "Archiving" My standard kvetch about geographically separated
backup/archiving in case the building burns down.
4. page 3. "Our environment does not demand continuous availability of a file
server; we can tolerate scheduled downtime during off hours....." When Juniper
needed to find a down time for backup, no time was available in the 24hr slot
that didn't impact somebody. Of course, replicated files solves a lot of this.
5. page 4. "Alpine uses Grapevine to implement file access control." And disk
space control, too.
6. page 5:
"The archive log **actually** includes only the log records that are
**actually** relevant to media recovery."
The archive log actually includes only the log records that are relevant to
media recovery.
7. page 5: "A volume can be moved from one server to another. This operation
requires a "volume quiesce"...This operation is supported by Juniper...." Is this
the belief that a Juniper pack can disappear and the system keeps running?
That's what I used to think too, until I saw the directory system fall apart when
a pack went off line, and the system refuse to come up when pack 0 was not
online, and.... I think Howard used to believe that too, but Jay didn't, and so in
the implementation.....
8. page 5.
"Two of a file's properties are its read list and its modify list. Each of these
contains (at most two) RNames, such as (CSL↑.pa ISL↑.pa). An Alpine client may
read a file if he is contained in (the closure of) one element of its read list. An
Alpine client may read or modify a file (including its read and modify lists) if
he is contained in (the closure of) one element of its modify list.
Another of a file's properties is its owner. An owner is an entity indentified by
an RName, such as "McCreight.pa" or "CedarImplementors↑.pa". The disk space
occupied by a file is charged against its owner's space quota."
I suggest:
One of a file's properties is its owner. An owner is an entity indentified by an
RName, such as "McCreight.pa" or "CedarImplementors↑.pa". The disk space
occupied by a file is charged against its owner's space quota.
"Two other file properties are its read list and its modify list. Each of these
contains (at most two) RNames, such as (CSL↑.pa Taft.pa), as well as fields
indicating access to the owner and world. An Alpine client may read a file if
he is contained in (the closure of) one element of its read list. An Alpine client
may read or modify a file (including its read and modify lists) if he is contained
in (the closure of) one element of its modify list.
(I changed the example to contain an individual rname and included owner and
world.)
9. page 5: typo: "Disk space quotas for owners are maintained on for volume
groups rather than individual volumes."
10. page 5: awkward: "A client may request to create a file on a volume group".
11. page 6: "The memos [Ivy]<Alpine>Doc>FileStoreDesign*.bravo"
*start*
01018 00024 US
Date: 31 Aug. 1981 5:55 pm PDT (Monday)
From: kolling.PA
Subject: two algorithms
To: mbrown
cc: kolling
In the cold light of day, I began to think again about the difference between
your algorithm with the buffers on one list and a use bit, and mine with buffers
in different states separated on different lists. I think I picked the latter because
I found it clearer to think in terms of as well as because I was dealing with
smaller lists, so I compared rough sketches of the code and, for the most likely
cases, ignoring duplicate stuff and small things, this is the difference between
the two: (I don't guarantee I haven't dropped something on the floor here,
especially since my D0 booted while I was working the more detailed version of
this out.)
Mine Yours
----- -----
space:
1 extra word per owner buffer.
GetPage:
unlink.
ReleasePage:
unlink.
Phase One:
unlink
link
Since (un)linking is relatively expensive, I'm going ahead with your scheme.
KK
*start*
00540 00024 US
Date: 31 Aug. 1981 6:27 pm PDT (Monday)
From: kolling.PA
Subject: filestoredesign5
To: mbrown
cc: kolling
I haven't had a chance to read this whole memo yet, but the reason there are
separate calls to change the access lists and the space quota is not because the
internals of managing the access to these fields is different, but because these
tend not to be things that will naturally happen at the same time, so it avoids
the user specifying fields he/she doesn't care about or changing them
inadvertantly.
KK
*start*
02262 00024 US
Date: 2 Sept. 1981 11:05 am PDT (Wednesday)
From: MBrown.PA
Subject: Re: Alpine overview
In-reply-to: Your message of 31 Aug. 1981 4:05 pm PDT (Monday)
To: kolling
cc: MBrown
2. Some servers may not choose to replicate the log (use mirrored disks for log.)
In this case, a crash that occurs after commit but before the intentions are carried
out, and that destroys the single copy of the log, wipes out committed
information in the system.
Also, a workstation-based system will surely not take regular backups or save
its log (though one can imagine fixing this if the transaction rate is low.)
3. Noted. This applies even more empatically to backup than to archive, since
many files are never archived (I'm not sure how to archive a database.)
4. I find it hard to believe that Ron can't take the server down for 15 minutes at
6am. If pack copy is implemented carefully, this is about how long it should
take (the disk can be read in 5 minutes at full speed, but you need a check pass,
...) Not all packs are copied every time backup is taken.
There are alternatives that provide continuous availability, and someday we
should implement one just for the experience, but I don't believe we'll do it next
year.
5. Grapevine's contribution to disk space control is minimal, and I felt that it
would add confusion to mention it here.
7. Right. What I meant was simply that you can spin down a Juniper pack at
any instant, move it to another machine, spin it up, and it can recover to a
transaction consistent state. It is another question whether you can find the
files, etc.
8. I want to avoid confusion, and make the scheme sound as simple as possible.
Now, "world" can be thought of as an RName, since if you allow "world", other
RNames are not needed. (Grapevine uses "*" or something.) "owner" can also be
thought of as an RName, but one that does not take up one of the two slots. We
can introduce a notation for owner ("+"?), or just have the caller supply the
owner RName and check for a match when setting up the access lists.
The question would get a little more complex if we had a "change owner" call.
We have not provided one yet, and I suggest we do until someone explains why
it is required.
--mark
*start*
00328 00024 US
Date: 2 Sept. 1981 12:17 pm PDT (Wednesday)
From: Kolling.PA
Subject: Re: Alpine overview
In-reply-to: Your message of 2 Sept. 1981 11:05 am PDT (Wednesday)
To: MBrown
cc: kolling
Yes, "Ron can take the server down for 15 minutes at 6am;" it was two hours at 6
am that had people gnashing their teeth.
*start*
00373 00024 US
Date: 2 Sept. 1981 12:26 pm PDT (Wednesday)
From: MBrown.PA
Subject: Re: Alpine overview
In-reply-to: Your message of 2 Sept. 1981 12:17 pm PDT (Wednesday)
To: Kolling
cc: MBrown
Right. We'll have to do better than two hours. Eventually disks may get so
large that online backup is a necessity, but that day will not arrive for awhile.
--mark
*start*
01666 00024 US
Date: 3 Sept. 1981 12:02 pm PDT (Thursday)
From: Taft.PA
Subject: Re: Alpine overview
In-reply-to: MBrown's message of 29 Aug. 1981 12:16 am PDT (Saturday)
To: MBrown
cc: Kolling, Taft
Some random comments:
2. Why a new file server?
a) Juniper should at least be mentioned: why are we designing a new file server
rather than just pushing Juniper into the Pilot/D-machine world?
b) "Nearly all CSL application programs will deal primarily with database systems
instead of file systems." I think this is overstating the case. I view a database as
a way of organizing information, but a file as a way of storing the data itself
(particularly if there is a lot of it).
3. Alpine's scope
a) "(Note that an archive system would reduce the problem of disk space quotas
for users.)" I don't understand this statement.
b) Archiving: I agree with your conclusion. But I should point out that the
Alpine file organization (low-level naming by unique IDs, and high-level
naming by a location-independent UFS) eliminates a lot of the problems that
would make it difficult to implement a satisfactory archive system for IFS.
4. Alpine's implementation strategies
a) "The time required to restore lost information from log tapes is bounded by
periodically copying entire volumes to backup packs..." Are the volumes taken
off-line during this operation? If so, I don't like it! Can't we do some sort of
on-line replication?
5. Alpine objects
a) Having one log per file server (rather than per volume) seems to conflict with
the notion that volumes are self-contained and (re)movable. Maybe one log per
volume group would be better ??
Ed
*start*
04841 00024 US
Date: 3 Sept. 1981 4:41 pm PDT (Thursday)
From: Taft.PA
Subject: Access control design
To: Kolling
cc: MBrown, Taft
Here are my comments on Alpine AccessControl design, version 6.
"Some comments"
2. Having a separate access control list for privileged access to each Alpine server
is a good idea. This is how we control gateways now (that is, there are groups
called Twinkle.internet, SoEasy.internet, etc.), and it seems to work well. It
nicely decentralizes adminstration of multiple servers.
5. There is a small argument to be made for separating modifications to a file's
access control lists from all other file modifications. This permits an owner to
write-protect a file from himself (to prevent blunders), or to give other clients
write access without control over the access control lists themselves.
I assume you don't want to associate any new access control lists with each file.
so I propose the following changes: (a) only the owner or the file's creator is
allowed to change access control lists; (b) it is permissible to exclude the owner
from the modify list.
(By the way, I assume that AlpineWheels is on all these lists implicitly and does
not actually consume one of the two RName slots!)
12. "The owner data base file will be entered in the root directory." Meaning a
Pilot Directory package root directory? I didn't think we were planning to use
the Directory package on Alpine volumes.
