...
I understand your concernings but I really don't think they are
relevant... This checks will be much faster then the procedure itself
and it would conform to normal munmap behaviour. From man page:
"The address start must be a multiple of the page size. All pages
containing a part of the indicated range are unmapped, and subsequent
references to these pages will generate SIGSEGV. It is not an error if
the indicated range does not contain any mapped pages."
XENO_ASSERT will not be a procedure, it will be a macro containing the
check for the assertion and the code to be executed on failure.
Yes, I understood, I just don't agree...
The point is just to control if this code will make it into the kernel or
not - via CONFIG_XENO_OPT_DEBUG. Kind of #ifdef CONFIG_XENO_OPT_DEBUG,
just more comfortable. A typical bug-free driver will not need such
checks
I'm not very sure about that. It can work for most situations but there
could be one situation where it would crash just because it was chosen
to have a little smaller code size and a little speed gain... I would
not like to think in the consequences of a crash in the driver while a
RT-system is working on real world...
as it will just pass those values already use for or returned by
rtdm_mmap_to_user. This is not just about speed, it's also about code size.
I think that if there was an extra parameter for user_info, it would
also verify for validity. BTW, I think there is missing some
documentation about the user_info parameter. I had to remember our
conversation and look at the code to understand that I should record
"current" on this parameter on the moment I called mmap and passing it
again on munmap. And it would be good to see the rtdm_user_info_t
defined as struct task_struct on the documentation.
This is intentionally opaque to the driver developer. As you receive a
rtdm_user pointer via the device handler invocation (as documented in
rtdm_mmap_to_user - feel free to improve my description!), you don't
need to (and you actually shouldn't) deal with task_struct or even mm
directly.
Sorry, I didn't understand the description in documentation and continue
not understanding it. In which device handler invocation I did receive
this pointer? I didn't find reference for it anywhere else...
Moreover, I have an experimental (and unreleased) Xenomai skin with RTDM
support here which maps rtdm_user_info_t to a different type.
I have not written the user-space program yet, so you'll have to wait
until monday, when I'll be able to test it, hopefully. But it seems
to be working... I changed my driver design. I do the mmap's on
driver initialization and just pass the returned addresses on the
IOCTL, so I can do them in a RT-context. The problem is that even if
the user call an IOCTL to
Hmm, I guess there is still some lacking documentation about what is
possible with RTDM. If you call an IOCTL from RT context, you end up in
the _rt-handler the driver of a device may have registered (if there is
no _rt-handler at all, the _nrt one is invoked, but this is likely not
relevant here).
I assume that you were wondering how to call rtdm_mmap_to_user from this
real-time handler, right?
No. I know it is not possible from the moment. I think I did not explain
myself very well. I was wondering how to define a RT mmap like ioctl. As
I know I could not use rtdm_mmap_to_user then, I thought in another way
of doing it. So I mmaped on driver initialization. On the IOCTL I just
passed the already known addresses to the user requesting it. I would
have to explain you how these buffers work on V4L2. It is a bit long
explanation but I can explain it on other message if you wish.
I had a look at a V4L2 tutorial, and I think I grabbed the idea. This
idea does *not* include any mmap during runtime, just once after device
opening and buffer setup. I think you shouldn't follow the syntactic
V4L2 interface, rather grab it's overall model.
Yes, but it would be simpler for new developers used to the V4L2 API to
migrate to a similar RT-Video interface. Anyway, as I've already said, I
don't think some users would do any mmap during runtime, but maybe for
some reconfiguration for some reason and they would like to make this in
a RT-context. I can't really imagine a practical situation from the
moment, but I think it is still possible to happen...
That's also why I cannot follow your desire for a transparent rt-mmap
implementation. I tried to define such a wrapper, but I didn't find it
useful, rather problematic as a lot of restrictions from the normal mmap
had to be defined.
Don't forget that normal mmap is a very generic interface, covering also
a lot of use cases (files e.g.) we will never see with RTDM. As it is
the standard interface for mapping, V4L2 likely does this split-up of
buffer allocation (via IOCTL) and mapping (via mmap). With
rtdm_mmap_to_user, this is not required! There are a few DRM Linux
driver unifying those steps as well, BTW.
