path: root/kio/DESIGN

DESIGN:
=======

libkio uses kioslaves (separate processes) that handle a given protocol.
Launching those slaves is taken care of by the kdeinit/klauncher tandem,
which are notified by DCOP.

Connection is the most low-level class, the one that encapsulates the pipe.

SlaveInterface is the main class for transferring anything to the slave
and Slave, which inherits SlaveInterface, is the sub class that Job should handle.

A slave inherits SlaveBase, which is the other half of SlaveInterface.

The scheduling is supposed to be on a two level basis. One is in the daemon
and one is in the application. The daemon one (as opposite to the holy one? :)
will determine how many slaves are ok for this app to be opened and it will
also assign tasks to actually existing slaves.
The application will still have some kind of a scheduler, but it should be
a lot simpler as it doesn't have to decide anything besides which
task goes to which pool of slaves (related to the protocol/host/user/port)
and move tasks around.
Currently a design study to name it cool is in scheduler.cpp but in the
application side. This is just to test other things like recursive jobs
and signals/slots within SlaveInterface. If someone feels brave, the scheduler
is yours!
On a second thought: at the daemon side there is no real scheduler, but a
pool of slaves. So what we need is some kind of load calculation of the
scheduler in the application and load balancing in the daemon. 

A third thought: Maybe the daemon can just take care of a number of 'unused'
slaves. When an application needs a slave, it can request it from the daemon. 
The application will get one, either from the pool of unused slaves, 
or a new one will be created. This keeps things simple at the daemon level.
It is up to the application to give the slaves back to the daemon.
The scheduler in the application must take care not to request too many
slaves and could implement priorities.

Thought on usage:
* Typically a single slave-type is used exclusively in one application. E.g.
http slaves are used in a web-browser. POP3 slaves used in a mail program.

* Sometimes a single program can have multiple roles. E.g. konqueror is
both a web-browser and a file-manager. As a web-browser it primarily uses
http-slaves as a file-manager file-slaves.

* Selecting a link in konqueror: konqueror does a partial download of
the file to check the mimetype (right??) then the application is
started which downloads the complete file. In this case it should 
be able to pass the slave which does the partial download from konqueror
to the application where it can do the complete download.

Do we need to have a hard limit on the number of slaves/host?
It seems so, because some protocols are about to fail if you
have two slaves running in parralel (e.g. POP3) 
This has to be implemented in the daemon because only at daemon
level all the slaves are known. As a consequence slaves must 
be returned to the daemon before connecting to another host.
(Returning the slaves back to the daemon after every job is not
strictly needed and only causes extra overhead) 

Instead of actually returning the slave to the daemon, it could 
be enough to ask 'recycling permission' from the daemon: the 
application asks the daemon whether it is ok to use a slave for 
another host. The daemon can then update its administration of
which slave is connected to which host.

The above does of course not apply to hostless protocols (like file).
(They will never change host).

Apart from a 'hard limit' on the number of slaves/host we can have
a 'soft limit'. E.g. upon connection to a HTTP 1.1 server, the web-
server tells the slave the number of parallel connections allowed.
THe simplest solution seems to be to treat 'soft limits' the same
as 'hard limits'. This means that the slave has to communicate the
'soft limit' to the daemon.

Jobs using multiple slaves.

If a job needs multiple slaves in parallel (e.g. copying a file from 
a web-server to a ftp-server or browsing a tar-file on a ftp-site)
we must make sure to request the daemon for all slaves together since 
otherwise there is a risk of deadlock. 

(If two applications both need a 'pop3' and a 'ftp' slave for a single 
job and only a single slave/host is allowed for pop3 and ftp, we must 
prevent giving the single pop3 slave to application #1 and the single 
ftp slave to application #2. Both applications will then wait till the 
end of times till they get the other slave so that they can start the 
job. (This is a quite unlikely situation, but nevertheless possible))


File Operations:
listRecursive is implemented as listDir and tqfinding out if in the result
 is a directory. If there is, another listDir job is issued. As listDir
 is a readonly operation it fails when a directory isn't readable
  .. but the main job goes on and discards the error, because
bIgnoreSubJobsError is true, which is what we want (David)

del is implemented as listRecursive, removing all files and removing all
 empty directories. This basically means if one directory isn't readable
 we don't remove it as listRecursive didn't tqfind it. But the del will later
 on try to remove it's parent directory and fail. But there are cases when
 it would be possible to delete the dir in chmod the dir before. On the
 other hand del("/") shouldn't list the whole file system and remove all
 user owned files just to tqfind out it can't remove everything else (this
 basically means we have to take care of things we can remove before we try)

 ... Well, rm -rf / refuses to do anything, so we should just do the same:
 use a listRecursive with bIgnoreSubJobsError = false. If anything can't
 be removed, we just abort. (David)
 
 ... My concern was more that the fact we can list / doesn't mean we can
 remove it. So we shouldn't remove everything we could list without checking
 we can. But then the question arises how do we check whether we can remove it?
 (Stephan)