"What AccessControl expects from FileStore"
"FileStore is assumed to cache the permissions ... in its open file table." I agree
with this statement with regard to open files. But I believe that AccessControl
should also do a certain amount of caching of [ClientID, FileID, access] triples
(or something), so that successive openings of a popular file don't have to
consult Grapevine every time.
Not allowing the sum of all quotas to exceed the available space is nice in
principle, but impractical. It leads to underutilization of available space (since
many users don't use their entire quota); and in any event it's difficult to
administer, especially in an environment where users don't pay for storage. (I
do agree, though, that AccessControl should maintain a running total of quotas,
so an administrator can determine how badly the available space is
overcommitted.)
"What AccessControl expects from Locks"
(2) "... it cannot just break the lock, instead it must abort the unlucky
transaction." I thought that in Alpine deadlocked transactions were always
aborted. Am I mistaken?
"Interface to AccessControl"
IsClientValidForThisFileStore: I don't think this is structured quite right. I think
we need a separate Authenticate operation that simply authenticates an RName
(i.e., that verifies that the RName is valid, and that the client is who he says he
is). Furthermore, this should probably be put in a separate interface, since the
mechanisms for authentication are likely to change over time, as RPC and our
facilities for authentication and encryption mature. The FileStore and/or RPC
should do the authentication; and AccessControl should deal only with access
control, not authentication.
CheckAccessFile: there is a potentially nasty recursion here. If we read the
access control list properties using the general mechanisms, then we have to
open the file first; but opening the file requires access to be checked ... We can
get around this either by some sort of trap-door in the implementation for
reading properties, or by having AccessControl open the file while assuming the
identity of some privileged client who can always open files without having to
read their access control lists.
EnumerateAllOwners: let's get together and settle on a convention for doing
enumerations in Alpine. If enumerating a data structure locks it up, then
passing in an enumeration procedure that is called for each item is equivalent to
using your "continuationKey", and seems a lot cleaner. But I don't think that in
general we can afford to lock up data structures while enumerating them, and
the Pilot-style "stateless enumeration" may be preferable.
One thing that's not clear to me from the implementation notes is whether the
internal data structures are modified under the caller's transaction, under a
private transaction, or under no transaction at all (i.e., manipulating raw
FilePageMgr storage). Similarly, are the locks for manipulating file properties
acquired under the caller's transaction or under a private transaction?
General comment: there aren't any SIGNALs in the interface, though there are a
number of places where it would be natural to use them (instead of return codes
that the caller must test). Since AccessControl is an internal interface (not
exported via RPC), I don't see any reason not to use SIGNALs in the
conventional ways.
Ed
*start*
07753 00024 US
Date: 3 Sept. 1981 6:22 pm PDT (Thursday)
From: kolling.PA
Subject: Re: Access control design
In-reply-to: Taft's message of 3 Sept. 1981 4:41 pm PDT (Thursday)
To: Taft
cc: Kolling, MBrown
There is yet another version of AC (design 7), which addresses some of your
points. I wasn't planning to release it until I got further into the
implementation, but it's on
[Juniper]<kolling>dumpaltodisks>accesscontroldesign7.bravo (in a state of flux) if
you want to peek.
1. "Having a separate access control list for privileged access to each Alpine
server is a good idea."
I propose to do this by having FS tell me the list name when it does
InitializeVolatileStructure.
2. "There is a small argument to be made for separating modifications to a file's
access control lists from all other file modifications. This permits an owner to
write-protect a file from himself (to prevent blunders), or to give other clients
write access without control over the access control lists themselves."
Yes, I left this out because I was tight on space for AccessLists.
3. I assume you don't want to associate any new access control lists with each
file. So I propose the following changes: (a) only the owner or the file's creator
is allowed to change access control lists; (b) it is permissible to exclude the owner
from the modify list.
Generally, I don't know the file's creator unless we keep this as a file
property; there is probably room for this in the leader page. Just the
owner seems restrictive.... I don't know what to do about this. Since we
are just trying to put a vestigial implementation in, I vote to leave it as it
is -- what do you both think?
4. "(By the way, I assume that AlpineWheels is on all these lists implicitly and
does not actually consume one of the two RName slots!)"
Yes, AlpineWheels, Owner, World: BOOLEAN. (See, Mark, I told you the
overview memo should say that explicitly.)
5. "The owner data base file will be entered in the root directory." Meaning a
Pilot Directory package root directory? I didn't think we were planning to use
the Directory package on Alpine volumes.
I thought the root directory avoided the use of the Directory package; am
I wrong?
6. "FileStore is assumed to cache the permissions ... in its open file table." I agree
with this statement with regard to open files. But I believe that AccessControl
should also do a certain amount of caching of [ClientID, FileID, access] triples
(or something), so that successive openings of a popular file don't have to
consult Grapevine every time.
Maybe. But an RName is a fat thing.... How likely are we to see open
requests for the same file frequently enough so that we would catch
it in the small cache we could afford?)
7. Not allowing the sum of all quotas to exceed the available space is nice in
principle, but impractical. It leads to underutilization of available space (since
many users don't use their entire quota); and in any event it's difficult to
administer, especially in an environment where users don't pay for storage. (I
do agree, though, that AccessControl should maintain a running total of quotas,
so an administrator can determine how badly the available space is
overcommitted.)
Also, I really don't quite know the space used up since Pilot is snarfing
some up as well. Most of the work is in keeping track of the sum of the
quotas, rather than checking against the space, so it's close to immaterial
in terms of code which decision we make, especially as I would like to
keep hooks in in case we do do this later. Let's make a policy decision:
shall we allow overcommitting the disk space or not?
8. "... it cannot just break the lock, instead it must abort the unlucky
transaction." I thought that in Alpine deadlocked transactions were always
aborted. Am I mistaken?
Right, I meant I couldn't survive Juniper-type broken read locks.
9. IsClientValidForThisFileStore: I don't think this is structured quite right. I
think we need a separate Authenticate operation that simply authenticates an
RName (i.e., that verifies that the RName is valid, and that the client is who he
says he is). Furthermore, this should probably be put in a separate interface,
since the mechanisms for authentication are likely to change over time, as RPC
and our facilities for authentication and encryption mature. The FileStore and/or
RPC should do the authentication; and AccessControl should deal only with
access control, not authentication.
Currently all the authentication calls the system makes are localized in
AC, except for this one from FS. Since authentication is so
closely connected with AC, I prefer to keep my hands on it (i.e., make FS
ask me instead of having a little hunk of code out in FS), but separating
it out into a separate interface is probably okay, let me see what the code
looks like when I get further along. Of course, it already is sort of in a
separate interface, since I'll be calling that interface to do the
authentication.
10. CheckAccessFile: there is a potentially nasty recursion here. If we read the
access control list properties using the general mechanisms, then we have to
open the file first; but opening the file requires access to be checked ... We can
get around this either by some sort of trap-door in the implementation for
reading properties, or by having AccessControl open the file while assuming the
identity of some privileged client who can always open files without having to
read their access control lists.
No recursion, AC does stuff in its "own highly priviledged name" where
necessary.
11. EnumerateAllOwners: let's get together and settle on a convention for doing
enumerations in Alpine. If enumerating a data structure locks it up, then
passing in an enumeration procedure that is called for each item is equivalent to
using your "continuationKey", and seems a lot cleaner. But I don't think that in
general we can afford to lock up data structures while enumerating them, and
the Pilot-style "stateless enumeration" may be preferable.
Addressed in new design, client gets to choose.
12. One thing that's not clear to me from the implementation notes is whether the
internal data structures are modified under the caller's transaction, under a
private transaction, or under no transaction at all (i.e., manipulating raw
FilePageMgr storage). Similarly, are the locks for manipulating file properties
acquired under the caller's transaction or under a private transaction?
We threw out the separate transaction idea because it got too complex and
people still had to block on writing to the log. I should have noted this is
client transaction oriented in the memo, but I got snow blind. AC calls
FS, not FPM. AC only reads file properties, in its own priviledged name,
under the client's transaction.
13. General comment: there aren't any SIGNALs in the interface, though there
are a number of places where it would be natural to use them (instead of return
codes that the caller must test). Since AccessControl is an internal interface (not
exported via RPC), I don't see any reason not to use SIGNALs in the
conventional ways.
Error handling is re-addressed in latest design, a la Roy's memo; I ERROR
on my "programming" errors and on FS "programming" errors, but I still
have "client-of-FS" things ("owner already exists", etc.) coming back as
return codes since I guessed that would be easier for FS to decide how to
send them back to the user, but that's certainly open to change if FS
would prefer SIGNALS (or ERRORs).
Time fleets: as I have started the implementation, it would be a good idea to
wrap up the design changes.....
Karen
*start*
01375 00024 US
Date: 4 Sept. 1981 8:42 am PDT (Friday)
From: MBrown.PA
Subject: Re: Access control design
In-reply-to: kolling's message of 3 Sept. 1981 6:22 pm PDT (Thursday)
To: kolling
cc: Taft, MBrown
I haven't had a chance to more than glance at this exchange of messages, but I
note the following:
3. You could substitute "a user who is currently enabled to create files for this
owner" for "creator", and I think the effect would not be far wrong. The extra
degree of redundancy suggested by Ed seems like a good thing, and should
probably extend to file deletion (i.e. it should be possible to make it impossible to
delete a file without first setting some property to the correct state.)
