aboutsummaryrefslogtreecommitdiff
path: root/ipc/sem.c
AgeCommit message (Collapse)Author
2013-09-30ipc/sem.c: update sem_otime for all operationsManfred Spraul
In commit 0a2b9d4c7967 ("ipc/sem.c: move wake_up_process out of the spinlock section"), the update of semaphore's sem_otime(last semop time) was moved to one central position (do_smart_update). But since do_smart_update() is only called for operations that modify the array, this means that wait-for-zero semops do not update sem_otime anymore. The fix is simple: Non-alter operations must update sem_otime. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Reported-by: Jia He <jiakernel@gmail.com> Tested-by: Jia He <jiakernel@gmail.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-30ipc/sem.c: synchronize the proc interfaceManfred Spraul
The proc interface is not aware of sem_lock(), it instead calls ipc_lock_object() directly. This means that simple semop() operations can run in parallel with the proc interface. Right now, this is uncritical, because the implementation doesn't do anything that requires a proper synchronization. But it is dangerous and therefore should be fixed. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-30ipc/sem.c: optimize sem_lock()Manfred Spraul
Operations that need access to the whole array must guarantee that there are no simple operations ongoing. Right now this is achieved by spin_unlock_wait(sem->lock) on all semaphores. If complex_count is nonzero, then this spin_unlock_wait() is not necessary, because it was already performed in the past by the thread that increased complex_count and even though sem_perm.lock was dropped inbetween, no simple operation could have started, because simple operations cannot start when complex_count is non-zero. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Rik van Riel <riel@redhat.com> Reviewed-by: Davidlohr Bueso <davidlohr@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-30ipc/sem.c: fix race in sem_lock()Manfred Spraul
The exclusion of complex operations in sem_lock() is insufficient: after acquiring the per-semaphore lock, a simple op must first check that sem_perm.lock is not locked and only after that test check complex_count. The current code does it the other way around - and that creates a race. Details are below. The patch is a complete rewrite of sem_lock(), based in part on the code from Mike Galbraith. It removes all gotos and all loops and thus the risk of livelocks. I have tested the patch (together with the next one) on my i3 laptop and it didn't cause any problems. The bug is probably also present in 3.10 and 3.11, but for these kernels it might be simpler just to move the test of sma->complex_count after the spin_is_locked() test. Details of the bug: Assume: - sma->complex_count = 0. - Thread 1: semtimedop(complex op that must sleep) - Thread 2: semtimedop(simple op). Pseudo-Trace: Thread 1: sem_lock(): acquire sem_perm.lock Thread 1: sem_lock(): check for ongoing simple ops Nothing ongoing, thread 2 is still before sem_lock(). Thread 1: try_atomic_semop() <<< preempted. Thread 2: sem_lock(): static inline int sem_lock(struct sem_array *sma, struct sembuf *sops, int nsops) { int locknum; again: if (nsops == 1 && !sma->complex_count) { struct sem *sem = sma->sem_base + sops->sem_num; /* Lock just the semaphore we are interested in. */ spin_lock(&sem->lock); /* * If sma->complex_count was set while we were spinning, * we may need to look at things we did not lock here. */ if (unlikely(sma->complex_count)) { spin_unlock(&sem->lock); goto lock_array; } <<<<<<<<< <<< complex_count is still 0. <<< <<< Here it is preempted <<<<<<<<< Thread 1: try_atomic_semop() returns, notices that it must sleep. Thread 1: increases sma->complex_count. Thread 1: drops sem_perm.lock Thread 2: /* * Another process is holding the global lock on the * sem_array; we cannot enter our critical section, * but have to wait for the global lock to be released. */ if (unlikely(spin_is_locked(&sma->sem_perm.lock))) { spin_unlock(&sem->lock); spin_unlock_wait(&sma->sem_perm.lock); goto again; } <<< sem_perm.lock already dropped, thus no "goto again;" locknum = sops->sem_num; Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Mike Galbraith <bitbucket@online.de> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: <stable@vger.kernel.org> [3.10+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-24ipc: fix race with LSMsDavidlohr Bueso
