Feature #7431


ZFS Channel Programs

Added by Chris Williamson almost 7 years ago. Updated over 6 years ago.

Start date:
Due date:
% Done:


Estimated time:
Gerrit CR:
External Bug:


ZFS channel programs (ZCP) adds support for performing compound ZFS administrative actions via Lua scripts in a sandboxed environment (with time and memory limits).

This initial commit includes both base support for running ZCP scripts, and a small initial library of API calls which support getting properties and listing, destroying, and promoting datasets.

Testing: in addition to the included unit tests, channel programs have been in use at Delphix for several months for batch destroying filesystems. The dsl_destroy_snaps_nvl() call has also been replaced with

For reference, the new zfs-program manpage is included below.

ZFS-PROGRAM(1M)                       1M                       ZFS-PROGRAM(1M)

     zfs program – executes ZFS channel programs

     zfs program [-t timeout] [-m memory-limit] pool script

     The ZFS channel program interface allows ZFS administrative operations to
     be run programmatically as a Lua script. The entire script is executed
     atomically, with no other administrative operations taking effect
     concurrently. A library of ZFS calls is made available to channel program
     scripts. Channel programs may only be run with root privileges.

     A modified version of the Lua 5.2 interpreter is used to run channel
     program scripts. The Lua 5.2 manual can be found at:


     The channel program given by script will be run on pool, and any attempts
     to access or modify other pools will cause an error.

     -t timeout
         Execution time limit, in milliseconds. If a channel program executes
         for longer than the provided timeout, it will be stopped and an error
         will be returned.  The default timeout is 1000 ms, and can be set to
         a maximum of 10000 ms.

     -m memory-limit
         Memory limit, in bytes. If a channel program attempts to allocate
         more memory than the given limit, it will be stopped and an error
         returned. The default memory limit is 10 MB, and can be set to a
         maximum of 100 MB.

     All remaining argument strings will be passed directly to the Lua script
     as described in the LUA INTERFACE section below.

     A channel program can be invoked either from the command line, or via a
     library call to lzc_channel_program().

     Arguments passed to the channel program are converted to a Lua table. If
     invoked from the command line, extra arguments to the Lua script will be
     accessible as an array stored in the argument table with the key 'argv':

           args = ...
           argv = args["argv"]
           -- argv == {1="arg1", 2="arg2", ...}

     If invoked from the libZFS interface, an arbitrary argument list can be
     passed to the channel program, which is accessible via the same "..." 
     syntax in Lua:

           args = ...
           -- args == {"foo"="bar", "baz"={...}, ...}

     Note that because Lua arrays are 1-indexed, arrays passed to Lua from the
     libZFS interface will have their indices incremented by 1. That is, the
     element in arr[0] in a C array passed to a channel program will be stored
     in arr[1] when accessed from Lua.

   Return Values
     Lua return statements take the form:

           return ret0, ret1, ret2, ...

     Return statements returning multiple values are permitted internally in a
     channel program script, but attempting to return more than one value from
     the top level of the channel program is not permitted and will throw an
     error.  However, tables containing multiple values can still be returned.
     If invoked from the command line, a return statement:

           a = {foo="bar", baz=2}
           return a

     Will be output formatted as:

           Channel program fully executed with return value:
                   baz: 2
                   foo: 'bar'

   Fatal Errors
     If the channel program encounters a fatal error while running, a non-zero
     exit status will be returned. If more information about the error is
     available, a singleton list will be returned detailing the error:

           error: "error string, including Lua stack trace" 

     If a fatal error is returned, the channel program may have not executed
     at all, may have partially executed, or may have fully executed but
     failed to pass a return value back to userland.

     If the channel program exhausts a time or memory limit, a fatal error
     will be generated and the program will be stopped, leaving the program
     partially executed. No attempt is made to reverse or undo any operations
     already performed. If a channel program attempts to return too large a
     value, the program will fully execute but exit with a nonzero status code
     and no return value.

     Note: ZFS API functions do not generate Fatal Errors when correctly
     invoked, they return an error code and the channel program continues
     executing. See the ZFS API section below for function-specific details on
     error return codes.

   Lua to C Value Conversion
     When invoking a channel program via the libZFS interface, it is necessary
     to translate arguments and return values from Lua values to their C
     equivalents, and vice-versa.