That is something I could change in my driver... Making the mmap on the
VIDIOC_REQBUFS ioctl request. But, as a driver could return a different
number of buffers than the requested, the user could not be satisfied
with the returned buffers and the mmap would have been useless... I
still don't know how will the final design be like, but I'm fine with
your already provided rtdm_mmap_to_user solution. But I still want to
keep the API as closer as possible to V4L2 for facilitating the users to
migrate to real-time while taking advantage of the already available
V4L2 documentation and examples all over the web and just doing the
minimal needed modifications.
Well, the trick is to return -ENOSYS for those
IOCTL codes that can only be handled by the _nrt-handler. Xenomai will
then switch your RT task to secondary mode, restart the IOCTL, and the
mmap can safely be executed.
But as I've said, it is not the behaviour I want :)
Well, maybe you do not have any arguments for rtdm_mmap_to_user that
should be influenced by the user's IOCTL.
That is my case.
In this case your current driver design is ok as well. I just wanted
to underline that it is not necessarily the only way.
But I couldn't find other way of doing it in a RT-context.
As explained before, it doesn't make sense to mmap in time-constraint
contexts, even if we are only talking about standard Linux. No
high-speed capturing application will do this - especially under Linux.
I know that. I was just wondering if someone would need to reconfigure
the system for some reason in a rt-context...
munmap, it will still be possible to him to continue using the
provided address and this would result in a problem. But, as in all
situations, there
When rtdm_munmap is executed, the virtual address range becomes invalid
for the user. Thus any further access to it will raise a segfault.
That's the only problem, but it will not influence the driver integrity.
Yes, that is the problem. Since I only mark as unused on the munmap
IOCTL, it would be possible to the user to continue using that address
even after the munmap IOCTL call. It I was using a really rtdm_munmap,
it wouldn't be possible. It would be a better behaviour, but it would
not be run on RT-context. That is the trade-off.
If you avoid mmap in RT, there will also be no need for munmap in this
context. Just release it on closure.
Not that there is a need for munmaping, but it is useful. If the user
frees some memory when not using it anymore but is still using the file
descriptor, that memory will be then available for use by another process.
are trade-offs and I prefer to rely on the user, while providing a
RT-MMAP-IOCTL. Of course it isn't really a mmap, but if the user
don't mess with the pointers, it will work like if it was...
The user can only access the window you mapped in and only as long as it
is mapped.
In my case, it is always mapped to make possible the RT-IOCTLs.
And if you map it read-only, there is even no chance to
destroy potential management structures of the hardware or the driver
within this range.
I do not want to make it read-only because it will probably be very
usefull to the user to write on it. The user may want to capture a frame
and do some image processing routines on the same memory area when it is
possible, avoiding to copy that memory region.
I see, this makes sense. Then just prepare enough buffers for the
capturing, preprocessing, and potentially forwarding steps so that you
can continue to capture even if the user still hogs on a few previously
filled buffers. In the end it's just a cyclic process, and the only
question is how many spare buffers have to be created to keep it running.
This option is given to the user on V4L2 interface. That is not the
question. The problem arises on the follow situation (I'll try to be
more didactic this time):
1 - User A maps the buffer via ioctl.
2 - As the driver doesn't really do a mmap on ioctl, it just returns the
addresses to be used for the user.
3 - User A frees the buffer.
4 - The driver will mark that region as free but won't do any munmap for
keep doing it on RT-context.
5 - User B request buffer and mmap that same memory region via ioctl.
6 - Same goes on driver.
7 - A continue using the address he/she requested. If a munmap was
really called, this would result in SEGFAULT. But, in this case, he/she
will be able to continue messing with that area.
8 - User B will have unexpected behaviour.
That is the trade-off of using this approach for making it possible to
deal with buffers in a RT-context...
Best Regards,
Rodrigo.
_______________________________________________________
Yahoo! Acesso Grátis - Internet rápida e grátis. Instale o discador agora!
http://br.acesso.yahoo.com