5. By "root directory" I mean "root file of logical volume" in the current release.
We can impose any structure we want on this file; it should be simple enough
not to require the ComSoft directory package. For Trinity, there will be a
primitive "root directory" facility that allows multiple "root files" without recourse
to the ComSoft package.
7. I think that overcommitting quotas should be an administrative decision, not
imposed by our implementation. We should make it easy for the administrator to
tell what he is doing.
8. We don't support Juniper style broken locks; we'll use another mechanism to
validate remote caches.
More to come ...
--mark
*start*
01199 00024 US
Date: 4 Sept. 1981 1:04 pm PDT (Friday)
From: Kolling.PA
Subject: Re: Access control design
In-reply-to: MBrown's message of 4 Sept. 1981 8:42 am PDT (Friday)
To: MBrown
cc: kolling, Taft
How about this for the file access problem:
two access lists: default minimum
--------------- ------- ---------
read world owner, AW
modify owner, OwnerCreateListWhenFileWasCreated null
These control access to the file itself. Access to the access lists is also controlled
by these lists, but in addition, we always also permit access to the owner, AW,
and the current ownerCreateList members.
I'm not sure what you mean by "The extra degree of redundancy .... should
probably extend to file deletion (i.e. it should be possible to make it impossible to
delete a file without first setting some property to the correct state.)" Currently
file deletion, like truncation and extending, requires file modify priviledges, so
this minimum null setting would protect against it. Is that what you want?
There isn't much difference between deleting a file and scribbling all over it....
I'll put in some way to make overcommitting quotas an option.
Karen
*start*
00730 00024 US
Date: 4 Sept. 1981 1:58 pm PDT (Friday)
From: kolling.PA
Subject: Re: Access control design
In-reply-to: MBrown's message of 4 Sept. 1981 8:42 am PDT (Friday)
To: MBrown
cc: kolling, Taft
Shoot. I just realized the following: In the old scheme, all FS had to do to read
or write any file property was check its open file table, just like it handles read
or write page requests. In the new scheme, if the open file table check says no,
FS needs to recognize a file property which is an access list, so it can call AC to
see if it can do it anyway. Do we want this extra stuff when all this is going to
disappear anyway? This null minimum setting sounds nice, but does it ever
really get used?
Karen
*start*
01586 00024 US
Date: 4 Sept. 1981 3:21 pm PDT (Friday)
From: Taft.PA
Subject: Re: Access control design
In-reply-to: kolling's message of 3 Sept. 1981 6:22 pm PDT (Thursday)
To: kolling
cc: Taft, MBrown
1. Caching access permissions: I didn't think this through or explain it very
carefully. The case I'm really concerned about is a single client accessing many
different files that have the same access control lists. If I retrieve 100 files that
are all readable by CSL↑.pa, I don't want Alpine asking Grapevine each time.
IsMemberClosure is quite slow for large lists (I think testing for membership in
CSL↑.pa takes several seconds), and in any case we don't want to swamp
Grapevine.
I think the right thing to do is to implement a simple cache in front of
IsMemberClosure. When IsMemberClosure[name, member] returns "yes", put the
pair [name, member] in the cache. Before calling IsMemberClosure[name,
member], check the cache for presence of that pair and just return "yes" if so.
The cache can be of fixed size and relatively small (say, 25 or 50 entries); it
should be managed LRU, and there should also be a timer that flushes the cache
periodically (say, once per hour) so that the cache doesn't get too far out of sync
with Grapevine.
2. I don't see any difficulty with FileStore giving special treatment to access
control properties and calling AccessControl before modifying them (rather than
just looking in the OpenFileMap). I think that being able to write-protect a file
against yourself is sufficiently useful that we should do this initially.
Ed
*start*
01154 00024 US
Date: 4 Sept. 1981 3:38 pm PDT (Friday)
From: kolling.PA
Subject: Re: Access control design
In-reply-to: Taft's message of 4 Sept. 1981 3:21 pm PDT (Friday)
To: Taft
cc: kolling, MBrown
1. I thought you meant stashing the [ClientID, FileID, access] away. Yes,
stashing [ClientID, accesslistname] is a lot more useful and quite desirable. I'll
put it in. (By the way, IsMember is a pig particularly for lists of more than one
level; Andrew explained why to me at one point but I forget. I think they
btree down for each entry as opposed to a direct search thru the one level list.
IsMember[CSLOnly↑] + IsMember[CSLTemp↑] goes like the wind compared to the
equivalent IsMember[CSL↑].)
2. Mark isn't around, so I don't know how he feels about the access stuff.
3. In rereading your memo, I see that I didn't properly answer your question
about CheckAccessFile. In my head I had an arrow from AC to FP directly, but
this is wrong, I think, or at least incomplete. I don't think I know the answer
yet, because I don't know as much as I thought I did about how the various
boxes interact, so I'm now looking at it.
Karen
*start*
00620 00024 US
Date: 4 Sept. 1981 4:12 pm PDT (Friday)
From: kolling.PA
Subject: Here's Mark's solution for CheckAccessFile.
To: Taft
cc: mbrown, kolling
FS.OpenFile makes an OpenFileID at the beginning of OpenFile.
FS calls AC.CheckAccessFile[openFileID blah blah].
AC calls FileProperties.something to read the access list it wants[openFileID].
FP calls into FS to read the leader page, which makes the locking and looking in
the log happen right.
I'll go ahead with the extra stuff for the access lists as you suggest, unless Mark
says differently. I think this resolves all the questions raised?
Karen
*start*
05668 00024 US
Date: 4 Sept. 1981 5:15 pm PDT (Friday)
From: MBrown.PA
Subject: Re: OpenFileMap and LogMap, version 0
To: Taft
cc: Kolling, MBrown
I like the style, organization and general content of this memo very much. As
usual, I have my own opinions on some of the details:
OpenFileMap:
1. I think that an OpenFileMap.Handle should be longer than 2 words. Reason:
we want it to be very difficult for a client to "guess" the Handle that someone
may be assigned on opening a file; a Handle is a "capability" in the technical
sense (not quite since a Handle only works if someone has actually opened a file;
we don't open files implicitly when an unknown Handle is presented!)
As you suggest, we can use one word of the Handle to index into the internal
data structure. The other words can hold a random number generated at Open
time.
2. The current FileStore interface includes a type called Handle, representing
one client's session with one server. The design given there binds a Handle into
the OpenFileHandle. This seems to imply that a FileStore.Handle, not a ClientID,
should be stored in the OpenFileMap entry. I am assuming that the ClientID can
be gotten by following the FileStore.Handle. I was also assuming that any
Handle that corresponded to an entry in the OpenFileMap was valid;
ValidateHandle would not take ClientID as a parameter.
Your way of structuring things, which seems viable also, seems to imply that
calls that take a FileStore.OpenFileHandle also take a FileStore.Handle, and that
on each call we first validate the Handle, then check consistency between the
ClientID in the Handle and the ClientID in the OpenFileHandle. My point (1) is
an alternative solution: make the OpenFileHandle hard to forge, and just take it
on faith that if you are passed an OpenFileHandle that looks valid, it is valid for
the client who passed it to you.
3. A hash table does seem to be the preferred data structure. Since the entries
are relatively large compared to the size of a pointer, hashing with separate
chains for resolving collisions is a good method. Since you have freedom to
choose hash codes, you can arrange for no collisions as long as the number of
open files is less than the number of hash buckets.
LogMap:
1. The log map is specific to files; a data manager that writes log records and
defers updates will have its own version of the log map.
2. I don't understand why you assume that the log map represents only one state
of a FileStore. I was assuming all along that for any page of a file there were
two possible views: the view seen by all transactions that have done only reads,
and the view seen by the single transaction (if any) that has done a write. I
don't see that it is any harder to implement this view than the simpler one you
suggest. I don't feel strongly that allowing reads to go "underneath"
uncommitted writes is necessary, so if you can explain why it is harder, we can
flush it. I thought it came for free with the deferred update philosophy we are
using.
My idea was: in response to the "prepare to commit" command, a FileStore would
upgrade all update locks to write locks. Then the transaction commits. At the
instant it commits (signified, in the no crash case, by a change of state in the
volatile transaction representation), the log map has in effect changed: all readers
will see the new data recorded in the log. Before the commit, only the writing
transaction saw the new data. The write lock before commit assures consistency.
After commit, locks on the updated pages are released (or downgraded to read in
case of checkpoint.)
After commit, reading one of the just written pages from ANY transaction will
"hit" in the log map and will read the new data from the log. As a side effect, it
will write the new data to FilePageMgr and then erase the log map entry. In no
case will we allow two updates to the same page from different transactions to
exist in the log map. One of the two must be committed, and we will force it
through FilePageMgr before making the second entry. (This is a VERY unlikely
case!)
3. Update before commit is a separate issue. If update is done before commit,
then a log map entry might be made but it would be only for the purposes of
undo or recovery. It would not "hit" on a log map lookup, since the truth is in
the file already (as well as in the log.)
