+ remove-the-likelypid-check-in-copy_process.patch added to -mm tree
Quoted 8 lines · Sukadev Bhattiprolu
(quoted diff) +0
-4@@ -1237,26 +1237,24 @@ static struct task_struct *copy_process( } } - if (likely(p->pid)) { - add_parent(p); - tracehook_init_task(p); -
Review static struct task_struct *copy_process( · line 1240
Quoted 10 lines
(quoted diff) +1
-9- if (thread_group_leader(p)) { - pid_t pgid = process_group(current); - pid_t sid = process_session(current); - - p->signal->tty = current->signal->tty; - p->signal->pgrp = pgid; - set_signal_session(p->signal, process_session(current)); - attach_pid(p, PIDTYPE_PGID, find_pid(pgid)); - attach_pid(p, PIDTYPE_SID, find_pid(sid)); + add_parent(p);
This means that idle threads (except "swapper") are visible to for_each_process() and do_each_thread(). Looks dangerous and somewhat strange to me.
Could you explain this change?
Oleg.
Quoted 7 lines · Oleg Nesterov
Sukadev Bhattiprolu wrote:
This means that idle threads (except "swapper") are visible to
for_each_process()
anddo_each_thread(). Looks dangerous and somewhat strange to me.Could you explain this change?
Good catch. I've been so busy pounding reviewing this patches into something that made sense that I missed the fact that we care about this for more than just the NULL pointer that would occur if we didn't do this.
Still it would be good if we could find a way to remove this rare special case.
Any chance we can undo what we don't want done for_idle, or create a factor of copy_process that only does as much as fork_idle should do, and make copy_process a wrapper that does the rest.
I doubt it is significant anywhere but it would be nice to remove a branch that except at boot up never happens.
Eric
Oleg Nesterov
Acked-by Eric W. BiedermanQuoted 9 lines · Eric W. Biederman
Quoted 6 lines · Oleg Nesterov
Sukadev Bhattiprolu wrote:
This means that idle threads (except "swapper") are visible to
for_each_process()
anddo_each_thread(). Looks dangerous and somewhat strange to me.
Quoted 6 lines
Quoted 1 line
Could you explain this change?
Good catch. I've been so busy pounding reviewing this patches into something that made sense that I missed the fact that we care about this for more than just the NULL pointer that would occur if we didn't do this.
Why it is bad to have a NULL pointer for idle thread? (Sorry for stupid question, I can't track the code changes these days).
Quoted 9 lines
Still it would be good if we could find a way to remove this rare special case.
Any chance we can undo what we don't want done
for_idle, or create a factor ofcopy_processthat only does as much asfork_idleshould do, and makecopy_processa wrapper that does the rest.I doubt it is significant anywhere but it would be nice to remove a branch that except at boot up never happens.
... or at cpu-hotplug. Probably you are right, but I am not sure.
The "if (p->pid)" check in essence implements CLONE_UNHASHED flag, it may be useful.
Btw. Looking at http://marc.theaimsgroup.com/?l=linux-mm-commits,
Subject: Explicitly set pgid and sid of init process
From: Sukadev Bhattiprolu <[email protected]>
Explicitly set pgid and sid of init process to 1.
---
init/main.c | 1 +
1 file changed, 1 insertion(+)
diff -puN init/main.c~explicitly-set-pgid-and-sid-of-init-process init/main.c --- a/init/main.c~explicitly-set-pgid-and-sid-of-init-process +++ a/init/main.c
@@ -783,6 +783,7 @@ static int __init init(void * unused)
*/
init_pid_ns.child_reaper = current;
+ __set_special_pids(1, 1);
cad_pid = task_pid(current);
smp_prepare_cpus(max_cpus);
Nice changelog :)
The patch looks good, except __set_special_pids(1, 1) should be no-op. This is a child forked by swapper. copy_process() was changed by
use-task_pgrp-task_session-in-copy_process.patch
, but signal->{pgrp,_session} get its value from INIT_SIGNALS ?
Could you explain this as well? Some other changes I missed?
Oleg.
Quoted 10 lines · Oleg Nesterov
Quoted 9 lines · Eric W. Biederman
Quoted 6 lines · Oleg Nesterov
Sukadev Bhattiprolu wrote:
This means that idle threads (except "swapper") are visible to
for_each_process()
anddo_each_thread(). Looks dangerous and somewhat strange to me.
Quoted 6 lines
Quoted 6 lines
Quoted 1 line
Could you explain this change?
