Skip to contentSkip to conversation

EXPORT_SYMBOL_GPL for __symbol_get replacing EXPORT_SYMBOL for deprecated inter_module_get

DISCUSSION Yuri Vilmanis 6 participants 6 messages Apr 13, 2005

lore
Share & tools
lore
#0 lore raw

I believe that, in general, new functions which replace deprecated functions which were exported as EXPORT_SYMBOL, should also be exported as EXPORT_SYMBOL, not as EXPORT_SYMBOL_GPL. The reason I say this is because deprecation of old functions breaks old modules and drivers that use them, and changing the level of GPL strictness of the function prevents these old modules being updated and used with newer kernels.

The case in point for me is ATI's binary openGL accelerated drivers (fglrx) - these used inter_module_get() to communicate with the agp gart module, for obvious reasons - this AGP communication is essential to the functionality of the driver. No, I don't like ATI only having closed-source drivers any more than you, but given the extremely competetive nature of high end video card sales, I can see why they want to do it this way. The point is that now, as of 2.6.11-? or 2.6.12-? (not sure of the exact revision), the inter_module_get() functions has been removed, and the functionality can only be got (as far as I can tell) through use of the symbol_get() function.

Am I take it to mean that no closed-source / binary-only driver may use AGP acceleration in the future, including ones that have in the past? Or if an alternate method of AGP accelerated access exists (I dont know the kernel well enough to know for sure), does this mean companies which have at least been respectful enough to have provided linux drivers must now rewrite them to use an entirely different methodology in their drivers, write entirely new agp drivers which will then probably also be closed-source, and are likely to cause further problems down the line, or GPL drivers which contain code important to the competetiveness of the company? I think this is not a good situation for linux users or hardware manufacturers, and indeed shows Linux itself in a bad light.

I respect and agree with the use of EXPORT_SYMBOL_GPL to maintain new *software* functionality added by kernel developers, but when this prevents propriatary drivers from using advanced *hardware* features of common hardware connectivity devices like an AGP port, its silliness, plain and simple (I think this applies equally to other devices which are required for the connection of the device for which the propriatary driver was created, such as USB, PCI, serial, AGP, PCIExpress, firewire, IDA/SATA etcetc where DMA, ATA, fast serial (as provided by some Maxim chips with accelarated slew-rate) or other advanced advanced transfer modes are provided and enabled by the *harware*).

Also, deprecating non-GPL-strict functionality and replacing it with GPL-strict funtionality is just as bad, in my view, as going through the kernel and randomly changing instance of EXPORT_SYMBOL to EXPORT_SYMBOL_GPL - it has the same effect, of breaking established propriatary drivers, and forcing companies which do not feel that they can GPL the source and maintain their competetive advantage, to rewrite their drivers to use alternate methods (which they therefore have no certainty will be available in the future either), or dropping support altogether (which I'm sure nobody really wants). If Linux had a near-monopoly market share, it might be possible to get away with such strong-arm tactics in the interest of spreading the FSF distribution philosophy - but then again, look what similar tactics did to public opinion of the distribution philosophy of a certain Redmond company which did have near-monopoly market share.... I shudder at the thought of the point being applicable to Linux, but I think it must be made in order to prevent it becoming creepingly (and creepily) more valid in the future.

The case of inter_module_get deprecation and replacement with symbol_get embodies both sides: it makes a non-GPL export into a GPL export (yes, its now called by a different name, and the backend has changed, but the first change is cosmetic and the second change is internal), and it also prevents closed-source drivers from using functionality of a device that they have no choice but to connect through, and require for competetive performance.

To sum up:
1) Please make symbol_get a non-GPL export, so that those of us with hardware that needs it can continue to use it (eg anyone with an ATI video card that wants to use the shaders and extensions that are the point in favour over NVidia offerings)
2) I think the use of GPL symbol exports should be more carefully monitored where it affects existing closed-source drivers and designed-in hardware functionality like AGP, to avoid alienating companies which have been polite enough to come half way and supported their hardware under this great OS.

After all, I think we all want Linux to win by being the best - not by going around shouting "Thou shalt use the GPL, it ownz your propriatary distribution model and if you dont use it youre a moron and we wont let you use Linux" at passing strangers (not that this is likely to work anyway).