4. I feel that the log map should be implemented in virtual memory. The log
map should have a reasonable size. It is a sign of TROUBLE if the server is
deferring huge quantities of work (has a huge log map.) The optimization of
writing "new" pages without deferring (the "file high water mark" idea) will
allow us to avoid generating a huge log map for a file transfer. Database
transactions are short and update a few pages (I would say 5 pages typical, 100
pages ok but unusual, we would never expect to process such transactions at a
high rate.)
The log map is consulted on each read operation. BTree lookups are slow (even
if the pages are all in main store), so I feel we don't want to do one on every
read. In particular, the readers of a file that has not been updated recently
should not pay for lengthy log map searches.
I have no concrete design proposal. I think a hybrid structure might be best: a
hash table keyed on FileID, where the entries are balanced trees (red-black or
whatever). Then a file with little update activity has a tiny tree to search, and
a file with lots of update activity has a big tree to search, which only seems
fair. I have a balanced tree package that could be adapted to the purpose with
little work.
--mark
*start*
01120 00024 US
Date: 8 Sept. 1981 1:33 pm PDT (Tuesday)
From: MBrown.PA
Subject: FS version 5
To: Kolling
---------------------------
Date: 4 Sept. 1981 6:12 pm PDT (Friday)
From: kolling.PA
Subject: FS version 5
To: mbrown
cc: kolling
As you requested, I put off reading your latest FileStore memo to get started
implementing AccessContol. Tonight I scanned the FS memo and found:
1. It believes DeleteFile needs owner "create/delete" permission. I think this is
the way we had stuff set up originally and then you asked me to change it so
Delete only needed file modify.....
2. ReadNextOwner needs the new param to specify locking the whole file or
walking thru reading and releasing read locks (latter most;y avoids the noted
problem of conflicts).
3. Note on WriteOwnerAccessLists that we OR in minimums without complaint
and don't write "empty" lists.
4. Defaults are missing in places on this page.
5. I have some debugging routines (EnumerateOwnerDatabase,
ReportAccessControlStatistics) but maybe they shouldn't be noted here.
------------------------------------------------------------
*start*
01828 00024 US
Date: 8 Sept. 1981 2:14 pm PDT (Tuesday)
From: MBrown.PA
Subject: Re: Access control design
To: Kolling
cc: Taft, MBrown
Here are more thoughts on some of the points discussed in last week's exchange
of messages.
"What AccessControl expects from FileStore"
"AccessControl should maintain a running total of quotas, so an administrator can
determine how badly the available space is overcommitted." Such a running
total is a resource that is difficult to maintain in a transaction-consistent way
without serializing all transactions that update it. We have gone to some trouble
in the design to avoid such bottlenecks. I suggest that the administrator should
run a little program that enumerates the owners and gathers whatever statistics it
desires. Since the server is not keeping this total, it cannot enforce the "do not
overcommit quotas" restriction, but we seem to agree that overcommitting quotas
is ok anyway.
"Some comments"
Capability to delete files: the design that I have advocated is that file
lengthening/shortening should be treated as equivalent to any other file
modification, and require only file modify permission (even though the
length-changing operations debit or credit the owner database.) But file deletion
should require the same capability that the creator had. This is impractical since
we cannot store away the contents of the owner create list at the time a file is
created, so we compromise and simply require the client doing the deleting to be
on the owner create list at the time of the deletion.
Ability to protect file against owner modify: as Ed says, this seems worth the
trouble to do from the start. The solution of allowing the modify minimum to be
null, and allowing access list modify only to members of owner create, seems
fine to me.
--mark
*start*
00793 00024 US
Date: 8 Sept. 1981 2:34 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: Access control design
In-reply-to: MBrown's message of 8 Sept. 1981 2:14 pm PDT (Tuesday)
To: MBrown
cc: Kolling, Taft
Sigh, am I tired of reorganizing this memo on a daily basis and then having to
start over on the code from scratch..... The way it stands now, null permission
for file modify can be set, and foo = {read/modify} permission for the access lists
themselves is file foo + (owner, AlpineWheels, ownerCreateListNow). Unless
there is an uproar, that's how I plan to leave it.
I will reflect file deletion requiring owner create permission, but why a client
needs special permission just to throw away the leader page after having
previously thrown away the entire file.....
Karen
*start*
01569 00024 US
Date: 8 Sept. 1981 2:53 pm PDT (Tuesday)
From: Taft.PA
Subject: Handles, authentication, etc.
To: MBrown, Kolling
cc: Birrell, Boggs, Nelson, Taft
I've talked with Andrew about how an RPC implementor can determine who
called it. The problem, briefly stated, is that if RPC is doing the connection
management and the secure communication (if any), then the implementor (the
guy called via RPC) will have trouble identifying of the client who called it --
it knows only that the call originated from SOME authentic client!
One way to deal with this is for the implementor, at authentication time, to
manufacture and pass to the client some sort of capability (it must be more than
a simple handle, since it must be difficult to forge).
However, since RPC is doing connection management and is already maintaining
ConnectionIDs for its own purposes, it seems straightforward to make this
mechanism available for client/implementor identification as well. A simple
(though slightly inelegant) implementation is this: if a procedure in a remotely
exported interface is declared as taking an argument of type RPC.ConnectionID,
then when that procedure is called, the caller's ConnectionID argument is
ignored. Instead, the RPC on the callee's side substitutes the ConnectionID
corresponding to the (secure) connection over which the call came.
It's an open issue whether the ConnectionID should be generated by the
implementor or the RPC. Speaking as an implementor, I would rather generate it
myself; but it's not really that important.
Ed
*start*
00738 00024 US
Date: 8 Sept. 1981 3:16 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: Access control design
In-reply-to: MBrown's message of 8 Sept. 1981 2:14 pm PDT (Tuesday)
To: MBrown
cc: Kolling, Taft
Quotas: the only transactions that need be serialized to maintain a consistent
handle on the quota space left are the ones that do AddOwner or
ChangeOwnerSpaceQuota. These transactions are rare birds. If the
quota-space-left is kept in the first page of the owner database file or somesuch
and such a transaction locks that page exclusively, I think this serialization will
happen properly with no other effort, and this will give the system administrator
more flexibility than our not supplying this feature.
Karen
*start*
00447 00024 US
Date: 8 Sept. 1981 3:19 pm PDT (Tuesday)
From: Taft.PA
Subject: Re: Access control design
In-reply-to: kolling's message of 8 Sept. 1981 2:34 pm PDT (Tuesday)
To: kolling
cc: MBrown, Taft
I'm with Karen on this one: why distinguish between the capability to delete a
file and the capability to destroy all the information within it? I think having
modify permission also control deleting the file is quite adequate.
Ed
*start*
00249 00024 US
Date: 8 Sept. 1981 3:34 pm PDT (Tuesday)
From: kolling.PA
Subject: authentication
To: mbrown
cc: kolling
I'm confused about rpc authentication and how this affects clientIDs, etc.
Sometime when you have a few minutes.....
*start*
00574 00024 US
Date: 8 Sept. 1981 4:06 pm PDT (Tuesday)
From: kolling.PA
Subject: AccessLists
To: mbrown
cc: kolling
I noticed that in FS version 5, CreateFile etc. take an AccessList, which is a LIST
of RNames. Is that what you plan to hand off to me or will this be changed to
allow the client of FS to say something like:
[owner, AlpineWheels, world: BOOLEAN, xxxxx: LIST OF RNames]
or somesuch? If you plan to leave it like this, how shall AC recognize "world",
etc.?
Let me know what you decide about quotas and DeleteFile before you leave.
Thanks.
*start*
00536 00024 US
Date: 8 Sept. 1981 4:14 pm PDT (Tuesday)
From: MBrown.PA
Subject: Re: Access control design
In-reply-to: kolling's message of 8 Sept. 1981 3:16 pm PDT (Tuesday)
To: kolling
cc: MBrown, Taft
blush, I overlooked that. yes, the contention is not too bad over the allocation
totals.
i don't feel strongly about the delete protection thing. since Ed likes your
design, go with it.
i wasn't trying to throw a monkey wrench into your design, it just seems that
way! now i understand it all much better.
--mark
*start*
00639 00024 US
Date: 8 Sept. 1981 4:19 pm PDT (Tuesday)
From: MBrown.PA
Subject: Re: AccessLists
In-reply-to: Your message of 8 Sept. 1981 4:06 pm PDT (Tuesday)
To: kolling
cc: mbrown
I was hoping that there would be some conventional way for a client to say
owner and world, just as there is in Grapevine (the "*" RName, etc.) This would
avoid having to learn some tiresome record constructor ...
I don't think that a client gets any say about AlpineWheels, does he? If so, I
would favor having him say it in a way that is compatible with the way for
owner and world.
See what Ed thinks, then do what seems right.
--mark
*start*
03110 00024 US
Date: 8 Sept. 1981 4:39 pm PDT (Tuesday)
From: Birrell.pa
Subject: Re: Handles, authentication, etc.
In-reply-to: Taft's message of 8 Sept. 1981 2:53 pm PDT (Tuesday)
To: Taft
cc: MBrown, Kolling, Birrell, Boggs, Nelson
Some slight modifications of your description....