Currently, IPC mechanisms do security and auditing related checks under RCU. However, since security modules can free the security structure, for example, through selinux_[sem,msg_queue,shm]_free_security(), we can race if the structure is freed before other tasks are done with it, creating a use-after-free condition. Manfred illustrates this nicely, for instance with shared mem and selinux: -> do_shmat calls rcu_read_lock() -> do_shmat calls shm_object_check(). Checks that the object is still valid - but doesn't acquire any locks. Then it returns. -> do_shmat calls security_shm_shmat (e.g. selinux_shm_shmat) -> selinux_shm_shmat calls ipc_has_perm() -> ipc_has_perm accesses ipc_perms->security shm_close() -> shm_close acquires rw_mutex & shm_lock -> shm_close calls shm_destroy -> shm_destroy calls security_shm_free (e.g. selinux_shm_free_security) -> selinux_shm_free_security calls ipc_free_security(&shp->shm_perm) -> ipc_free_security calls kfree(ipc_perms->security) This patch delays the freeing of the security structures after all RCU readers are done. Furthermore it aligns the security life cycle with that of the rest of IPC - freeing them based on the reference counter. For situations where we need not free security, the current behavior is kept. Linus states: "... the old behavior was suspect for another reason too: having the security blob go away from under a user sounds like it could cause various other problems anyway, so I think the old code was at least _prone_ to bugs even if it didn't have catastrophic behavior." I have tested this patch with IPC testcases from LTP on both my quad-core laptop and on a 64 core NUMA server. In both cases selinux is enabled, and tests pass for both voluntary and forced preemption models. While the mentioned races are theoretical (at least no one as reported them), I wanted to make sure that this new logic doesn't break anything we weren't aware of. Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Davidlohr Bueso <davidlohr@hp.com> Acked-by: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-09-11ipc: rename ids->rw_mutexDavidlohr Bueso
Since in some situations the lock can be shared for readers, we shouldn't be calling it a mutex, rename it to rwsem. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc/sem.c: rename try_atomic_semop() to perform_atomic_semop(), docu updateManfred Spraul
Cleanup: Some minor points that I noticed while writing the previous patches 1) The name try_atomic_semop() is misleading: The function performs the operation (if it is possible). 2) Some documentation updates. No real code change, a rename and documentation changes. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc/sem.c: replace shared sem_otime with per-semaphore valueManfred Spraul
sem_otime contains the time of the last semaphore operation that completed successfully. Every operation updates this value, thus access from multiple cpus can cause thrashing. Therefore the patch replaces the variable with a per-semaphore variable. The per-array sem_otime is only calculated when required. No performance improvement on a single-socket i3 - only important for larger systems. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc/sem.c: always use only one queue for alter operationsManfred Spraul
There are two places that can contain alter operations: - the global queue: sma->pending_alter - the per-semaphore queues: sma->sem_base[].pending_alter. Since one of the queues must be processed first, this causes an odd priorization of the wakeups: complex operations have priority over simple ops. The patch restores the behavior of linux <=3.0.9: The longest waiting operation has the highest priority. This is done by using only one queue: - if there are complex ops, then sma->pending_alter is used. - otherwise, the per-semaphore queues are used. As a side effect, do_smart_update_queue() becomes much simpler: no more goto logic. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc/sem: separate wait-for-zero and alter tasks into seperate queuesManfred Spraul
Introduce separate queues for operations that do not modify the semaphore values. Advantages: - Simpler logic in check_restart(). - Faster update_queue(): Right now, all wait-for-zero operations are always tested, even if the semaphore value is not 0. - wait-for-zero gets again priority, as in linux <=3.0.9 Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc/sem.c: cacheline align the semaphore structuresManfred Spraul
