xref: /external/libmtp/README

Building and Installing
-----------------------

See the "INSTALL" file.


Heritage
--------

libmtp is based on several ancestors:

* libptp2 by Mariusz Woloszyn was the starting point used
  by Richard A. Low for the initial starter port. You can
  find it at http://libptp.sourceforge.net/

* libgphoto2 by Mariusz Woloszyn and Marcus Meissner was
  used at a later stage since it was (is) more actively
  maintained. libmtp tracks the PTP implementation in
  libgphoto2 and considers it an upstream project. We will
  try to submit anything generally useful back to libgphoto2
  and not make double efforts. In practice this means we
  use ptp.c, ptp.h and ptp-pack.c verbatim from the libgphoto2
  source code. If you need to change things in these files,
  make sure it is so general that libgphoto2 will want to
  merge it to their codebase too. You find libgphoto2 as part
  of gPhoto: http://gphoto.sourceforge.net/

* libnjb was a project that Richard and Linus were working
  on before libmtp. When Linus took Richards initial port
  and made an generic C API he re-used the philosophy and
  much code from libnjb. Many of the sample programs are for
  example taken quite literally from libnjb. You find it here:
  http://libnjb.sourceforge.net/


Contacting and Contributing
---------------------------

See the project page at http://libmtp.sourceforge.net/
We always need your help. There is a mailinglist and a
bug report system there.

People who want to discuss MTP devices in fora seem to
hang out on the forums at AnythingbutiPod:
http://www.anythingbutipod.com/forum/


Compiling programs for libmtp
-----------------------------

libmtp has support for the pkg-config script by adding a libmtp.pc
entry in $(prefix)/lib/pkgconfig. To compile a libmtp program,
"just" write:

gcc -o foo `pkg-config --cflags --libs libmtp` foo.c

This also simplifies compilation using autoconf and pkg-config: just
write e.g.

PKG_CHECK_MODULES(MTP, libmtp)
AC_SUBST(MTP_CFLAGS)
AC_SUBST(MTP_LIBS)

To have libmtp LIBS and CFLAGS defined. Needless to say, this will
only work if you have pkgconfig installed on your system, but most
people have nowadays.

If your library is installed in e.g. /usr/local you may have to tell
this to pkgconfig by setting the PKG_CONFIG_PATH thus:

export PKG_CONFIG_PATH=/usr/local/lib/pkgconfig


Documentation
-------------

Read the API documentation that can be generated with doxygen.
It will be output in doc/html if you have Doxygen properly
installed. (It will not be created unless you have Doxygen!)

For information about the Media Transfer Protocol, see:
http://en.wikipedia.org/wiki/Media_Transfer_Protocol

The official 1.0 specification for MTP was released by the
USB Implementers Forum in may, 2008. Prior to this, only a
proprietary Microsoft version was available, and quite a few
devices out there still use some aspects of the Microsoft
version, which deviates from the specified standard. You can
find the official specification here:
http://www.usb.org/developers/devclass_docs/MTP_1.0.zip


The Examples
------------

In the subdirectory "examples" you find a number of
command-line tools, illustrating the use of libmtp in very
simple terms.

Please do not complain about the usability or documentation
of these examples, they look like they do for two reasons:

1. They are examples, not tools. If they were intended for
   day-to-day usage by commandline freaks, I would have
   called them "tools" not "examples".

2. The MTP usage paradigm is that a daemon should hook
   the device upon connection, and that it should be
   released by unplugging. GUI tools utilizing HAL (hald)
   and D-Bus do this much better than any commandline
   program ever can. (See below on bugs.) Specificationwise
   this is a bug, however it is present in many, many
   devices.

That said, if you want to pick up and maintain the examples,
please volunteer.