For "ConnectionID", Bruce and I have been using the less over-used term
"ConversationID". It isn't really a connectionID anyway, because there can be
multiple independent calls simultaneously in a single conversation (through a
single interface). Our analog of "connectionID" is really [calling-host,
calling-process].
Since the conversationID appears as an argument of the procedures in the DEFs
file, the caller will need to provide one. I propose to use this value as the
conversationID for the call (which implies to both hosts which encryption key to
use), and then to pass the conversationID as argument to the call to the
implementor. So the argument given by the user isn't "ignored".
Some extra advantages of this scheme are that implementors can use the
conversationID instead of a handle to locate instance data (particularly if we can
allow the implementor to allocate the conversationID), and that within a
particular interface not all procedures need to be encrypted. It also means that a
single conversation can use multiple DEFs files.
The problem with allowing the implementor to allocate the conversationID is that
in the RPC runtime I need to map conversationID to encryption key (in the
caller and the callee host). This requires that the conversationID be unique
within a single host (mds?). If the implementor wants 32 bits, I'd need an extra
16 for uniqueness (on the wire, not as client-percieved arguments), which
extends our already over-large per-packet length overhead. The advantage of
allowing the implementor to allocate them is that it gives the implementor much
freedom: he can use pointers, array indexes, or hash keys depending on what is
best for the particular application. I think the advantages outweigh the
disadvantages, but bear in mind that there is very little chance these days that
we'll be able to squeeze 256 words of argument (or result) into a single packet.
Packet length may also become the dominant performance factor on a Dorado
using the 3MB Ethernet (but not the 10MB Ethernet).
Bruce is still considering whether this scheme causes any undue problems for
the stub generator; otherwise, it seems to be the best we have.
On an independent topic, Bruce and I now think that providing (almost) the full
semantics of SIGNALs and ERRORs is not difficult, so it should appear in an
early second version (or possibly first version). The "almost" is because, if a
signal/error is raised by the implementor in the server and that signal/error isn't
declared on the DEFs file, we have no reasonable way to represent the signal on
the calling host. Therefore we propose to translate any such signals/errors into a
single ERROR declared publicly in the RPC runtime. This agrees with the
exception methodology advocated by Roy and me, anyway.
Andrew
*start*
04796 00024 US
Date: 8 Sept. 1981 5:17 pm PDT (Tuesday)
From: Taft.PA
Subject: Re: Handles, authentication, etc.
In-reply-to: Birrell's message of 8 Sept. 1981 4:39 pm PDT (Tuesday)
To: Birrell
cc: Taft, MBrown, Kolling, Boggs, Nelson
Thank you for correcting my misrepresentations.
I've changed my mind about wanting to be able to assign ConversationIDs
myself. I now realize that all the state information kept by the Alpine server is
associated with either a transaction or an open file, identified by an ID or handle
passed as a separate argument. So the ConversationID will be used only for
validation and not for lookup; its only required property is local uniqueness.
Therefore, I'm happy to let RPC be the source of ConversationIDs.
With regard to SIGNALs and ERRORs, my present feelings on the subject are
summarized in the attached message (which is a reply to a message from Mark
which, for fairness' sake, I've excerpted also). If I understand your position
correctly, I think we are having a violent agreement. In any case, I certainly
agree that the only SIGNALs that cross machine boundaries should be ones
declared in the remote interface.
-------------------
Date: 25 Aug. 1981 4:56 pm PDT (Tuesday)
From: MBrown.PA
Subject: File Store document, version 5
To: Kolling, Taft
cc: MBrown
I am now trying to flesh out the treatment of errors in the interface. I am in
general agreement with the stylistic conventions suggested in Roy's memo
"Guidelines for Signalling in Cedar"
(ivy]Levin>Cedar>SignallingGuidelines.bravo.) But we need to figure out how
to apply these ideas to our system.
One general principle that seems useful: if an error detected by a procedure is
likely to the the result of a sequence of calls outside the immediate scope of the
caller, the error should be raised as an ERROR and not by a return code. Failure
of credentials (except on Login itself) and transaction aborts are examples of this
type of error.
Abstraction failures represent a fairly well-delimited class. For instance, a create
owner failure due to the owner database being full is an abstraction failure, as is
a hard disk error. Is a SetLength or CreateFile failure due to out of disk quota an
abstraction failure? I guess so.
I am assuming that we shall not require SIGNALs in our interface, only
ERRORs. Remote RESUME is too much to ask for right now. Andrew seems
willing to provide remote ERROR in a timely fashion.
A remote ERROR will unwind the remote call before we see the error locally.
This means that what Roy calls "Client programming errors" (e.g. RName too long
or containing an invalid character) must return enough information to diagnose
the problem that the remote machine found.
The type of error for which the correct interface style is unclear is "read page
off end of file", "owner already exists in call to create owner", etc. Offhand I am
inclined to report these as ERRORs although they do not fit neatly into any of
Roy's categories (they come closest to being calling errors, but they clearly
should be documented in the interface, which calling errors are not.) Maybe
this means that I am wrong and these should really be reported by return codes.
--mark
-------------------
Date: 26 Aug. 1981 5:18 pm PDT (Wednesday)
From: Taft.PA
Subject: SIGNAL, etc.
In-reply-to: MBrown's message of 25 Aug. 1981 4:56 pm PDT (Tuesday)
To: MBrown
cc: Kolling, Taft
Your suggested treatment of SIGNALs across a remote interface seems like the
only reasonable one from the point of view of robustness of the server --
independent of whether or not RPC is eventually able to support SIGNALs in all
their generality.
I think you are effectively proposing that RETURN WITH ERROR be the only
allowed use of SIGNALs across a remote interface. It seems like the standard
Mesa semantics of RETURN WITH ERROR are exactly what we want. That is, it
unwinds the callee's stack; but catching the SIGNAL, passing parameters, etc.,
are done in the conventional way from the client's point of view. Can we get
the RPC guys to give us these semantics, perhaps even with the same syntax?
As you say, for debugging purposes this treatment may be somewhat of a
nuisance. But I think there is an easy solution. The top-level procedures in the
server (i.e., the ones exporting the remote interface) will have to catch all client
programming errors and abstraction failures coming up from below and turn
them into RETURN WITH ERROR on remotely exported SIGNALs. The trick is to
have the top-level procedures catch these errors conditionally, based on some
global debugging switch. If you turn on the debugging switch, these errors will
not be caught, and the server will land in CoPilot with the server's stack still
intact.
Ed
-------------------
*start*
00420 00024 US
Date: 8 Sept. 1981 5:44 pm PDT (Tuesday)
From: Taft.PA
Subject: OpenFileMap and LogMap, version 1
To: MBrown, Kolling
cc: Taft
This version addresses all Mark's comments and also incorporates the latest
thinking about RPC client validation. There are no other important changes.
I propose next to produce a complete interface, and also to think about a possible
TransactionMap interface.
Ed
*start*
01455 00024 US
Date: 8 Sept. 1981 5:49 pm PDT (Tuesday)
From: Birrell.pa
Subject: Re: Handles, authentication, etc.
In-reply-to: Taft's message of 8 Sept. 1981 5:17 pm PDT (Tuesday)
To: Taft
cc: Birrell, MBrown, Kolling, Boggs, Nelson
But you presumably still need to be able to map ConversationID's into user
names?
I may not have been clear about exceptions. What I think is not difficult is to
provide is precisely the local semantics of SIGNALs and ERRORs, including use
of RESUME and the order of events for unwinding. (I propose to do this by a
call-back style of arrangement.) The only variation will be the handling of
signals/errors coming out of an interface but not defined in that interface, which
will all be converted into RPCSignals.UndefinedError. If the caller ends up in
the debugger with an uncaught signal, the server call stack will still exist
(subject to the extent to which you employ RETURN WITH ERROR). Of course,
the server may apply a timeout to the call quite rapidly . . . .
It's possible that I'm wrong about the easiness of this, so as a fall-back position
the SDD work provides an existence proof of supporting ERRORs only, with the
slightly strange unwind sequence.
One thing I'm not sure about is whether RPCSignals.UndefinedError should be
raised in the server (where it will presumably be uncaught) or dutifully passed
back to the caller and raised there. Maybe we need a debugging switch.
Andrew
*start*
01260 00024 US
Date: 9 Sept. 1981 3:18 pm PDT (Wednesday)
From: kolling.PA
Subject: FS interface
To: Taft
cc: kolling
There's one thing we need to clear up about the FS interface. AC carries around
AccessLists that look something like:
owner, world, AlpineWheels: BOOLEAN, plus two RNames.
FS defines an AccessList as a LIST of RNames. So we need some way for the FS
client to say (a) world, (b) owner, and (c) although we may not document this
in the interface, AlpineWheels. Mark pointed me to the Grapevine conventions
and asked me to run them past you. What do you think of allowing the FS LIST
to contain the "distinguished" names Owner, World (also *), and AlpineWheels
as well as the two other RNames. Caps don't matter, of course. Note that we do
need the AlpineWheels field because it is typically turned off for file modify to
prevent wipe outs. The FS stuff should probably note that the following
minimums are ORed in, and the defaults are as follows (disregarding AWs):
File access: default minimum
read world owner
modify owner null
File access list access is setting of appropriate file access plus owner and current
owner-create list.