     There is a correspondence between nvlist values in C and Lua tables. A
     Lua table which is returned from the channel program will be recursively
     converted to an nvlist, with table values converted to their natural

           string -> string
           number -> int64
           boolean -> boolean_value
           nil -> boolean (no value)
           table -> nvlist

     Likewise, table keys are replaced by string equivalents as follows:

           string -> no change
           number -> signed decimal string ("%lld")
           boolean -> "true" | "false" 

     Any collision of table key strings (for example, the string "true" and a
     true boolean value) will cause a fatal error.

     Lua numbers are represented internally as signed 64-bit integers.

     The following Lua built-in base library functions are available:

           assert                  rawlen
           collectgarbage          rawget
           error                   rawset
           getmetatable            select
           ipairs                  setmetatable
           next                    tonumber
           pairs                   tostring
           rawequal                type

     All functions in the coroutine, string, and table built-in submodules are
     also available. A complete list and documentation of these modules is
     available in the Lua manual.

     The following functions base library functions have been disabled and are
     not available for use in channel programs:


     In addition, the exponentiation (^) operator has been disabled.

   Function Arguments
     Each API function takes a fixed set of required positional arguments and
     optional keyword arguments. For example, the destroy function takes a
     single positional string argument (the name of the dataset to destroy)
     and an optional "defer" keyword boolean argument. When using parentheses
     to specify the arguments to a Lua function, only positional arguments can
     be used:


     To use keyword arguments, functions must be called with a single argument
     that is a Lua table containing entries mapping integers to positional
     arguments and strings to keyword arguments:

           zfs.sync.destroy({1="rpool@snap", defer=true})

     The Lua language allows curly braces to be used in place of parenthesis
     as syntactic sugar for this calling convention:

           zfs.sync.snapshot{"rpool@snap", defer=true}

   Function Return Values
     If an API function succeeds, it returns 0. If it fails, it returns an
     error code and the channel program continues executing. API functions do
     not generate Fatal Errors except in the case of an unrecoverable internal
     file system error.

     In addition to returning an error code, some functions also return extra
     details describing what caused the error. This extra description is given
     as a second return value, and will always be a Lua table, or Nil if no
     error details were returned. Different keys will exist in the error
     details table depending on the function and error case. Any such function
     may be called expecting a single return value:

           errno = zfs.sync.promote(dataset)

     Or, the error details can be retrieved:

           errno, details = zfs.sync.promote(dataset)
           if (errno == EEXIST) then
               assert(details ~= Nil)
               list_of_conflicting_snapshots = details

     The following global aliases for API function error return codes are
     defined for use in channel programs:

           EPERM     ECHILD      ENODEV      ENOSPC
           ENOENT    EAGAIN      ENOTDIR     ESPIPE
           ESRCH     ENOMEM      EISDIR      EROFS
           EINTR     EACCES      EINVAL      EMLINK
           EIO       EFAULT      ENFILE      EPIPE
           ENXIO     ENOTBLK     EMFILE      EDOM
           E2BIG     EBUSY       ENOTTY      ERANGE
           EBADF     EXDEV       EFBIG

   API Functions
     For detailed descriptions of the exact behavior of any zfs administrative
     operations, see the main zfs(1) manual page.

         Record a debug message in the zfs_dbgmsg log. A log of these messages
         can be printed via mdb's "::zfs_dbgmsg" command, or can be monitored
         live by running:

                 dtrace -n 'zfs-dbgmsg{trace(stringof(arg0))}'

         msg (string)
             Debug message to be printed.

     zfs.get_prop(dataset, property)
         Returns two values. First, a string, number or table containing the
         property value for the given dataset. Second, a string containing the
         source of the property (i.e. the name of the dataset in which it was
         set or nil if it is readonly). Throws a Lua error if the dataset is
         invalid or the property doesn't exist. Note that Lua only supports
         int64 number types whereas ZFS number properties are uint64. This
         means very large values (like guid) may wrap around and appear

         dataset (string)
             Filesystem or snapshot path to retrieve properties from.

         property (string)
             Name of property to retrieve. All filesystem, snapshot and volume
             properties are supported except for 'mounted' and 'iscsioptions.'
             Also supports the 'written@snap' and 'written#bookmark'
             properties and the '<user|group><quota|used>@id' properties,
             though the id must be in numeric form.

     zfs.sync submodule
         The sync submodule contains functions that modify the on-disk state.
         They are executed in "syncing context".

         The available sync submodule functions are as follows:

         zfs.sync.destroy(dataset, [defer=true|false])
             Destroy the given dataset. Returns 0 on successful destroy, or a
             nonzero error code if the dataset could not be destroyed (for
             example, if the dataset has any active children or clones).

             dataset (string)
                 Filesystem or snapshot to be destroyed.