-Yuri

PS - This was not meant to be a rant, start a flamewar or even tread on anyones toes, really.... I'm just a bit annoyed that I can't do my shader development while running a nice bleeding-edge kernel, and apprehensive of not being able to do my shader development in a few months/years when this kernel is officially OldHat (TM, Inc, or whatever)

#1 lore raw
Quoted 9 lines · Yuri Vilmanis

I believe that, in general, new functions which replace deprecated functions which were exported as EXPORT_SYMBOL, should also be exported as EXPORT_SYMBOL, not as EXPORT_SYMBOL_GPL. The reason I say this is because deprecation of old functions breaks old modules and drivers that use them, and changing the level of GPL strictness of the function prevents these old modules being updated and used with newer kernels.

The case in point for me is ATI's binary openGL accelerated drivers (fglrx) - these used inter_module_get() to communicate with the agp gart module,

except that the AGP gart module no longer uses inter_module_*

Quoted 3 lines

for
obvious reasons - this AGP communication is essential to the functionality of the driver.

this is thus non-obvious..

Quoted 1 line

The case of inter_module_get deprecation and replacement with symbol_get

inter_module_get is absolutely not replaced by symbol_get. They are entirely different usages and only have a few letters in common in the name.

#2 lore raw
In reply to Yuri Vilmanis
Quoted 10 lines

The case in point for me is ATI's binary openGL accelerated drivers (fglrx) - these used inter_module_get() to communicate with the agp gart module, for obvious reasons - this AGP communication is essential to the functionality of the driver. No, I don't like ATI only having closed-source drivers any more than you, but given the extremely competetive nature of high end video card sales, I can see why they want to do it this way. The point is that now, as of 2.6.11-? or 2.6.12-? (not sure of the exact revision), the inter_module_get() functions has been removed, and the functionality can only be got (as far as I can tell) through use of the symbol_get() function.

you might have done some research on what happened before ranting away like a crazy.....

AGP dropped the drm_agp interface in favour of the DRM modules calling the AGP backend directly, a patch for the ATI drivers has been sailing around for ages I even wrote one myself for someone I think but I've lost it now ... all the AGP symbols needed are exported with EXPORT_SYMBOL.

you might have noticed no-one else complaining around here which is usually a good sign that there is nothing worth complaining about :-)

Dave.

#3 lore raw
In reply to Yuri Vilmanis
Quoted 9 lines · Yuri Vilmanis

The case in point for me is ATI's binary openGL accelerated drivers (fglrx) - these used inter_module_get() to communicate with the agp gart module, for obvious reasons - this AGP communication is essential to the functionality of the driver. No, I don't like ATI only having closed-source drivers any more than you, but given the extremely competetive nature of high end video card sales, I can see why they want to do it this way.
....
Am I take it to mean that no closed-source / binary-only driver may use AGP acceleration in the future, including ones that have in the past?

They can use the in-kernel GART driver just fine. Of course, they choose to take a bastardised version from some ancient tree, mangle it to all hell, strip off the GPL MODULE_VERSION, and weld it to their own driver, but that's their decision. Which is btw, whats partly causing your problem. (They still would've needed to change some code, but the AGP side of the fence would be taken care of).

Dave

#4 lore raw
Quoted 12 lines · Dave Jones
Quoted 9 lines · Yuri Vilmanis

The case in point for me is ATI's binary openGL accelerated drivers (fglrx) - these used inter_module_get() to communicate with the agp gart module, for obvious reasons - this AGP communication is essential to the functionality of the driver. No, I don't like ATI only having closed-source drivers any more than you, but given the extremely competetive nature of high end video card sales, I can see why they want to do it this way.
....
Am I take it to mean that no closed-source / binary-only driver may use AGP acceleration in the future, including ones that have in the past?

Quoted 8 lines

They can use the in-kernel GART driver just fine. Of course, they choose to take a bastardised version from some ancient tree, mangle it to all hell, strip off the GPL MODULE_VERSION, and weld it to their own driver, but that's their decision. Which is btw, whats partly causing your problem. (They still would've needed to change some code, but the AGP side of the fence would be taken care of).