As now each semaphore has its own spinlock and parallel operations are possible, give each semaphore its own cacheline. On a i3 laptop, this gives up to 28% better performance: #semscale 10 | grep "interleave 2" - before: Cpus 1, interleave 2 delay 0: 36109234 in 10 secs Cpus 2, interleave 2 delay 0: 55276317 in 10 secs Cpus 3, interleave 2 delay 0: 62411025 in 10 secs Cpus 4, interleave 2 delay 0: 81963928 in 10 secs -after: Cpus 1, interleave 2 delay 0: 35527306 in 10 secs Cpus 2, interleave 2 delay 0: 70922909 in 10 secs <<< + 28% Cpus 3, interleave 2 delay 0: 80518538 in 10 secs Cpus 4, interleave 2 delay 0: 89115148 in 10 secs <<< + 8.7% i3, with 2 cores and with hyperthreading enabled. Interleave 2 in order use first the full cores. HT partially hides the delay from cacheline trashing, thus the improvement is "only" 8.7% if 4 threads are running. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.com> Cc: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc: remove unused functionsDavidlohr Bueso
We can now drop the msg_lock and msg_lock_check functions along with a bogus comment introduced previously in semctl_down. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc: move locking out of ipcctl_pre_down_nolockDavidlohr Bueso
This function currently acquires both the rw_mutex and the rcu lock on successful lookups, leaving the callers to explicitly unlock them, creating another two level locking situation. Make the callers (including those that still use ipcctl_pre_down()) explicitly lock and unlock the rwsem and rcu lock. Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-07-09ipc: close open coded spin lock callsDavidlohr Bueso
Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Rik van Riel <riel@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-26ipc/sem.c: Fix missing wakeups in do_smart_update_queue()Manfred Spraul
do_smart_update_queue() is called when an operation (semop, semctl(SETVAL), semctl(SETALL), ...) modified the array. It must check which of the sleeping tasks can proceed. do_smart_update_queue() missed a few wakeups: - if a sleeping complex op was completed, then all per-semaphore queues must be scanned - not only those that were modified by *sops - if a sleeping simple op proceeded, then the global queue must be scanned again And: - the test for "|sops == NULL) before scanning the global queue is not required: If the global queue is empty, then it doesn't need to be scanned - regardless of the reason for calling do_smart_update_queue() The patch is not optimized, i.e. even completing a wait-for-zero operation causes a rescan. This is done to keep the patch as simple as possible. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-09ipc,sem: fix semctl(..., GETNCNT)Rik van Riel
The semctl GETNCNT returns the number of semops waiting for the specified semaphore to become nonzero. After commit 9f1bc2c9022c ("ipc,sem: have only one list in struct sem_queue"), the semops waiting on just one semaphore are waiting on that semaphore's list. In order to return the correct count, we have to walk that list too, in addition to the sem_array's list for complex operations. Signed-off-by: Rik van Riel <riel@redhat.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-09ipc,sem: fix semctl(..., GETZCNT)Rik van Riel
The semctl GETZCNT returns the number of semops waiting for the specified semaphore to become zero. After commit 9f1bc2c9022c ("ipc,sem: have only one list in struct sem_queue"), the semops waiting on just one semaphore are waiting on that semaphore's list. In order to return the correct count, we have to walk that list too, in addition to the sem_array's list for complex operations. This bug broke dbench; it works again with this patch applied. Signed-off-by: Rik van Riel <riel@redhat.com> Reported-by: Kent Overstreet <koverstreet@google.com> Tested-by: Kent Overstreet <koverstreet@google.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: simplify rcu_read_lock() in semctl_nolock()Linus Torvalds
This trivially combines two rcu_read_lock() calls in both sides of a if-statement into one single one in front of the if-statement. Split out as an independent cleanup from the previous commit. Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: simplify semtimedop/semctl_main() common error path handlingLinus Torvalds
With various straight RCU lock/unlock movements, one common exit path pattern had become rcu_read_unlock(); goto out_wakeup; and in fact there were no cases where we wanted to exit to out_wakeup _without_ releasing the RCU read lock. So replace that pattern with "goto out_rcu_wakeup", and remove the old out_wakeup. Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: move sem_obtain_lock() rcu locking into the only callerLinus Torvalds
sem_obtain_lock() was another of those functions that returned with the RCU lock held for reading in the success case. Move the RCU locking to the caller (semtimedop()), making it more obvious. We already did RCU locking elsewhere in that function. Side note: why does semtimedop() re-do the semphore lookup after the sleep, rather than just getting a reference to the semaphore it already looked up originally? Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: fix double sem unlock in semctl error pathLinus Torvalds