New Devices
-----------

If you happen upon a device which libmtp claims it cannot
autodetect, please submit the vendor ID and device ID
(these can be obtained from the "lsusb" and "lsusb -n"
commands run as root) as a bug, patch or feature request
on the Sourceforge bug tracker at our homepage. If it
gives a sensible  output from "mtp-detect" then please attach
the result as well as it teach us some stuff about your
device. If you've done some additional hacking, join our
mailinglist and post your experiences there.

If you want to be able to hack some more and you're not
afraid of C hacking, add an entry for your device's
vendor/product ID and a descriptive string to the database
in the file src/music-players.h.

If you want to poke around to see if your device has some
special pecularities, you can test some special device
flags (defined in src/device-flags.h) by inserting them
together with your device entry in src/music-players.h.
Flags can be tested in isolation or catenated with "|"
(binary OR). If relatives to your device use a certain
flag, chances are high that a new device will need it
too, typically from the same manufacturer.

The most common flag that needs to be set is the
DEVICE_FLAG_UNLOAD_DRIVER that detach any Linux kernel
drivers that may have attached to the device making
MTP access impossible. This is however not expected to
really work: this is a problem being tracked as of
now (2007-08-04). See the "last resort" solutions below
if you really need to get your dual-mode device to work
with MTP.

Another flag which is easy to identify is the
DEVICE_FLAG_NO_ZERO_READS, which remedies connection
timeouts when getting files, and some timeouts on e.g.
successive "mtp-connect" calls.

If you are a device vendor, please consider assigning one
of your employees as a contact person for libmtp, have them
sign up to the libmtp development list and answer questions
and post new device ID:s as they are released to our
mailing list. By the way: do you have spare devices you
can give us? Send them to Richard (Mac support) or Linus
(Linux support). (So far nobody did that except for Microsoft
who sent us a Zune by proxy!)

If your device is very problematic we are curious of how it
works under Windows, so we enjoy reading USB packet sniffs
that reveal the low-level traffic carried out between
Windows Media Player and your device. This can be done
using e.g.:

* USBsnoop:
  http://benoit.papillault.free.fr/usbsnoop/

* The trial version of HHD Softwares software-only
  USB monitor. You need to get a copy of version 2.37 since
  the newer trial versions won't let you carry out the
  needed packet sniffs. (As of 2007-03-10 a copy can be found
  at: http://www.cobbleware.com/files/usb-monitor-237.exe)

There are other USB monitors as well, some more expensive
alternatives use hardware and even measure electronic
characteristics of the traffic (which is far too much
detail for us).

Device sniffs are an easy read since the PTP/MTP protocol
is nicely structured. All commands will have a structure such
as this in the log, we examplify with a object list request:

PTP REQEUST:
000120: Bulk or Interrupt Transfer (UP), 03.09.2007 12:49:25.9843750 +0.0
Pipe Handle: 0x863ce234 (Endpoint Address: 0x2)
Send 0x20 bytes to the device:
 20 00 00 00 01 00 05 98 23 00 00 00 27 03 00 10    ......?#...'...
 Length      TYPE  CMD   Trans#      Param1

 00 00 00 00 02 DC 00 00 00 00 00 00 00 00 00 00   .....Ü..........
 Param2      Param3      Param4      Param5

[OPTIONAL] DATA PHASE:
000121: Bulk or Interrupt Transfer (UP), 03.09.2007 12:49:26.0 +0.0156250
Pipe Handle: 0x863ce214 (Endpoint Address: 0x81)
Get 0x1a bytes from the device:
 1A 00 00 00 02 00 05 98 23 00 00 00 01 00 00 00   .......?#.......
 Length      TYPE  CMD   Trans#      DATA

 27 03 00 10 02 DC 04 00 00 30                     '....Ü...0

RESPONSE:
000122: Bulk or Interrupt Transfer (UP), 03.09.2007 12:49:26.0 +0.0
Pipe Handle: 0x863ce214 (Endpoint Address: 0x81)
Get 0xc bytes from the device:
 0C 00 00 00 03 00 01 20 23 00 00 00               ....... #...
 Length      TYPE  CODE  Trans#

* One send (OUT to the device), two reads (IN from the device).