 ... I was wrong, rm -rf /, even as a user, lists everything and removes
 everything it can (don't try this at home!). I don't think we can do
 better, unless we add a protocol-dependent "canDelete(path)", which is
 _really_ not easy to implement, whatever protocol. (David)


Lib docu
========

mkdir: ...

rmdir: ...

chmod: ...

special: ...

stat: ...

get is implemented as TransferJob. Clients get 'data' signals with the data.
A data block of zero size indicates end of data (EOD)

put is implemented as TransferJob. Clients have to connect to the 
'dataReq' signal. The slave will call you when it needs your data.

mimetype: ...

file_copy: copies a single file, either using CMD_COPY if the slave 
           supports that or get & put otherwise.

file_move: moves a single file, either using CMD_RENAME if the slave
           supports that, CMD_COPY + del otherwise, or eventually
           get & put & del.

file_delete: delete a single file. 

copy: copies a file or directory, recursively if the latter

move: moves a file or directory, recursively if the latter

del: deletes a file or directory, recursively if the latter

PROGRESS DISPLAYING :
=====================
Taj brought up the idea of deligating all progress informations to an extern
GUI daemon which could be provided in several implementations - examples
are popup dialogs (most are annoyed by them, like me :) or a kicker applet
or something completely different. This would also remove the dependency on
libkdeui (I hope).
Conclusion: kio_uiserver is this single GUI daemon, but the dependency on
libkdeui couldn't be removed (for many reasons, including Job::showErrorDialog())

A. progress handling
---------------------
There will be two ways how the application can display progress :

1. regular apps will use NetAccess for all kio operations and will not  care
    about progress handling :
 - NetAccess creates Job
 - NetAccess creates JobObserver that will connect to the Job's signals and
    pass them via dcop to the running GUI Progress Server

2. apps that want to do some handling with progress dialogs like Caitoo or
    KMail  :
 - app creates Job
 - app creates a progress dialog : this should be a ProgressBase descendant
    e.g. tqStatusProgress or custom progress dialog
 - app calls progress->setJob( job ) in order to connect job's signals with
    progress dialog slots

B. customized progress dialogs
-------------------------------
 This will be similar to what we had before.

 - ProgressBase class that all other dialogs will inherit.
    will contain an initialization method setJob( KIO::Job*) for apps of the
    second class (see A.2 above), that will connect job's signals to dialog's
    slots

 - DefaultProgress ( former KIOSimpleProgressDialog ) that will be used for    
    regular progress dialogs created by GUI Progress Server

 - tqStatusProgress ( former KIOLittleProgressDialog ) that can be used for      
    embedding in status bar

C. GUI Progress Server
-----------------------
 This is a special progress server.
 - createProgress() will either create a DefaultProgress dialog or add new entry
    in a ListProgress ( an all-jobs-in-one progress dialog )
 - after receiving signals from the JobObserver via DCOP it will call
    appropriate method of progress dialog ( either in DefaultProgress or ListProgress )
 - ListProgres can be a Caitoo style dialog, kicker applet or both in one.

D. Some notes
--------------
 1. most of the apps will not care at all about the progress display
 2. user will be able to choose whether he wants to see separate progress
    dialogs or all-in-one ListProgress dialog
 3. developers can create their custom progress dialogs that inherit
    ProgressBase and do any manipulation with a dialog if they use a second
    approach ( see A.2 above )


Streaming 
---------

 1. We currently support a streaming "GET": e.g. file:/tmp/test.gz#gzip:/
    works. The following should also work: file:/tmp/test.gz.gz#gzip:/#gzip:/
    The current approach makes a TrasnferJob for gzip:/ and then adds a
    subjob for "file:/tmp/test.gz.gz#gzip:/" which itself adds a subjob
    for "file:/tmp/test.gz.gz".
 2. This doesn't extend very well to PUT, because there the order should
    basically be the other way around, but the "input" to the job as a whole
    should go to the "gzip:/" job, not to the "file:/tmp/test.gz.gz."
    It would probably be easier to implement such a job in the way the 
    current "CopyJob" is done. Have a Job and make all sub-urls sub-jobs of
    this Job.
 3. As a result of 1. COPY FROM an url like file:/tmp/test.gz#gzip:/ should
    work. COPY TO does not, because that would require PUT.
    

Resuming
--------

A rough note for now, just to have this somewhere :
(PJ=put-job, GJ=get-job)

PJ can't resume:
PJ-->app: canResume(0)  (emitted by dataReq)
GJ-->app: data()
PJ-->app: dataReq()
app->PJ: data()

PJ can resume but GJ can't resume:
PJ-->app: canResume(xx)
app->GJ: start job with "resume=xxx" metadata.
GJ-->app: data()
PJ-->app: dataReq()
app->PJ: data()

PJ can resume and GJ can resume:
PJ-->app: canResume(xx)
app->GJ: start job with "resume=xxx" metadata.
GJ-->app: canResume(xx)
GJ-->app: data()
PJ-->app: dataReq()
app->PJ: canResume(xx)
app->PJ: data()

So when the slave supports resume for "put" it has to check after the first 
dataRequest() whether it has got a canResume() back from the app. If it did 
it must resume. Otherwise it must start from 0.