Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
On Fri, Nov 9, 2018 at 9:43 PM chouryzhou(周威) wrote: > > > > > > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it > > > will be a static > > > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by > > > me) with > > > no namespace-ization. You will get the same one in all processes, > > > everything is > > > the same as without this patch. > > > > except, as far as I can tell, binder_init_ns() would never have been > > called on it so the mutex and list heads are not initialized so its > > completely broken. Am I missing something? How do those fields get > > initialized in this case? > > > @@ -5832,8 +5888,12 @@ static int __init binder_init(void) > > goto err_init_binder_device_failed; > > } > > > > - return ret; > > + ret = binder_init_ns(_ipc_ns); > > + if (ret) > > + goto err_init_namespace_failed; > > > > + return ret; > > They are initialized here. Ok, This init_ipc_ns is a global declared in msgutil.c if SYSVIPC || POSIX_MQUEUE. This seems kinda hacky, but now I finally see why the dependancy... msgutil.c is the only file we can count on if !IPC_NS && (SYSVIPC || POSIX_MQUEUE). There must be a cleaner way to do this, I really don't like this dependency... wouldn't it be cleaner to do: #ifndef CONFIG_IPC_NS static struct ipc_namespace binder_ipc_ns; #define ipcns (_ipc_ns) #else #define ipcns (current->nsproxy->ipc_ns) #endif (and make the initialization of binder_ipc_ns conditional on IPC_NS) This gets us the same thing without the incestuous dependency on the msgutil.c version of init_ipc_ns...and then binder doesn't rely on SYSVIPC or POSIX_MQUEUE directly. > > - choury - >
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
On Fri, Nov 9, 2018 at 9:43 PM chouryzhou(周威) wrote: > > > > > > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it > > > will be a static > > > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by > > > me) with > > > no namespace-ization. You will get the same one in all processes, > > > everything is > > > the same as without this patch. > > > > except, as far as I can tell, binder_init_ns() would never have been > > called on it so the mutex and list heads are not initialized so its > > completely broken. Am I missing something? How do those fields get > > initialized in this case? > > > @@ -5832,8 +5888,12 @@ static int __init binder_init(void) > > goto err_init_binder_device_failed; > > } > > > > - return ret; > > + ret = binder_init_ns(_ipc_ns); > > + if (ret) > > + goto err_init_namespace_failed; > > > > + return ret; > > They are initialized here. Ok, This init_ipc_ns is a global declared in msgutil.c if SYSVIPC || POSIX_MQUEUE. This seems kinda hacky, but now I finally see why the dependancy... msgutil.c is the only file we can count on if !IPC_NS && (SYSVIPC || POSIX_MQUEUE). There must be a cleaner way to do this, I really don't like this dependency... wouldn't it be cleaner to do: #ifndef CONFIG_IPC_NS static struct ipc_namespace binder_ipc_ns; #define ipcns (_ipc_ns) #else #define ipcns (current->nsproxy->ipc_ns) #endif (and make the initialization of binder_ipc_ns conditional on IPC_NS) This gets us the same thing without the incestuous dependency on the msgutil.c version of init_ipc_ns...and then binder doesn't rely on SYSVIPC or POSIX_MQUEUE directly. > > - choury - >
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
> > > I still don't understand the dependencies on SYSVIPC or POSIX_MQUEUE. > > > It seems like this mechanism would work even if both are disabled -- > > > as long as IPC_NS is enabled. Seems cleaner to change init/Kconfig and > > > allow IPC_NS if CONFIG_ANDROID_BINDER_IPC and change this line to > > > "#ifndef CONFIG_IPC_NS" > > > > Let me explain it in detail. If SYSIPC and IPC_NS are both defined, > > current->nsproxy->ipc_ns will save the ipc namespace variables. We just use > > it. If SYSIPC (or POSIX_MQUEUE) is defined while IPC_NS is not set, > > current->nsproxy->ipc_ns will always refer to init_ipc_ns in ipc/msgutil.c, > > which is also fine to us. But if neither SYSIPC nor POSIX_MQUEUE is set > > (IPC_NS can't be set in this situation), there is no > > current->nsproxy->ipc_ns. > > We make a fack init_ipc_ns here and use it. > > Yes, I can read the code. I'm wondering specifically about SYSVIPC and > POSIX_MQUEUE. Even with your code changes, binder has no dependency on > these configs. Why are you creating one? The actual dependency with > your changes is on "current->nsproxy->ipc_ns" being initialized for > binder -- which is dependent on CONFIG_IPC_NS being enabled, isn't it? > > If SYSVIPC or POSIX_MQUEUE are enabled, but IPC_NS is disabled, does this > work? If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will be a static reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by me) with no namespace-ization. You will get the same one in all processes, everything is the same as without this patch. - choury -