             [optional] defer (boolean)
                 Valid only for destroying snapshots. If set to true, and the
                 snapshot has holds or clones, allows the snapshot to be
                 marked for deferred deletion rather than failing.

             Promote the given clone to a filesystem. Returns 0 on successful
             promotion, or a nonzero error code otherwise. If EEXIST is
             returned, the second return value will be an array of the clone's
             snapshots whose names collide with snapshots of the parent

             dataset (string)
                 Clone to be promoted.

     zfs.check submodule
         For each function in the zfs.sync submodule, there is a corresponding
         zfs.check function which performs a "dry run" of the same operation.
         Each takes the same arguments as its zfs.sync counterpart and returns
         0 if the operation would succeed, or a non-zero error code if it
         would fail, along with any other error details. That is, each has the
         same behavior as the corresponding sync function except for actually
         executing the requested change. For example, zfs.check.destroy("fs")
         returns 0 if zfs.sync.destroy("fs") would successfully destroy the

         The available zfs.check functions are:

         zfs.check.destroy(dataset, [defer=true|false])


     zfs.list submodule
         The zfs.list submodule provides functions for iterating over datasets
         and properties. Rather than returning tables, these functions act as
         Lua iterators, and are generally used as follows:

               for child in zfs.list.children("rpool") do

         The available zfs.list functions are:

             Iterate through all clones of the given snapshot.

             snapshot (string)
                 Must be a valid snapshot path in the current pool.

             Iterate through all snapshots of the given dataset. Each snapshot
             is returned as a string containing the full dataset name, e.g.

             dataset (string)
                 Must be a valid filesystem or volume.

             Iterate through all direct children of the given dataset. Each
             child is returned as a string containing the full dataset name,
             e.g. "pool/fs/child".

             dataset (string)
                 Must be a valid filesystem or volume.

             Iterate through all user properties for the given dataset.

             dataset (string)
                 Must be a valid filesystem, snapshot, or volume.

             Returns an array of strings, the names of the valid system (non-
             user defined) properties for the given dataset. Throws a Lua
             error if the dataset is invalid.

             dataset (string)
                 Must be a valid filesystem, snapshot or volume.

   Example 1
     The following channel program recursively destroys a filesystem and all
     its snapshots and children in a naive manner. Note that this does not
     involve any error handling or reporting.

           function destroy_recursive(root)
               for child in zfs.list.children(root) do
               for snap in zfs.list.snapshots(root) do

   Example 2
     A more verbose and robust version of the same channel program, which
     properly detects and reports errors, and also takes the dataset to
     destroy as a command line argument, would be as follows:

           succeeded = {}
           failed = {}

           function destroy_recursive(root)
               for child in zfs.list.children(root) do
               for snap in zfs.list.snapshots(root) do
                   err = zfs.sync.destroy(snap)
                   if (err ~= 0) then
                       failed[snap] = err
                       succeeded[snap] = err
               err = zfs.sync.destroy(root)
               if (err ~= 0) then
                   failed[root] = err
                   succeeded[root] = err

           args = ...
           argv = args["argv"]


           results = {}
           results["succeeded"] = succeeded
           results["failed"] = failed
           return results

   Example 3
     The following function performs a forced promote operation by attempting
     to promote the given clone and destroying any conflicting snapshots.

           function force_promote(ds)
              errno, details = zfs.check.promote(ds)
              if (errno == EEXIST) then
                  assert(details ~= Nil)
                  for i, snap in ipairs(details) do
                      zfs.sync.destroy(ds .. "@" .. snap)
              elseif (errno ~= 0) then
                  return errno
              return zfs.sync.promote(ds)

Related issues

Related to illumos gate - Bug #15913: zfs: get_special_prop() should also check for ZFS_TYPE_SNAPSHOTIn ProgressToomas Soome

Actions #1

Updated by Electric Monk over 6 years ago

  • Status changed from New to Closed

git commit dfc115332c94a2f62058ac7f2bce7631fbd20b3d

commit  dfc115332c94a2f62058ac7f2bce7631fbd20b3d
Author: Chris Williamson <>
Date:   2017-06-26T13:45:06.000Z

    7431 ZFS Channel Programs
    Reviewed by: Matthew Ahrens <>
    Reviewed by: George Wilson <>
    Reviewed by: John Kennedy <>
    Reviewed by: Dan Kimmel <>
    Approved by: Garrett D'Amore <>

Actions #2

Updated by Toomas Soome 11 days ago

  • Related to Bug #15913: zfs: get_special_prop() should also check for ZFS_TYPE_SNAPSHOT added

Also available in: Atom PDF