Fix another ipc locking buglet introduced by the scalability patches: when semctl_down() was changed to delay the semaphore locking, one error path for security_sem_semctl() went through the semaphore unlock logic even though the semaphore had never been locked. Introduced by commit 16df3674efe3 ("ipc,sem: do not hold ipc lock more than necessary") Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: move the rcu_read_lock() from sem_lock_and_putref() into callersLinus Torvalds
This is another ipc semaphore locking cleanup, trying to make the locking more straightforward. We move the rcu read locking into the callers of sem_lock_and_putref(), which in general means that we now mostly do the rcu_read_lock() and rcu_read_unlock() in the same function. Mostly. We still have the ipc_addid/newary/freeary mess, and things like ipcctl_pre_down_nolock(). Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: sem_putref() does not need the semaphore lock any moreLinus Torvalds
ipc_rcu_putref() uses atomics for the refcount, and the games to lock and unlock the semaphore just to try to keep the reference counting working are no longer useful. Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-04ipc: move rcu_read_unlock() out of sem_unlock() and into callersLinus Torvalds
The IPC locking is a mess, and sem_unlock() unlocks not only the semaphore spinlock, it also drops the rcu read lock. Unlike sem_lock(), which just gets the spin-lock, and expects the caller to get the rcu read lock. This all makes things very hard to follow, and it's very confusing when you take the rcu read lock in one function, and then release it in another. And it has caused actual bugs: the sem_obtain_lock() function ended up dropping the RCU read lock twice in one error path, because it first did the sem_unlock(), and then did a rcu_read_unlock() to match the rcu_read_lock() it had done. This is just a totally mindless "remove rcu_read_unlock() from sem_unlock() and add it immediately after each caller" (except for the aforementioned bug where we did too many rcu_read_unlock(), and in find_alloc_undo() where we just got the rcu_read_lock() to correct for the fact that sem_unlock would immediately drop it again). We can (and should) clean things up further, but this fixes the bug with the minimal amount of subtlety. Reviewed-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Rik van Riel <riel@redhat.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-02ipc: fix GETALL/IPC_RM race for sysv semaphoresAl Viro
We can step on WARN_ON_ONCE() in sem_getref() if a semaphore is removed just as we are about to call sem_getref() from semctl_main(); results are not pretty. We should fail with -EIDRM, same as if IPC_RM happened while we'd been doing allocation there. This also expands sem_getref() at its only callsite (and fixed there), while sem_getref_and_unlock() is simply killed off - it has no callers at all. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-01ipc,sem: fine grained locking for semtimedopRik van Riel
Introduce finer grained locking for semtimedop, to handle the common case of a program wanting to manipulate one semaphore from an array with multiple semaphores. If the call is a semop manipulating just one semaphore in an array with multiple semaphores, only take the lock for that semaphore itself. If the call needs to manipulate multiple semaphores, or another caller is in a transaction that manipulates multiple semaphores, the sem_array lock is taken, as well as all the locks for the individual semaphores. On a 24 CPU system, performance numbers with the semop-multi test with N threads and N semaphores, look like this: vanilla Davidlohr's Davidlohr's + Davidlohr's + threads patches rwlock patches v3 patches 10 610652 726325 1783589 2142206 20 341570 365699 1520453 1977878 30 288102 307037 1498167 2037995 40 290714 305955 1612665 2256484 50 288620 312890 1733453 2650292 60 289987 306043 1649360 2388008 70 291298 306347 1723167 2717486 80 290948 305662 1729545 2763582 90 290996 306680 1736021 2757524 100 292243 306700 1773700 3059159 [davidlohr.bueso@hp.com: do not call sem_lock when bogus sma] [davidlohr.bueso@hp.com: make refcounter atomic] Signed-off-by: Rik van Riel <riel@redhat.com> Suggested-by: Linus Torvalds <torvalds@linux-foundation.org> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Chegu Vinod <chegu_vinod@hp.com> Cc: Jason Low <jason.low2@hp.com> Reviewed-by: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Stanislav Kinsbursky <skinsbursky@parallels.com> Tested-by: Emmanuel Benisty <benisty.e@gmail.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-01ipc,sem: have only one list in struct sem_queueRik van Riel
