Device API

From The Open Source Backup Wiki (Amanda, MySQL Backup, BackupPC)
Revision as of 15:36, 25 May 2008 by Dustin (talk | contribs) (significant reorganization begun)
Jump to navigationJump to search

The Device API is a clean interface between Amanda and data-storage systems. It provides a tape-like model -- a sequence of bytestreams on each volume, identified only by their on-volume file number -- even of non-tape devices.

Background

Device API Features

The Device API also adds a number of new features not previously available in Amanda. These include:

  • Device properties allow drivers to describe themselves (and the devices and media they control), as well as allow arbitrary user settings to propagate down to the driver.
  • Smart locking means unrelated accesses can be performed without issue, while conflicting accesses wait for one another.
  • Appending to volumes is supported for capable devices.
  • A device can describe its supported blocksize range to the Amanda core, instead of the other way around.
  • Deleting parts of volumes (without erasing the whole thing) is supported for capable devices.
  • The amount of free space on a device can be reported to the Amanda core, where the device supports it (e.g., VFS Device)

History

The Device API is designed to replace the ancient tapeio subsystem (sometimes called the vtape API or Virtual Tape API), originally introduced to support virtual tapes. The original design of the tapeio system (located in tape-src/) was to abstract tape-related functionality into a separate library. Thus, different devices (tape, vtape and RAIT) were all phrased in terms of tape operations: Rewind, fast-forward, read a block, etc. Operations were assumed to have the same semantics as a UNIX tape device. Furthermore, the API revealed a file descriptor, with the assumption that other parts of Amanda could perform operations such as stat() or dup2() on it.

The Device API clears up a large number of limitations and device assumptions that the tapeio system made, allowing native support for new devices (such as CDs and DVDs) as well as new device-related functionality (such as parallel access, partial recycling, and appending).

The Device API does not change (or even address) media formats.

Limitations of tapeio

Important limitations inherent in the design of tapeio include:

  • Exposes a UNIX file descriptor
  • Operates in terms of UNIX operations
  • Does not store information in a portable (yet device-specific) way.
  • Not reentrant: Access to the device is not permitted from multiple threads or processess.
  • Assumes devices can be opened and closed at will.
  • RAIT driver performs operations in series (rather than parallel).

Tapeio Operations

Tapeio included the following operations:

  • tape_access(), tape_open(), tape_stat(), tapefd_close(), tapefd_read(), tapefd_write(): Act like their UNIX equivalents
  • tapefd_rewind(), tapefd_unload(): Do what you expect.
  • tapefd_fsf(): Seeks forward a certain number of filemarks.
  • tapefd_weof(): Writes a filemark.
  • tapefd_resetofs(): Workaround for some buggy kernels.
  • tapefd_status(): Prints status to stdout (!).

In addition, all tapefd_ commands have matching tape_ commands, which work on an unopened device.

What the Device API is not

The Device API does not distinguish between random-access and linear-access media: The seek operation may take a long time for some devices, or it may be instantaneous for others. It does, however, distinguish between concurrent devices and exclusive devices: Concurrent devices may be accessed by multiple readers (and sometimes multiple writers) simultaneously, while exclusive devices cannot.

Although the Device API deals in on-medium headers and blocks, it is otherwise agnostic to media format issues. Changes to the Amanda media format, including split vs traditional formats, can be made without change to the Device API.

Specification

The Device API is implemented in the device-src/ directory. It relies heavily on GLib's type system, so developers unfamiliar with the GObject system are encouraged to consult the relevant documentation

Instantiation

Broadly speaking, the C interface revolves around a single glib virtual class -- Device -- which is actually implemented as one of many possible subclasses. Subclasses are identified with device-name prefixes, e.g., tape: for the Tape Device or s3: for the S3 Device. device_open takes a device name, selects a Device subclass based on the prefix, instantiates the subclass, and passes the remainder of the device name to the subclass's factory function.

Properties

Device properties provide a bidirectional means of communication between devices and their users. A device provides values for some properties, which other parts of Amanda can use to adjust their behavior to suit the device. For example, Amanda will only attempt to append to a volume if the device's properties indicate that it supports this activity. Some devices have additional properties that can be set to control its activity. For example, the S3 Device requires that the users' keys be given via properties.

The Amanda configuration interface for properties is not completely developed. Currently, global device_property parameters are parsed by device_get_startup_properties_from_config(), which means that individual properties cannot be specified for individual devices.

Standard Properties

The Device API includes a number of standard properties, and allows devices to declare and register other custom properties as may be appropriate. Standard properties include (TODO: list these below, and list individual properties on pages for individual devices):

  • Concurrency: Does the device support concurrent readers and/or writers?
  • Streaming: Does the device require (or desire) streaming data?
  • Compression: Does the device support it, and what compression rate has been yielded.
  • Blocksize: What block size(s) are supported by the device? This property is settable for devices that support a variety of block sizes, or can be left unset for variable block size.
  • Device UUID: Returns a unique identifier for this piece of hardware.
  • Media access mode: What is the access paradigm for this volume? Can be Read-Only, Write-Once-Read-Many (WORM), Read-Write, or Write-Once-Read-Never.
  • Feature support: Does the device support partial deletion or appending to volumes?

