backports: backport drvdata = NULL core driver fixes
PATCH 1/1 backports: backport drvdata = NULL core driver fixes
From: "Luis R. Rodriguez" <[email protected]>
The Linux kernel had tons of code which at times cleared the drvdata upon probe failure or release. There are however a bunch of drivers that didn't clear this.
Commit 0998d063 implmented clearing this upon device_release_driver() and dealt with probe failure on driver_probe_device(). After this the kernel was cleaned up separately with *tons* of patches to remove all these driver specific settings given that the clearing is now done internally by the device core.
Instead of ifdef'ing code back in for older code where it was properly in place backport this by piggy backing the new required code upon the calls used in place. There is a small race here upon device_release_driver() but we can live with that theoretical race.
Due to the way we hack this backport we can't use a separate namespace as we have with other symbols.
mcgrof@frijol ~/linux-stable (git::master)$ git describe --contains \ 0998d0631001288a5974afc0b2a5f568bcdecb4d
v3.6-rc1~99^2~14^2~17
commit 0998d0631001288a5974afc0b2a5f568bcdecb4d
Author: Hans de Goede <[email protected]> Date: Wed May 23 00:09:34 2012 +0200
device-core: Ensure drvdata = NULL when no driver is bound
1) drvdata is for a driver to store a pointer to driver specific data
2) If no driver is bound, there is no driver specific data associated with the device
3) Thus logically drvdata should be NULL if no driver is bound.
But many drivers don't clear drvdata on device_release, or set drvdata early on in probe and leave it set on probe error. Both of which results in a dangling pointer in drvdata.
This patch enforce for drvdata to be NULL after device_release or on probe failure.
Tested with ckmake against next-20130618:
1 2.6.24 [ OK ] 2 2.6.25 [ OK ] 3 2.6.26 [ OK ] 4 2.6.27 [ OK ] 5 2.6.28 [ OK ] 6 2.6.29 [ OK ] 7 2.6.30 [ OK ] 8 2.6.31 [ OK ] 9 2.6.32 [ OK ] 10 2.6.33 [ OK ] 11 2.6.34 [ OK ] 12 2.6.35 [ OK ] 13 2.6.36 [ OK ] 14 2.6.37 [ OK ] 15 2.6.38 [ OK ] 16 2.6.39 [ OK ] 17 3.0.79 [ OK ] 18 3.1.10 [ OK ] 19 3.10-rc1 [ OK ] 20 3.2.45 [ OK ] 21 3.3.8 [ OK ] 22 3.4.46 [ OK ] 23 3.5.7 [ OK ] 24 3.6.11 [ OK ] 25 3.7.10 [ OK ] 26 3.8.13 [ OK ] 27 3.9.3 [ OK ]
real 32m2.332s user 860m23.688s sys 121m20.840s
--- backport/backport-include/linux/device.h | 19 +++++++++++++++++++ 1 file changed, 19 insertions(+)
backport/backport-include/linux/device.h +19 -0
@@ -176,4 +176,23 @@ extern int dev_set_name(struct device *dev, const char *name, ...) __attribute__((format(printf, 2, 3)));#endif +#if LINUX_VERSION_CODE <=KERNEL_VERSION (3 ,6 ,0 ) +#define driver_probe_device (__drv, __dev) \ +({ \ + int ret; \ + ret = (driver_probe_device)(__drv, __dev); \ + if (ret) \ + dev_set_drvdata(__dev, NULL); \ + return ret; \ +}) + +#define device_release_driver (__dev) \ +({ \ + (device_release_driver)(__dev); \ + device_lock(__dev); \ + dev_set_drvdata(__dev, NULL); \ + device_unlock(__dev); \ +}) +#endif /* LINUX_VERSION_CODE <= KERNEL_VERSION(3,6,0) */ +#endif /* __BACKPORT_DEVICE_H */
Signature · 1 line
1.7.10.4
Quoted 39 lines · Luis R. Rodriguez
From: "Luis R. Rodriguez" <[email protected]>
The Linux kernel had tons of code which at times cleared the drvdata upon probe failure or release. There are however a bunch of drivers that didn't clear this.
Commit
0998d063implmented clearing this upondevice_release_driver()and dealt with probe failure ondriver_probe_device(). After this the kernel was cleaned up separately with *tons* of patches to remove all these driver specific settings given that the clearing is now done internally by the device core.Instead of ifdef'ing code back in for older code where it was properly in place backport this by piggy backing the new required code upon the calls used in place. There is a small race here upon
device_release_driver()but we can live with that theoretical race.Due to the way we hack this backport we can't use a separate namespace as we have with other symbols.
mcgrof@frijol ~/linux-stable (git::master)$ git describe --contains \ 0998d0631001288a5974afc0b2a5f568bcdecb4dv3.6-rc1~99^2~14^2~17
commit
0998d0631001288a5974afc0b2a5f568bcdecb4dAuthor: Hans de Goede <[email protected]> Date: Wed May 23 00:09:34 2012 +0200device-core: Ensure drvdata = NULL when no driver is bound
1) drvdata is for a driver to store a pointer to driver specific data
2) If no driver is bound, there is no driver specific data associated with the device
3) Thus logically drvdata should be NULL if no driver is bound.But many drivers don't clear drvdata on
device_release, or set drvdata early on in probe and leave it set on probe error. Both of which results in a dangling pointer in drvdata.