* All three byte chunks commands are
  sent/recieved/recieeved by the function  ptp_transaction()
  in the file ptp.c.

* It boils down to ptp_usb_sendreq(), optionally ptp_usb_senddata()
  or ptp_usb_getdata() and finally ptp_usb_getresp() in the file
  libusb-glue.c. Notice ptp_usb_sendreq() and ptp_usb_getresp()
  are ALWAYS called. The TYPE field correspond to this, so the
  TYPES in this case are "COMMAND" (0x0001), "DATA" (0x0002),
  and "RESPONSE" (0x0003).

* Notice that the byte order is little endian, so you need to read
  each field from right to left.

* This COMMAND has:
  CMD 0x99805, we see in ptp.h that this is PTP_OC_MTP_GetObjPropList.
  Transaction# 0x00000023.
  REQUEST parameters 0x10000327, 0x00000000, 0x0000DC02, 0x00000000
    0x00000000, in this case it means "get props for object 0x10000327",
    "any format", "property 0xDC02" (PTP_OPC_ObjectFormat), then two
    parameters that are always zero (no idea what they mean or their
    use).

* The DATA has:
  CMD 0x99805, we see in ptp.h that this is PTP_OC_MTP_GetObjPropList.
  Transaction# 0x00000023.
  Then comes data 0x00000001, 0x10000327, 0xDC02, 0x0004, 0x3000
  Which means in this case, (and this is the tricky part) "here
  you have 1 property", "for object 0x10000327", "it is property
  0xDC02" (PTP_OPC_ObjectFormat), "which is of type 0x0004"
  (PTP_DTC_UINT16), "and set to 0x3000" (PTP_OFC_Undefined, it
  is perfectly valid to have undefined object formats, since it
  is a legal value defining this).

* This RESPONSE has:
  CMD 0x99805, we see in ptp.h that this is PTP_OC_MTP_GetObjPropList.
  Return Code ("RC") = 0x2001, PTP_RC_OK, all went fine.
  Transaction# 0x00000023.

If you want to compare the Windows behaviour with a similar
operation using libmtp you can go into the src/libusb-glue.c
file and uncomment the row that reads:

//#define ENABLE_USB_BULK_DEBUG

(I.e. remove the two //.)

This will make libmtp print out a hex dump of every bulk USB
transaction. The bulk transactions contain all the PTP/MTP layer
data, which is usually where the problems appear.


Dual-mode devices does not work - last resort:
----------------------------------------------

Some devices that are dual-mode are simply impossible to get
to work under Linux because the usb-storage(.ko) kernel
module hook them first, and refuse to release them, even
when we specify the DEVICE_FLAG_UNLOAD_DRIVER flag. (Maybe
it DOES release it but the device will immediately be probed
at the USB mass storage interface AGAIN because it
enumerates.)

Here is what some people do:

 1. Plug in the device.
 2. USB-mass storage folder will open automatically.
 3. Unmount the device.
 4. Run mtp-detect. It will most likely fail the first time.
 5. Run mtp-detect again, it might work this time, or fail. Keep running
    till it works. 99% it works by the third try.
 6. Once mtp-detect gives you an "Ok", open either Rhythmbox or Gnomad2,
    everything should work.

Linux: Try this, if you have a recent 2.6.x Linux kernel,
run (as root) something like:

> rmmod usb_storage ; mtp-detect

You can run most any command or a client like gnomad2 or
Amarok immediately after the rmmod command. This works
sometimes. Another way:

* Edit /etc/modprobe.d/blacklist

* Add the line "blacklist usb-storage"

* Reboot.

Now none of you USB disks, flash memory sticks etc will be
working (you just disabled them all). However you *can* try
your device, and it might have started working because there
is no longer a USB mass storage driver that tries to hook onto
the mass storage interface of your device.

