Documentation Source Text

  Although it is less common, a call to [sqlite3_prepare()] or
  [sqlite3_prepare_v2()] may also return SQLITE_LOCKED if it cannot obtain
  a read-lock on the sqlite_master table of each attached database. These
  APIs need to read the schema data contained in the sqlite_master table
  in order to compile SQL statements to [sqlite3_stmt*] objects.

<p>
  This page presents a technique using the SQLite [sqlite3_unlock_notify()]
  API to create a versions of [sqlite3_step()] and [sqlite3_prepare_v2()] that 
  block until the required locks are available instead of returning 
  SQLITE_LOCKED immediately, for use in multi-threaded applications. If the
  sqlite3_blocking_step() or sqlite3_blocking_prepare_v2() functions presented 
  to the left return SQLITE_LOCKED, this indicates that to block would 
  deadlock the system.

<p>
  The [sqlite3_unlock_notify()] API, which is only available if the library is
  compiled with the pre-processor symbol [SQLITE_ENABLE_UNLOCK_NOTIFY] defined,
  is [sqlite3_unlock_notify | documented here]. This page is no substitute for
  reading the full API documentation!

<p>
  The [sqlite3_unlock_notify()] interface is designed for use in systems
  that have a separate thread assigned to each [database connection].  There
  is nothing in the implementation that prevents a single thread from running
  multiple database connections.  However, the [sqlite3_unlock_notify()]
  interface only works on a single connection at a time, so the lock

  connection that the unlock-notify callback is waiting on, in this case
  connection Y, is known as the "blocking connection".

<p>
  If a call to sqlite3_step() that attempts to write to a database table
  returns SQLITE_LOCKED, then more than one other connection may be holding 
  a read-lock on the database table in question. In this case SQLite simply
  selects one of those other connections at random and issues the 
  unlock-notify callback when that connection's transaction is finished.
  Whether the call to sqlite3_step() was blocked by one or many connections, 
  when the corresponding unlock-notify callback is issued it is not 
  guaranteed that the required lock is available, only that it may be.

<p>
  When the unlock-notify callback is issued, it is issued from within a
  call to sqlite3_step() (or sqlite3_close()) associated with the blocking 
  connection. It is illegal to invoke any sqlite3_XXX() API functions from
  within an unlock-notify callback. The expected use is that the unlock-notify
  callback will single some other waiting thread or schedule some action
  to take place later.

<p>
  The algorithm used by the sqlite3_blocking_step() function is as follows:

<ol>
  <li><p> Call sqlite3_step() on the supplied statement handle. If the call

       SQLITE_ROW and then the next SQLITE_LOCKED), the statement handle may 
       be reset at this point without affecting the results of the query
       from the point of view of the caller. If sqlite3_reset() were not
       called at this point, the next call to sqlite3_step() would return
       SQLITE_MISUSE.

  <li><p> Return to step 1.


<p>
  The algorithm used by the sqlite3_blocking_prepare_v2() function is similar,
  except that step 4 (reseting the statement handle) is omitted.

</ol>

<p><b>Writer Starvation</b>

<p>
  Based on the description above, it could be concluded that if there are
  sufficient database readers reading the same table often enough, it is
  possible that the table will never become unlocked and that a connection
  waiting for a write-lock on the table will wait indefinitely. This 
  phenomena is known as writer-starvation.

<p>
  SQLite helps applications avoid this scenario. After any attempt to
  obtain a write-lock on a table fails (because one or more other 
  connections are holding read-locks), all attempts to open new transactions
  on the shared-cache fail until one of the following is true:

<ul>
  <li> The current writer concludes its transaction, OR
  <li> The number of open read-transactions on the shared-cache drops to zero.
</ul>

<p>
  Failed attempts to open new read-transactions return SQLITE_LOCKED to the
  caller. If the caller then calls [sqlite3_unlock_notify()] to register for
  an unlock-notify callback, the blocking connection is the connection that
  currently has an open write-transaction on the shared-cache. This prevents
  writer-starvation as, if no new read-transactions may be opened and 
  assuming all existing read-transactions are eventually concluded, the 
  writer will eventually have an opportunity to obtain the required
  write-lock.

<p><b>The pthreads API</b>

  <p> By the time [sqlite3_unlock_notify()] is invoked by

  <p> This way, it doesn't matter if the unlock-notify callback has already
      been invoked, or is being invoked, when the wait_for_unlock_notify() 
      thread begins blocking.

<p><b>Possible Enhancements</b>

  <p> The code on this page could be improved in at least two ways:

  <ul>
    <li> It could manage thread priorities.
    <li> It could handle a special case of SQLITE_LOCKED that can occur
         when dropping a table or index.
  </ul>

    statements complete execution in the meantime, re-attempting the "DROP
    TABLE" or "DROP INDEX" statement will return another SQLITE_LOCKED 
    error. In the implementation of sqlite3_blocking_step() shown to the
    left, this could cause an infinite loop.

  <p>
    The caller could distinguish between this special "DROP TABLE|INDEX" 
    case and other cases by using extended error codes. When it is appropriate
    to call [sqlite3_unlock_notify()], the extended error code is
    SQLITE_LOCKED_SHAREDCACHE. Otherwise, in the "DROP TABLE|INDEX" case,
    it is just plain SQLITE_LOCKED. Another solution might be to limit
    the number of times that any single query could be reattempted (to say 
    100). Although this might be less efficient than one might wish, the 
    situation in question is not likely to occur often.

<div style="clear:both"></div>