Owner access: default minimum
create owner owner
read/set owner owner
Karen
*start*
01694 00024 US
Date: 9 Sept. 1981 4:30 pm PDT (Wednesday)
From: Taft.PA
Subject: Re: FS interface
In-reply-to: Your message of 9 Sept. 1981 3:18 pm PDT (Wednesday)
To: kolling
cc: MBrown, Taft
I agree with the notion that clients of FileStore (and perhaps of AccessControl
also) should traffic in lists of RNames uniformly, and that the optimized internal
representation of certain RNames should be hidden. I also have no problem with
introducing distinguished "RNames" that are interpreted by Alpine rather than
being passed on to Grapevine -- though we should limit the extent to which we
do this, since such "RNames" can't be members of Grapevine groups.
I think the "uniform" approach can be pushed a bit further, as follows: instead of
having a distinguished RName "Owner", how about letting the
FileAccessList.owner bit stand for the actual owner's name. For example, if the
owner of file Foo is "Taft.pa", and I ask to change the read access list of Foo to
["Taft.pa", "Kolling.pa"], then "Taft.pa" should be represented internally by
setting the owner bit, and only "Kolling.pa" should actually be stored as a
string. Obviously, the inverse thing should happen when converting a
FileAccessList to a list of RNames.
A similar argument can be made for AlpineWheels. I think we've agreed that
each server will have an independently-settable RName for local wheels.
Suppose we have a server named "S" whose wheel RName is "S.Alpine". Then a
reference to "S.Alpine" in a client-supplied list of RNames should operate on the
FileAccessList.alpineWheels bit.
There is no similar argument with respect to "World" or "*", since they don't
stand for any particular RName.
Ed
*start*
00467 00024 US
Date: 11 Sept. 1981 7:27 pm PDT (Friday)
From: kolling.PA
Subject: AccessControlDesign7
To: Taft
cc: MBrown, kolling
AccessControlDesign7.bravo incorporates the latest changes and lives on
Alpine>doc. It has solutions for all the problems I know about (I think), except
two I am currently ruminating on (a slight problem with releasing locks, and all
the handshaking that goes on during read/write properties that are access lists).
Karen
*start*
01092 00024 US
Date: 14 Sept. 1981 9:04 am PDT (Monday)
From: Taft.PA
Subject: AccessControl
To: Kolling
cc: MBrown, Taft
It seems somewhat redundant for AddOwner to take the space quota and access
lists as arguments given that there also exist ChangeOwnerSpaceQuota and
ChangeOwnerAccessList procedures. A cleaner factoring of function would be
for AddOwner simply to create the new owner with zero quota and empty access
lists. The client may then call the ChangeXXX procedures as necessary. (I don't
think there are any access permission problems here: the client must be an
AlpineWheel to call AddOwner and therefore will certainly be able to call
ChangeXXX.)
Also, is there enough flexibility in the AccessControl design for us to consider
associating additional information with each owner entry? I'm thinking that it
might be worthwhile to have some default file properties. That is, when a file is
created for an owner, the file's properties (e.g., access control lists) are initialized
to values specified in the owner entry rather than to system-wide defaults.
Ed
*start*
00753 00024 US
Date: 14 Sept. 1981 10:50 am PDT (Monday)
From: Kolling.PA
Subject: Re: AccessControl
In-reply-to: Taft's message of 14 Sept. 1981 9:04 am PDT (Monday)
To: Taft
cc: Kolling, MBrown
I think it would be a bother to have to call two different procedures to do one
thing, i.e., create a new owner. Presumably AddOwner and
ChangeOwnerSpaceQuota will not duplicate code unnecessarily.
We kicked around tucking file defaults in the owner entries at one point but
decided against it, as I recall for two reasons: (1) not tying the owner and file
stuff together any more than is necessary (already unfortunately we have crept
away from this objection somewhat), and (2) it would bump the owner record
space over one page size.
Karen
*start*
00309 00024 US
Date: 14 Sept. 1981 11:19 am PDT (Monday)
From: Kolling.PA
Subject: Re: AccessControl
In-reply-to: Kolling's message of 14 Sept. 1981 10:50 am PDT (Monday)
To: Kolling
cc: Taft, MBrown
Also, having two calls is less optimal internally, since I have to snarf up the
owner page twice.
*start*
02261 00024 US
Date: 14 Sept. 1981 12:04 pm PDT (Monday)
From: Kolling.PA
Subject: some access control problems
To: mbrown, taft
cc: kolling
1. Awhile back I asked for the ability to release read locks on pages of the owner
database file, for use by enumerations in the unfrozen mode. There is a
problem, I think, with that however. Since locks are in the name of the
transaction, suppose an AlpineWheel (in AlpineWheel mode) tries to do
something directly with the owner database file and calls an unfrozen
enumeration at the same time. Theoretically he/she could be depending on data
that gets written over. What to do? (See related question #3, which might give
us a way out.)
2. When AC goes after a page of the owner file, it first write locks it, then gets
it in some fashion (from buffer or read), then looks to see if it is really the page
it wants (it might not be because of collisions, etc.). If it isn't the desired page,
AC steps to the "next" page, leaving a write locked page behind it, etc. This is
normally fine, as the owner file is set up to be sparse. But if the owner file gets
really full, searches for non-existent owners or searches for empty pages will
lock many pages.
(a) We could say, tough bunny, the system administrator should call
ReorganizeOwnerDatabase.
(b) Alternatively, I could simply have a call on Locks to release the write
lock because I promised I didn't dirty the page. I could avoid getting zapped by
an intervening reorganize (no other writes can cause me any trouble) by not
releasing the lock on page n until I have page n+1. (I don't want to get a read
lock initially and then upgrade it, because of efficiency, interactions become
more complex.)
If it weren't for the problem in #1, I would vote for (b), because it only takes a
few lines of code and two years from now system performance will degrade and
no one will realize why it is happening. As it is, I'm not sure which is
preferable.
3. Also, what will OpenFile do if it sees an open request with all the same
parameters (if the clientName is the same, is the FS.Handle returned by FS.Create
the same?)? I'm just about to read Ed's recent memo and your comments on it, so
perhaps you have already answered this.....
Karen
*start*
02215 00024 US
Date: 14 Sept. 1981 12:21 pm PDT (Monday)
From: Taft.PA
Subject: Re: AccessControl
In-reply-to: Kolling's message of 14 Sept. 1981 11:19 am PDT (Monday)
To: Kolling
cc: Taft, MBrown
I think it's silly for us to consider efficiency in this case -- after all, how often
do you change the owner data base?
This particular issue is unimportant enough that I'm not going to push it any
more. But in general, I think the interfaces exported by Alpine should be lean
and compact: provide one way to do every required function, and let the client
implement combinations of functions as he sees fit.
On another topic, I've discovered a problem with the way EnumerateAllOwners
is specified. I presume you intended that the FileStore remote interface would
include ContinuationKey as an exported type and initialContinuationKey as an
exported variable. (Actually exporting the definitions from AccessControl would
introduce an uncomfortable compilation dependency.) Unfortunately, RPC
doesn't support exported variables, so there's no way for the client to get his
hands on initialContinuationKey except by some sort of LOOPHOLE.
It seems to me that in this case the ContinuationKey can be done away with
altogether; instead, use the owner name itself as the key. That is, replace the
current EnumerateAllOwners with something like:
GetNextOwner: PROCEDURE [ ... previousOwner: OwnerName ← NIL ... ]
RETURNS [ ... nextOwner: OwnerName, ... & its properties ... ];
To start the enumeration, the client passes previousOwner=NIL; and
GetNextOwner returns nextOwner=NIL when passed the "last" owner in the data
base.
Of course, this means that the implementation has to effectively look up
previousOwner in order to find the right place in the enumeration. I can't
imagine that efficiency will be an issue when enumerating the owner data base;
but if it turns out to be, there are some obvious optimizations (e.g., the
implementation can remember the name of the owner that it gave out most
recently, along with its physical position in the data base).
This is the Pilot-style "stateless enumeration", which I have come around to for
various reasons that I will be happy to explain.
Ed
*start*
00389 00024 US
Date: 14 Sept. 1981 12:30 pm PDT (Monday)
From: Kolling.PA
Subject: Re: AccessControl
In-reply-to: Taft's message of 14 Sept. 1981 12:21 pm PDT (Monday)
To: Taft
cc: Kolling, MBrown
I'll chance over to your form of Enumerate; I had it set up with the
continuationKey to avoid the double read, but caching the last thing will almost
always avoid that extra read.
*start*
00375 00024 US
Date: 14 Sept. 1981 12:32 pm PDT (Monday)
From: Kolling.PA
Subject: Re: AccessControl
In-reply-to: Taft's message of 14 Sept. 1981 12:21 pm PDT (Monday)
To: Taft
cc: Kolling, MBrown
My objection to spliting add owner is not so much the efficiency, as the fact
that I think the current calls directly correspond to the things a user will want
to do.
*start*
00920 00024 US
Date: 14 Sept. 1981 12:43 pm PDT (Monday)
From: Taft.PA
Subject: Re: some access control problems
In-reply-to: Kolling's message of 14 Sept. 1981 12:04 pm PDT (Monday)
To: Kolling
cc: mbrown, taft
1. "Unfrozen" enumeration seems like a bad idea, for the reason you state.