If not even blacklisting works (check with
"lsmod | grep usb-storage"), there is some problem with
something else and you may need to remove or rename the file
/lib/modules/<VERSION>/kernel/drivers/usb/storage/usb-storage.ko
manually.

If you find the PerfectSolution(TM) to this dilemma, so you
can properly switch for individual devices whether to use it
as USB mass storage or not, please tell us how you did it. We
know we cannot use udev, because udev is called after-the-fact:
the device is already configured for USB mass storage when
udev is called.

On Mac OS there is another ugly hack:

1. Open up a terminal window
2. Type:
sudo mv /System/Library/Extensions/IOUSBMassStorageClass.kext
/System/Library/Extensions/IOUSBMassStorageClass.kext.disabled

and when prompted enter your password.

3. Restart.

To reverse this change, just reverse the filenames:

sudo mv /System/Library/Extensions/
IOUSBMassStorageClass.kext.disabled /System/Library/Extensions/
IOUSBMassStorageClass.kext

and restart.


Calendar and contact support:
-----------------------------

The Creative Zen series can read VCALENDAR2 (.ics) files
and VCard (.vcf) files from programs like for example
Evolution with the following limitations/conditions:

- The file must be in DOS (CR/LF) format, use the unix2dos
  program to convert if needed

- Repeat events in calendar files do not seem to be supported,
  entries will only appear once.

- Calendar (.ics) files should be stored in the folder "My Organizer"
  when sent to the device (this directory should be autodetected
  for use with calendar files, otherwise use the option
  -f "My Organizer" to sendfile for this) Apparently this file can
  also contain tasklists.

- Contact (.vcf) files should be stored in the folder "My Contacts"
  when sent to the device. (-f "My Contacts")

- Some devices are picky about the name of the calendar and
  contact files. For example the Zen Microphoto wants:

  Calendar: My Organizer/6651416.ics
  Contacts: My Organizer/6651416.vcf


Syncing in with Evolution and Creative Devices
----------------------------------------------

Evolution can easily export .ics an .vcf files, but you currently
need some command-line hacking to get you stuff copied over in
one direction host -> device. The examples/ directory contains a script
created for the Creative Zen Microphoto by Nicolas Tetreault.


It's Not Our Bug!
-----------------

Some MTP devices have strange pecularities. We try to work around
these whenever we can, sometimes we cannot work around it or we
cannot test your solution.

* Generic MTP/PTP disconnect misbehaviour: we have noticed that
  Windows Media Player apparently never close the session to an MTP
  device. There is a daemon in Windows that "hooks" the device
  by opening a PTP session to any MTP device, whenever it is
  plugged in. This daemon proxies any subsequent transactions
  to/from the device and will never close the session, thus
  Windows simply does not close sessions at all.

  Typical sign of this illness: broken pipes on closing sessions,
  on the main transfer pipes(s) or the interrupt pipe:

    Closing session
    usb_clear_halt() on INTERRUPT endpoint: Broken pipe
    OK.

  This means that device manufacturers doesn't notice any problems
  with devices that do not correctly handle closing PTP/MTP
  sessions, since Windows never do it. The proper way of closing
  a session in Windows is to unplug the device, simply put.

  Since libmtp actually tries to close sessions, some devices
  may fail since the close session functionality has never been
  properly tested, and "it works with Windows" is sort of the
  testing criteria at some companies.

  You can get Windows-like behaviour on Linux by running a HAL-aware
  libmtp GUI client like Rhythmbox or Gnomad2, which will "hook"
  the device when you plug it in, and "release" it if you unplug
  it.

  If this bug in your device annoys you, contact your device
  manufacturer and ask them to test their product with some libmtp
  program.

