Passing an empty string as a filename for obc_transfer_get will be
similar to passing a NULL filename. This means a temporary file will be
created to store the content of the transfer.
NULL and "" are not exactly equivalent though: in case of NULL the file
will be automatically removed immediately after being open, which means
that the transfer initiator should also open the file to prevent it from
being removed (to be used from the modules). In this case, the filename
will not be exposed in D-Bus.
On the other hand, if "" is given, the file will be removed only in case
of error. So after success the transfer initiator should decide whether
the file should be removed or not.
This change is convenient in order to expose the same API in D-Bus.
There is no reason to have inconsistent behavior between GetFile and
PutFile, in FileTransfer D-Bus API.
Before this change, PutFile reported success immediately after queueing
the transfer, even though the D-Bus signature includes the async flag.
Relying on a internal policy (based on transfer type) to decide if a
transfer should be exposed or not in D-Bus has some limitations. The
simplest possible alternative to this is to expose all transfers in
D-Bus, assuming the overhead is not significant.
The authorization mechanism is entirely removed from the session, and
thus transfers are automatically started (once popped from the queue)
without confirmation and without any name/filename change.
The number of transferred bytes is exposed in D-Bus using a specific
property for this purpose.
Internally, the value of this property does not necessarily match the
internal progress counter. In order to avoid D-Bus overhead, the
property will be updated once per second.
With some only kernels auto assigning psm seem to fail, leaving the psm
with wrong value:
obexd[2930]: plugins/bluetooth.c:start() listening on channel 9
obexd[2930]: bluetooth: unable to listen in psm 65535
obexd[2930]: plugins/bluetooth.c:start() listening on channel 10
obexd[2930]: bluetooth: unable to listen in psm 65535
After this patch the modules are responsible for creating the transfers,
and these objects must be queued using the session API.
This way the transfer initiator has full access to the transfer object,
in case for example it wants to access some member variable.
Transfer API now takes const buffers (both params and contents) and
internally copies the memory as necessary. This new API is safer to use,
which is convenient if the modules would start using it directly.
The transfer-creating functions (obc_transfer_get and obc_transfer_put)
no longer register the transfer automatically.
This separation makes it possible that the modules would create the
transfers and then pass the object to the session, which would be
responsible for the registration.
The creation process has been internally split into two steps: creation
and D-Bus registration. This is easier to understand and it also allows
to expose these two-steps in the transfer API.
obc_transfer_get() and obc_transfer_put() should only assume ownership
of the given params only in case of success. Otherwise some erros might
result in a double free of such memory.
Writing anything to syslog in the signal handler can cause a deadlock
with any ongoing syslog write.
This is also aligned with what BlueZ and other projects has been doing.
It is safer to remove the transfer from the internal queue (including
session->p) before calling the transfer callback. This makes sure the
callback will not manipulate the session in a way that the transfer is
removed more than once.
This was previously protected with session->p->id != 0 checks, but once
the new callbacks have been adopted in session API, this logic can be
removed.
Operations involving a transfer object will receive a pointer to such
transfer in the callback.
Note that the ownership of this object is not changed in any way,
meaning that the session is still responsible for it. However this
pointer can be useful during the execution of the callback, in order to
access data members of the transfer.
Trivial changes in buffer getters in both session and transfer,
regarding the access of transfer parameters:
- const qualifiers added, to avoid unwanted frees
- Buffers are now returned as void* instead of guint8*
This simplify the API a bit by not having to call obc_transfer_set_file
to open the file.
In addition to that split transfer creation/registration function so
GET/PUT can have more specific logic and different paramenters.
A new enum type is used to distinguish put and get transfers.
This is more convenient since it is done when registering the transfer,
and not when it is actually started. The main benefits would be:
- Some actions can be taken during creation, such as opening files.
- session.c gets simplified.
- The size of a put transfer can be exposed in D-Bus, while queued.
- The transfer operation (put or get) can be exposed in D-Bus.
None of these D-Bus changes are included in this patch.
The errorcode field is set but never used, so it can safely be removed.
In addition there is no need for such a field, because errors can be
propagated using the available callback.
