Next config, open and close¶

The filter’s ".config", ".config_complete", ".get_ready", ".after_fork" and ".open" methods may only call the next ".config", ".config_complete", ".get_ready", ".after_fork" and ".open" method in the chain (optionally for ".config").

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

"next_ops"¶

The filter’s other methods like ".prepare", ".get_size", ".pread" etc ― always called in the context of a connection ― are passed a pointer to "struct nbdkit_next_ops" which contains a comparable set of accessors to plugin methods that can be called during a connection. The "next_ops" parameter is stable between ".prepare" and ".finalize"; intermediate functions (such as ".pread") receive the same value for convenience.

It is possible for a filter to issue (for example) extra "next_ops->pread" calls in response to a single ".pwrite" call.

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

There is also a "next_ops->reopen" function which is used by nbdkit-retry-filter(1) to close and reopen the underlying plugin. It should be used with caution because it is difficult to use safely.

Other considerations¶

You can modify parameters when you call the "next" function. However be careful when modifying strings because for some methods (eg. ".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 "nbdkit_strdup_intern" 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 "next" function then the corresponding method in the plugin will never be called. In particular, your ".open" method, if you have one, must call the ".next" method.

".name"¶

 const char *name;

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

".longname"¶

 const char *longname;

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

".description"¶

 const char *description;

An optional multi-line description of the filter.

".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.

".unload"¶

 void unload (void);

This may be called once just before the filter is unloaded from memory. Note that it's not guaranteed that ".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 "SHUTDOWN" in nbdkit-plugin(3).

".config"¶

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

This intercepts the plugin ".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, ".config" should call "nbdkit_error" with an error message and return "-1".

".config_complete"¶

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

This intercepts the plugin ".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, ".config_complete" should call "nbdkit_error" with an error message and return "-1".

".config_help"¶

 const char *config_help;

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

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

".thread_model"¶

 int (*thread_model) (void);

Filters may tighten (but not relax) the thread model of the plugin, by defining this callback. Note that while plugins use a compile-time definition of "THREAD_MODEL", filters do not need to declare a model at compile time; instead, this callback is called after ".config_complete" and before any connections are created. See "THREADS" in nbdkit-plugin(3) 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 ".config" and ".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 "NBDKIT_THREAD_MODEL_PARALLEL". This ensures the filter doesn't slow down other filters or plugins.

If there is an error, ".thread_model" should call "nbdkit_error" with an error message and return "-1".

".get_ready"¶

 int (*get_ready) (nbdkit_next_get_ready *next, void *nxdata,
                   int thread_model);

This intercepts the plugin ".get_ready" method and can be used by the filter to get ready to serve requests.

The "thread_model" parameter informs the filter about the final thread model chosen by nbdkit after considering the results of ".thread_model" of all filters in the chain after ".config_complete". This does not need to be passed on to "next", as the model can no longer be altered at this point.

If there is an error, ".get_ready" should call "nbdkit_error" with an error message and return "-1".

".after_fork"¶

 int (*after_fork) (nbdkit_next_after_fork *next, void *nxdata);

This intercepts the plugin ".after_fork" method and can be used by the filter to start background threads (although these have limited usefulness since they will not be able to access the plugin).

If there is an error, ".after_fork" should call "nbdkit_error" with an error message and return "-1".

".preconnect"¶

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

This intercepts the plugin ".preconnect" method and can be used to filter access to the server.

If there is an error, ".preconnect" should call "nbdkit_error" with an error message and return "-1".

".list_exports"¶

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

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

The "readonly" parameter matches what is passed to <.preconnect> and ".open", and may be changed by the filter when calling into the plugin. The "is_tls" parameter informs the filter whether TLS negotiation has been completed by the client, but is not passed on to "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, ".list_exports" should call "nbdkit_error" with an error message and return "-1".

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 "NULL". In this case it calls "nbdkit_error" as required. "errno" will be set to a suitable value.

 void nbdkit_exports_free (struct nbdkit_exports *);

Frees an existing exports list.

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 "i"'th export.

".default_export"¶

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

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

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

".open"¶

 void * (*open) (nbdkit_next_open *next, void *nxdata,
                 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 ".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 "NBDKIT_HANDLE_NOT_NEEDED" which is a static non-NULL pointer.

If there is an error, ".open" should call "nbdkit_error" with an error message and return "NULL".

This callback is optional, but if provided, it should call "next", passing "readonly" and "exportname" possibly modified according to how the filter plans to use the plugin ("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 "exportname" string is only guaranteed to be available during the call (different than the lifetime for the return of "nbdkit_export_name" 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 "nbdkit_strdup_intern" is useful for this task. The "exportname" and "is_tls" parameters are provided so that filters do not need to use the plugin-only interfaces of "nbdkit_export_name" and "nbdkit_is_tls".