Having only one list in struct sem_queue, and only queueing simple semaphore operations on the list for the semaphore involved, allows us to introduce finer grained locking for semtimedop. Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Chegu Vinod <chegu_vinod@hp.com> Cc: Emmanuel Benisty <benisty.e@gmail.com> Cc: Jason Low <jason.low2@hp.com> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Stanislav Kinsbursky <skinsbursky@parallels.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-01ipc,sem: open code and rename sem_lockRik van Riel
Rename sem_lock() to sem_obtain_lock(), so we can introduce a sem_lock() later that only locks the sem_array and does nothing else. Open code the locking from ipc_lock() in sem_obtain_lock() so we can introduce finer grained locking for the sem_array in the next patch. [akpm@linux-foundation.org: propagate the ipc_obtain_object() errno out of sem_obtain_lock()] Signed-off-by: Rik van Riel <riel@redhat.com> Acked-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Chegu Vinod <chegu_vinod@hp.com> Cc: Emmanuel Benisty <benisty.e@gmail.com> Cc: Jason Low <jason.low2@hp.com> Cc: Michel Lespinasse <walken@google.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Stanislav Kinsbursky <skinsbursky@parallels.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-05-01ipc,sem: do not hold ipc lock more than necessaryDavidlohr Bueso
Instead of holding the ipc lock for permissions and security checks, among others, only acquire it when necessary. Some numbers.... 1) With Rik's semop-multi.c microbenchmark we can see the following results: Baseline (3.9-rc1): cpus 4, threads: 256, semaphores: 128, test duration: 30 secs total operations: 151452270, ops/sec 5048409 + 59.40% a.out [kernel.kallsyms] [k] _raw_spin_lock + 6.14% a.out [kernel.kallsyms] [k] sys_semtimedop + 3.84% a.out [kernel.kallsyms] [k] avc_has_perm_flags + 3.64% a.out [kernel.kallsyms] [k] __audit_syscall_exit + 2.06% a.out [kernel.kallsyms] [k] copy_user_enhanced_fast_string + 1.86% a.out [kernel.kallsyms] [k] ipc_lock With this patchset: cpus 4, threads: 256, semaphores: 128, test duration: 30 secs total operations: 273156400, ops/sec 9105213 + 18.54% a.out [kernel.kallsyms] [k] _raw_spin_lock + 11.72% a.out [kernel.kallsyms] [k] sys_semtimedop + 7.70% a.out [kernel.kallsyms] [k] ipc_has_perm.isra.21 + 6.58% a.out [kernel.kallsyms] [k] avc_has_perm_flags + 6.54% a.out [kernel.kallsyms] [k] __audit_syscall_exit + 4.71% a.out [kernel.kallsyms] [k] ipc_obtain_object_check 2) While on an Oracle swingbench DSS (data mining) workload the improvements are not as exciting as with Rik's benchmark, we can see some positive numbers. For an 8 socket machine the following are the percentages of %sys time incurred in the ipc lock: Baseline (3.9-rc1): 100 swingbench users: 8,74% 400 swingbench users: 21,86% 800 swingbench users: 84,35% With this patchset: 100 swingbench users: 8,11% 400 swingbench users: 19,93% 800 swingbench users: 77,69% [riel@redhat.com: fix two locking bugs] [sasha.levin@oracle.com: prevent releasing RCU read lock twice in semctl_main] [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com> Signed-off-by: Rik van Riel <riel@redhat.com> Reviewed-by: Chegu Vinod <chegu_vinod@hp.com> Acked-by: Michel Lespinasse <walken@google.com> Cc: Rik van Riel <riel@redhat.com> Cc: Jason Low <jason.low2@hp.com> Cc: Emmanuel Benisty <benisty.e@gmail.com> Cc: Peter Hurley <peter@hurleysoftware.com> Cc: Stanislav Kinsbursky <skinsbursky@parallels.com> Tested-by: Sedat Dilek <sedat.dilek@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2013-03-05get rid of union semop in sys_semctl(2) argumentsAl Viro
just have the bugger take unsigned long and deal with SETVAL case (when we use an int member in the union) explicitly. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2013-03-03make HAVE_SYSCALL_WRAPPERS unconditionalAl Viro
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-09-06userns: Convert ipc to use kuid and kgid where appropriateEric W. Biederman
- Store the ipc owner and creator with a kuid - Store the ipc group and the crators group with a kgid. - Add error handling to ipc_update_perms, allowing it to fail if the uids and gids can not be converted to kuids or kgids. - Modify the proc files to display the ipc creator and owner in the user namespace of the opener of the proc file. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
2011-11-02ipc/sem.c: remove private structures from public header fileManfred Spraul