Good catch. I've been so busy pounding reviewing this patches into something that made sense that I missed the fact that we care about this for more than just the NULL pointer that would occur if we didn't do this.
Err. I meant NULL pointer dereference.
Quoted 53 lines
Why it is bad to have a NULL pointer for idle thread? (Sorry for stupid question, I can't track the code changes these days).
Quoted 9 lines
Still it would be good if we could find a way to remove this rare special case.
Any chance we can undo what we don't want done
for_idle, or create a factor ofcopy_processthat only does as much asfork_idleshould do, and makecopy_processa wrapper that does the rest.I doubt it is significant anywhere but it would be nice to remove a branch that except at boot up never happens.
... or at cpu-hotplug. Probably you are right, but I am not sure.
The "if (
p->pid)" check in essence implementsCLONE_UNHASHEDflag, it may be useful.Btw. Looking at http://marc.theaimsgroup.com/?l=linux-mm-commits,
Subject: Explicitly set pgid and sid of init process
From: Sukadev Bhattiprolu <[email protected]>Explicitly set pgid and sid of init process to 1.
Signed-off-by Sukadev Bhattiprolu <[email protected]>Cc Cedric Le Goater <[email protected]>Cc Dave Hansen <[email protected]>Cc Serge Hallyn <[email protected]>Cc Eric Biederman <[email protected]>Cc Herbert Poetzl <[email protected]>Cc <[email protected]>Signed-off-by Andrew Morton <[email protected]>---
init/main.c| 1 +
1 file changed, 1 insertion(+)diff -puN
init/main.c~explicitly-set-pgid-and-sid-of-init-processinit/main.c--- a/init/main.c~explicitly-set-pgid-and-sid-of-init-process +++ a/init/main.c @@ -783,6 +783,7 @@ static int __init init(void * unused) */ init_pid_ns.child_reaper = current; + __set_special_pids(1, 1); cad_pid = task_pid(current); smp_prepare_cpus(max_cpus);
Quoted 8 lines
Nice changelog :)
The patch looks good, except
__set_special_pids(1, 1)should be no-op. This is a child forked by swapper.copy_process()was changed by
use-task_pgrp-task_session-in-copy_process.patch
, but signal->{pgrp,_session} get its value fromINIT_SIGNALS?Could you explain this as well? Some other changes I missed?
As I recall the patch series started with modifying attach_pid to take a struct pid pointer instead of a pid_t value. It means fewer hash table looks ups and it should help in implementing the pid namespace.
Well the initial kernel process does not have a struct pid so when it's children start doing:
attach_pid(p, PIDTYPE_PGID, task_group(p)); attach_pid(p, PIDTYPE_SID, task_session(p));
We will get an oops.
So a dummy unhashed struct pid was added for the idle threads. Allowing several special cases in the code to be removed.
With that chance the previous special case to force the idle thread init session 1 pgrp 1 no longer works because attach_pid no longer looks at the pid value but instead at the struct pid pointers.
So we had to add the __set_special_pids() to continue to keep init in session 1 pgrp 1. Since /sbin/init calls setsid() that our setting the sid and the pgrp may not be strictly necessary. Still is better to not take any chances.
Anyway the point of removing the likely(pid) check was that it didn't look necessary any longer. But as you have correctly pointed putting it on the task list and incrementing the process count for the idle threads is probably still a problem. So while we are much better we still have some use for the if (likely(p->pid)) special case.
Is that enough to bring you up to speed?
Eric
Quoted 28 lines · Eric W. Biederman
Quoted 11 lines · Oleg Nesterov
--- a/init/main.c~explicitly-set-pgid-and-sid-of-init-process +++ a/init/main.c @@ -783,6 +783,7 @@ static int __init init(void * unused) */ init_pid_ns.child_reaper = current; + __set_special_pids(1, 1); cad_pid = task_pid(current); smp_prepare_cpus(max_cpus);Quoted 8 lines
Nice changelog :)
The patch looks good, except
__set_special_pids(1, 1)should be no-op. This is a child forked by swapper.copy_process()was changed by
use-task_pgrp-task_session-in-copy_process.patch
, but signal->{pgrp,_session} get its value fromINIT_SIGNALS?Could you explain this as well? Some other changes I missed?
As I recall the patch series started with modifying
attach_pidto take a struct pid pointer instead of apid_tvalue. It means fewer hash table looks ups and it should help in implementing the pid namespace.
