=head1 NAME

nbdkit-filter - how to write nbdkit filters

=head1 SYNOPSIS

 #include <nbdkit-filter.h>

 static int
 myfilter_config (nbdkit_next_config *next, void *nxdata,
                  const char *key, const char *value)
 {
   if (strcmp (key, "myparameter") == 0) {
     // ...
     return 0;
   }
   else {
     // pass through to next filter or plugin
     return next (nxdata, key, value);
   }
 }

 static struct nbdkit_filter filter = {
   .name              = "filter",
   .config            = myfilter_config,
   /* etc */
 };

 NBDKIT_REGISTER_FILTER(filter)

When this has been compiled to a shared library, do:

 nbdkit [--args ...] --filter=./myfilter.so plugin [key=value ...]

When debugging, use the I<-fv> options:

 nbdkit -fv --filter=./myfilter.so plugin [key=value ...]

=head1 DESCRIPTION

One or more nbdkit filters can be placed in front of an nbdkit plugin
to modify the behaviour of the plugin.  This manual page describes how
to create an nbdkit filter.

Filters can be used for example to limit requests to an offset/limit,
add copy-on-write support, or inject delays or errors (for testing).

Different filters can be stacked:

     NBD     ┌─────────┐    ┌─────────┐          ┌────────┐
  client ───▶│ filter1 │───▶│ filter2 │── ─ ─ ──▶│ plugin │
 request     └─────────┘    └─────────┘          └────────┘

Each filter intercepts plugin functions (see L<nbdkit-plugin(3)>) and
can call the next filter or plugin in the chain, modifying parameters,
calling before the filter function, in the middle or after.  Filters
may even short-cut the chain.  As an example, to process its own
parameters the filter can intercept the C<.config> method:

 static int
 myfilter_config (nbdkit_next_config *next, void *nxdata,
                  const char *key, const char *value)
 {
   if (strcmp (key, "myparameter") == 0) {
     // ...
     // here you would handle this key, value
     // ...
     return 0;
   }
   else {
     // pass through to next filter or plugin
     return next (nxdata, key, value);
   }
 }

 static struct nbdkit_filter filter = {
   // ...
   .config            = myfilter_config,
   // ...
 };

The call to C<next (nxdata, ...)> calls the C<.config> method of the
next filter or plugin in the chain.  In the example above any
instances of C<myparameter=...> on the command line would not be seen
by the plugin.

To see example filters:
L<https://gitlab.com/nbdkit/nbdkit/tree/master/filters>

Filters must be written in C.

Unlike plugins, where we provide a stable ABI guarantee that permits
operation across version differences, filters can only be run with the
same version of nbdkit that they were compiled with.  The reason for
this is two-fold: the filter API includes access to struct
nbdkit_next_ops that is likely to change if new callbacks are added
(old nbdkit cannot safely run new filters that access new methods);
and if we added new methods then an old filter would not see them and
so they would be passed unmodified through the filter, and in some
cases that leads to data corruption (new nbdkit cannot safely run old
filters unaware of new methods).  Therefore, unlike plugins, you
should not expect to distribute filters separately from nbdkit.

=head1 C<#include E<lt>nbdkit-filter.hE<gt>>

All filters should start by including this header file.

=head1 C<struct nbdkit_filter>

All filters must define and register one C<struct nbdkit_filter>,
which contains the name of the filter and pointers to plugin methods
that the filter wants to intercept.

 static struct nbdkit_filter filter = {
   .name              = "filter",
   .longname          = "My Filter",
   .description       = "This is my great filter for nbdkit",
   .config            = myfilter_config,
   /* etc */
 };

 NBDKIT_REGISTER_FILTER(filter)

The C<.name> field is the name of the filter.  This is the only field
which is required.

=head1 NEXT PLUGIN

F<nbdkit-filter.h> defines some function types (C<nbdkit_next_config>,
C<nbdkit_next_config_complete>, C<nbdkit_next_preconnect>,
C<nbdkit_next_list_exports>, C<nbdkit_next_default_export>,
C<nbdkit_next_open>) and a structure called C<struct nbdkit_next_ops>.
These abstract the next plugin or filter in the chain.  There is also
an opaque pointer C<backend>, C<context> or C<nxdata> which must be
passed along when calling these functions.  The value of C<backend> is
stable between C<.after_fork>, C<.preconnect>, C<.list_exports>, and
C<.default_export>, and can also be obtained by using
C<nbdkit_context_get_backend> on the C<context> parameter to C<.open>.

Meanwhile, if the filter does not use C<nbdkit_context_set_next>, the
value of C<next> passed to C<.prepare> has a stable lifetime that
lasts to the corresponding C<.finalize>, with all intermediate
functions (such as C<.pread>) receiving the same value for
convenience.  Functions where C<nxdata> is not reused are C<.config>,
C<.config_complete>, and C<.get_ready>, which are all called during
initialization outside any connections.  The value of C<backend>
passed to C<.after_fork> also occurs without connections, but is
shared with C<.preconnect>, C<.list_exports>, and C<.default_export>,
and can also be obtained from the C<context> passed to C<.open>, and
has a lifetime that lasts to C<.cleanup> for use by
C<nbdkit_next_context_open>.  In turn, the value of C<context> passed
to C<.open> has a lifetime that lasts until the matching C<.close> for
use by C<nbdkit_context_get_backend> and C<nbdkit_context_set_next>.

