Re: [PATCH] Potential reference count races
On 10/28/2012 17:44, Max TenEyck Woodbury wrote:
On 10/28/2012 02:40 AM, Nikolay Sivov wrote:
On 10/28/2012 04:59, max(a)mtew.isa-geek.net wrote:
From: Max TenEyck Woodbury <max(a)mtew.isa-geek.net>
I have been looking at the Microsoft COM and related documentations and noticed that they emphatically recommend using the Interlocked... functions when manipulating reference counts. I managed to set up a search that showed where many of the reference count updates occur and was somewhat surprised at how often this advice was not followed. It doesn't mean it always has to be followed. True in a limited sense, but there is a good reason behind their recommendation. Unless there is a good reason not to do so, this particular piece of advice should be followed. COM objects in wine follow this recommendation in general, even object itself is not thread safe. This doesn't mean however that you need this every time you have some kind of refcount of any sort.
While I have not converted every reference count update to use the Interlocked... functions, this set of patches fixes a fair number of them.
These are not associated with any particular bug report; they are simply a general precaution against operations that are known to be associated with race conditions. This precaution doesn't work in general. It's not enough to atomically update refcount to make an implementation thread safe. Also not everything is supposed to be thread safe in a first place.
First, explain what does not have to be thread safe. Anything really, COM objects in particular if you were talking about them. I believe that application may try to use multiple threads anywhere, so everything that can be made thread safe, should be. No. Using interlocks on the reference count updates is a necessary step for thread safety. You are correct that it is not sufficient, but it is necessary. Again, it depends.
Changes like this:
- for (i=0;i<howmuch;i++) + for (i=0;i<howmuch;++i) TRACE("notify at %d to %p\n", notify[i].dwOffset,notify[i].hEventNotify); are not helpful at all. The post increment and decrement operation are specified as saving the original value for use in the evaluation of the expression they are part of and modifying the underlying stored value. In expressions like this, that saved value is then discarded. The optimization phase of the compilation usually removes both the save and discard operations. Sure, but I don't think it's enough to justify such changes all over the place, in existing code.
Murphy's law suggests that this might cause problems at some point when least desired.
Please don't do this.
Specifying the unnecessary use of a temporary store is a bad habit to have. You should tell the compiler exactly what needs to be done, not throw in extraneous operations.
So, the use of a post operator where a prefix operator is sufficient, while almost certainly harmless, is still technically a mistake.
At the minimum, these changes provide examples of the correct way these statements should be coded. So, I have to disagree, although very mildly, that changes like that are not helpful. This kind of change does not deserve a separate patch, but should be perfectly acceptable as an adjunct to other patches.
Nikolay Sivov <bunglehead(a)gmail.com> writes:
On 10/28/2012 17:44, Max TenEyck Woodbury wrote:
Specifying the unnecessary use of a temporary store is a bad habit to have. You should tell the compiler exactly what needs to be done, not throw in extraneous operations.
So, the use of a post operator where a prefix operator is sufficient, while almost certainly harmless, is still technically a mistake.
You must be joking. -- Alexandre Julliard julliard(a)winehq.org
On 10/28/2012 12:07 PM, Alexandre Julliard wrote:
Nikolay Sivov <bunglehead(a)gmail.com> writes:
On 10/28/2012 17:44, Max TenEyck Woodbury wrote:
Specifying the unnecessary use of a temporary store is a bad habit to have. You should tell the compiler exactly what needs to be done, not throw in extraneous operations.
So, the use of a post operator where a prefix operator is sufficient, while almost certainly harmless, is still technically a mistake.
You must be joking.
Joking? Only sort of. Just being my usual wryly pedantic self... :-)
On 10/28/2012 12:06 PM, Nikolay Sivov wrote:
On 10/28/2012 17:44, Max TenEyck Woodbury wrote:
On 10/28/2012 02:40 AM, Nikolay Sivov wrote:
On 10/28/2012 04:59, max(a)mtew.isa-geek.net wrote:
From: Max TenEyck Woodbury <max(a)mtew.isa-geek.net>
I have been looking at the Microsoft COM and related documentations and noticed that they emphatically recommend using the Interlocked... functions when manipulating reference counts. I managed to set up a search that showed where many of the reference count updates occur and was somewhat surprised at how often this advice was not followed.
It doesn't mean it always has to be followed.
True in a limited sense, but there is a good reason behind their recommendation. Unless there is a good reason not to do so, this particular piece of advice should be followed.
COM objects in wine follow this recommendation in general, even object itself is not thread safe. This doesn't mean however that you need this every time you have some kind of refcount of any sort.
It may or may not be necessary every time, but it should be demonstrated that it is not necessary rather than assumed that it is not. This is a 'race condition' after all, and they are known to be rare and difficult to isolate. I think it is good practice to assume there could be a race problem rather than otherwise.
While I have not converted every reference count update to use the Interlocked... functions, this set of patches fixes a fair number of them.
These are not associated with any particular bug report; they are simply a general precaution against operations that are known to be associated with race conditions.