Quoted 5 lines
Well the initial kernel process does not have a struct pid so when it's children start doing:
attach_pid(p, PIDTYPE_PGID, task_group(p)); attach_pid(p, PIDTYPE_SID, task_session(p));We will get an oops.
So far this is the only reason to have init_struct_pid. Because the boot CPU (swapper) forks, right?
Quoted 11 lines
So a dummy unhashed struct pid was added for the idle threads. Allowing several special cases in the code to be removed.
With that chance the previous special case to force the idle thread init session 1 pgrp 1 no longer works because
attach_pidno longer looks at the pid value but instead at the struct pid pointers.So we had to add the
__set_special_pids()to continue to keep init in session 1 pgrp 1. Since/sbin/initcallssetsid()that our setting the sid and the pgrp may not be strictly necessary. Still is better to not take any chances.
Yes, yes, I see. But my (very unclear, sorry) question was: shouldn't we change INIT_SIGNALS then? /sbin/init inherits ->pgrp == ->_session == 1, in that case __set_special_pids(1,1) does nothing.
Quoted 4 lines
Anyway the point of removing the likely(pid) check was that it didn't look necessary any longer. But as you have correctly pointed putting it on the task list and incrementing the process count for the idle threads is probably still a problem.
Yes. Note also that the parent doing fork_idle() is not always swapper, it is just wrong to do attach_pid(PIDTYPE_PGID/PIDTYPE_SID) in this case.
example: arch/x86_64/kernel/smpboot.c:do_boot_cpu()
Quoted 2 lines
So while we are much better we
still have some use for the if (likely(p->pid)) special case.
Yes, I think this change should be dropped for now.
Quoted 1 line
Is that enough to bring you up to speed?
Thanks for your explanations!
Oleg.
Quoted 9 lines · Oleg Nesterov
Quoted 6 lines
Well the initial kernel process does not have a struct pid so when it's children start doing:
attach_pid(p, PIDTYPE_PGID, task_group(p)); attach_pid(p, PIDTYPE_SID, task_session(p));We will get an oops.
So far this is the only reason to have
init_struct_pid. Because the boot CPU (swapper) forks, right?
Damn. I am afraid I was not clear again :) Not init_struct_pid, but
+ .pids = { \
+ [PIDTYPE_PID] = INIT_PID_LINK(PIDTYPE_PID), \
+ [PIDTYPE_PGID] = INIT_PID_LINK(PIDTYPE_PGID), \
+ [PIDTYPE_SID] = INIT_PID_LINK(PIDTYPE_SID), \
+ }, \for INIT_TASK().
Quoted 7 lines
Quoted 7 lines
So a dummy unhashed struct pid was added for the idle threads. Allowing several special cases in the code to be removed.
With that chance the previous special case to force the idle thread init session 1 pgrp 1 no longer works because
attach_pidno longer looks at the pid value but instead at the struct pid pointers.
Quoted 8 lines
Quoted 4 lines
So we had to add the
__set_special_pids()to continue to keep init in session 1 pgrp 1. Since/sbin/initcallssetsid()that our setting the sid and the pgrp may not be strictly necessary. Still is better to not take any chances.Yes, yes, I see. But my (very unclear, sorry) question was: shouldn't we change
INIT_SIGNALSthen?/sbin/initinherits ->pgrp == ->_session == 1, in that case__set_special_pids(1,1)does nothing.
... and thus /sbin/init remains attached to the .pids above, no?
Oleg.
Quoted 33 lines · Oleg Nesterov
Quoted 9 lines · Oleg Nesterov
Quoted 6 lines
Well the initial kernel process does not have a struct pid so when it's children start doing:
attach_pid(p, PIDTYPE_PGID, task_group(p)); attach_pid(p, PIDTYPE_SID, task_session(p));We will get an oops.
So far this is the only reason to have
init_struct_pid. Because the boot CPU (swapper) forks, right?Damn. I am afraid I was not clear again :) Not
init_struct_pid, but+ .pids = { \
+ [PIDTYPE_PID] =INIT_PID_LINK(PIDTYPE_PID), \
+ [PIDTYPE_PGID] =INIT_PID_LINK(PIDTYPE_PGID), \
+ [PIDTYPE_SID] =INIT_PID_LINK(PIDTYPE_SID), \
+ }, \for
INIT_TASK().Quoted 12 lines
Quoted 11 lines
So a dummy unhashed struct pid was added for the idle threads. Allowing several special cases in the code to be removed.