include/linux/sem.h contains several structures that are only used within ipc/sem.c. The patch moves them into ipc/sem.c - there is no need to expose the structures to the whole kernel. No functional changes, only whitespace cleanups and 80-char per line fixes. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-02ipc/sem.c: handle spurious wakeupsManfred Spraul
semtimedop() does not handle spurious wakeups, it returns -EINTR to user space. Most other schedule() users would just loop and not return to user space. The patch adds such a loop to semtimedop() Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Reported-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Mike Galbraith <efault@gmx.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-11-02ipc/sem.c: fix return code race with semop vs. semop +semctl(IPC_RMID)Manfred Spraul
sys_semtimedop() may return -EIDRM although the semaphore operation completed successfully: thread 1: thread 2: semtimedop(), sleeps semop(): * acquires sem_lock() semtimedop() woken up due to timeout sem_lock() loops * notices that thread 2 could be completed. * performs the operations that thread 2 is sleeping on. * marks the semaphore operation as IN_WAKEUP * drops sem_lock(), does wakeup, sets return code to 0 * thread delayed due to interrupt, whatever * returns to user space * thread still delayed semctl(IPC_RMID) * acquires sem_lock() * ipc_rmid(), ipcp->deleted=1 * drops sem_lock() * thread finally continues - but seem_lock() now fails due to ipcp->deleted == 1 * returns -EIDRM instead of 0 The fix is trivial: Always use the return code in queue.status. In real world, the race probably doesn't matter: If the semaphore array is destroyed, the app is probably not interested if the last operation succeeded or was already cancelled. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Mike Galbraith <efault@gmx.de> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-25ipc/sem.c: fix race with concurrent semtimedop() timeouts and IPC_RMIDManfred Spraul
If a semaphore array is removed and in parallel a sleeping task is woken up (signal or timeout, does not matter), then the woken up task does not wait until wake_up_sem_queue_do() is completed. This will cause crashes, because wake_up_sem_queue_do() will read from a stale pointer. The fix is simple: Regardless of anything, always call get_queue_result(). This function waits until wake_up_sem_queue_do() has finished it's task. Addresses https://bugzilla.kernel.org/show_bug.cgi?id=27142 Reported-by: Yuriy Yevtukhov <yuriy@ucoz.com> Reported-by: Harald Laabs <kernel@dasr.de> Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Acked-by: Eric Dumazet <eric.dumazet@gmail.com> Cc: <stable@kernel.org> [2.6.35+] Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2011-07-20ipc,rcu: Convert call_rcu(free_un) to kfree_rcu()Lai Jiangshan
The rcu callback free_un() just calls a kfree(), so we use kfree_rcu() instead of the call_rcu(free_un). Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Manfred Spraul <manfred@colorfullife.com> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
2011-03-31Fix common misspellingsLucas De Marchi
Fixes generated by 'codespell' and manually reviewed. Signed-off-by: Lucas De Marchi <lucas.demarchi@profusion.mobi>
2011-03-23userns: user namespaces: convert several capable() callsSerge E. Hallyn
CAP_IPC_OWNER and CAP_IPC_LOCK can be checked against current_user_ns(), because the resource comes from current's own ipc namespace. setuid/setgid are to uids in own namespace, so again checks can be against current_user_ns(). Changelog: Jan 11: Use task_ns_capable() in place of sched_capable(). Jan 11: Use nsown_capable() as suggested by Bastian Blank. Jan 11: Clarify (hopefully) some logic in futex and sched.c Feb 15: use ns_capable for ipc, not nsown_capable Feb 23: let copy_ipcs handle setting ipc_ns->user_ns Feb 23: pass ns down rather than taking it from current [akpm@linux-foundation.org: coding-style fixes] Signed-off-by: Serge E. Hallyn <serge.hallyn@canonical.com> Acked-by: "Eric W. Biederman" <ebiederm@xmission.com> Acked-by: Daniel Lezcano <daniel.lezcano@free.fr> Acked-by: David Howells <dhowells@redhat.com> Cc: James Morris <jmorris@namei.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-10-01sys_semctl: fix kernel stack leakageDan Rosenberg