Perhaps it should be flushed. If the client REALLY doesn't care about the
consistency of the enumeration with respect to anything else (as may sometimes
be the case), he can do it under the "nil" transaction.
2. I concur with your strategy (b). Releasing locks based on internal knowledge
about the state of the abstraction is perfectly ok -- I just don't want such a
capability exported to the client.
3. Each OpenFile creates a distinct handle. I don't think I'd want it any other
way -- otherwise there would be horrible confusion if a single client ran
multiple programs that happened to open the same file.
Ed
*start*
00375 00024 US
Date: 14 Sept. 1981 12:49 pm PDT (Monday)
From: Kolling.PA
Subject: Re: some access control problems
In-reply-to: Taft's message of 14 Sept. 1981 12:43 pm PDT (Monday)
To: Taft
cc: Kolling, mbrown
Yes, but releasing locks for 2 has the same problem as releasing locks for
unfrozen enumeration; in fact, it is worse, because it releases write locks.
*start*
02265 00024 US
Date: 14 Sept. 1981 1:02 pm PDT (Monday)
From: Nelson.PA
Subject: ConversationIDs
To: Birrell, MBrown, Taft, Kolling, Boggs
cc: Nelson
I have now read through the msgs from last week. Andrew's final one stated our
position accurately: From the RpcRuntime's point of view, (resumeable) signals
and callback procs are already handled. From Lupine's view, the hair has not
settled, but it looks promising. Only transparent syntax and semantics are
planned, not anything funny. The speed of signals, however, is another matter.
On conversationIDs, as I understand it, Lupine's job is simply to look for one in
every client argument list. If there is one, in addition to being marshaled in
the normal way, it will also be passed to the client's RpcRuntime at the start of
the call. I propose the following syntactic arrangments:
Assume that the ID is called Conversation.ID (separate interface).
Lupine will allow any number of TYPE defs such as
SecureChannel: TYPE = Conversation.ID;
RpcChannel: TYPE = SecureChannel;
If an interface proc has an arg list of the form
GetPage: PROC [ ..., channel: RpcChannel, ...] RETURNS [...];
then the 'channel' arg will get passed to the RpcRuntime as the ID. Note that it
need not be the first arg, altho this leaves open what happens if there is more
than one--either error, or use the first (prob. error). Or we could decide that it
must be first. Also note that it is not allowed to be embeded in anything; it
must be a top-level argument.
If desired, we could get more mileage from IDs by declaring them like
Conversation.ID: TYPE = RECORD [
id: PRIVATE ...,
handle: REF (or LONG POINTER) ];
where 'handle' is for the client. Not clear exactly what this buys; Lupine can't
send handle↑ because the referent type is unknown. Perhaps this is a bad
suggestion.
On RpcSignals.UndefinedError, I believe it should pass through by default, but
there should be an RPC.ExportInterface optional argument that can override this.
Go for local and remote transparency when in doubt, but leave the needed hooks
for robust server operation.
I would also rename this ERROR to RpcSignals.UncaughtRemoteSignal. Less
confusing to passersby (and me) who stare at the debugger.
Bruce
*start*
00433 00024 US
Date: 14 Sept. 1981 1:06 pm PDT (Monday)
From: Kolling.PA
Subject: mind slip
To: taft,mbrown
cc: kolling
"of course" (now that I think of it, sigh), the transaction where an AlpineWheel
munges the owner file will get screwed up regardless of whether or not AC
releases the locks, since the locks don't protect the client and AC from each other
anyway, so I don't suppose it matters from that point of view.
*start*
00442 00024 US
Date: 14 Sept. 1981 1:10 pm PDT (Monday)
From: Kolling.PA
Subject: no transaction
To: taft, mbrown
cc: kolling
I haven't been able to find the latest documentation on the "nil" transaction. Are
we actually going to support this? If so, I assume AC should be prepared for
calls with some distinguishing NIL trans and what it should do with these is
carry them out immediately (i.e., action + phase 1 and 2)?
Karen
*start*
00710 00024 US
Date: 14 Sept. 1981 1:35 pm PDT (Monday)
From: Taft.PA
Subject: Re: some access control problems
In-reply-to: Kolling's message of 14 Sept. 1981 12:49 pm PDT (Monday)
To: Kolling
cc: Taft, mbrown
I claim that allowing AccessControl to release locks internally (yes, even write
locks) is qualitatively different from providing such a capability to the clients.
Maintaining the consistency of the AccessControl abstraction is the only thing
that matters. The hash probe path by which the implementation locates an
owner record is not part of the abstraction, so it doesn't need to be locked. You
need to lock only those records that you actually read out and give to the client.
Ed
*start*
00472 00024 US
Date: 14 Sept. 1981 3:16 pm PDT (Monday)
From: kolling.PA
Subject: releasing write locks is a no-no.
To: taft, mbrown
cc: kolling
Here's an reason AC can't release writelocks whilst searching down a hash
chain: suppose the page it is releasing the lock for (because it thinks it isn't
interesting) is actually a page it wrote previously in this transaction. Gorp. So I
guess we'll just have to degrade as the owner database files become full.
*start*
01390 00024 US
Date: 14 Sept. 1981 5:34 pm PDT (Monday)
From: Taft.PA
Subject: FileStore memo(s) and defs files
To: MBrown, Kolling
cc: Taft
I've revised the FileStoreDesign memo and split it into two parts: "FileStore public
interfaces", intended for clients, and "FileStore interface internals", intended for
implementors. These documents are stored on [Ivy]<Alpine>doc>FileStore6.bravo
and FileStoreInternal0.bravo.
I've also split up the interface into several pieces (for reasons described in
"interface internals"). The current partitioning is quite lopsided and may not be
quite right, but we can shuffle things around easily enough.
Also, I have produced Mesa defs files, and stored them on [Ivy]<Alpine>defs>.
Included are two semi-hypothetical interfaces, Authenticate and RPC, intended to
be implemented by Grapevine and RPC respectively. (Semi-hypothetical in that
I have talked with Andrew about them, but they haven't actually been adopted
by the RPC design group, and no doubt will change.) The set of defs files is
described by AlpineDefs.df; we should endeavor to keep this DF file up-to-date
so we can quickly update our individual disks with the latest stuff.
These defs files are intended to be compilable now, but I haven't yet set up a
Cedar world on my machine, and the Cedar compiler won't run on top of vanilla
CoPilot. That's my next task.
Ed
*start*
01831 00024 US
Date: 14 Sept. 1981 5:40 pm PDT (Monday)
From: Taft.PA
Subject: FYI: Alpine assumptions about RPC and Grapevine
To: Birrell, Nelson, Boggs
cc: MBrown, Kolling, Schroeder, Taft
---------------------------
Date: 14 Sept. 1981 5:34 pm PDT (Monday)
From: Taft.PA
Subject: FileStore memo(s) and defs files
To: MBrown, Kolling
cc: Taft
I've revised the FileStoreDesign memo and split it into two parts: "FileStore public
interfaces", intended for clients, and "FileStore interface internals", intended for
implementors. These documents are stored on [Ivy]<Alpine>doc>FileStore6.bravo
and FileStoreInternal0.bravo.
I've also split up the interface into several pieces (for reasons described in
"interface internals"). The current partitioning is quite lopsided and may not be
quite right, but we can shuffle things around easily enough.
Also, I have produced Mesa defs files, and stored them on [Ivy]<Alpine>defs>.
Included are two semi-hypothetical interfaces, Authenticate and RPC, intended to
be implemented by Grapevine and RPC respectively. (Semi-hypothetical in that
I have talked with Andrew about them, but they haven't actually been adopted
by the RPC design group, and no doubt will change.) The set of defs files is
described by AlpineDefs.df; we should endeavor to keep this DF file up-to-date
so we can quickly update our individual disks with the latest stuff.
These defs files are intended to be compilable now, but I haven't yet set up a
Cedar world on my machine, and the Cedar compiler won't run on top of vanilla
CoPilot. That's my next task.
Ed
------------------------------------------------------------
There is a section at the beginning of each of the two memos describing how the
client and server (respectively) intend to make use of the Authenticate and RPC
facilities.
*start*
00261 00024 US
Date: 15 Sept. 1981 10:36 am PDT (Tuesday)
From: Taft.PA
Subject: FileStore defs files
To: MBrown, Kolling
cc: Taft
... have all been compiled successfully. The errors were all minor, so I have not
bothered to print new listings.
Ed
*start*
00796 00024 US
Date: 15 Sept. 1981 11:37 am PDT (Tuesday)
From: Kolling.PA
Subject: FileStore6.bravo
To: taft
cc: mbrown, kolling
I've only skimmed this, so here are a few top of the head comments:
Access control:
1. Are we going to allow the "distinguished names" Owner and World, in
addition to *?
2. "A client who is the owner of a file is permitted to change the file's access
control lists". So can anyone in the file modify access list and in the owner
create list (I know this is awkward, but it is what fell out of our discussion...)
Owner data base:
3. Clients may be interested in setting the owner create lists. Also, there are
more than two pieces of info associated with the owner name. Also owner
RNames do not have to be Grapevine recognizable (i.e., FooDocs).
