- 04 Nov, 2013 2 commits
-
-
Eric Dumazet authored
[ Upstream commit c9eeec26 ] When TCP Small Queues was added, we used a sysctl to limit amount of packets queues on Qdisc/device queues for a given TCP flow. Problem is this limit is either too big for low rates, or too small for high rates. Now TCP stack has rate estimation in sk->sk_pacing_rate, and TSO auto sizing, it can better control number of packets in Qdisc/device queues. New limit is two packets or at least 1 to 2 ms worth of packets. Low rates flows benefit from this patch by having even smaller number of packets in queues, allowing for faster recovery, better RTT estimations. High rates flows benefit from this patch by allowing more than 2 packets in flight as we had reports this was a limiting factor to reach line rate. [ In particular if TX completion is delayed because of coalescing parameters ] Example for a single flow on 10Gbp link controlled by FQ/pacing 14 packets in flight instead of 2 $ tc -s -d qd qdisc fq 8001: dev eth0 root refcnt 32 limit 10000p flow_limit 100p buckets 1024 quantum 3028 initial_quantum 15140 Sent 1168459366606 bytes 771822841 pkt (dropped 0, overlimits 0 requeues 6822476) rate 9346Mbit 771713pps backlog 953820b 14p requeues 6822476 2047 flow, 2046 inactive, 1 throttled, delay 15673 ns 2372 gc, 0 highprio, 0 retrans, 9739249 throttled, 0 flows_plimit Note that sk_pacing_rate is currently set to twice the actual rate, but this might be refined in the future when a flow is in congestion avoidance. Additional change : skb->destructor should be set to tcp_wfree(). A future patch (for linux 3.13+) might remove tcp_limit_output_bytes Signed-off-by:
Eric Dumazet <edumazet@google.com> Cc: Wei Liu <wei.liu2@citrix.com> Cc: Cong Wang <xiyou.wangcong@gmail.com> Cc: Yuchung Cheng <ycheng@google.com> Cc: Neal Cardwell <ncardwell@google.com> Acked-by:
Neal Cardwell <ncardwell@google.com> Signed-off-by:
David S. Miller <davem@davemloft.net> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Eric Dumazet authored
[ Upstream commits 6d36824e730f247b602c90e8715a792003e3c5a7, 02cf4ebd, and parts of 7eec4174 ] After hearing many people over past years complaining against TSO being bursty or even buggy, we are proud to present automatic sizing of TSO packets. One part of the problem is that tcp_tso_should_defer() uses an heuristic relying on upcoming ACKS instead of a timer, but more generally, having big TSO packets makes little sense for low rates, as it tends to create micro bursts on the network, and general consensus is to reduce the buffering amount. This patch introduces a per socket sk_pacing_rate, that approximates the current sending rate, and allows us to size the TSO packets so that we try to send one packet every ms. This field could be set by other transports. Patch has no impact for high speed flows, where having large TSO packets makes sense to reach line rate. For other flows, this helps better packet scheduling and ACK clocking. This patch increases performance of TCP flows in lossy environments. A new sysctl (tcp_min_tso_segs) is added, to specify the minimal size of a TSO packet (default being 2). A follow-up patch will provide a new packet scheduler (FQ), using sk_pacing_rate as an input to perform optional per flow pacing. This explains why we chose to set sk_pacing_rate to twice the current rate, allowing 'slow start' ramp up. sk_pacing_rate = 2 * cwnd * mss / srtt v2: Neal Cardwell reported a suspect deferring of last two segments on initial write of 10 MSS, I had to change tcp_tso_should_defer() to take into account tp->xmit_size_goal_segs Signed-off-by:
Yuchung Cheng <ycheng@google.com> Acked-by:
-
- 18 Oct, 2013 38 commits
-
-
Greg Kroah-Hartman authored
-
Linn Crosetto authored
commit 30e46b57 upstream. Type SETUP_PCI, added by setup_efi_pci(), may advertise a ROM size larger than early_memremap() is able to handle, which is currently limited to 256kB. If this occurs it leads to a NULL dereference in parse_setup_data(). To avoid this, remap the setup_data header and allow parsing functions for individual types to handle their own data remapping. Signed-off-by:
Linn Crosetto <linn@hp.com> Link: http://lkml.kernel.org/r/1376430401-67445-1-git-send-email-linn@hp.comAcked-by:
Yinghai Lu <yinghai@kernel.org> Reviewed-by:
Pekka Enberg <penberg@kernel.org> Signed-off-by:
H. Peter Anvin <hpa@linux.intel.com> Cc: Paul Gortmaker <paul.gortmaker@windriver.com> Signed-off-by:
-
Davidlohr Bueso authored
commit 4271b05a upstream. This fixes a race in both msgrcv() and msgsnd() between finding the msg and actually dealing with the queue, as another thread can delete shmid underneath us if we are preempted before acquiring the kern_ipc_perm.lock. Manfred illustrates this nicely: Assume a preemptible kernel that is preempted just after msq = msq_obtain_object_check(ns, msqid) in do_msgrcv(). The only lock that is held is rcu_read_lock(). Now the other thread processes IPC_RMID. When the first task is resumed, then it will happily wait for messages on a deleted queue. Fix this by checking for if the queue has been deleted after taking the lock. Signed-off-by:
Davidlohr Bueso <davidlohr@hp.com> Reported-by:
Manfred Spraul <manfred@colorfullife.com> Cc: Rik van Riel <riel@redhat.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> Signed-off-by:
-
Manfred Spraul authored
commit 0e8c6656 upstream. In commit 0a2b9d4c ("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:
Jia He <jiakernel@gmail.com> Tested-by:
-
Manfred Spraul authored
commit d8c63376 upstream. 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 authored
commit 6d07b68c upstream. 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 authored
commit 5e9d5275 upstream. 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:
-
Davidlohr Bueso authored
commit 53dad6d3 upstream. 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:
-
Davidlohr Bueso authored
commit 20b8875a upstream. No remaining users, we now use ipc_obtain_object_check(). Signed-off-by:
Davidlohr Bueso <davidlohr.bueso@hp.com> Cc: Sedat Dilek <sedat.dilek@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by:
-
Davidlohr Bueso authored
commit 7a25dd9e upstream. This function was replaced by a the lockless shm_obtain_object_check(), and no longer has any users. Signed-off-by:
-
Davidlohr Bueso authored
commit 32a27500 upstream. After previous cleanups and optimizations, this function is no longer heavily used and we don't have a good reason to keep it. Update the few remaining callers and get rid of it. Signed-off-by:
-
Davidlohr Bueso authored
commit 530fcd16 upstream. When !CONFIG_MMU there's a chance we can derefence a NULL pointer when the VM area isn't found - check the return value of find_vma(). Also, remove the redundant -EINVAL return: retval is set to the proper return code and *only* changed to 0, when we actually unmap the segments. Signed-off-by:
-
Davidlohr Bueso authored
commit 05603c44 upstream. As suggested by Andrew, add a generic initial locking scheme used throughout all sysv ipc mechanisms. Documenting the ids rwsem, how rcu can be enough to do the initial checks and when to actually acquire the kern_ipc_perm.lock spinlock. I found that adding it to util.c was generic enough. Signed-off-by:
Sedat Dilek <sedat.dilek@gmail.com> Cc: Rik van Riel <riel@redhat.com> Cc: Manfred Spraul <manfred@colorfullife.com> Signed-off-by:
-
Davidlohr Bueso authored
commit 4718787d upstream. There is only one user left, drop this function and just call ipc_unlock_object() and rcu_read_unlock(). Signed-off-by:
-
Davidlohr Bueso authored
commit d9a605e4 upstream. 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 authored
commit c2c737a0 upstream. Similar to other system calls, acquire the kern_ipc_perm lock after doing the initial permission and security checks. [sasha.levin@oracle.com: dont leave do_shmat with rcu lock held] Signed-off-by:
Sasha Levin <sasha.levin@oracle.com> Signed-off-by:
-
Davidlohr Bueso authored
commit f42569b1 upstream. Clean up some of the messy do_shmat() spaghetti code, getting rid of out_free and out_put_dentry labels. This makes shortening the critical region of this function in the next patch a little easier to do and read. Signed-off-by:
-
Davidlohr Bueso authored
commit 2caacaa8 upstream. With the *_INFO, *_STAT, IPC_RMID and IPC_SET commands already optimized, deal with the remaining SHM_LOCK and SHM_UNLOCK commands. Take the shm_perm lock after doing the initial auditing and security checks. The rest of the logic remains unchanged. Signed-off-by:
-
Davidlohr Bueso authored
commit c97cb9cc upstream. While the INFO cmd doesn't take the ipc lock, the STAT commands do acquire it unnecessarily. We can do the permissions and security checks only holding the rcu lock. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
-
Davidlohr Bueso authored
commit 68eccc1d upstream. Similar to semctl and msgctl, when calling msgctl, the *_INFO and *_STAT commands can be performed without acquiring the ipc object. Add a shmctl_nolock() function and move the logic of *_INFO and *_STAT out of msgctl(). Since we are just moving functionality, this change still takes the lock and it will be properly lockless in the next patch. Signed-off-by:
-
Davidlohr Bueso authored
commit 3b1c4ad3 upstream. Now that sem, msgque and shm, through *_down(), all use the lockless variant of ipcctl_pre_down(), go ahead and delete it. [akpm@linux-foundation.org: fix function name in kerneldoc, cleanups] Signed-off-by:
-
Davidlohr Bueso authored
commit 79ccf0f8 upstream. Instead of holding the ipc lock for the entire function, use the ipcctl_pre_down_nolock and only acquire the lock for specific commands: RMID and SET. Signed-off-by:
-
Davidlohr Bueso authored