=head2 Next config, open and close

The filter’s C<.config>, C<.config_complete>, C<.get_ready>,
C<.after_fork>, C<.preconnect>, C<.list_exports>, C<.default_export>
and C<.open> methods may only call the next C<.config>,
C<.config_complete>, C<.get_ready>, C<.after_fork>, C<.preconnect>,
C<.list_exports>, C<.default_export> and C<.open> method in the chain
(optionally for C<.config> and C<.open>).

The filter’s C<.close> method is called when an old connection closed,
and this has no C<next> parameter because it cannot be
short-circuited.

=head2 C<nbdkit_next>

The filter generally needs to call into the underlying plugin, which
is done via a pointer to C<struct nbdkit_next_ops>, also available as
the typedef C<nbdkit_next>.  The most common behavior is to create a
next context per connection by calling the C<next_open> parameter
during C<.open>, at which point the next context will be automatically
provided to the filter’s other methods like C<.prepare>, C<.get_size>,
C<.pread> etc.  The C<nbdkit_next> struct contains a comparable set of
accessors to plugin methods that can be called during a connection.
When using automatic registration, the C<next> parameter is stable
between C<.prepare> and C<.finalize>, and nbdkit automatically
prepares, finalizes, and closes the next context at the right point in
the filter connection lifecycle.

Alternatively, the filter can manage plugin contexts manually, whether
to multiplex multiple client connections through a single context into
the plugin, or to open multiple plugin contexts to perform retries or
otherwise service a single client connection more efficiently.  In
this mode of operation, the filter uses C<nbdkit_next_context_open> to
open a plugin context using the C<backend> parameter passed to
C<.after_fork>, C<.preconnect>, C<.list_exports>, C<.default_export>,
or obtained from using C<nbdkit_context_get_backend> on the C<context>
parameter to C<.open>.  The resulting next context has a lifecycle
under manual control, where the filter must use C<next-E<gt>prepare
(next)> before using any other function pointers within the next
context, and must reclaim the memory using C<next-E<gt>finalize
(next)> and C<nbdkit_next_context_close> when done.  A filter using
manual lifecycle management may use C<nbdkit_context_set_next> to
associate the next context into the current connection, which lets
nbdkit then pass that context as the C<next> parameter to future
connection-related functions like C<.pread> and take over lifecycle
responsibility.

=head3 C<nbdkit_context_get_backend>

=head3 C<nbdkit_next_context_open>

=head3 C<nbdkit_next_context_close>

=head3 C<nbdkit_context_set_next>

 nbdkit_backend *nbdkit_context_get_backend (nbdkit_context *context);

Obtains the backend pointer from the C<context> parameter to C<.open>,
matching the backend pointer available to C<.after_fork>,
C<.preconnect>, C<.list_exports>, and C<.default_export>.  This
backend pointer has a stable lifetime from the time of C<.after_fork>
until C<.cleanup>.

 nbdkit_next *nbdkit_next_context_open (nbdkit_backend *backend,
                                        int readonly, const char *exportname,
                                        int shared);

This function attempts to open a new context into the plugin in
relation to the filter's current C<backend>.  The C<readonly> and
C<exportname> parameters behave the same as documented in C<.open>.
The resulting context will be under the filter's manual lifecycle
control unless the filter associates it into the connection with
C<nbdkit_context_set_next>.  The filter should be careful to not
violate any threading model restrictions of the plugin if it opens
more than one context.

If C<shared> is false, this function must be called while servicing an
existing client connection, and the new context will share the same
connection details (export name, tls status, and shorter interned
string lifetimes) as the current connection, and thus should not be
used after the client connection ends.  Conversely, if C<shared> is
true, this function may be called outside of a current client
connection (such as during C<.after_fork>), and the resulting context
may be freely shared among multiple client connections.  In shared
mode, it will not be possible for the plugin to differentiate content
based on the client export name, the result of the plugin calling
L<nbdkit_is_tls(3)> will depend solely whether I<--tls=require> was on
the command line, the lifetime of interned strings (via
L<nbdkit_strdup_intern(3)> and friends) lasts for the life of the
filter, and the filter must take care to not expose potentially-secure
information from the backend to an insecure client.

 void nbdkit_next_context_close (nbdkit_next *next);

This function closes a context into the plugin.  If the context has
previously been prepared, it should first be finalized before using
this function.  This function does not need to be called for a plugin
context that has been associated with the filter connection via
C<nbdkit_context_set_next> prior to the C<.close> callback.

 nbdkit_next *nbdkit_context_set_next (nbdkit_context *context,
                                       nbdkit_next *next);

This function associates a plugin context with the filter's current
connection context, given by the C<context> parameter to C<.open>.
Once associated, this plugin context will be given as the C<next>
parameter to all other connection-specific callbacks.  If associated
during C<.open>, nbdkit will take care of preparing the context prior
to C<.prepare>; if still associated before C<.finalize>, nbdkit will
take care of finalizing the context, and also for closing it.  A
filter may also pass C<NULL> for C<next>, to remove any association;
if no plugin context is associated with the connection, then filter
callbacks such as C<.pread> will receive C<NULL> for their C<next>
parameter.

This function returns the previous context that had been associated
with the connection prior to switching the association to C<next>;
this result will be C<NULL> if there was no previous association.  The
filter assumes manual responsibility for any remaining lifecycle
functions that must be called on the returned context.