Quoted 5 lines
This patch enforce for drvdata to be NULL after
device_releaseor on probe failure.Signed-off-by Hans de Goede <[email protected]>Signed-off-by Greg Kroah-Hartman <[email protected]>
This is not a very good idea. Although setting drvdata to NULL allowed a lot of code to be removed, it also exposed a bunch of hidden bugs -- drivers were using the drvdata value even after their remove function returned.
Several commits have been applied to fix those bugs. Finding and backporting them and their dependencies will not be easy.
I suggest this patch not be applied.
Alan Stern
Quoted 44 lines · Alan Stern
Quoted 42 lines · Luis R. Rodriguez
From: "Luis R. Rodriguez" <[email protected]>
The Linux kernel had tons of code which at times cleared the drvdata upon probe failure or release. There are however a bunch of drivers that didn't clear this.
Commit
0998d063implmented clearing this upondevice_release_driver()and dealt with probe failure ondriver_probe_device(). After this the kernel was cleaned up separately with *tons* of patches to remove all these driver specific settings given that the clearing is now done internally by the device core.Instead of ifdef'ing code back in for older code where it was properly in place backport this by piggy backing the new required code upon the calls used in place. There is a small race here upon
device_release_driver()but we can live with that theoretical race.Due to the way we hack this backport we can't use a separate namespace as we have with other symbols.
mcgrof@frijol ~/linux-stable (git::master)$ git describe --contains \ 0998d0631001288a5974afc0b2a5f568bcdecb4dv3.6-rc1~99^2~14^2~17
commit
0998d0631001288a5974afc0b2a5f568bcdecb4dAuthor: Hans de Goede <[email protected]> Date: Wed May 23 00:09:34 2012 +0200device-core: Ensure drvdata = NULL when no driver is bound
1) drvdata is for a driver to store a pointer to driver specific data
2) If no driver is bound, there is no driver specific data associated with the device
3) Thus logically drvdata should be NULL if no driver is bound.But many drivers don't clear drvdata on
device_release, or set drvdata early on in probe and leave it set on probe error. Both of which results in a dangling pointer in drvdata.This patch enforce for drvdata to be NULL after
device_releaseor on probe failure.
Quoted 7 lines
Quoted 2 lines
Signed-off-by Hans de Goede <[email protected]>Signed-off-by Greg Kroah-Hartman <[email protected]>This is not a very good idea. Although setting drvdata to NULL allowed a lot of code to be removed, it also exposed a bunch of hidden bugs -- drivers were using the drvdata value even after their remove function returned.
Eek, have we not SmPL'ify'd a proof yet to ensure code like this no longer exists? Julia? :)
Quoted 1 line
Several commits have been applied to fix those bugs.
I actualy found *tons* !
Quoted 1 line
Finding and backporting them and their dependencies will not be easy.
May the SmPL proof be with us.
Quoted 1 line
I suggest this patch not be applied.
Thanks for the review Alan!
Luis
Quoted 7 lines · Luis R. Rodriguez
Quoted 4 lines
This is not a very good idea. Although setting drvdata to NULL allowed a lot of code to be removed, it also exposed a bunch of hidden bugs -- drivers were using the drvdata value even after their remove function returned.
Eek, have we not SmPL'ify'd a proof yet to ensure code like this no longer exists? Julia? :)
Come to think of it, perhaps we should require *proof* with SmPL like this in future to avoid regressions ?
Luis
Quoted 7 lines · Luis R. Rodriguez
Quoted 5 lines · Luis R. Rodriguez
Quoted 4 lines
This is not a very good idea. Although setting drvdata to NULL allowed a lot of code to be removed, it also exposed a bunch of hidden bugs -- drivers were using the drvdata value even after their remove function returned.
Quoted 5 lines
Quoted 2 lines
Eek, have we not SmPL'ify'd a proof yet to ensure code like this no longer exists? Julia? :)
Come to think of it, perhaps we should require *proof* with SmPL like this in future to avoid regressions ?