The semctl syscall has several code paths that lead to the leakage of uninitialized kernel stack memory (namely the IPC_INFO, SEM_INFO, IPC_STAT, and SEM_STAT commands) during the use of the older, obsolete version of the semid_ds struct. The copy_semid_to_user() function declares a semid_ds struct on the stack and copies it back to the user without initializing or zeroing the "sem_base", "sem_pending", "sem_pending_last", and "undo" pointers, allowing the leakage of 16 bytes of kernel stack memory. The code is still reachable on 32-bit systems - when calling semctl() newer glibc's automatically OR the IPC command with the IPC_64 flag, but invoking the syscall directly allows users to use the older versions of the struct. Signed-off-by: Dan Rosenberg <dan.j.rosenberg@gmail.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-07-20ipc/sem.c: bugfix for semop() not reporting successful operationManfred Spraul
The last change to improve the scalability moved the actual wake-up out of the section that is protected by spin_lock(sma->sem_perm.lock). This means that IN_WAKEUP can be in queue.status even when the spinlock is acquired by the current task. Thus the same loop that is performed when queue.status is read without the spinlock acquired must be performed when the spinlock is acquired. Thanks to kamezawa.hiroyu@jp.fujitsu.com for noticing lack of the memory barrier. Addresses https://bugzilla.kernel.org/show_bug.cgi?id=16255 [akpm@linux-foundation.org: clean up kerneldoc, checkpatch warning and whitespace] Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Reported-by: Luca Tettamanti <kronos.it@gmail.com> Tested-by: Luca Tettamanti <kronos.it@gmail.com> Reported-by: Christoph Lameter <cl@linux-foundation.org> Cc: Maciej Rutecki <maciej.rutecki@gmail.com> Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-27ipc/sem.c: use ERR_CASTJulia Lawall
Use ERR_CAST(x) rather than ERR_PTR(PTR_ERR(x)). The former makes more clear what is the purpose of the operation, which otherwise looks like a no-op. The semantic patch that makes this change is as follows: (http://coccinelle.lip6.fr/) // <smpl> @@ type T; T x; identifier f; @@ T f (...) { <+... - ERR_PTR(PTR_ERR(x)) + x ...+> } @@ expression x; @@ - ERR_PTR(PTR_ERR(x)) + ERR_CAST(x) // </smpl> Signed-off-by: Julia Lawall <julia@diku.dk> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-27ipc/sem.c: update description of the implementationManfred Spraul
ipc/sem.c begins with a 15 year old description about bugs in the initial implementation in Linux-1.0. The patch replaces that with a top level description of the current code. A TODO could be derived from this text: The opengroup man page for semop() does not mandate FIFO. Thus there is no need for a semaphore array list of pending operations. If - this list is removed - the per-semaphore array spinlock is removed (possible if there is no list to protect) - sem_otime is moved into the semaphores and calculated on demand during semctl() then the array would be read-mostly - which would significantly improve scaling for applications that use semaphore arrays with lots of entries. The price would be expensive semctl() calls: for(i=0;i<sma->sem_nsems;i++) spin_lock(sma->sem_lock); <do stuff> for(i=0;i<sma->sem_nsems;i++) spin_unlock(sma->sem_lock); I'm not sure if the complexity is worth the effort, thus here is the documentation of the current behavior first. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Chris Mason <chris.mason@oracle.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Jens Axboe <jens.axboe@oracle.com> Cc: Nick Piggin <npiggin@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-27ipc/sem.c: move wake_up_process out of the spinlock sectionManfred Spraul
The wake-up part of semtimedop() consists out of two steps: - the right tasks must be identified. - they must be woken up. Right now, both steps run while the array spinlock is held. This patch reorders the code and moves the actual wake_up_process() behind the point where the spinlock is dropped. The code also moves setting sem->sem_otime to one place: It does not make sense to set the last modify time multiple times. [akpm@linux-foundation.org: repair kerneldoc] [akpm@linux-foundation.org: fix uninitialised retval] Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Chris Mason <chris.mason@oracle.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Jens Axboe <jens.axboe@oracle.com> Cc: Nick Piggin <npiggin@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2010-05-27ipc/sem.c: optimize update_queue() for bulk wakeup callsManfred Spraul