=head2 Using C<nbdkit_next>

Regardless of whether the plugin context is managed automatically or
manually, it is possible for a filter to issue (for example) extra
C<next-E<gt>pread> calls in response to a single C<.pwrite> call.

The C<next> parameter serves two purposes: it serves as the struct
to access the pointers to all the plugin connection functions, and it
serves as the opaque data that must be passed as the first parameter
to those functions.  For example, calling the plugin's can_flush
functionality would be done via

 next->can_flush (next)

Note that the semantics of the functions in C<struct nbdkit_next_ops>
are slightly different from what a plugin implements: for example,
when a plugin's C<.pread> returns -1 on error, the error value to
advertise to the client is implicit (via the plugin calling
L<nbdkit_set_error(3)> or setting C<errno>), whereas
C<next-E<gt>pread> exposes this via an explicit parameter,
allowing a filter to learn or modify this error if desired.

Use of C<next-E<gt>prepare> and C<next-E<gt>finalize> is only
needed when manually managing the plugin context lifetime.

=head2 Other considerations

You can modify parameters when you call the C<next> function.  However
be careful when modifying strings because for some methods
(eg. C<.config>) the plugin may save the string pointer that you pass
along.  So you may have to ensure that the string is not freed for the
lifetime of the server; you may find L<nbdkit_strdup_intern(3)>
helpful for avoiding a memory leak while still obeying lifecycle
constraints.

Note that if your filter registers a callback but in that callback it
doesn't call the C<next> function then the corresponding method in the
plugin will never be called.  In particular, your C<.open> method, if
you have one, B<must> call the C<next> method if you want the
underlying plugin to be available to all further C<nbdkit_next> use.

=head1 CALLBACKS

C<struct nbdkit_filter> has some static fields describing the filter
and optional callback functions which can be used to intercept plugin
methods.

=head2 C<.name>

 const char *name;

This field (a string) is required, and B<must> contain only ASCII
alphanumeric characters or non-leading dashes, and be unique amongst
all filters.

=head2 C<.longname>

 const char *longname;

An optional free text name of the filter.  This field is used in error
messages.

=head2 C<.description>

 const char *description;

An optional multi-line description of the filter.

=head2 C<.load>

 void load (void);

This is called once just after the filter is loaded into memory.  You
can use this to perform any global initialization needed by the
filter.

=head2 C<.unload>

 void unload (void);

This may be called once just before the filter is unloaded from
memory.  Note that it's not guaranteed that C<.unload> will always be
called (eg. the server might be killed or segfault), so you should try
to make the filter as robust as possible by not requiring cleanup.
See also L<nbdkit-plugin(3)/SHUTDOWN>.

=head2 C<.config>

 int (*config) (nbdkit_next_config *next, void *nxdata,
                const char *key, const char *value);

This intercepts the plugin C<.config> method and can be used by the
filter to parse its own command line parameters.  You should try to
make sure that command line parameter keys that the filter uses do not
conflict with ones that could be used by a plugin.

If there is an error, C<.config> should call L<nbdkit_error(3)> with
an error message and return C<-1>.

=head2 C<.config_complete>

 int (*config_complete) (nbdkit_next_config_complete *next, void *nxdata);

This intercepts the plugin C<.config_complete> method and can be used
to ensure that all parameters needed by the filter were supplied on
the command line.

If there is an error, C<.config_complete> should call L<nbdkit_error(3)>
with an error message and return C<-1>.

=head2 C<.config_help>

 const char *config_help;

This optional multi-line help message should summarize any
C<key=value> parameters that it takes.  It does I<not> need to repeat
what already appears in C<.description>.

If the filter doesn't take any config parameters you should probably
omit this.

=head2 C<.thread_model>

 int (*thread_model) (void);

Filters may tighten (but not relax) the thread model of the plugin, by
defining this callback.

See L<nbdkit-plugin(3)/THREADS> for a discussion of thread models.

The final thread model used by nbdkit is the smallest (ie. most
serialized) out of all the filters and the plugin, and applies for all
connections.  Requests for a model larger than permitted by the plugin
are silently ignored. It is acceptable for decisions made during
C<.config> and C<.config_complete> to determine which model to
request.

This callback is optional; if it is not present, the filter must be
written to handle fully parallel requests, including when multiple
requests are issued in parallel on the same connection, similar to a
plugin requesting C<NBDKIT_THREAD_MODEL_PARALLEL>.  This ensures the
filter doesn't slow down other filters or plugins.

If there is an error, C<.thread_model> should call L<nbdkit_error(3)>
with an error message and return C<-1>.

=head2 C<.dump_plugin>

 void (*dump_plugin) (void);

This optional callback is called when the
S<C<nbdkit null --filter=filtername --dump-plugin>> command is used.
It should print any additional informative C<key=value> fields to
stdout as needed.  Prefixing the keys with the name of the filter will
avoid conflicts.

=head2 C<.get_ready>

 int (*get_ready) (int final_thread_model);

This optional callback is reached if the plugin C<.get_ready> method
succeeded (if the plugin failed, nbdkit has already exited), and can
be used by the filter to get ready to serve requests.

The C<final_thread_model> parameter informs the filter about the final
thread model chosen by nbdkit after considering the results of
C<.thread_model> of all filters in the chain after
C<.config_complete>.

If there is an error, C<.get_ready> should call L<nbdkit_error(3)> with
an error message and return C<-1>.