Dave

As a practical matter, one can make or modify the source-code of a driver to use any symbols available in System.map. One can even make a "preloader" program that gets the right stuff for the correct kernel, puts it into the module, then has the standard module loader load it.

There is way too much effort being applied to hiding kernel symbols. As long as you have the tools available to build a kernel, you have the tools available to use any symbol.

Here is the kernel offset of 'sys_call_table' and 'sys_ni_syscall'

TABLE := $(shell grep sys_call_table /boot/System.map | cut -f1 -d' ') NISYS := $(shell grep sys_ni_syscall /boot/System.map | cut -f1 -d' ')
DEFS = -D__KERNEL__ -DMODULE -DMAJOR_NR=$(MAJR) -DCONFIG_SMP
DEFS += -DMODNAME=$(NAME) -DTABLE=0x$(TABLE) -DNISYS=0x$(NISYS)

You just initialize your module pointers to these values and you have access to these objects. Simple.

Although I haven't tried it, I think one can even borrow a __mod_licensexxx by using /proc/kallsyms. The point being that trying to prevent one from using existing kernel code in kernel modules will, eventually, fail completely unless we only get binaries with no source-code. Even in that case, many symbols within /proc/kallsyms are useful.

Cheers,
Dick Johnson
Penguin : Linux version 2.6.11 on an i686 machine (5537.79 BogoMips).
Notice : All mail here is now cached for review by Dictator Bush.
98.36% of all statistics are fiction.

#5 lore raw
Quoted 49 lines · Richard B. Johnson
Quoted 18 lines · Dave Jones
Quoted 1 line · Yuri Vilmanis

The case in point for me is ATI's binary openGL accelerated drivers

(fglrx) -

Quoted 1 line

these used inter_module_get() to communicate with the agp gart module,

for

Quoted 1 line

obvious reasons - this AGP communication is essential to the

functionality of

Quoted 1 line

the driver. No, I don't like ATI only having closed-source drivers any

more

Quoted 1 line

than you, but given the extremely competetive nature of high end video

card

Quoted 3 lines

sales, I can see why they want to do it this way.
....
Am I take it to mean that no closed-source / binary-only driver may

use AGP

Quoted 1 line

acceleration in the future, including ones that have in the past?

Quoted 8 lines

They can use the in-kernel GART driver just fine. Of course, they choose to take a bastardised version from some ancient tree, mangle it to all hell, strip off the GPL MODULE_VERSION, and weld it to their own driver, but that's their decision. Which is btw, whats partly causing your problem. (They still would've needed to change some code, but the AGP side of the fence would be taken care of).

Dave

As a practical matter, one can make or modify the source-code of a driver to use any symbols available in System.map. One can even make a "preloader" program that gets the right stuff for the correct kernel, puts it into the module, then has the standard module loader load it.

There is way too much effort being applied to hiding kernel symbols. As long as you have the tools available to build a kernel, you have the tools available to use any symbol.

Here is the kernel offset of 'sys_call_table' and 'sys_ni_syscall'

TABLE := $(shell grep sys_call_table /boot/System.map | cut -f1 -d' ') NISYS := $(shell grep sys_ni_syscall /boot/System.map | cut -f1 -d' ')
DEFS = -D__KERNEL__ -DMODULE -DMAJOR_NR=$(MAJR) -DCONFIG_SMP
DEFS += -DMODNAME=$(NAME) -DTABLE=0x$(TABLE) -DNISYS=0x$(NISYS)

You just initialize your module pointers to these values and you have access to these objects. Simple.

Quoted 7 lines

Although I haven't tried it, I think one can even borrow a __mod_licensexxx by using /proc/kallsyms. The point being that trying to prevent one from using existing kernel code in kernel modules will, eventually, fail completely unless we only get binaries with no source-code. Even in that case, many symbols within /proc/kallsyms are useful.

Yeah yeah.. yet another brilliant idea from the peanut gallery. GPL_
symbols aren't meant to be some amazing protection from criminals that
don't care about licenses. Hell, people pirate none GPL software too.
The symbols are meant to help HONEST people comply with the license.

Please for gods sake, shut up.
Sean