This precaution doesn't work in general. It's not enough to atomically update refcount to make an implementation thread safe. Also not everything is supposed to be thread safe in a first place.
First, explain what does not have to be thread safe.
Anything really, COM objects in particular if you were talking about them.
I think you are talking about the apartment model here, which forces thread serialization. Despite that, the Microsoft documentation still recommends interlocked operations on reference counts...
I believe that application may try to use multiple threads anywhere, so everything that can be made thread safe, should be.
No.
What do you mean 'No'. That is an opinion, If you disagree, please explain why.
Using interlocks on the reference count updates is a necessary step for thread safety. You are correct that it is not sufficient, but it is necessary.
Again, it depends.
Yes, it might depend on many factors, most of which will make it safe to not interlock, but it is a lot of work checking that all those factors necessary to not use the interlock are in fact in place. Further, if enough of those factors get changed, you have a problem. This is what make code fragile. Fragile can be avoided with steps like this.
Changes like this:
- for (i=0;i<howmuch;i++) + for (i=0;i<howmuch;++i) TRACE("notify at %d to %p\n", notify[i].dwOffset,notify[i].hEventNotify);
are not helpful at all.
The post increment and decrement operation are specified as saving the original value for use in the evaluation of the expression they are part of and modifying the underlying stored value. In expressions like this, that saved value is then discarded. The optimization phase of the compilation usually removes both the save and discard operations.
Sure, but I don't think it's enough to justify such changes all over the place, in existing code.
I agree that it is not enough to justify a separate set of patches, but as part of another set of changes, I think it is justified. After all, how else are examples of bad code going to be removed.
On 10/29/12 01:25, Max TenEyck Woodbury wrote:
On 10/28/2012 12:06 PM, Nikolay Sivov wrote:
On 10/28/2012 17:44, Max TenEyck Woodbury wrote:
On 10/28/2012 02:40 AM, Nikolay Sivov wrote:
On 10/28/2012 04:59, max(a)mtew.isa-geek.net wrote:
From: Max TenEyck Woodbury <max(a)mtew.isa-geek.net>
I have been looking at the Microsoft COM and related documentations and noticed that they emphatically recommend using the Interlocked... functions when manipulating reference counts. I managed to set up a search that showed where many of the reference count updates occur and was somewhat surprised at how often this advice was not followed. It doesn't mean it always has to be followed. True in a limited sense, but there is a good reason behind their recommendation. Unless there is a good reason not to do so, this particular piece of advice should be followed. COM objects in wine follow this recommendation in general, even object itself is not thread safe. This doesn't mean however that you need this every time you have some kind of refcount of any sort. It may or may not be necessary every time, but it should be demonstrated that it is not necessary rather than assumed that it is not. This is a 'race condition' after all, and they are known to be rare and difficult to isolate. I think it is good practice to assume there could be a race problem rather than otherwise.
No, it's a good practice to understand what you're changing instead of blindly assuming that everything with 'ref' in its name needs Interlocked* functions. You're patch even changes things in jscript that are not COM objecst, in a middle of code that is not thread safe anyway. It's documented that those parts of the code are not thread safe. Even worse, you were changing lines that even have a comment about no need for atomic operations (thanks for Michael, who predicted that someone may send patches like you).
While I have not converted every reference count update to use the Interlocked... functions, this set of patches fixes a fair number of them.
These are not associated with any particular bug report; they are simply a general precaution against operations that are known to be associated with race conditions. This precaution doesn't work in general. It's not enough to atomically update refcount to make an implementation thread safe. Also not everything is supposed to be thread safe in a first place.
First, explain what does not have to be thread safe. Anything really, COM objects in particular if you were talking about them.
I think you are talking about the apartment model here, which forces thread serialization. Despite that, the Microsoft documentation still recommends interlocked operations on reference counts...
Just to make it clear: using Interlocked* functions in generic COM object is a good practice. But lack of it is not always a bug and you shouldn't change it without understanding what you are doing.
I believe that application may try to use multiple threads anywhere, so everything that can be made thread safe, should be.
No. What do you mean 'No'. That is an opinion, If you disagree, please explain why.
See above.
Changes like this:
- for (i=0;i<howmuch;i++) + for (i=0;i<howmuch;++i) TRACE("notify at %d to %p\n", notify[i].dwOffset,notify[i].hEventNotify); are not helpful at all.
The post increment and decrement operation are specified as saving the original value for use in the evaluation of the expression they are part of and modifying the underlying stored value. In expressions like this, that saved value is then discarded. The optimization phase of the compilation usually removes both the save and discard operations. Sure, but I don't think it's enough to justify such changes all over the place, in existing code.
I agree that it is not enough to justify a separate set of patches, but as part of another set of changes, I think it is justified. After all, how else are examples of bad code going to be removed.
This is not a bad coding. You're changing a lot of my code, which uses my coding style and this style is to use postfix incrementation. I'm not saying it's better of worse because it's not, so there is no reason to change it. Your argument about requirement for temporary storage is an example of bad way to think about the code. Jacek
participants (4)
-
Alexandre Julliard -
Jacek Caban -
Max TenEyck Woodbury -
Nikolay Sivov