=head2 C<.after_fork>

 int (*after_fork) (nbdkit_backend *backend);

This optional callback is reached after the plugin C<.after_fork>
method has succeeded (if the plugin failed, nbdkit has already
exited), and can be used by the filter to start background threads.
The C<backend> parameter is valid until C<.cleanup>, for creating manual
contexts into the backend with C<nbdkit_next_context_open>.

If there is an error, C<.after_fork> should call L<nbdkit_error(3)> with
an error message and return C<-1>.

=head2 C<.cleanup>

 int (cleanup) (nbdkit_backend *backend);

This optional callback is reached once after all client connections
have been closed, but before the underlying plugin C<.cleanup> or any
C<.unload> callbacks.  It can be used by the filter to gracefully
close any background threads created during C<.after_fork>, as well as
close any manual contexts into C<backend> previously opened with
C<nbdkit_next_context_open>.

Note that it's not guaranteed that C<.cleanup> will always be called
(eg. the server might be killed or segfault), so you should try to
make the filter as robust as possible by not requiring cleanup.  See
also L<nbdkit-plugin(3)/SHUTDOWN>.

=head2 C<.preconnect>

 int (*preconnect) (nbdkit_next_preconnect *next, nbdkit_backend *nxdata,
                    int readonly);

This intercepts the plugin C<.preconnect> method and can be used to
filter access to the server.

If there is an error, C<.preconnect> should call L<nbdkit_error(3)> with
an error message and return C<-1>.

=head2 C<.list_exports>

 int (*list_exports) (nbdkit_next_list_exports *next, nbdkit_backend *nxdata,
                      int readonly, int is_tls,
                      struct nbdkit_exports *exports);

This intercepts the plugin C<.list_exports> method and can be used to
filter which exports are advertised.

The C<readonly> parameter matches what is passed to <.preconnect> and
C<.open>, and may be changed by the filter when calling into the
plugin.  The C<is_tls> parameter informs the filter whether TLS
negotiation has been completed by the client, but is not passed on to
C<next> because it cannot be altered.

It is possible for filters to transform the exports list received back
from the layer below.  Without error checking it would look like this:

 myfilter_list_exports (...)
 {
   size_t i;
   struct nbdkit_exports *exports2;
   struct nbdkit_export e;
   char *name, *desc;

   exports2 = nbdkit_exports_new ();
   next_list_exports (nxdata, readonly, exports);
   for (i = 0; i < nbdkit_exports_count (exports2); ++i) {
     e = nbdkit_get_export (exports2, i);
     name = adjust (e.name);
     desc = adjust (e.desc);
     nbdkit_add_export (exports, name, desc);
     free (name);
     free (desc);
   }
   nbdkit_exports_free (exports2);
 }

If there is an error, C<.list_exports> should call L<nbdkit_error(3)> with
an error message and return C<-1>.

=head3 Allocating and freeing nbdkit_exports list

Two functions are provided to filters only for allocating and freeing
the list:

 struct nbdkit_exports *nbdkit_exports_new (void);

Allocates and returns a new, empty exports list.

On error this function can return C<NULL>.  In this case it calls
L<nbdkit_error(3)> as required.  C<errno> will be set to a suitable
value.

 void nbdkit_exports_free (struct nbdkit_exports *);

Frees an existing exports list.

=head3 Iterating over nbdkit_exports list

Two functions are provided to filters only to iterate over the exports
in order:

 size_t nbdkit_exports_count (const struct nbdkit_exports *);

Returns the number of exports in the list.

 struct nbdkit_export {
   char *name;
   char *description;
 };
 const struct nbdkit_export nbdkit_get_export (const struct nbdkit_exports *,
                                               size_t i);

Returns a copy of the C<i>'th export.

=head2 C<.default_export>

 const char *default_export (nbdkit_next_default_export *next,
                             nbdkit_backend *nxdata,
                             int readonly, int is_tls)

This intercepts the plugin C<.default_export> method and can be used to
alter the canonical export name used in place of the default C<"">.

The C<readonly> parameter matches what is passed to <.preconnect> and
C<.open>, and may be changed by the filter when calling into the
plugin.  The C<is_tls> parameter informs the filter whether TLS
negotiation has been completed by the client, but is not passed on to
C<next> because it cannot be altered.

=head2 C<.open>

 void * (*open) (nbdkit_next_open *next, nbdkit_context *context,
                 int readonly, const char *exportname, int is_tls);

This is called when a new client connection is opened and can be used
to allocate any per-connection data structures needed by the filter.
The handle (which is not the same as the plugin handle) is passed back
to other filter callbacks and could be freed in the C<.close>
callback.

Note that the handle is completely opaque to nbdkit, but it must not
be NULL.  If you don't need to use a handle, return
C<NBDKIT_HANDLE_NOT_NEEDED> which is a static non-NULL pointer.

If there is an error, C<.open> should call L<nbdkit_error(3)> with an
error message and return C<NULL>.

This callback is optional, but if provided, it should call C<next>,
passing C<readonly> and C<exportname> possibly modified according to
how the filter plans to use the plugin (C<is_tls> is not passed,
because a filter cannot modify it).  Typically, the filter passes the
same values as it received, or passes readonly=true to provide a
writable layer on top of a read-only backend.  However, it is also
acceptable to attempt write access to the plugin even if this filter
is readonly, such as when a file system mounted read-only still
requires write access to the underlying device in case a journal needs
to be replayed for consistency as part of the mounting process.