The filter should generally call "next" as its first step, to allocate from the plugin outwards, so that ".close" running from the outer filter to the plugin will be in reverse. Skipping a call to "next" is acceptable if the filter will not access "next_ops" during any of the remaining callbacks reached on the same connection.

".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 ".open" if all filters call "next" first), although this order technically does not matter since the callback cannot report failures or access the underlying plugin.

".prepare"¶

".finalize"¶

  int (*prepare) (struct nbdkit_next_ops *next_ops, void *nxdata,
                  void *handle, int readonly);
  int (*finalize) (struct nbdkit_next_ops *next_ops, void *nxdata,
                   void *handle);

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

For example if you need to scan the underlying disk to check for a partition table, you could do it in your ".prepare" method (calling the plugin's ".get_size" and ".pread" methods via "next_ops"). Or if you need to cleanly update superblock data in the image on close you can do it in your ".finalize" method (calling the plugin's ".pwrite" method). Doing these things in the filter's ".open" or ".close" method is not possible.

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

Note that nbdkit performs sanity checking on requests made to the underlying plugin; for example, "next_ops->pread" cannot be called on a given connection unless "next_ops->get_size" has first been called at least once in the same connection (to ensure the read requests are in bounds), and "next_ops->pwrite" further requires an earlier successful call to "next_ops->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 ".prepare".

There is no "next_ops->prepare" or "next_ops->finalize". Unlike other filter methods, prepare and finalize are not chained through the "next_ops" structure. Instead the core nbdkit server calls the prepare and finalize methods of all filters. Prepare methods are called starting with the filter closest to the plugin and proceeding outwards (matching the order of ".open" if all filters call "next" before doing anything locally), and only when an outer filter did not skip the "next" call during ".open". Finalize methods are called in the reverse order of prepare methods, with the outermost filter first (and matching the order of ".close"), and only if the prepare method succeeded.

If there is an error, both callbacks should call "nbdkit_error" with an error message and return "-1". An error in ".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 ".finalize" forces the client to disconnect.

".get_size"¶

 int64_t (*get_size) (struct nbdkit_next_ops *next_ops, void *nxdata,
                      void *handle);

This intercepts the plugin ".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 ≥ 0. If there is an error, ".get_size" should call "nbdkit_error" with an error message and return "-1". This function is only called once per connection and cached by nbdkit. Similarly, repeated calls to "next_ops->get_size" will return a cached value.

".export_description"¶

 const char *export_description (struct nbdkit_next_ops *next_ops,
                                 void *nxdata, void *handle);

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

".can_write"¶

".can_flush"¶

".is_rotational"¶

".can_trim"¶

".can_zero"¶

".can_fast_zero"¶

".can_extents"¶

".can_fua"¶

".can_multi_conn"¶

".can_cache"¶

 int (*can_write) (struct nbdkit_next_ops *next_ops, void *nxdata,
                   void *handle);
 int (*can_flush) (struct nbdkit_next_ops *next_ops, void *nxdata,
                   void *handle);
 int (*is_rotational) (struct nbdkit_next_ops *next_ops,
                       void *nxdata,
                       void *handle);
 int (*can_trim) (struct nbdkit_next_ops *next_ops, void *nxdata,
                  void *handle);
 int (*can_zero) (struct nbdkit_next_ops *next_ops, void *nxdata,
                  void *handle);
 int (*can_fast_zero) (struct nbdkit_next_ops *next_ops, void *nxdata,
                       void *handle);
 int (*can_extents) (struct nbdkit_next_ops *next_ops, void *nxdata,
                     void *handle);
 int (*can_fua) (struct nbdkit_next_ops *next_ops, void *nxdata,
                 void *handle);
 int (*can_multi_conn) (struct nbdkit_next_ops *next_ops, void *nxdata,
                        void *handle);
 int (*can_cache) (struct nbdkit_next_ops *next_ops, void *nxdata,
                   void *handle);

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

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

"NBDKIT_ZERO_NONE"

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

"NBDKIT_ZERO_EMULATE"

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

"NBDKIT_ZERO_NATIVE"

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

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

The difference between ".can_fua" values may affect choices made in the filter: when splitting a write request that requested FUA from the client, if "next_ops->can_fua" returns "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 "NBDKIT_FUA_EMULATE", it is more efficient to only flush once after all sub-requests have completed (often by passing "NBDKIT_FLAG_FUA" on to only the final sub-request, or by dropping the flag and ending with a direct call to "next_ops->flush").