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
> > > I still don't understand the dependencies on SYSVIPC or POSIX_MQUEUE. > > > It seems like this mechanism would work even if both are disabled -- > > > as long as IPC_NS is enabled. Seems cleaner to change init/Kconfig and > > > allow IPC_NS if CONFIG_ANDROID_BINDER_IPC and change this line to > > > "#ifndef CONFIG_IPC_NS" > > > > Let me explain it in detail. If SYSIPC and IPC_NS are both defined, > > current->nsproxy->ipc_ns will save the ipc namespace variables. We just use > > it. If SYSIPC (or POSIX_MQUEUE) is defined while IPC_NS is not set, > > current->nsproxy->ipc_ns will always refer to init_ipc_ns in ipc/msgutil.c, > > which is also fine to us. But if neither SYSIPC nor POSIX_MQUEUE is set > > (IPC_NS can't be set in this situation), there is no > > current->nsproxy->ipc_ns. > > We make a fack init_ipc_ns here and use it. > > Yes, I can read the code. I'm wondering specifically about SYSVIPC and > POSIX_MQUEUE. Even with your code changes, binder has no dependency on > these configs. Why are you creating one? The actual dependency with > your changes is on "current->nsproxy->ipc_ns" being initialized for > binder -- which is dependent on CONFIG_IPC_NS being enabled, isn't it? > > If SYSVIPC or POSIX_MQUEUE are enabled, but IPC_NS is disabled, does this > work? If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will be a static reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by me) with no namespace-ization. You will get the same one in all processes, everything is the same as without this patch. - choury -
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
> > > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will > > be a static > > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by > > me) with > > no namespace-ization. You will get the same one in all processes, > > everything is > > the same as without this patch. > > except, as far as I can tell, binder_init_ns() would never have been > called on it so the mutex and list heads are not initialized so its > completely broken. Am I missing something? How do those fields get > initialized in this case? > @@ -5832,8 +5888,12 @@ static int __init binder_init(void) > goto err_init_binder_device_failed; > } > > - return ret; > + ret = binder_init_ns(_ipc_ns); > + if (ret) > + goto err_init_namespace_failed; > > + return ret; They are initialized here. - choury -
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
> > > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will > > be a static > > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by > > me) with > > no namespace-ization. You will get the same one in all processes, > > everything is > > the same as without this patch. > > except, as far as I can tell, binder_init_ns() would never have been > called on it so the mutex and list heads are not initialized so its > completely broken. Am I missing something? How do those fields get > initialized in this case? > @@ -5832,8 +5888,12 @@ static int __init binder_init(void) > goto err_init_binder_device_failed; > } > > - return ret; > + ret = binder_init_ns(_ipc_ns); > + if (ret) > + goto err_init_namespace_failed; > > + return ret; They are initialized here. - choury -
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
On Fri, Nov 9, 2018 at 8:43 PM chouryzhou(周威) wrote: > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will > be a static > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by me) > with > no namespace-ization. You will get the same one in all processes, everything > is > the same as without this patch. except, as far as I can tell, binder_init_ns() would never have been called on it so the mutex and list heads are not initialized so its completely broken. Am I missing something? How do those fields get initialized in this case? > > - choury - >
Re: Re: [PATCH V3] binder: ipc namespace support for android binder(Internet mail)
On Fri, Nov 9, 2018 at 8:43 PM chouryzhou(周威) wrote: > > If IPC_NS is disabled, "current-nsporxy->ipc_ns" will also exists, it will > be a static > reference of "init_ipc_ns" (in ipc/msgutil.c, not defined in binder.c by me) > with > no namespace-ization. You will get the same one in all processes, everything > is > the same as without this patch. except, as far as I can tell, binder_init_ns() would never have been called on it so the mutex and list heads are not initialized so its completely broken. Am I missing something? How do those fields get initialized in this case? > > - choury - >