The C<exportname> string is only guaranteed to be available during the
call (different than the lifetime for the return of
L<nbdkit_export_name(3)> used by plugins).  If the filter needs to use
it (other than immediately passing it down to the next layer) it must
take a copy, although L<nbdkit_strdup_intern(3)> is useful for this
task.  The C<exportname> and C<is_tls> parameters are provided so that
filters do not need to use the plugin-only interfaces of
L<nbdkit_export_name(3)> and L<nbdkit_is_tls(3)>.

The filter should generally call C<next> as its first step, to
allocate from the plugin outwards, so that C<.close> running from the
outer filter to the plugin will be in reverse.  Skipping a call to
C<next> is acceptable if the filter will not access C<nbdkit_next>
during any of the remaining callbacks reached on the same connection.
The C<next> function is provided for convenience; the same
functionality can be obtained manually (other than error checking) by
using the following:

 nbdkit_context_set_next (context, nbdkit_next_context_open
    (nbdkit_context_get_backend (context), readonly, exportname, false));

The value of C<context> in this call has a lifetime that lasts until
the counterpart C<.close>, and it is this value that may be passed to
C<nbdkit_context_get_backend> to obtain the C<backend> parameter used
to open a plugin context with C<nbdkit_next_context_open>, as well as
the C<context> parameter used to associate a plugin context into the
current connection with C<nbdkit_context_set_next>.

=head2 C<.close>

 void (*close) (void *handle);

This is called when the client closes the connection.  It should clean
up any per-connection resources used by the filter.  It is called
beginning with the outermost filter and ending with the plugin (the
opposite order of C<.open> if all filters call C<next> first),
although this order technically does not matter since the callback
cannot report failures or access the underlying plugin.

=head2 C<.prepare>

=head2 C<.finalize>

  int (*prepare) (nbdkit_next *next, void *handle, int readonly);
  int (*finalize) (nbdkit_next *next, void *handle);

These two methods can be used to perform any necessary operations just
after opening the connection (C<.prepare>) or just before closing the
connection (C<.finalize>).

For example if you need to scan the underlying disk to check for a
partition table, you could do it in your C<.prepare> method (calling
the plugin's C<.get_size> and C<.pread> methods via C<next>).  Or
if you need to cleanly update superblock data in the image on close
you can do it in your C<.finalize> method (calling the plugin's
C<.pwrite> method).  Doing these things in the filter's C<.open> or
C<.close> method is not possible without using manual context
lifecycle management.

For C<.prepare>, the value of C<readonly> is the same as was passed to
C<.open>, declaring how this filter will be used.

Note that nbdkit performs sanity checking on requests made to the
underlying plugin; for example, C<next-E<gt>pread> cannot be
called on a given connection unless C<next-E<gt>get_size> has
first been called at least once in the same connection (to ensure the
read requests are in bounds), and C<next-E<gt>pwrite> further
requires an earlier successful call to C<next-E<gt>can_write>.  In
many filters, these prerequisites will be automatically called during
the client negotiation phase, but there are cases where a filter
overrides query functions or makes I/O calls into the plugin before
handshaking is complete, where the filter needs to make those
prerequisite calls manually during C<.prepare>.

While there are C<next-E<gt>prepare> and C<next-E<gt>finalize>
functions, these are different from other filter methods, in that any
plugin context associated with the current connection (via the C<next>
parameter to C<.open>, or via C<nbdkit_context_set_next>, is prepared
and finalized automatically by nbdkit, so they are only used during
manual lifecycle management.  Prepare methods are called starting with
the filter closest to the plugin and proceeding outwards (matching the
order of C<.open> if all filters call C<next> before doing anything
locally), and only when an outer filter did not skip the C<next> call
during C<.open>.  Finalize methods are called in the reverse order of
prepare methods, with the outermost filter first (and matching the
order of C<.close>), and only if the prepare method succeeded.

If there is an error, both callbacks should call L<nbdkit_error(3)> with
an error message and return C<-1>.  An error in C<.prepare> is
reported to the client, but leaves the connection open (a client may
try again with a different export name, for example); while an error
in C<.finalize> forces the client to disconnect.

=head2 C<.get_size>

 int64_t (*get_size) (nbdkit_next *next, void *handle);

This intercepts the plugin C<.get_size> method and can be used to read
or modify the apparent size of the block device that the NBD client
will see.

The returned size must be E<ge> 0.  If there is an error, C<.get_size>
should call L<nbdkit_error(3)> with an error message and return C<-1>.
This function is only called once per connection and cached by nbdkit.
Similarly, repeated calls to C<next-E<gt>get_size> will return a
cached value.

=head2 C<.export_description>

 const char *export_description (nbdkit_next *next, void *handle);

This intercepts the plugin C<.export_description> method and can be
used to read or modify the export description that the NBD client
will see.

=head2 C<.block_size>

 int block_size (nbdkit_next *next, void *handle, uint32_t *minimum,
                 uint32_t *preferred, uint32_t *maximum);

This intercepts the plugin C<.block_size> method and can be used to
read or modify the block size constraints that the NBD client will
see.