Properties Interface

Most code related to device-independent properties is in property.h and property.c. Since property values are passed around as a GValue, all value types must be registered in the GLib type system.

Every abstract property has a DevicePropertyBase, which refers to a DevicePropertyId. The Base structure is the same for all properties of all devices; it holds information about the property outside other context. Each property has a unique DevicePropertyId, though the ID of a property is only guaranteed within a single run of a program: Between runs, IDs might change, so it's best to refer to properties by name outside of a particular program.

When a device starts up, it creates a DeviceProperty structure for each supported property. This structure refers to the DevicePropertyBase, but also holds information about when the property may be accessed: Not all properties may be set or gotten at any time. The PropertyAccessFlags field access holds this information: it is the bitwise OR of any of the various PROPERTY_ACCESS_GET_ and PROPERTY_ACCESS_SET_ values defined. Note that there are five distinct "time periods" as far as access is concerned:

  • BEFORE_START: Before device_start() has been called; i.e., before any permanant action may have been taken.
  • BETWEEN_FILE_WRITE: When in write mode, but not actually writing a file. Specifically, after device_start() has been called with ACCESS_WRITE or ACCESS_APPEND, but outside of a pair of calls to device_start_file() and device_finish_file().
  • INSIDE_FILE_WRITE: In the middle of writing a file. Specifically, after device_start_file() has been called, but before the coresponding call to device_finish_file().
  • BETWEEN_FILE_READ: When in read mode, but not actually reading a file. This means after device_start() has been called with ACCESS_READ, but either before a call to device_seek_file, or after a call to device_read_block returns EOF (provided no intervening call to device_seek_file() or device_seek_block() since the EOF).
  • INSIDE_FILE_READ: When in read mode, and while actually reading a file. This means after a call to device_seek_block(), but before a call to device_read_block() returns EOF.

Adding a new property requires modifying both property.h and property.c, as C code in property.c is required to register the properties and generate their IDs. See those files for the symbolic names for the standard properties described above.


Using the Device class

To make use of the Device API, call one of the device_... functions defined in device-src/device.h. Functions of interest:

/* This is how you get a new Device. Pass in a device name like
 * file:/path/to/storage, and (assuming everything goes OK) you will get
 * back a nice happy Device* that you can do operations on. Note that you
 * must device_start() it before you can do anything besides talk about
 * properties or read the label. device_name remains the responsibility
 * of the caller. */
Device* 	device_open	(char * device_name);

/* This instructs the device to read the label on the current
 * volume. It may called automatically during device_open() or
 * device_start().  If you need the most up-to-date information (for
 * example, after setting properties that may affect the result),
 * call the function explicitly.  Otherwise, consult
 * device->volume_label, which may be NULL. */
ReadLabelStatusFlags        device_read_label (Device * self);

/* This tells the Device that it's OK to start reading and writing
 * data. Before you call this, you can only call
 * device_property_{get, set} and device_read_label. You can only call
 * this a second time if you call device_finish() first.
 *
 * You should pass a label and timestamp if and only if you are
 * opening in WRITE mode (not READ or APPEND). The label and timestamp
 * remain the caller's responsibility in terms of memory management. The
 * passed timestamp may be NULL, in which case it will be filled in with 
 * the current time. */
gboolean 	device_start	(Device * self,
                                 DeviceAccessMode mode, char * label,
                                 char * timestamp);

/* This undoes device_start, returning you to the NULL state. Do this
 * if you want to (for example) change access mode.
 * 
 * Note to subclass implementors: Call this function first from your
 * finalization function. */
gboolean 	device_finish	(Device * self);

/* But you can't write any data until you call this function, too.
 * This function does not take ownership of the passed dumpfile_t; you must
 * free it yourself. */
gboolean        device_start_file       (Device * self,
                                         const dumpfile_t * jobInfo);

guint           device_write_min_size   (Device * self);
guint           device_write_max_size   (Device * self);
guint           device_read_max_size   (Device * self);

/* Does what you expect. size had better be inside the block size
 * range, or this function will write nothing.
 *
 * The short_block parameter needs some additional explanation: If
 * short_block is set to TRUE, then this function will accept a write
 * smaller than the minimum block size, subject to the following
 * caveats:
 * % The block may be padded with NULL bytes, which will be present on
 *   restore.
 * % device_write_block will automatically call device_finish_file()
 *   after writing this short block.
 * It is permitted to use short_block with a block that is not short;
 * in this case, it is equivalent to calling device_write() and then
 * calling device_finish_file(). */
gboolean 	device_write_block	(Device * self,
                                         guint size,
                                         gpointer data,
                                         gboolean short_block);

/* This will drain the given fd (reading until EOF), and write the
 * resulting data out to the device using maximally-sized blocks. */
gboolean 	device_write_from_fd	(Device * self,
					int fd);

