The CyrusDB interface

[Bron Gondwana]

Hi All,

By popular request, a primer on the cyrusdb interface.

The cyrusdb interface is a common API for accessing key-value datastores
from within cyrus code.  It's part of the libcyrus shared object (but
not libcyrus_min), used by all cyrus binaries.  The use of the cyrusdb
API abstracts the details of the underlying data store, allowing sites
to use different database types depending on their needs, and allowing
Cyrus developers to use a common API for data storage needs.
The entire cyrusdb source lives in the lib/ directory of the cyrus-imapd
repository, in the following files:
lib/cyrusdb.h
lib/cyrusdb.c
lib/cyrusdb_flat.c
lib/cyrusdb_quotalegacy.c
lib/cyrusdb_skiplist.c
lib/cyrusdb_sql.c
lib/cyrusdb_twoskip.c

cyrusdb.h:

* interface definitions (all access to cyrusdb databases happens through
  the functions defined here)* the "struct cyrusdb_backend" data structure which defines the
  interface implemented by each backend.* constants for flags to the cyrusdb_open call, and return codes.
  Cyrusdb functions all return their own CYRUSDB_* error codes, which
  are not compatible with the r = IMAP_* return codes used throughout
  much of the rest of the codebase.
cyrusdb.c:

* implementations of the wrapper functions around the backends,
  including default implementations of some functions which are common
  to many backends but overridden by some.* a wrapper to initialise and cleanup the state of each backend (if
  needed) during Cyrus set up / tear down.
cyrusdb_*.c:

* the actual implementations of each backend!  We'll look at some in a
  sec.
There are also some tools to work with and support cyrus databases:

imap/ctl_cyrusdb:

* performs maintenance on the cyrusdb subsystem.  This is called in
  two places:* START: "ctl_cyrusdb -r" (recovery).  This is the ONLY PLACE that code
  is guaranteed to be run at startup on every cyrus installation, so
  you'll find quite a lot of detritus has built up in this codepath over
  the years.* EVENTS: "ctl_cyrusdb -c" (checkpoint).  This is run regularly
  (period=180 at FastMail, examples in the codebase have period=5 or
  period=30). Both this codepath and cyr_expire tend to run periodically
  on cyrus systems, and cleanup code is spread between those two
  locations.
imap/cvt_cyrusdb:

* used for converting a database between versions.  This is often used
  to prepare for upgrade, particularly in the past when Cyrus supported
  berkeley DB which didn't upgrade cleanly across OS versions, it was
  common to use cvt_cyrusdb to turn databases into a very portable
  format (flat or skiplist) before upgrading, upgrade the OS, convert
  back to the fast format (berkeley) and then restart.
imap/cyr_dbtool:

* once known as brontool, this is the first piece of Cyrus code I ever
  wrote!  It's a fairly dumb wrapper around the CyrusDB interface, and
  able to be used to read, write, or iterate any parts of a database.
  Its interactive mode is not special-character clean, but it can also
  be used in batch mode, which uses IMAP atom-string literal8 for
  input/output, and hence can roundtrip data reliably.
There are also tools like: ctl_conversationsdb, dav_reconstruct and
ctl_mboxlist which can be used to manage individual databases through a
more specific interface which understands the context as well as just
the raw key/value.
How to use CyrusDB:
===

Assuming that cyrus_init() has been called, which calls cyrusdb_init(),
you can assume that databases will work in any Cyrus code.
The first step is to open a database.  Databases have a filename - this
might be a literal filename on the backend, a directly containing data,
or an opaque token used by the backend to locate a  datasource.
    int flags = 0;
    struct db *mydb = NULL;
    int r = cyrusdb_open("skiplist", "/tmp/database.db", flags, &mydb);    if (!r) return mydb;  // if (r == CYRUSDB_OK) { ... }
    /* XXX: error handling */

Accepted flags:

CYRUSDB_CREATE - if the named database doesn't exist, create a
blank database.
CYRUSDB_MBOXSORT - use the abomination called improved_mboxlist_sort
which re-orders a couple of characters to allow "foo.bar" to sort before
"foo bar", for perfectly good reasons, but we're going to fix it a
better way.  Not every engine supports arbitrary collation, and if many
engines corrupt horribly if the same database is opened with different
choices for this flag.  Ouch.
CYRUSDB_CONVERT - if set and the database fails to open, attempt a magic
detection on the file content and try to convert the database to the
requested backend type before opening it.  In-place upgrades!  If this
is NOT set, then instead the magic detection will still be performed,
but the open database will be returned using the correct engine for that
database rather than converted.  Magic detection only currently works
for single-file database formats.
CYRUSDB_NOCOMPACT - if the database format supports automatic
compacting, don't use it.  Handy for when you want to read without
causing any possible issues (e.g. read-only filesystem during recovery)
or when performance is critical and you don't want to risk waiting while
a recompact happens.
All the remaining functions take that "struct db" pointer.

There's also a cyrusdb_lockopen() interface which takes a transaction
pointer and returns with the transaction already active.  This isn't
actually being used yet, but is intended to allow slightly more
efficient single-operation database use.  Right now, open returns an
unlocked database, but may need to lock as part of the setup, so keeping
that lock would avoid one extra unlock/lock cycle.
Reading, writing, transactions:
===

CyrusDB supports both transactional and non-transactional access.
Transactions are always exclusive.  This is arguably a deficiency in the
interface, particularly since many engines implement a non-exclusive
(read) lock internally anyway.
There are now 4 interfaces to read data.  Two of which are original
cyrusdb and two of which are more recently added.
original:
* cyrusdb_fetch() - fetch a single value by exact key.
* cyrusdb_foreach() - given a prefix, iterate over all the keys with
  that prefix (including exactly that key) in order.
newer:
* cyrusdb_fetchnext() - given an exact key, fetch the following key and
  value (regardless of whether the key exists), e.g given keys "f" and
  "g", fetchnext "foo" would return "g", as would fetchnext "f".  This
  can be used to implement foreach (indeed, the skips do exactly that).* cyrusdb_forone() - given an exact key, act like "cyrusdb_foreach" but
  only for that one key.  This is a convenience wrapper around fetch to
  allow doing things like:
    r = cyrusdb_forone(mydb, "folder", 6, p, cb, rock, &tid);
    if (!r) r = cyrusdb_foreach(mydb, "folder.", 7, p, cb, rock, &tid);
Which does precisely "folder" and its children without visiting any
other keys that have "folder" as a prefix.
Since the cyrusdb interface always takes both a pointer and a length,
it's also possible to use:
    char *key = "folder.";
    r = cyrusdb_forone(mydb, key, 6, p, cb, rock, &tid);
    if (!r) r = cyrusdb_foreach(mydb, key, 7, p, cb, rock, &tid);

You may have noticed that 'tid' at the end.  Every function for acting
on the database takes as its last argument a "struct txn **".  You can
pass one of three things to this:
NULL - non-transactional request.  Do whatever you need for internal
locking, but starts with an unlocked database and ends with an unlocked
database.  NOTE: at least skiplist and twoskip implement a hack where if
the database IS locked for a non-transactional read request, they will
act as if you'd passed the current transaction in for the NULL case.
This is a hack around layering violations and kind of sucks.
&NULL - e.g:

   struct txn *tid = NULL;
   const char *data = NULL;
   size_t datalen = 0;
   int r = cyrusdb_fetch(mydb, key, keylen, &data, &datalen, &tid);

After calling this, tid will have an opaque value allocated by the
database backend, which must be passed to all further cyrusdb
operations on that database until either cyrusdb_commit() or
cyrusdb_abort() are called.
&tid - e.g.

    if (r == CYRUSDB_NOTFOUND) {
        r = cyrusdb_store(mydb, key, keylen, "DEFAULT", 7, &tid); // set
        a default value    }

Given an existing transaction, perform this call in the context of the
transaction.
OK, foreach.  Foreach is very tricky, because it takes TWO callbacks.
The callbacks have an identical signature, but different return codes!
typedef int foreach_p(void *rock,
                      const char *key, size_t keylen,
                      const char *data, size_t datalen);

typedef int foreach_cb(void *rock,
                       const char *key, size_t keylen,
                       const char *data, size_t datalen);