=head2 C<.can_write>

=head2 C<.can_flush>

=head2 C<.is_rotational>

=head2 C<.can_trim>

=head2 C<.can_zero>

=head2 C<.can_fast_zero>

=head2 C<.can_extents>

=head2 C<.can_fua>

=head2 C<.can_multi_conn>

=head2 C<.can_cache>

 int (*can_write) (nbdkit_next *next, void *handle);
 int (*can_flush) (nbdkit_next *next, void *handle);
 int (*is_rotational) (nbdkit_next *next, void *handle);
 int (*can_trim) (nbdkit_next *next, void *handle);
 int (*can_zero) (nbdkit_next *next, void *handle);
 int (*can_fast_zero) (nbdkit_next *next, void *handle);
 int (*can_extents) (nbdkit_next *next, void *handle);
 int (*can_fua) (nbdkit_next *next, void *handle);
 int (*can_multi_conn) (nbdkit_next *next, void *handle);
 int (*can_cache) (nbdkit_next *next, void *handle);

These intercept the corresponding plugin methods, and control feature
bits advertised to the client.

Of note, the semantics of C<.can_zero> callback in the filter are
slightly different from the plugin, and must be one of three success
values visible only to filters:

=over 4

=item C<NBDKIT_ZERO_NONE>

Completely suppress advertisement of write zero support (this can only
be done from filters, not plugins).

=item C<NBDKIT_ZERO_EMULATE>

Inform nbdkit that write zeroes should immediately fall back to
C<.pwrite> emulation without trying C<.zero> (this value is returned
by C<next-E<gt>can_zero> if the plugin returned false in its
C<.can_zero>).

=item C<NBDKIT_ZERO_NATIVE>

Inform nbdkit that write zeroes should attempt to use C<.zero>,
although it may still fall back to C<.pwrite> emulation for C<ENOTSUP>
or C<EOPNOTSUPP> failures (this value is returned by
C<next-E<gt>can_zero> if the plugin returned true in its
C<.can_zero>).

=back

Remember that most of the feature check functions return merely a
boolean success value, while C<.can_zero>, C<.can_fua> and
C<.can_cache> have three success values.

The difference between C<.can_fua> values may affect choices made in
the filter: when splitting a write request that requested FUA from the
client, if C<next-E<gt>can_fua> returns C<NBDKIT_FUA_NATIVE>, then
the filter should pass the FUA flag on to each sub-request; while if
it is known that FUA is emulated by a flush because of a return of
C<NBDKIT_FUA_EMULATE>, it is more efficient to only flush once after
all sub-requests have completed (often by passing C<NBDKIT_FLAG_FUA>
on to only the final sub-request, or by dropping the flag and ending
with a direct call to C<next-E<gt>flush>).

If there is an error, the callback should call L<nbdkit_error(3)> with an
error message and return C<-1>.  These functions are called at most
once per connection and cached by nbdkit. Similarly, repeated calls to
any of the C<nbdkit_next> counterparts will return a cached value; by
calling into the plugin during C<.prepare>, you can ensure that later
use of the cached values during data commands like <.pwrite> will not
fail.

=head2 C<.pread>

 int (*pread) (nbdkit_next *next,
               void *handle, void *buf, uint32_t count, uint64_t offset,
               uint32_t flags, int *err);

This intercepts the plugin C<.pread> method and can be used to read or
modify data read by the plugin.

The parameter C<flags> exists in case of future NBD protocol
extensions; at this time, it will be 0 on input, and the filter should
not pass any flags to C<next-E<gt>pread>.

If there is an error (including a short read which couldn't be
recovered from), C<.pread> should call L<nbdkit_error(3)> with an error
message B<and> return -1 with C<err> set to the positive errno value
to return to the client.

=head2 C<.pwrite>

 int (*pwrite) (nbdkit_next *next,
                void *handle,
                const void *buf, uint32_t count, uint64_t offset,
                uint32_t flags, int *err);

This intercepts the plugin C<.pwrite> method and can be used to modify
data written by the plugin.

This function will not be called if C<.can_write> returned false; in
turn, the filter should not call C<next-E<gt>pwrite> if
C<next-E<gt>can_write> did not return true.

The parameter C<flags> may include C<NBDKIT_FLAG_FUA> on input based
on the result of C<.can_fua>.  In turn, the filter should only pass
C<NBDKIT_FLAG_FUA> on to C<next-E<gt>pwrite> if
C<next-E<gt>can_fua> returned a positive value.

If there is an error (including a short write which couldn't be
recovered from), C<.pwrite> should call L<nbdkit_error(3)> with an error
message B<and> return -1 with C<err> set to the positive errno value
to return to the client.

=head2 C<.flush>

 int (*flush) (nbdkit_next *next,
               void *handle, uint32_t flags, int *err);

This intercepts the plugin C<.flush> method and can be used to modify
flush requests.

This function will not be called if C<.can_flush> returned false; in
turn, the filter should not call C<next-E<gt>flush> if
C<next-E<gt>can_flush> did not return true.

The parameter C<flags> exists in case of future NBD protocol
extensions; at this time, it will be 0 on input, and the filter should
not pass any flags to C<next-E<gt>flush>.

If there is an error, C<.flush> should call L<nbdkit_error(3)> with an
error message B<and> return -1 with C<err> set to the positive errno
value to return to the client.

=head2 C<.trim>

 int (*trim) (nbdkit_next *next,
              void *handle, uint32_t count, uint64_t offset,
              uint32_t flags, int *err);

This intercepts the plugin C<.trim> method and can be used to modify
trim requests.

This function will not be called if C<.can_trim> returned false; in
turn, the filter should not call C<next-E<gt>trim> if
C<next-E<gt>can_trim> did not return true.

