sis_malloc / sis_free
Quoted 1 line · Christoph Pittracher
Quoted 1 line
Hello!
Hi everyone,
Quoted 1 line
Quoted 1 line
Yes, the sisfb/drm code has some design lacks.
Well, well. I have by now come to the conclusion that this isn't really the case...
Quoted 8 lines
Quoted 3 lines
Another lack is the necessary memory offset between framebuffer/drm and the X driver (see http://www.webit.at/~twinny/linuxsis630.shtml for details).
Oh, okay. Should that memory offset be a source of processor time waste for allocating memory when switching through FB and X, for an example? If so, sharing the code (instead of duplicating it) is the ideal approach.
This has nothing whatsoever with processor time to do. It's to keep DRM/DRI from overwriting X's screen and off-screen buffers.
Quoted 6 lines
Quoted 3 lines
I don't think that it would be a problem to duplicate the code. the
sis_malloc/sis_freefunctions seems quite stable. I don't think that there will be big updates for this code.Hmm, but as you stated, I do not think code duplication should be the best approach.
No, it would not. You've obviously not understood the problems.
It's not done by simply duplicating the sis_malloc/sis_free functions as these also require
- detecting the type and size of the memory (of many different SiS chipsets),
- initializing a heap (based on which you later allocate/free the memory) taking into account a possibe TurboQueue, a possible AGP command queue and the HWCursor memory area.
Believe me: I know the code by heart right now, and moving the memory management out of the framebuffer driver will require HUGE parts of the code copied (not speaking about maintainance issues).
Finally: I intend to implement (real) 2D accelleration in the framebuffer driver. Thus, I will need a >common< memory management (ie. ONE SINGLE heap) for both the framebuffer driver as well as the X driver. Otherwise these two will again overwrite each others video memory, very possibly leading into problems with the accellerator engines (which use parts of the video RAM for buffering).
This can only be avoided by keeping the memory management inside the framebuffer driver. I don't think it's good to be forced to load the DRM module (if that one then contains memory management) for only being able to use the framebuffer driver...
Quoted 5 lines
Quoted 5 lines
Thomas Winischhofer <[email protected]> is working on that SiS stuff for about 2 months. I think it would be best if you contact him and ask what he thinks about that. I know that he said it would be a good idea to seperate the sisfb and
sis_drmcode but he doesn't have enough time to do it...
It's not a time issue. As said, the concept isn't that bad for future development. I consider it wise enough to keep the drivers separated, as this also allows separate usage (framebuffer only, X only) without memory clashes.
Quoted 3 lines
If I have such time, I'll contact him. But for the moment, if the code does not compile (still with 2.4.18-rc2-ac1) I'll duplicate the code to get it working until a better solution raises.
You don't have to do that. If you don't like a graphical console, simple start the framebuffer driver with "mode=none" (or as a kernel parameter "mode:none"). In this case, sisfb will only initialize the memory but leave the console alone.
I don't recommend changing anything that big in the framebuffer driver at the moment. I am currently in close contact with SiS who help me fixing the (still) remaining problems with it - based on the existing code. In the very only case you'd like to reinclude a number of changes for a number of times - well, go ahead :)
Thomas
Signature · 3 lines
Thomas Winischhofer Vienna/Austria mailto:[email protected] *** http://www.webit.com/tw
Quoted 4 lines · Thomas Winischhofer
Quoted 1 line
Quoted 1 line
Yes, the sisfb/drm code has some design lacks.
Well, well. I have by now come to the conclusion that this isn't really the case...
Hm, now it is clear, thanks for the explanation. :)
Quoted 11 lines
Quoted 8 lines
Quoted 3 lines
Another lack is the necessary memory offset between framebuffer/drm and the X driver (see http://www.webit.at/~twinny/linuxsis630.shtml for details).
Oh, okay. Should that memory offset be a source of processor time waste for allocating memory when switching through FB and X, for an example? If so, sharing the code (instead of duplicating it) is the ideal approach.
This has nothing whatsoever with processor time to do. It's to keep DRM/DRI from overwriting X's screen and off-screen buffers.
(hm, you still lose CPU time when deallocating memory when
switching back to console (fb) and switching to X, it they are not really shared, as with copy_from_user() and copy_to_user()...)
Quoted 7 lines
Quoted 6 lines
Quoted 3 lines
I don't think that it would be a problem to duplicate the code. the
sis_malloc/sis_freefunctions seems quite stable. I don't think that there will be big updates for this code.Hmm, but as you stated, I do not think code duplication should be the best approach.
Quoted 7 lines
No, it would not. You've obviously not understood the problems.
It's not done by simply duplicating the sis_malloc/sis_free functions as these also require
- detecting the type and size of the memory (of many different SiS chipsets),
Yes, with those DRAM detection routines.
Quoted 3 lines
- initializing a heap (based on which you later allocate/free the memory) taking into account a possibe TurboQueue, a possible AGP command queue and the HWCursor memory area.