The following series of patches tries to fix the spinlock contention reported by Chris Mason - his benchmark exposes problems of the current code: - In the worst case, the algorithm used by update_queue() is O(N^2). Bulk wake-up calls can enter this worst case. The patch series fix that. Note that the benchmark app doesn't expose the problem, it just should be fixed: Real world apps might do the wake-ups in another order than perfect FIFO. - The part of the code that runs within the semaphore array spinlock is significantly larger than necessary. The patch series fixes that. This change is responsible for the main improvement. - The cacheline with the spinlock is also used for a variable that is read in the hot path (sem_base) and for a variable that is unnecessarily written to multiple times (sem_otime). The last step of the series cacheline-aligns the spinlock. This patch: The SysV semaphore code allows to perform multiple operations on all semaphores in the array as atomic operations. After a modification, update_queue() checks which of the waiting tasks can complete. The algorithm that is used to identify the tasks is O(N^2) in the worst case. For some cases, it is simple to avoid the O(N^2). The patch adds a detection logic for some cases, especially for the case of an array where all sleeping tasks are single sembuf operations and a multi-sembuf operation is used to wake up multiple tasks. A big database application uses that approach. The patch fixes wakeup due to semctl(,,SETALL,) - the initial version of the patch breaks that. [akpm@linux-foundation.org: make do_smart_update() static] Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Chris Mason <chris.mason@oracle.com> Cc: Zach Brown <zach.brown@oracle.com> Cc: Jens Axboe <jens.axboe@oracle.com> Cc: Nick Piggin <npiggin@suse.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16ipc: remove unreachable code in sem.cAmerigo Wang
This line is unreachable, remove it. [akpm@linux-foundation.org: remove unneeded initialisation of `err'] Signed-off-by: WANG Cong <amwang@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16ipc/sem.c: optimize single sops when semval is zeroManfred Spraul
If multiple simple decrements on the same semaphore are pending, then the current code scans all decrement operations, even if the semaphore value is already 0. The patch optimizes that: if the semaphore value is 0, then there is no need to scan the q->alter entries. Note that this is a common case: It happens if 100 decrements by one are pending and now an increment by one increases the semaphore value from 0 to 1. Without this patch, all 100 entries are scanned. With the patch, only one entry is scanned, then woken up. Then the new rule triggers and the scanning is aborted, without looking at the remaining 99 tasks. With this patch, single sop increment/decrement by 1 are now O(1). (same as with Nick's patch) Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Pierre Peiffer <peifferp@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16ipc/sem.c: optimize single semop operationsManfred Spraul
sysv sem has the concept of semaphore arrays that consist out of multiple semaphores. Atomic operations that affect multiple semaphores are supported. The patch optimizes single semaphore operation calls that affect only one semaphore: It's not necessary to scan all pending operations, it is sufficient to scan the per-semaphore list. The idea is from Nick Piggin version of an ipc sem improvement, the implementation is different: The code tries to keep as much common code as possible. As the result, the patch is simpler, but optimizes fewer cases. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Pierre Peiffer <peifferp@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16ipc/sem.c: add a per-semaphore pending listManfred Spraul
Based on Nick's findings: sysv sem has the concept of semaphore arrays that consist out of multiple semaphores. Atomic operations that affect multiple semaphores are supported. The patch is the first step for optimizing simple, single semaphore operations: In addition to the global list of all pending operations, a 2nd, per-semaphore list with the simple operations is added. Note: this patch does not make sense by itself, the new list is used nowhere. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Pierre Peiffer <peifferp@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-16ipc/sem.c: optimize if semops failManfred Spraul
Reduce the amount of scanning of the list of pending semaphore operations: If try_atomic_semop failed, then no changes were applied. Thus no need to restart. Additionally, this patch correct an incorrect comment: It's possible to wait for arbitrary semaphore values (do a dec by <x>, wait-for-zero, inc by <x> in one atomic operation) Both changes are from Nick Piggin, the patch is the result of a different split of the individual changes. Signed-off-by: Manfred Spraul <manfred@colorfullife.com> Cc: Nick Piggin <npiggin@suse.de> Cc: Pierre Peiffer <peifferp@gmail.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>