The parameter C<flags> may include C<NBDKIT_FLAG_FUA> on input based
on the result of C<.can_fua>.  In turn, the filter should only pass
C<NBDKIT_FLAG_FUA> on to C<next-E<gt>trim> if
C<next-E<gt>can_fua> returned a positive value.

If there is an error, C<.trim> should call L<nbdkit_error(3)> with an
error message B<and> return -1 with C<err> set to the positive errno
value to return to the client.

=head2 C<.zero>

 int (*zero) (nbdkit_next *next,
              void *handle, uint32_t count, uint64_t offset, uint32_t flags,
              int *err);

This intercepts the plugin C<.zero> method and can be used to modify
zero requests.

This function will not be called if C<.can_zero> returned
C<NBDKIT_ZERO_NONE> or C<NBDKIT_ZERO_EMULATE>; in turn, the filter
should not call C<next-E<gt>zero> if C<next-E<gt>can_zero> returned
C<NBDKIT_ZERO_NONE>.

On input, the parameter C<flags> may include C<NBDKIT_FLAG_MAY_TRIM>
unconditionally, C<NBDKIT_FLAG_FUA> based on the result of
C<.can_fua>, and C<NBDKIT_FLAG_FAST_ZERO> based on the result of
C<.can_fast_zero>.  In turn, the filter may pass
C<NBDKIT_FLAG_MAY_TRIM> unconditionally, but should only pass
C<NBDKIT_FLAG_FUA> or C<NBDKIT_FLAG_FAST_ZERO> on to
C<next-E<gt>zero> if the corresponding C<next-E<gt>can_fua> or
C<next-E<gt>can_fast_zero> returned a positive value.

Note that unlike the plugin C<.zero> which is permitted to fail with
C<ENOTSUP> or C<EOPNOTSUPP> to force a fallback to C<.pwrite>, the
function C<next-E<gt>zero> will not fail with C<err> set to
C<ENOTSUP> or C<EOPNOTSUPP> unless C<NBDKIT_FLAG_FAST_ZERO> was used,
because otherwise the fallback has already taken place.

If there is an error, C<.zero> should call L<nbdkit_error(3)> with an
error message B<and> return -1 with C<err> set to the positive errno
value to return to the client.  The filter should not fail with
C<ENOTSUP> or C<EOPNOTSUPP> unless C<flags> includes
C<NBDKIT_FLAG_FAST_ZERO> (while plugins have automatic fallback to
C<.pwrite>, filters do not).

=head2 C<.extents>

 int (*extents) (nbdkit_next *next,
                 void *handle, uint32_t count, uint64_t offset, uint32_t flags,
                 struct nbdkit_extents *extents,
                 int *err);

This intercepts the plugin C<.extents> method and can be used to
modify extent requests.

This function will not be called if C<.can_extents> returned false; in
turn, the filter should not call C<next-E<gt>extents> if
C<next-E<gt>can_extents> did not return true.

It is possible for filters to transform the extents list received back
from the layer below.  Without error checking it would look like this:

 myfilter_extents (..., uint32_t count, uint64_t offset, ...)
 {
   size_t i;
   struct nbdkit_extents *extents2;
   struct nbdkit_extent e;
   int64_t size;

   size = next->get_size (next);
   extents2 = nbdkit_extents_new (offset + shift, size);
   next->extents (next, count, offset + shift, flags, extents2, err);
   for (i = 0; i < nbdkit_extents_count (extents2); ++i) {
     e = nbdkit_get_extent (extents2, i);
     e.offset -= shift;
     nbdkit_add_extent (extents, e.offset, e.length, e.type);
   }
   nbdkit_extents_free (extents2);
 }

If there is an error, C<.extents> should call L<nbdkit_error(3)> with an
error message B<and> return -1 with C<err> set to the positive errno
value to return to the client.

=head3 Allocating and freeing nbdkit_extents list

Two functions are provided to filters only for allocating and freeing
the map:

 struct nbdkit_extents *nbdkit_extents_new (uint64_t start, uint64_t end);

Allocates and returns a new, empty extents list.  The C<start>
parameter is the start of the range described in the list, and the
C<end> parameter is the offset of the byte beyond the end.  Normally
you would pass in C<offset> as the start and the size of the plugin as
the end, but for filters which adjust offsets, they should pass in the
adjusted offset.

On error this function can return C<NULL>.  In this case it calls
L<nbdkit_error(3)> and/or L<nbdkit_set_error(3)> as required.  C<errno> will
be set to a suitable value.

 void nbdkit_extents_free (struct nbdkit_extents *);

Frees an existing extents list.

=head3 Iterating over nbdkit_extents list

Two functions are provided to filters only to iterate over the extents
in order:

 size_t nbdkit_extents_count (const struct nbdkit_extents *);

Returns the number of extents in the list.

 struct nbdkit_extent {
   uint64_t offset;
   uint64_t length;
   uint32_t type;
 };
 struct nbdkit_extent nbdkit_get_extent (const struct nbdkit_extents *,
                                         size_t i);

Returns a copy of the C<i>'th extent.

=head3 Reading the full extents from the plugin

A convenience function is provided to filters only which makes one or
more requests to the underlying plugin until we have a full set of
extents covering the region C<[offset..offset+count-1]>.

 struct nbdkit_extents *nbdkit_extents_full (
                             nbdkit_next *next,
                             uint32_t count, uint64_t offset,
                             uint32_t flags, int *err);

Note this allocates a new C<struct nbdkit_extents> which the caller
must free.  C<flags> is passed through to the underlying plugin, but
C<NBDKIT_FLAG_REQ_ONE> is removed from the set of flags so that the
plugin returns as much information as possible (this is usually what
you want).