With the sisfb_poh_allocate() (which I by found itself the
complicated part), the display struct, the *huge* heap initializer which depends on all the rest, eheh, the duplication would cost too much.
Quoted 10 lines
Believe me: I know the code by heart right now, and moving the memory management out of the framebuffer driver will require HUGE parts of the code copied (not speaking about maintainance issues).
Finally: I intend to implement (real) 2D accelleration in the framebuffer driver. Thus, I will need a >common< memory management (ie. ONE SINGLE heap) for both the framebuffer driver as well as the X driver. Otherwise these two will again overwrite each others video memory, very possibly leading into problems with the accellerator engines (which use parts of the video RAM for buffering).
Ahn, then you are speaking of moving SIS common heap functions,
command queue determination, queue area size and whatsoever common outside to a common area? This is really a good approach. :)
Quoted 4 lines
This can only be avoided by keeping the memory management inside the framebuffer driver. I don't think it's good to be forced to load the DRM module (if that one then contains memory management) for only being able to use the framebuffer driver...
In this current way, yes. But one can compile, let's say, the SIS
FB support as a module and the SIS DRM support as a builtin, which would break the driver. I threw a patch in the list with a brief description for Documentation/Configure.help (which is lacking) and the correct dependency checking for Config.in (that will make FB and DRM support be both modules or builtin).
Quoted 14 lines
It's not a time issue. As said, the concept isn't that bad for future development. I consider it wise enough to keep the drivers separated, as this also allows separate usage (framebuffer only, X only) without memory clashes.
Quoted 3 lines
If I have such time, I'll contact him. But for the moment, if the code does not compile (still with 2.4.18-rc2-ac1) I'll duplicate the code to get it working until a better solution raises.
You don't have to do that. If you don't like a graphical console, simple start the framebuffer driver with "mode=none" (or as a kernel parameter "mode:none"). In this case, sisfb will only initialize the memory but leave the console alone.
Quoted 5 lines
I don't recommend changing anything that big in the framebuffer driver at the moment. I am currently in close contact with SiS who help me fixing the (still) remaining problems with it - based on the existing code. In the very only case you'd like to reinclude a number of changes for a number of times - well, go ahead :)
Well, now that I've got into the code, neither I do. :)
Any idea on when you'll split allocation codes from the FB code?
(Or will you work on doing just one common module that'll give support for both FB and DRM?)
Regards,
Cesar Suga <[email protected]>
Quoted 6 lines · Jean Paul Sartre
Quoted 2 lines
This has nothing whatsoever with processor time to do. It's to keep DRM/DRI from overwriting X's screen and off-screen buffers.
(hm, you still lose CPU time when deallocating memory when
switching back to console (fb) and switching to X, it they are not really shared, as withcopy_from_user()andcopy_to_user()...)
AFAIK Alloc/Dealloc is not done when switching back and forth to/from VT. Why do you assume this? (Remember: We are talking about video RAM here)
Quoted 7 lines
Quoted 5 lines
It's not done by simply duplicating the sis_malloc/sis_free functions as these also require
- detecting the type and size of the memory (of many different SiS chipsets),
Yes, with those DRAM detection routines.
And these are HUGE (check init.c - half that file deals with that)
Quoted 3 lines
With the
sisfb_poh_allocate()(which I by found itself the
complicated part), the display struct, the *huge* heap initializer which depends on all the rest, eheh, the duplication would cost too much.
Exactly :)
Quoted 3 lines
Ahn, then you are speaking of moving SIS common heap functions,
command queue determination, queue area size and whatsoever common outside to a common area? This is really a good approach. :)
I am actually thinking of this - but
1) I haven't developed a new concept yet (which has to satisfy framebuffer, DRM and X), and
2) it's not on the top of my list. There are by far more important things to be done first (see below).
Quoted 3 lines
In this current way, yes. But one can compile, let's say, the SIS
FB support as a module and the SIS DRM support as a builtin, which would break the driver.
I'd never compile the DRM "module" into the kernel - what for...? (But you're right, of course. That combination will cause trouble.)
Quoted 1 line
Well, now that I've got into the code, neither I do. :)
:)))
Quoted 3 lines
Any idea on when you'll split allocation codes from the FB code?
(Or will you work on doing just one common module that'll give support for both FB and DRM?)
I figure leaving these routines inside the fb code is the best choice for now. I don't see any harm with that at the moment. (Apart from the compilation combination you mentioned, of course.)