With that chance the previous special case to force the idle thread init session 1 pgrp 1 no longer works because
attach_pidno longer looks at the pid value but instead at the struct pid pointers.So we had to add the
__set_special_pids()to continue to keep init in session 1 pgrp 1. Since/sbin/initcallssetsid()that our setting the sid and the pgrp may not be strictly necessary. Still is better to not take any chances.
Quoted 5 lines
Quoted 3 lines
Yes, yes, I see. But my (very unclear, sorry) question was: shouldn't we change
INIT_SIGNALSthen?/sbin/initinherits ->pgrp == ->_session == 1, in that case__set_special_pids(1,1)does nothing.... and thus
/sbin/initremains attached to the .pids above, no?
The problem is that we dynamically allocate the struct pid for pid_t == 1 when we fork init.
Which means we don't have access to it at compile time so we can no longer make INIT_SIGNALS set ->gprp == ->session == 1.
Eric
Quoted 40 lines · Eric W. Biederman
Quoted 37 lines · Oleg Nesterov
Quoted 9 lines · Oleg Nesterov
Quoted 6 lines
Well the initial kernel process does not have a struct pid so when it's children start doing:
attach_pid(p, PIDTYPE_PGID, task_group(p)); attach_pid(p, PIDTYPE_SID, task_session(p));We will get an oops.
So far this is the only reason to have
init_struct_pid. Because the boot CPU (swapper) forks, right?Damn. I am afraid I was not clear again :) Not
init_struct_pid, but+ .pids = { \
+ [PIDTYPE_PID] =INIT_PID_LINK(PIDTYPE_PID), \
+ [PIDTYPE_PGID] =INIT_PID_LINK(PIDTYPE_PGID), \
+ [PIDTYPE_SID] =INIT_PID_LINK(PIDTYPE_SID), \
+ }, \for
INIT_TASK().Quoted 15 lines
Quoted 11 lines
So a dummy unhashed struct pid was added for the idle threads. Allowing several special cases in the code to be removed.
With that chance the previous special case to force the idle thread init session 1 pgrp 1 no longer works because
attach_pidno longer looks at the pid value but instead at the struct pid pointers.So we had to add the
__set_special_pids()to continue to keep init in session 1 pgrp 1. Since/sbin/initcallssetsid()that our setting the sid and the pgrp may not be strictly necessary. Still is better to not take any chances.Yes, yes, I see. But my (very unclear, sorry) question was: shouldn't we change
INIT_SIGNALSthen?/sbin/initinherits ->pgrp == ->_session == 1, in that case__set_special_pids(1,1)does nothing.
Quoted 7 lines
Quoted 1 line
... and thus
/sbin/initremains attached to the .pids above, no?The problem is that we dynamically allocate the struct pid for
pid_t== 1 when we fork init.Which means we don't have access to it at compile time so we can no longer make
INIT_SIGNALSset ->gprp == ->session == 1.
Yes! I meant we should change INIT_SIGNALS(), currently it does
#define INIT_SIGNALS(sig) {
...
.pgrp = 1,
{ .__session = 1},and this confuses (I think) set_special_pids(1,1) above. Because __set_special_pids() still deals with pid_t, not "struct pid".
Unless I missed something, we should kill these 2 initializations above.
Oleg.
Quoted 7 lines · Oleg Nesterov
Yes! I meant we should change
INIT_SIGNALS(), currently it does#define INIT_SIGNALS(sig) { ... .pgrp = 1, { .__session = 1},
Quoted 5 lines
and this confuses (I think)
set_special_pids(1,1)above. Because__set_special_pids()still deals withpid_t, not "struct pid".Unless I missed something, we should kill these 2 initializations above.
Got it. I agree we should initialize those fields to 0.
Sukadev you want to get that?
Eric
Eric W. Biederman [[email protected]] wrote:
| Oleg Nesterov <[email protected]> writes: | | > Yes! I meant we should change INIT_SIGNALS(), currently it does | > | > #define INIT_SIGNALS(sig) { | > ... | > .pgrp = 1, | > { .__session = 1}, | > | > and this confuses (I think) set_special_pids(1,1) above. Because | > __set_special_pids() still deals with pid_t, not "struct pid". | > | > Unless I missed something, we should kill these 2 initializations | > above. | | Got it. I agree we should initialize those fields to 0. | | Sukadev you want to get that?
Sure. Will do that.
Thanks Oleg for your detailed review/comments.
Suka