The difference is this:  foreach_p is called with the database locked,
always - even if called without a transaction.  foreach_p returns 1 or
0.  0 means "skip this record", 1 means "process this record".  This is
useful to pre-filter records when called without a transaction, because
otherwise you lock and unlock all the time.
NULL for foreach_p is treated like a test which always returns '1', so
you can pass NULL if you don't need filtering.
If foreach_p returns 1, then with an unlocked transaction, the database
is now unlocked BEFORE calling foreach_cb, the callback.  foreach_cb
returns a CYRUSDB_ response.  If zero, the foreach will continue.  If
non-zero, the foreach will abort and return the non-zero response.  This
is both useful for error cases, and useful for short-circuiting, if you
only care that a key exists, you can do something like:
static int exists_cb(void *rock __attribute__((unused)), [...])
{
    return CYRUSDB_DONE; /* one is enough */
}

and then use exists_cb as your foreach_cb and check if the return code
is CYRUSDB_DONE to know if the foreach found a key.
If foreach is called with a transaction pointer, then it is your
responsibility as the caller to also pass that pointer (and a pointer to
the database) in that rock, so that callees can make further operations
within the same transaction.  A foreach with a transaction does NOT
unlock before calling its callback.
Writing:
===

There are three write operations:

cyrusdb_create
cyrusdb_store
cyrusdb_delete

cyrusdb_store will either create or overwrite an existing key.
cyrusdb_create will abort if the key already exists.  cyrusdb_delete
takes a flag 'force' which just makes it return CYRUSDB_OK (0) rather
than CYRUSDB_NOTFOUND if they key doesn't exist.  Strangely, 'force' is
after &tid, making it the only cyrusdb API that does that, but hey -
keeps you on your toes.
&tid behaves exactly the same for the write APIs.  If not passed, then
the database engine will behave as it if creates a writable transaction,
does the operation, then commits all within the cyrusdb_* call.
Gotchas!
===

* NULL is permitted in both keys and values, though 'flat' and
  'quotalegacy' have 8-bit cleanliness issues.
* zero-length keys are not supported

* zero-length values are theoretically supported, but a little
  interesting.  Certainly, pass "" rather than NULL as the value when
  writing or things will get weird.  I'm pretty sure at least the *skip
  databases assert on these kinds of weirdness.
* unlocked foreach: this is the land of the gotcha!  They key and data
  pointers (const char *) passed to your foreach_cb are only valid
  *UNTIL YOU TOUCH THE DATABASE*.  A common cause of rare and hard to
  diagnose bugs is writing something to the same database in the same
  process (*OR EVEN READING FROM IT AGAIN*).  I cannot emphasise this
  enough.  If you want to zero-copy access that data, you need to access
  it first, before touching that DB again.  Otherwise the map in which
  the data was a pointer may have been replaced as the next read found a
  new file and mapped it in!
also: if you're implementing a backend.  Unlocked foreach must find future records created by the current callback.  Consider a database containing 4 keys:
A B C D

if you are at key B and insert a key BB, then it must be iterated over.
If you insert AA while at B, it must NOT be iterated over.
* Opening the same database multiple times.  In the bad old days,
  opening the same database multiple times in the same process led to
  locking bugs (fcntl is braindead).  Each database engine is
  responsible for making sure this doesn't happen.  Most engines keep a
  linked lists of open databases.  If you try to open the same database
  again, they will just return the existing opened copy and bump a
  refcount.  Beware.  If a database is locked and you try to lock again
  - thinking you were opening it brand new, it will assertion fail
  and/or error.

I think that covers about everything!  Cyrusdb is used just about
everywhere that sorted key-value databases give what's needed, including
mailboxes.db, annotations.db (global and per mailbox databases), seen
state (non-owner), subscriptions, cyrus.indexed.db for Xapian, and the
rather massive (and increasingly inaccurately named) user.conversations.
Future plans are to increase the usage of cyrusdb databases, possibly by
building an indexing layer on top and using that instead of the sqldb
interface used for sqlite databases by DAV code, and possibly also
moving other custom file formats into a cyrusdb to allow easier
stateless server builds on a distributed backend.

Bron.

--
  Bron Gondwana, CEO, FastMail Pty Ltd
  brong at fastmailteam.com


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.andrew.cmu.edu/pipermail/cyrus-devel/attachments/20180228/acaec446/attachment-0001.html>