On error this function can return C<NULL>.  In this case it calls
L<nbdkit_error(3)> and/or L<nbdkit_set_error(3)> as required.  C<*err> will
be set to a suitable value.

=head3 Enforcing alignment of an nbdkit_extents list

A convenience function is provided to filters only which makes it
easier to ensure that the client only encounters aligned extents.

 int nbdkit_extents_aligned (nbdkit_next *next,
                             uint32_t count, uint64_t offset,
                             uint32_t flags, uint32_t align,
                             struct nbdkit_extents *extents, int *err);

Calls C<next-E<gt>extents> as needed until at least C<align> bytes are
obtained, where C<align> is a power of 2.  Anywhere the underlying
plugin returns differing extents within C<align> bytes, this function
treats that portion of the disk as a single extent with zero and
sparse status bits determined by the intersection of all underlying
extents.  It is an error to call this function with C<count> or
C<offset> that is not already aligned.

=head2 C<.cache>

 int (*cache) (nbdkit_next *next,
               void *handle, uint32_t count, uint64_t offset,
               uint32_t flags, int *err);

This intercepts the plugin C<.cache> method and can be used to modify
cache requests.

This function will not be called if C<.can_cache> returned
C<NBDKIT_CACHE_NONE> or C<NBDKIT_CACHE_EMULATE>; in turn, the filter
should not call C<next-E<gt>cache> if
C<next-E<gt>can_cache> returned C<NBDKIT_CACHE_NONE>.

The parameter C<flags> exists in case of future NBD protocol
extensions; at this time, it will be 0 on input, and the filter should
not pass any flags to C<next-E<gt>cache>.

If there is an error, C<.cache> should call L<nbdkit_error(3)> with an
error message B<and> return -1 with C<err> set to the positive errno
value to return to the client.

=head1 ERROR HANDLING

If there is an error in the filter itself, the filter should call
L<nbdkit_error(3)> to report an error message.  If the callback is
involved in serving data, the explicit C<err> parameter determines the
error code that will be sent to the client; other callbacks should
return the appropriate error indication, eg. C<NULL> or C<-1>.

=head1 DEBUGGING

Run the server with I<-f> and I<-v> options so it doesn't fork and you
can see debugging information:

 nbdkit -fv --filter=./myfilter.so plugin [key=value [key=value [...]]]

To print debugging information from within the filter, call
L<nbdkit_debug(3)>.  Note that C<nbdkit_debug> only prints things when
the server is in verbose mode (I<-v> option).

=head2 Debug Flags

Debug Flags in filters work exactly the same way as plugins.  See
L<nbdkit-plugin(3)/Debug Flags>.

=head1 INSTALLING THE FILTER

The filter is a C<*.so> file and possibly a manual page.  You can of
course install the filter C<*.so> file wherever you want, and users
will be able to use it by running:

 nbdkit --filter=/path/to/filter.so plugin [args]

However B<if> the shared library has a name of the form
C<nbdkit-I<name>-filter.so> B<and if> the library is installed in the
C<$filterdir> directory, then users can be run it by only typing:

 nbdkit --filter=name plugin [args]

The location of the C<$filterdir> directory is set when nbdkit is
compiled and can be found by doing:

 nbdkit --dump-config

If using the pkg-config/pkgconf system then you can also find the
filter directory at compile time by doing:

 pkg-config nbdkit --variable=filterdir

=head1 PKG-CONFIG/PKGCONF

nbdkit provides a pkg-config/pkgconf file called C<nbdkit.pc> which
should be installed on the correct path when the nbdkit development
environment is installed.  You can use this in autoconf
F<configure.ac> scripts to test for the development environment:

 PKG_CHECK_MODULES([NBDKIT], [nbdkit >= 1.2.3])

The above will fail unless nbdkit E<ge> 1.2.3 and the header file is
installed, and will set C<NBDKIT_CFLAGS> and C<NBDKIT_LIBS>
appropriately for compiling filters.

You can also run pkg-config/pkgconf directly, for example:

 if ! pkg-config nbdkit --exists; then
   echo "you must install the nbdkit development environment"
   exit 1
 fi

You can also substitute the filterdir variable by doing:

 PKG_CHECK_VAR([NBDKIT_FILTERDIR], [nbdkit], [filterdir])

which defines C<$(NBDKIT_FILTERDIR)> in automake-generated Makefiles.

=head1 WRITING FILTERS IN C++

Instead of using C, it is possible to write filters in C++.  However
for inclusion in upstream nbdkit we would generally prefer filters
written in C.

Filters in C++ work almost exactly like those in C, but the way you
define the C<nbdkit_filter> struct is slightly different:

 namespace {
   nbdkit_filter create_filter() {
     nbdkit_filter filter = nbdkit_filter ();
     filter.name   = "myfilter";
     filter.config = myfilter_config;
     return filter;
   }
 }
 static struct nbdkit_filter filter = create_filter ();
 NBDKIT_REGISTER_FILTER(filter)

=head1 SEE ALSO

L<nbdkit(1)>,
L<nbdkit-plugin(3)>.

Standard filters provided by nbdkit:

__FILTER_LINKS__.

=head1 AUTHORS

Eric Blake

Richard W.M. Jones

=head1 COPYRIGHT

Copyright Red Hat