/* Call this when you are finished writing a file. This function will
 * write a filemark or the local equivalent, flush the buffers, and do
 * whatever dirty work needs to be done at such a point in time. */
gboolean 	device_finish_file	(Device * self);

/* For reading only: Seeks to the beginning of a particular
 * filemark. Only do this when reading; opening in
 * ACCESS_WRITE will start you out at the first file, and opening in
 * ACCESS_APPEND will automatically seek to the end of the medium.
 * 
 * If the requested file doesn't exist, this function will seek to the
 * next-numbered valid file. You can check where this function seeked to
 * by examining the file field of the Device structure. If the requested
 * file number is exactly one more than the last valid file, this
 * function returns a TAPEEND header.
 *
 * If an error occurs or if the requested file is two or more beyond the
 * last valid file, this function returns NULL.
 *
 * Example results for a volume that has only files 1 and 3:
 * 1 -> Seeks to file 1
 * 2 -> Seeks to file 3
 * 3 -> Seeks to file 3
 * 4 -> Returns TAPEEND
 * 5 -> Returns NULL
 *
 * The returned dumpfile_t is yours to keep, at no extra charge. */
dumpfile_t* 	device_seek_file	(Device * self,
					guint file);

/* After you have called device_seek_file (and /only/ after having
 * called device_seek_file), you can call this to seek to a particular
 * block inside the file. It works like SEEK_SET, only in blocks. */
gboolean 	device_seek_block	(Device * self,
					guint64 block);

/* After you have called device_seek_file and/or device_seek_block,
 * you can start calling this function. It always reads exactly one whole
 * block at a time, however big that might be. You must pass in a buffer and
 * specify its size. If the buffer is big enough, the read is
 * performed, and both *size and the return value are equal to the
 * number of bytes actually read. If the buffer is not big enough, then
 * no read is performed, the function returns 0, and *size is set
 * to the minimum buffer size required to read the next block. If an
 * error occurs, the function returns -1  and *size is left unchanged.
 * 
 * It is not an error if buffer == NULL and *size == 0. This should be
 * treated as a query as to the possible size of the next block. */
int 	device_read_block	(Device * self,
                                 gpointer buffer,
                                 int * size);

/* This is the reading equivalent of device_write_from_fd(). It will
 * read from the device from the current location until end of file,
 * and drains the results out into the specified fd. Returns FALSE if
 * there is a problem writing to the fd. */
gboolean 	device_read_to_fd	(Device * self,
					int fd);

/* Returns TRUE if the last device_read_block returned -1 because
 * there is no data left to read. */
gboolean        device_is_eof           (Device * self);

/* This function tells you what properties are supported by this
 * device, and when you are allowed to get and set them. The return
 * value is an array of DeviceProperty structs. The last struct in
 * the array is zeroed, so you know when the end is (check the
 * pointer element "base"). The return value from this function on any
 * given object (or physical device) should be invariant. */
const DeviceProperty * 	device_property_get_list	(Device * self);

/* These functions get or set a particular property. The val should be
 * compatible with the DevicePropertyBase associated with the given
 * DevicePropertyId, and this function should only be called when
 * DeviceProperty.access says it is OK. Otherwise you will get an
 * error and not the tasty property action you wanted. */
gboolean 	device_property_get	(Device * self,
                                         DevicePropertyId id,
                                         GValue * val);
gboolean 	device_property_set	(Device * self,
                                         DevicePropertyId id,
                                         GValue * val);

/* On devices that support it (check PROPERTY_PARTIAL_DELETION),
 * this will free only the space associated with a particular file. 
 * This way, you can apply a different retention policy to every file
 * on the volume, appending new data at the end and recycling anywhere
 * in the middle -- even simultaneously (via different Device
 * handles)! Note that you generally can't recycle a file that is presently in
 * use (being read or written).
 *
 * To use this, open the device as DEVICE_MODE_APPEND. But you don't
 * have to call device_start_file(), unless you want to write some
 * data, too. */
gboolean 	device_recycle_file	(Device * self,
					guint filenum);

Implementing a device

To add a new device to the Device API, do the following:

  • Implement a subclass of one of the existing Device classes named above. In particular, implement the various virtual functions provided for in device.h.
    • Using the FdDevice class for devices that have a unix file discriptor will spare you from implementing device_read_block() and device_write_block(), but there are other virtual functions in FdDevice that you (may) need to reimplement.
    • Note that in addition to the (mostly virtual) functions discussed above, there are some additional protected functions that require implementation.
  • Make sure your device calls device_add_property() to register both its dynamic and static properties.
  • Arrange to call the register_device() function to register the device, so it will recieve relevant calls to device_open(). The easiest way to do this is to write a initialization function, and call it from device_api_init().

Users new to GLib will need to read up on the GObject system first, to see how GLib provides virtual abstract classes in C and how to implement a new subclass. Additionally, the existing device drivers may serve as a useful template.

Future Directions

Future directions for the Device API include:

  • new devices
  • runtime loading of device modules as shared objects
  • more flexible device configuration