* Generic USB misbehaviour: some devices behave badly under MTP
  and USB mass storage alike, even down to the lowest layers
  of USB. You can always discuss such issues at the linux-usb
  mailing list if you're using Linux:
  http://www.linux-usb.org/mailing.html

  If you have a problem specific to USB mass storage mode, there
  is a list of strange behaving devices in the Linux kernel:
  http://lxr.linux.no/linux/drivers/usb/storage/unusual_devs.h
  You can discuss this too on the mentioned list, for understanding
  the quirks, see:
  http://www2.one-eyed-alien.net/~mdharm/linux-usb/target_offenses.txt

* Kernel bug on Linux. Linux 2.6.16 is generally speaking required
  to use any MTP device under USB 2.0. This is because the EHCI
  driver previously did not support zero-length writes to endpoints.
  It should work in most cases however, or if you connect it
  to an UHCI/OHCI port instead (yielding lower speed). But
  please just use a recent kernel.

* Zen models AVI file seeking problem: the Zens cannot parse the
  files for the runlength metadata. Do not transfer file with e.g.
  mtp-sendfile, use mtp-sendtr and set the length of the track to
  the apropriate number of seconds and it will work. In graphical
  clients, use a "track transfer" function to send these AVI files,
  the Zens need the metadata associated with tracks to play back
  movies properly. Movies are considered "tracks" in the MTP world.

* Some devices that disregard the metadata sent with the MTP
  commands will parse the files for e.g. ID3 metadata. Some still
  of these devices expect only ID3v2.3 metadata and will fail with
  a modern ID3v2,4 tag writer, like many of those found in Linux
  applications. Windows Media Player use ID3v2.3 only, so many
  manufacturers only test this version.

* The Zen Vision:M (possibly more Creative Zens) has a firmware bug
  that makes it drop the last two characters off a playlist name.
  It is fixed in later firmware.

* For Creative Technology devices, there are hard limits on how
  many files can be put onto the device. For a 30 GiB device (like
  the Zen Xtra) the limit is 6000, for a 60 GiB device the limit
  is 15000 files. For further Creative pecularities, see the
  FAQ sections at www.nomadness.net.

* Sandisk sansa c150 and probably several other Sandisk devices
  (and possibly devices from other manufacturers) have a dual
  mode with MTP and USB mass storage. The device will initially
  claim to be mass storage so udev will capture is and make the
  use of MTP mode impossible. One way of avoiding it could be to
  be to blacklist the "usb-storage" module in
  /etc/modprobe.c/blacklist with a row like this:
  "blacklist usb-storage". Some have even removed the
  "usb-storage.ko" (kernel module file) to avoid loading.

* Sandisk Sansa Fuze has three modes: auto, MTP or mass storage
  (MSC). Please set it to MTP to avoid problems with libmtp.

* The iriver devices (possibly all of them) cannot handle the
  enhanced GetObjectPropList MTP command (0x9805) properly. So
  they have been banned from using it.

* iriver devices have problems with older versions of libmtp and
  with new devices libmtp does not know of as of yet, since it
  has an oldstyle USB device controller that cannot handle zero
  writes. (Register your device with us!) All their devices are
  likely to need a special device flag in the src/libusb-glue.c
  database.

* The Samsung Yepp T9 has several strange characteristics, some
  that we've managed to work around. (For example it will return
  multiple PTP packages in a single transaction.)

* The early firmware for Philips HDD players is known to be
  problematic. Please upgrade to as new firmware as you can get.
  (Yes this requires some kind of Windows Installation I think.)

* Philips HDD 1630/16 or 1630/17 etc may lock themselves up,
  turning inresponsive due to internal corruption. This typically
  gives an error in opening the PTP session. Apparently you can
  do a "repair" with the firmware utility (Windows only) which
  will often fix this problem and make the device responsive
  again.

* Some devices that implement GetObjectPropList (0x9805) will
  not return the entire object list if you request a list for object
  0xffffffffu. (But they should.) So they may need the special
  DEVICE_FLAG_BROKEN_MTPGETOBJPROPLIST_ALL.