Is it a concurrency problem? SmPL is not so good at that in the general case. One would have to know a specific case where other functions of the driver can be invoked after remove.
julia
Quoted 12 lines · Julia Lawall
Quoted 10 lines · Luis R. Rodriguez
Quoted 7 lines · Luis R. Rodriguez
Quoted 4 lines
This is not a very good idea. Although setting drvdata to NULL allowed a lot of code to be removed, it also exposed a bunch of hidden bugs -- drivers were using the drvdata value even after their remove function returned.
Eek, have we not SmPL'ify'd a proof yet to ensure code like this no longer exists? Julia? :)
Quoted 6 lines
Quoted 2 lines
Come to think of it, perhaps we should require *proof* with SmPL like this in future to avoid regressions ?
Is it a concurrency problem? SmPL is not so good at that in the general case. One would have to know a specific case where other functions of the driver can be invoked after remove.
Thanks Julia. In that case I'm going to just leave this in place given that if there's a bug upstream we'll get it fixed as soon as a respective patch gets upstream as well. That is, we are not using old drivers, we use the same upstream drivers so if a regression was found in backports the fix must go upstream s well. This is one of the benefits of backporting -- the range of users and testers increases and we still benefit from the upstream bandwagon.
Luis
Quoted 7 lines · Luis R. Rodriguez
Thanks Julia. In that case I'm going to just leave this in place given that if there's a bug upstream we'll get it fixed as soon as a respective patch gets upstream as well. That is, we are not using old drivers, we use the same upstream drivers so if a regression was found in backports the fix must go upstream s well. This is one of the benefits of backporting -- the range of users and testers increases and we still benefit from the upstream bandwagon.
I don't understand. If you're not using old drivers, and you incorporate all the upstream patches, then what's the difference between a backport and the current kernel? In fact, if you're incorporating all the upstream driver patches, then why haven't you already got the drvdata change?
As one example of the sort of subtle problem exposed by the drvdata change, take a look at commit b2ca69907657.
For more examples, see commits bf90ff5f3b8f, 638b9e15233c, 51ef847df746, 289b076f89c2, 53636555b919, 99a6f73c495c, 003615302a16, 94ab71ce2889, 3124d1d71d3d, c27f3efc5608, 95940a04bfe8, 5c1a0f418d8d, db5c8b524444, 8bf769eb5f6e, 4295fe7791a1, fa919751a2d2, a9556040119a, 7bdce71822f4, and a1028f0abfb3. Admittedly, these are all related problems in a single subsystem, but it gives you a little idea of how far this goes.
Alan Stern
Quoted 10 lines · Alan Stern
Quoted 7 lines · Luis R. Rodriguez
Thanks Julia. In that case I'm going to just leave this in place given that if there's a bug upstream we'll get it fixed as soon as a respective patch gets upstream as well. That is, we are not using old drivers, we use the same upstream drivers so if a regression was found in backports the fix must go upstream s well. This is one of the benefits of backporting -- the range of users and testers increases and we still benefit from the upstream bandwagon.
Quoted 3 lines
I don't understand. If you're not using old drivers, and you incorporate all the upstream patches, then what's the difference between a backport and the current kernel?
The difference is that in older kernels driver_probe_device() and device_release_driver() wouldn't do the setting to NULL in case of probe fail or release, that's what the patch does. Given that the new drivers have the superfluous setting removed we'd need a way to get older kernels to clear it as well and there were two ways to do it -- revert the changes for all drivers this was cleared from or hack up the same callback used driver_probe_device() and device_release_driver() to do the appropriate clearing prior to calling the old kernel's routines.
Quoted 3 lines
In fact, if you're
incorporating all the upstream driver patches, then why haven't you already got the drvdata change?
Because older kernel's driver_probe_device() anddevice_release_driver() would be used.
Quoted 2 lines
As one example of the sort of subtle problem exposed by the drvdata change, take a look at commit
b2ca69907657.
Understood.
Quoted 7 lines
For more examples, see commits
bf90ff5f3b8f,638b9e15233c,51ef847df746,289b076f89c2,53636555b919,99a6f73c495c,003615302a16,94ab71ce2889,3124d1d71d3d,c27f3efc5608,95940a04bfe8,5c1a0f418d8d,db5c8b524444,8bf769eb5f6e,4295fe7791a1,fa919751a2d2,a9556040119a,7bdce71822f4, anda1028f0abfb3. Admittedly, these are all related problems in a single subsystem, but it gives you a little idea of how far this goes.
Sure, I understand. By pushing the newer drivers as-is *and* by providing the driver_probe_device() and device_release_driver() respective changes for older kernels it should get us similar behavior on older kernels. Mind you, there is a small race in the way I implemented this on device_release_driver() but that seems like a reasonable tradeoff against all other alternatives I could come up with.
As it is now, unless my port is incorrect, the newer drivers used on older systems should trigger a similar bug as if using the upstream kernel with the same drivers.
Luis