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

".pread"¶

 int (*pread) (struct nbdkit_next_ops *next_ops, void *nxdata,
               void *handle, void *buf, uint32_t count, uint64_t offset,
               uint32_t flags, int *err);

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

The parameter "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 "next_ops->pread".

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

".pwrite"¶

 int (*pwrite) (struct nbdkit_next_ops *next_ops, void *nxdata,
                void *handle,
                const void *buf, uint32_t count, uint64_t offset,
                uint32_t flags, int *err);

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

This function will not be called if ".can_write" returned false; in turn, the filter should not call "next_ops->pwrite" if "next_ops->can_write" did not return true.

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

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

".flush"¶

 int (*flush) (struct nbdkit_next_ops *next_ops, void *nxdata,
               void *handle, uint32_t flags, int *err);

This intercepts the plugin ".flush" method and can be used to modify flush requests.

This function will not be called if ".can_flush" returned false; in turn, the filter should not call "next_ops->flush" if "next_ops->can_flush" did not return true.

The parameter "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 "next_ops->flush".

If there is an error, ".flush" should call "nbdkit_error" with an error message and return -1 with "err" set to the positive errno value to return to the client.

".trim"¶

 int (*trim) (struct nbdkit_next_ops *next_ops, void *nxdata,
              void *handle, uint32_t count, uint64_t offset,
              uint32_t flags, int *err);

This intercepts the plugin ".trim" method and can be used to modify trim requests.

This function will not be called if ".can_trim" returned false; in turn, the filter should not call "next_ops->trim" if "next_ops->can_trim" did not return true.

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

If there is an error, ".trim" should call "nbdkit_error" with an error message and return -1 with "err" set to the positive errno value to return to the client.

".zero"¶

 int (*zero) (struct nbdkit_next_ops *next_ops, void *nxdata,
              void *handle, uint32_t count, uint64_t offset, uint32_t flags,
              int *err);

This intercepts the plugin ".zero" method and can be used to modify zero requests.

This function will not be called if ".can_zero" returned "NBDKIT_ZERO_NONE"; in turn, the filter should not call "next_ops->zero" if "next_ops->can_zero" returned "NBDKIT_ZERO_NONE".

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

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

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

".extents"¶

 int (*extents) (struct nbdkit_next_ops *next_ops, void *nxdata,
                 void *handle, uint32_t count, uint64_t offset, uint32_t flags,
                 struct nbdkit_extents *extents,
                 int *err);

This intercepts the plugin ".extents" method and can be used to modify extent requests.

This function will not be called if ".can_extents" returned false; in turn, the filter should not call "next_ops->extents" if "next_ops->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_ops->get_size (nxdata);
   extents2 = nbdkit_extents_new (offset + shift, size);
   next_ops->extents (nxdata, 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, ".extents" should call "nbdkit_error" with an error message and return -1 with "err" set to the positive errno value to return to the client.

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 "start" parameter is the start of the range described in the list, and the "end" parameter is the offset of the byte beyond the end. Normally you would pass in "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 "NULL". In this case it calls "nbdkit_error" and/or "nbdkit_set_error" as required. "errno" will be set to a suitable value.

 void nbdkit_extents_free (struct nbdkit_extents *);

Frees an existing extents list.

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 "i"'th extent.

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 (struct nbdkit_next_ops *next_ops,
                             nbdkit_backend *nxdata,
                             uint32_t count, uint64_t offset,
                             uint32_t flags, uint32_t align,
                             struct nbdkit_extents *extents, int *err);

Calls next_ops->extents as needed until at least "align" bytes are obtained, where "align" is a power of 2. Anywhere the underlying plugin returns differing extents within "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 "count" or "offset" that is not already aligned.

".cache"¶

 int (*cache) (struct nbdkit_next_ops *next_ops, void *nxdata,
               void *handle, uint32_t count, uint64_t offset,
               uint32_t flags, int *err);

This intercepts the plugin ".cache" method and can be used to modify cache requests.

This function will not be called if ".can_cache" returned "NBDKIT_CACHE_NONE" or "NBDKIT_CACHE_EMULATE"; in turn, the filter should not call "next_ops->cache" unless "next_ops->can_cache" returned "NBDKIT_CACHE_NATIVE".

The parameter "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 "next_ops->cache".

If there is an error, ".cache" should call "nbdkit_error" with an error message and return -1 with "err" set to the positive errno value to return to the client.

Debug Flags¶

Debug Flags in filters work exactly the same way as plugins. See "Debug Flags" in nbdkit-plugin(3).