*start*
00299 00024 US
Date: 15 Sept. 1981 12:11 pm PDT (Tuesday)
From: kolling.PA
Subject: AlpineEnvironment: RNames? Ropes?
To: taft
cc: mbrown, kolling
Are we actually going to define owner names and client names as RNames? I
think Mark wanted me to use Ropes; I'm a little fuzzy about this.
*start*
01659 00024 US
Date: 15 Sept. 1981 2:45 pm PDT (Tuesday)
From: Taft.PA
Subject: Re: FileStore6.bravo
In-reply-to: Kolling's message of 15 Sept. 1981 11:37 am PDT (Tuesday)
To: Kolling
cc: taft, mbrown
AccessControl:
1. I don't have any objection to having "Owner" be a distinguished RName
equivalent to the owner's name itself. Remember, though, that when the client
reads the property list back, "Owner"it will come back as "Taft.PA" or whatever.
This seems perfectly ok to me.
2. I guess we have a misunderstanding. The whole point of the more
complicated semantics for changes to access lists is so that an owner can give
other clients modify access to the file without giving them modify access to the
file's access lists. So I can give you modify access to one of my files, but that
doesn't enable you to give Mark modify access to that file.
You are right about anyone in the file's owner's create list being able to modify
the access lists. This isn't in the statement you quoted (under "access control" in
section 4) because the owner data base hasn't yet been described. However, the
complete semantics (as I understand them) are given with the definition of
WriteProperties, under "file properties" in section 5.
(I'll number the subsections in the next revision of the memo!)
Owner data base:
I'll make OwnerName be a distinct type to emphasize that it is not necessarily a
Grapevine RName.
Ropes vs. strings vs. CedarXString vs. REF TEXT vs.....?
I'm fuzzy about this too. Wouldn't it be nice if we could just say STRING and
let the compiler produce whatever underlying representation seems to be in favor
this week.
Ed
*start*
01726 00024 US
Date: 15 Sept. 1981 4:21 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: FileStore6.bravo
In-reply-to: Taft's message of 15 Sept. 1981 2:45 pm PDT (Tuesday)
To: Taft
cc: Kolling, mbrown
1. Ref the definition of WriteProperties, under "file properties" in section 5:
no way do we allow write owner. Owner is set once and for all at
CreateFile.
2. "I guess we have a misunderstanding. The whole point of the more
complicated semantics for changes to access lists is so that an owner can give
other clients modify access to the file without giving them modify access to the
file's access lists. So I can give you modify access to one of my files, but that
doesn't enable you to give Mark modify access to that file."
I thought the whole point was to prevent inadvertant overwriting of
"one's own files" by being able to set the file modify list to Null, and
therefore we needed to allow the access lists themselves to be modified by
more than the people in the modify list. I have no
problem with someone in the modify list handling that priviledge off to
someone else. On the other hand, your way is easier to
implement. I'll change over to your way, unless Mark objects. So
my present belief is that FS will do the following:
on read access list request: look in its openfiletable for permission.
on write access list request: call AC.CheckWriteAccessFileAccessLists.
Naturally, it doesn't have to do the latter if it is in CreateFile and has
createfile permission.
3. I also see ClientNames, maybe these should be = whatever OwnerNames are.
4. Do we want elements of AccessLists to be RNames? (I use them in my cache
[ClientName, element of access list].)
KK
*start*
00286 00024 US
Date: 15 Sept. 1981 4:43 pm PDT (Tuesday)
From: kolling.PA
Subject: by the way..
To: taft, mbrown
cc: kolling
I'm keeping an in flux copy of AccessControlDesign8.bravo
[Ivy]<alpine>Doc>DRAFT> and will move it to [Ivy]<alpine>Doc> if it ever
stabilizes.
Karen
*start*
00942 00024 US
Date: 15 Sept. 1981 4:54 pm PDT (Tuesday)
From: Taft.PA
Subject: Re: FileStore6.bravo
In-reply-to: kolling's message of 15 Sept. 1981 4:21 pm PDT (Tuesday)
To: kolling
cc: Taft, mbrown
1. Actually, I think moving a file from one owner to another (without destroying
and recreating the file) is a useful operation. In any event, I don't think
AccessControl needs to make any special provision for this. Can't FileStore
simply call ChangeSpaceViaOpenFileID[ ..., -nPages], then change the owner
property, then call ChangeSpaceViaOpenFileID[ ..., +nPages] ?
2. You're right: I've invented a new justification for treating access control lists
specially.
3, 4. I'll change AlpineEnvironment as follows:
ClientName: TYPE = RName;
OwnerName: TYPE = STRING; -- whatever flavor of string we adopt
Elements of AccessLists are always Grapevine RNames, and I don't see any point
in defining a new type for these.
Ed
*start*
00377 00024 US
Date: 15 Sept. 1981 4:55 pm PDT (Tuesday)
From: kolling.PA
Subject: transID?
To: mbrown
cc: kolling
Mark, in the buffer pool I have a field for the transID for each buffer. You said
I maybe shouldn't use the real TransID here, but some index into something
instead that was shorter and equivalent. What exactly is it that you want me to
use?
Karen
*start*
00504 00024 US
Date: 15 Sept. 1981 5:06 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: FileStore6.bravo/changeowner
In-reply-to: Taft's message of 15 Sept. 1981 4:54 pm PDT (Tuesday)
To: Taft
cc: kolling, mbrown
IF FS does the work, I have no objection. It also has to check the create file
permission for the new owner. The order should be:
CheckAccessOwnerCreateThenAllocate[ ..., +nPages]
Change the owner property.
ChangeSpaceViaOpenFileID[ ..., -nPages]
this catches no-nos first.
*start*
00286 00024 US
Date: 15 Sept. 1981 5:06 pm PDT (Tuesday)
From: MBrown.PA
Subject: Re: transID?
In-reply-to: Your message of 15 Sept. 1981 4:55 pm PDT (Tuesday)
To: kolling
cc: mbrown
Call it a TransactionMap.Handle, I'll get around to defining it later in the week.
--mark
*start*
00739 00024 US
Date: 15 Sept. 1981 5:15 pm PDT (Tuesday)
From: MBrown.PA
Subject: Re: FileStore6.bravo
In-reply-to: Taft's message of 15 Sept. 1981 4:54 pm PDT (Tuesday)
To: Taft
cc: kolling, mbrown
3, 4. Maybe we should define AlpineEnvironment.String = Rope.Ref, and then
define ClientName = OwenerName = AlpineEnvironment.String. The point is, we
want to traffic in whatever Cedar standardizes on, I think. A Cedar veneer over
GrapevineUser is already available. Alternatively, we could stick with STRING
(hopefully LONG STRING), but I don't see the advantage in this.
1. So to change owner, what credentials are required? Owner create to both old
and new owner? Or just modify file plus create for new owner?
--mark
*start*
00437 00024 US
Date: 15 Sept. 1981 5:20 pm PDT (Tuesday)
From: kolling.PA
Subject: Re: FileStore6.bravo/changeowner again
In-reply-to: MBrown's message of 15 Sept. 1981 5:15 pm PDT (Tuesday)
To: MBrown
cc: Taft, kolling
right, I forgot, the client doing change owner also needs file modify permission
for the original owner (change owner is the equivalent of delete). (Of course,
he needs file modify to do the write owner.)
*start*
00325 00024 US
Date: 15 Sept. 1981 5:30 pm PDT (Tuesday)
From: kolling.PA
Subject: errors vs. return codes
To: taft, mbrown
cc: kolling
Who's working in the FS box? Do you want AC to ERROR or use return codes
for client-of-FileStore errors such as ownerAlreadyExists on AddOwner,
insufficientPrivilege, etc.?
KK
*start*
01570 00024 US
Date: 15 Sept. 1981 6:58 pm PDT (Tuesday)
From: Taft.PA
Subject: Re: errors vs. return codes
In-reply-to: kolling's message of 15 Sept. 1981 5:30 pm PDT (Tuesday)
To: kolling
cc: taft, mbrown
Well, I seem to be working in the FileStore box right now. My preference is to
have ERRORs for unexpected events, particularly if they are uncommon, but to
have return codes for everything else.
So, for example, a procedure like CheckAccessFile, whose main purpose is to
check an access control list and return a yes/no answer, should return "yes" or
"no", but should raise ERRORs for unexpected events such as no such file, can't
talk to Grapevine, etc.
It's often difficult to distinguish between caller errors and client-of-caller errors.
However, in the case of FileStore, it seems to me that this distinction is almost
irrelevant. FileStore can't allow any ERRORs from internal interfaces to escape to
the client, since the client has no way to catch them. So it has to catch all
exceptions arising from internal interfaces anyway, whether they are reported by
ERRORs or by return codes. FileStore then deals with the exceptions in its own
way (raising its own ERRORs when appropriate, etc.)
The advantage of ERRORs, of course, is that FileStore can ENABLE just once to
catch them anywhere in a particular context, rather than having to check the
result of each call. The disadvantage is that ERRORs are quite slow when they
occur. These considerations are what led me to the guidelines I suggested at the
beginning of this message.
Ed