* Some (smaller) subset of devices cannot even get all the
  properties for a single object in one go, these need the
  DEVICE_FLAG_BROKEN_MTPGETOBJPROPLIST. Currently only the
  iriver devices seem to have this bug.

* The Toshiba Gigabeat S (and probably its sibling the
  Microsoft Zune and other Toshiba devices) will only display
  album information tags for a song in case there is also
  an abstract album (created with the album interface) with
  the exact same name.

* The Zen Vision:M has an older firmware which is very corrupt,
  it is incompatible with the Linux USB stack altogether. The
  kernel dmesg will look something like this, and you have to
  upgrade the firmware using Windows:
  usb 4-5: new high speed USB device using ehci_hcd and address 5
  usb 4-5: configuration #1 chosen from 1 choice
  usb 4-5: can't set config #1, error -110

* The Sirus Stiletto does not seem to allow you to copy any files
  off the device. This may be someone's idea of copy protection.

* The Samsung P2 assigns parent folder ID 0 to all unknown file
  types.(i.e. moves them to the root folder)

Lost symbols
------------

Shared libraries can be troublesome to users not experienced with
them. The following is a condensed version of a generic question
that has appeared on the libmtp mailing list from time to time.

> PTP: Opening session
> Queried Creative Zen Vision:M
> gnomad2: relocation error: gnomad2: undefined symbol:
> LIBMTP_Get_Storageinfo
> (...)
> Are these type of errors related to libmtp or something else?

The problem is of a generic nature, and related to dynamic library
loading. It is colloquially known as "dependency hell".
(http://en.wikipedia.org/wiki/Dependency_hell)

The gnomad2 application calls upon the dynamic linker in Linux to
resolve the symbol "LIBMTP_Get_Storageinfo" or any other symbol
(ELF symbol, or link point or whatever you want to call them, a
symbol is a label on a memory address that the linker shall
resolve from label to actual address.)
For generic information on this subject see the INSTALL file and
this Wikipedia page:

http://en.wikipedia.org/wiki/Library_(computing)

When Linux /lib/ld-linux.so.X is called to link the symbols compiled
into gnomad2 (or any other executable using libmtp), it examines the
ELF file for the libmtp.so.X file it finds first and cannot resolve
the symbol "LIBMTP_Get_Storageinfo" (or whichever symbol you have a
problem witj) from it, since it's probably not there. There are many
possible causes of this symbol breakage:

1) You installed precompiled libmtp and gnomad2 packages (RPMs, debs
    whatever) that do not match up. Typical cause: your gnomad2 package was
    built against a newer version of libmtp than what's installed on your
    machine. Another typical cause: you installed a package you found on
    the web, somewhere, the dependency resolution system did not protest
    properly (as it should) or you forced it to install anyway, ignoring
    some warnings.

2) You compiled libmtp and/or gnomad2 from source, installing both or
    either in /usr/local/lib and /usr/local/bin. This means at compile-time
    gnomad2 finds the libmtp library in /usr/local/lib but at runtime, it
    depends on the Linux system wide library loader (/lib/ld-linux.so.X) in
    order to resolve the symbols. This loader will look into the file
    /etc/ld.so.conf and/or the folder /etc/ld.so.conf.d in order to find
    paths to libraries to be used for resolving the symbols. If you have
    some older version of libmtp in e.g. /usr/lib (typically installed by a
    package manager) it will take precedence over the new version you just
    installed in /usr/local/lib and the newly compiled library in
    /usr/local/lib will *not* be used, resulting in this error message.

3) You really did install the very latest versions (as of writing libmtp
    0.1.5 and gnomad2 2.8.11) from source and there really is no
    pre-installed package of either on your machine. In that case I'm
    totally lost, I have no idea what's causing this.

Typical remedies:

1) If you don't want to mess around with your system and risk these
    situations, only use pre-packaged software that came with the
    distribution or its official support channels. If it still breaks,
    blame your distribution, they're not packaging correctly. Relying on
    properly packaged software and not installing things yourself *is* the
    Linux solution to the "dependency hell" problem.