16 bytes in 1 blocks are definitely lost in loss record 26 of 146
at 0x4A075B2: realloc (vg_replace_malloc.c:525)
by 0x3B5104B76D: g_realloc (in /lib64/libglib-2.0.so.0.3000.2)
by 0x3B51064A96: ??? (in /lib64/libglib-2.0.so.0.3000.2)
by 0x3B51065156: g_string_insert_len (in /lib64/libglib-2.0.so.0.3000.2)
by 0x3B510305BC: ??? (in /lib64/libglib-2.0.so.0.3000.2)
by 0x3B51031BE7: g_build_filename (in /lib64/libglib-2.0.so.0.3000.2)
by 0x416FEE: pbap_select (pbap.c:254)
by 0x406CCE: process_message (object.c:224)
by 0x3B5301D9A0: ??? (in /lib64/libdbus-1.so.3.5.6)
by 0x3B5300F92F: dbus_connection_dispatch (in /lib64/libdbus-1.so.3.5.6)
Commit c07ddfbd019d3545cce2d7ec694143cdc55a2167 introduced the freeing of
the active pending request on obc_session_shutdown without checking if
the request was already processed/terminated.
This structure is used for giving filters as input argument to
messages_get_messages_listing(). There is no need for the string members
to be modifiable in this context. And making them const plays well with
map_ap_get_string(), from which the values of these members are going to
be retrieved.
The string pointers here are simply references to data held in map_ap_t
*inparams.
If g_slist_free_full is not supported by glib, build fails with error:
plugins/messages.o: In function `get_folder_listing':
/home/fdanis/src/obexd/plugins/messages.c:199: undefined reference to `g_slist_free_full'
Functions obc_session_get and obc_session_pull nearly share the same
code, so the later can be achieved by just calling the first one.
The session api is not modified in this patch.
The terms "name", "filename" and "targetname" are used in session.h that
can be confusing. This patch proposes to follow the terminology in the
D-Bus api:
- Name: the remote name of the object being transferred
- Filename: the local filesystem name of a file being sent
- Targetfile: the local filesystem name of a file being
received
The terms can be quite misleading, so this patch proposes to follow the
terminology in the D-Bus api:
- Name: the remote name of the object being transferred
- Filename: the name of the file in local the filesystem
Both values can be NULL independently.
This fixes the problem of using the terms differently in get and put
operations. The result was that the properties "Name" and "Filename" were
swapped in D-Bus in the case of get.
Once the fields map to obex fields, the interpretation of the response
from the agent becomes more complicated. Depending on the transfer type,
either the name or the filename field must be updated.
Previous implementation of session_terminate_transfer assumed that the
transfer being terminated would always be the active one. However, it
should be possible to cancel any queued transfer using the D-Bus api.
Queueing the error won't remove the original packet created by transfer
from the queue so upon resume gobex will attempt to send it again.
To fix this we no longer create a error packet instead the session is
market as aborted and the error stored so when gobex finally resumes the
error is forward properly.
The authorization request of a queued transfer could fail, and this
needs to be reported to the transfer initiator. Otherwise it would
likely result in D-Bus timeouts.
Problem: Even if transfer is aborted file will be saved with partial
content.
Fix: In os_reset_session() os->cmd is checked for PUT before calling
driver->remove(), but os->cmd is never assigned.This fix sets os->cmd
wit respective opcode
There is no much of point to have a user_data if it is always the same
type, besides this code is very inefficient and cause a lookup in the list
of pending calls everytime a reply is received.
This fix solves the following issues in multiple file push.
Agent authorize is happening only for first file.
Transfer_completed signal is getting called only after last push
Incorrect value is sent for "total" and "transfered" signal from
second file onwards.
Issue: As reported by static code analyzer (Coverity), if buffer is NULL
and aparams is not NULL, then buffer gets dereferenced in string_read
function.
Fix: Application parameter is allocated only if maxlistcount is zero
during phonebooksize request and buffer is allocated in case of pull
phonebook request, hence modify logic to avoid application parameter
check, as it is tightly coupled with maxlistcount.