Or does anybody want a third module for just doing video memory management...? Yes? You, there in the second row? Please stand up.... :)
I will, however, deal closer with that issue as soon as I got the X driver somewhat fixed - and this might take a few weeks (communication with SiS is kinda, let's say, time-taking and they aren't really generous with information on their devices). And aside from this - I got my attorney's exam on March 6 which will keep me from spending as much time as I'd like on SiS drivers until then... :(
What has the X driver to do with this, you ask? Well: The X driver now (after a hard struggle :) ) relates to sisfb by using the exact same code basis for the core functions. For assumingly understood maintainance reasons, I want to make the core functions re-usable for both of the drivers. This is partly finished but not complete yet. (If you want to test the latest drivers [both sisfb and X] and some more information on SiS VGA devices see
http://www.webit.com/tw/linuxsis630.shtml)
Another - rather important - issue is that SiS very soon releases a new chip (SiS 650) and a new video bridge (302B/301LV/302LV) and support for these should also be included in BOTH drivers. I am currently working on this.
For now, I believe patching the config dependencies is the best thing to do. I'd also patch the documentation so that folks don't compile DRM into the kernel (which is a waste of resources, IMHO)
Thomas
Signature · 3 lines
Thomas Winischhofer Vienna/Austria mailto:[email protected] *** http://www.webit.com/tw
Quoted 7 lines · Thomas Winischhofer
Quoted 3 lines
(hm, you still lose CPU time when deallocating memory when
switching back to console (fb) and switching to X, it they are not really shared, as withcopy_from_user()andcopy_to_user()...)AFAIK Alloc/Dealloc is not done when switching back and forth to/from VT. Why do you assume this? (Remember: We are talking about video RAM here)
# From sis_drm.h:
typedef struct {
int context;
unsigned int offset;
unsigned int size;
unsigned int free;
} drm_sis_mem_t;
Ahm, sorry. eh. Well, as I said. I'm not too familiarized and I
thought the source disposition is not too obvious for me to understand. (no, I am no C expert, so..)
Quoted 2 lines
Quoted 1 line
Yes, with those DRAM detection routines.
And these are HUGE (check
init.c- half that file deals with that)
Hm, they actually convinced me that, for me, the better way to get
those things working with ease of expansion is how you stated (and I asked, for the 'code split'), the module for video memory management.
Quoted 10 lines
Quoted 3 lines
With the
sisfb_poh_allocate()(which I by found itself the
complicated part), the display struct, the *huge* heap initializer which depends on all the rest, eheh, the duplication would cost too much.Exactly :)
Quoted 3 lines
Ahn, then you are speaking of moving SIS common heap functions,
command queue determination, queue area size and whatsoever common outside to a common area? This is really a good approach. :)
Quoted 7 lines
I am actually thinking of this - but
1) I haven't developed a new concept yet (which has to satisfy framebuffer, DRM and X), and
2) it's not on the top of my list. There are by far more important things to be done first (see below).
I think a video memory management module would be *very*
important for new modules both for FB and DRM. So the modules would care about the command queues, setting modes, etc. - nothing related to video memory.
Quoted 6 lines
Quoted 3 lines
In this current way, yes. But one can compile, let's say, the SIS
FB support as a module and the SIS DRM support as a builtin, which would break the driver.I'd never compile the DRM "module" into the kernel - what for...? (But you're right, of course. That combination will cause trouble.)
One should be, at least, adviced in the Configure.help whilst
configuring his/her kernel.
Quoted 8 lines
Quoted 1 line
Well, now that I've got into the code, neither I do. :)
:)))
Quoted 3 lines
Any idea on when you'll split allocation codes from the FB code?
(Or will you work on doing just one common module that'll give support for both FB and DRM?)
Quoted 6 lines
I figure leaving these routines inside the fb code is the best choice for now. I don't see any harm with that at the moment. (Apart from the compilation combination you mentioned, of course.)
Or does anybody want a third module for just doing video memory management...? Yes? You, there in the second row? Please stand up.... :)
Yes, eh. Hope someone can help (I do not know yet how to manage
that).
Quoted 6 lines
I will, however, deal closer with that issue as soon as I got the X driver somewhat fixed - and this might take a few weeks (communication with SiS is kinda, let's say, time-taking and they aren't really generous with information on their devices). And aside from this - I got my attorney's exam on March 6 which will keep me from spending as much time as I'd like on SiS drivers until then... :(
Hm. And I'll only be able to contribute in, say, documentation.
Quoted 9 lines
What has the X driver to do with this, you ask? Well: The X driver now (after a hard struggle :) ) relates to sisfb by using the exact same code basis for the core functions. For assumingly understood maintainance reasons, I want to make the core functions re-usable for both of the drivers. This is partly finished but not complete yet. (If you want to test the latest drivers [both sisfb and X] and some more information on SiS VGA devices see
http://www.webit.com/tw/linuxsis630.shtml)
Quoted 4 lines
Another - rather important - issue is that SiS very soon releases a new chip (SiS 650) and a new video bridge (302B/301LV/302LV) and support for these should also be included in BOTH drivers. I am currently working on this.
Hope you code that support with the possibility of splitting
functions between modules ;)
Quoted 3 lines
For now, I believe patching the config dependencies is the best thing to do. I'd also patch the documentation so that folks don't compile DRM into the kernel (which is a waste of resources, IMHO)
Yes, at least for now.
Regards,
Cesar Suga <[email protected]>