2) Read about dynamically linked library handling until the stuff I wrote
    about in the previous list sounds like music to your ears, inspect
    your /lib, /usr/lib, /usr/local/lib, /etc/ld.so.conf and the
    /etc/ld.so.conf.d, remove all pre-packed versions using RPM, APT,
    YaST or whatever your distribution uses, compile libmtp and gnomad2
    (or whatever) from source only and you will be enlighted.

I don't know if this helps you, it's the best answer we can give.


API is obscure - I want plain files!
------------------------------------

PTP/MTP devices does not actually contain "files", they contain
objects. These objects have file names, but that is actually
just a name tag on the object.

Folders/directories aren't really such entities: they are just
objects too, albeit objects that can act as parent to other
objects. They are called "associations" and are created in atomic
fashion and even though there is an MTP command to get all the
associations of a certain association, this command is optional
so it is perfectly possible (and most common, actually) to create
devices where the "folders" (which are actually associations) have
no idea whatsoever of what files they are associated as parents to
(i.e. which files they contain). This is very easy for device
manufacturers to implement, all the association (i.e. finding out
which files are in a certain folder) has to be done by the MTP
Initiator / host computer.

Moving a file to a new folder is for example very simple in a
"real" file system. In PTP/MTP devices it is often not even possible,
some devices *may* be able to do that. But actually the only
reliable way of doing that is to upload the file to the host,
download it with the new parent, then delete the old file.
We have played with the idea of implementing this time consuming
function, perhaps we will.

Then the issue that in PTP/MTP it is legal for two files to have
exactly the same path as long as their object IDs differ. A
folder/association can contain two files with the exact same name.
(And on the Creative devices this even works, too, though most devices
implicitly fail at this.) Perhaps one could add some custom hook for
handling that, so they become  /Foo.mp3 and /Foo.mp3(1) or something
similar, but it's really a bit kludgy.

Playlists and albums aren't really files, thinking about
them as files like the hacks in libgphoto2 is really backwards. They are
called associations and are more like a symbolic link that links in a
star-shaped pattern to all the files that are part of the album/playlist.
Some devices (Samsung) thought that was too complicated and have a
different way of storing playlists in an UTF-16 encoded .spl-like file
instead! This is why playlists/albums must have their own structs and
functions.

Plain file access also assumes to be able to write files of an
undetermined size, which is simply not possible in a transactional
file system like PTP/MTP. (See further below.)


I Want Streaming!
-----------------

Streaming reads is easy. Just connect the output file descriptor from
LIBMTP_Get_File_To_File_Descriptor() (and a similar function for tracks)
wherever you want.

People have connected this to TCP sockets for streaming web servers
etc, works like a charm. Some devices will even survive if the callback
functions return non-zero and cancel the download. Some devices will
lock up and even require a reset if you do that. Devices are poorly
implemented so that's life. If you want to stream off a device, the
best idea is always to stream the entire file and discard the stuff
at the end you don't want. It will incur a delay if you e.g. want to
skip between tracks, sadly.

Then we get to the complicated things: streaming WRITES...

There is a function:
LIBMTP_Send_File_From_File_Descriptor() (and similar for tracks)
which will write a file to a device from a file descriptor, which may
be a socket or whatever.

HOWEVER: this requires a piece of metadata with the .filesize properly
set first.

This is not because we think it is funny to require that, the protocol
requires it. The reason is that PTP/MTP is a transactional file system
and it wants to be able to deny file transfer if the file won't fit on
the device, so the transaction never even starts, it's impossible to
start a transaction without giving file length.

People really want streaming so I tried a lot of hacks to see if they
would work, such as setting file size to 0xffffffffU or something other
unnaturally big and then aborting the file transfer when the stream ends.
It doesn't work: either the device crashes or the file simply disappears
since the device rolls back all failed transactions.

So this is an inherent limitation of the PTP/MTP protocol.