commit 8b8d52ac upstream. This is the third and final patchset that deals with reducing the amount of contention we impose on the ipc lock (kern_ipc_perm.lock). These changes mostly deal with shared memory, previous work has already been done for semaphores and message queues: http://lkml.org/lkml/2013/3/20/546 (sems) http://lkml.org/lkml/2013/5/15/584 (mqueues) With these patches applied, a custom shm microbenchmark stressing shmctl doing IPC_STAT with 4 threads a million times, reduces the execution time by 50%. A similar run, this time with IPC_SET, reduces the execution time from 3 mins and 35 secs to 27 seconds. Patches 1-8: replaces blindly taking the ipc lock for a smarter combination of rcu and ipc_obtain_object, only acquiring the spinlock when updating. Patch 9: renames the ids rw_mutex to rwsem, which is what it already was. Patch 10: is a trivial mqueue leftover cleanup Patch 11: adds a brief lock scheme description, requested by Andrew. This patch: Add shm_obtain_object() and shm_obtain_object_check(), which will allow us to get the ipc object without acquiring the lock. Just as with other forms of ipc, these functions are basically wrappers around ipc_obtain_object*(). Signed-off-by:
-
Manfred Spraul authored
commit bebcb928 upstream. The check if the queue is full and adding current to the wait queue of pending msgsnd() operations (ss_add()) must be atomic. Otherwise: - the thread that performs msgsnd() finds a full queue and decides to sleep. - the thread that performs msgrcv() first reads all messages from the queue and then sleeps, because the queue is empty. - the msgrcv() calls do not perform any wakeups, because the msgsnd() task has not yet called ss_add(). - then the msgsnd()-thread first calls ss_add() and then sleeps. Net result: msgsnd() and msgrcv() both sleep forever. Observed with msgctl08 from ltp with a preemptible kernel. Fix: Call ipc_lock_object() before performing the check. The patch also moves security_msg_queue_msgsnd() under ipc_lock_object: - msgctl(IPC_SET) explicitely mentions that it tries to expunge any pending operations that are not allowed anymore with the new permissions. If security_msg_queue_msgsnd() is called without locks, then there might be races. - it makes the patch much simpler. Reported-and-tested-by:
Vineet Gupta <Vineet.Gupta1@synopsys.com> Acked-by:
Rik van Riel <riel@redhat.com> Signed-off-by:
-
Manfred Spraul authored
commit 758a6ba3 upstream. 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 authored
commit d12e1e50 upstream. 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 authored
commit f269f40a upstream. 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 authored
commit 1a82e9e1 upstream. 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 authored
commit f5c936c0 upstream. 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 authored
commit 196aa013 upstream. Enforce that ipc_rcu_alloc returns a cacheline aligned pointer on SMP. Rationale: The SysV sem code tries to move the main spinlock into a seperate cacheline (____cacheline_aligned_in_smp). This works only if ipc_rcu_alloc returns cacheline aligned pointers. vmalloc and kmalloc return cacheline algined pointers, the implementation of ipc_rcu_alloc breaks that. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
-
Davidlohr Bueso authored
commit 9ad66ae6 upstream. 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 authored
commit 41a0d523 upstream. do_msgrcv() is the last msg queue function that abuses the ipc lock Take it only when needed when actually updating msq. Signed-off-by:
-
Davidlohr Bueso authored
commit 3dd1f784 upstream. do_msgsnd() is another function that does too many things with the ipc object lock acquired. Take it only when needed when actually updating msq. Signed-off-by:
-
Davidlohr Bueso authored
commit ac0ba20e upstream. While the INFO cmd doesn't take the ipc lock, the STAT commands do acquire it unnecessarily. We can do the permissions and security checks only holding the rcu lock. This function now mimics semctl_nolock(). Signed-off-by:
-
Davidlohr Bueso authored
commit a5001a0d upstream. Add msq_obtain_object() and msq_obtain_object_check(), which will allow us to get the ipc object without acquiring the lock. Just as with semaphores, these functions are basically wrappers around ipc_obtain_object*(). Signed-off-by:
-
Davidlohr Bueso authored
commit 2cafed30 upstream. Similar to semctl, when calling msgctl, the *_INFO and *_STAT commands can be performed without acquiring the ipc object. Add a msgctl_nolock() function and move the logic of *_INFO and *_STAT out of msgctl(). This change still takes the lock and it will be properly lockless in the next patch Signed-off-by:
-
Davidlohr Bueso authored
commit 15724ecb upstream. Instead of holding the ipc lock for the entire function, use the ipcctl_pre_down_nolock and only acquire the lock for specific commands: RMID and SET. Signed-off-by:
-
Davidlohr Bueso authored
commit 7b4cc5d8 upstream. 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:
-
