ipc,sem: sysv semaphore scalability
Include lkml in the CC: this time... *sigh*
---8<---
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
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
PATCH 1/7 ipc: remove bogus lock comment for ipc_checkid
From: Davidlohr Bueso <[email protected]>
There is no reason to be holding the ipc lock while reading ipcp->seq, hence remove misleading comment.
Also simplify the return value for the function.
--- ipc/util.h | 7 +------ 1 files changed, 1 insertions(+), 6 deletions(-)
ipc/util.h +1 -6
@@ -150,14 +150,9 @@ static inline int ipc_buildid(int id, int seq)return SEQ_MULTIPLIER * seq + id; } -/* -* Must be called with ipcp locked -*/ static inline int ipc_checkid (struct kern_ipc_perm *ipcp,int uid) { -if (uid /SEQ_MULTIPLIER != ipcp->seq) -return 1 ; -return 0 ; +return uid /SEQ_MULTIPLIER != ipcp->seq; }static inline void ipc_lock_by_ptr (struct kern_ipc_perm *perm)
Signature · 1 line
1.7.7.6
PATCH 2/7 ipc: introduce obtaining a lockless ipc object
In reply to Rik van RielFrom: Davidlohr Bueso <[email protected]>
Through ipc_lock() and therefore ipc_lock_check() we currently return the locked ipc object. This is not necessary for all situations an can, therefore, incur in unnecessary ipc lock contention.
Introduce, analogous, ipc_obtain_object() and ipc_obtain_object_check() functions that only lookup and return the ipc object.
Both these functions must be called within the RCU read critical section.
--- ipc/util.c | 71 +++++++++++++++++++++++++++++++++++++++++++++++++----------- ipc/util.h | 2 + 2 files changed, 60 insertions(+), 13 deletions(-)
ipc/util.c +58 -13
@@ -668,6 +668,28 @@ void ipc64_perm_to_ipc_perm (struct ipc64_perm *in, struct ipc_perm *out) }/** +* ipc_obtain_object +* @ids: ipc identifier set +* @id: ipc id to look for +* +* Look for an id in the ipc ids idr and return associated ipc object. +* +* Call inside the RCU critical section. +* The ipc object is *not* locked on exit. +*/ +struct kern_ipc_perm *ipc_obtain_object (struct ipc_ids *ids,int id) +{ +struct kern_ipc_perm *out; +int lid =ipcid_to_idx (id); + + out =idr_find (&ids->ipcs_idr, lid); +if (!out) +return ERR_PTR (-EINVAL ); + +return out; +} + +/** * ipc_lock - Lock an ipc structure without rw_mutex held * @ids: IPC identifier set * @id: ipc id to look for
@@ -680,27 +702,50 @@ void ipc64_perm_to_ipc_perm (struct ipc64_perm *in, struct ipc_perm *out)struct kern_ipc_perm *ipc_lock (struct ipc_ids *ids,int id) {struct kern_ipc_perm *out; -int lid =ipcid_to_idx (id); - +rcu_read_lock (); - out =idr_find (&ids->ipcs_idr, lid); -if (out ==NULL ) { -rcu_read_unlock (); -return ERR_PTR (-EINVAL ); - } + out =ipc_obtain_object (ids, id); +if (IS_ERR (out)) + goto err1;spin_lock (&out->lock); - +/* ipc_rmid() may have already freed the ID while ipc_lock * was spinning: here verify that the structure is still valid */ -if (out->deleted) { -spin_unlock (&out->lock); -rcu_read_unlock (); -return ERR_PTR (-EINVAL ); - } +if (out->deleted) + goto err0;return out; +err0: +spin_unlock (&out->lock); +err1: +rcu_read_unlock (); +return ERR_PTR (-EINVAL ); +} + +/** +* ipc_obtain_object_check +* @ids: ipc identifier set +* @id: ipc id to look for +* +* Similar to ipc_obtain_object() but also checks +* the ipc object reference counter. +* +* Call inside the RCU critical section. +* The ipc object is *not* locked on exit. +*/ +struct kern_ipc_perm *ipc_obtain_object_check (struct ipc_ids *ids,int id) +{ +struct kern_ipc_perm *out =ipc_obtain_object (ids, id); + +if (IS_ERR (out)) + goto out; + +if (ipc_checkid (out, id)) +return ERR_PTR (-EIDRM ); +out: +return out; }struct kern_ipc_perm *ipc_lock_check (struct ipc_ids *ids,int id)
ipc/util.h +2 -0
@@ -123,6 +123,7 @@ void ipc_rcu_getref(void *ptr);void ipc_rcu_putref (void *ptr);struct kern_ipc_perm *ipc_lock (struct ipc_ids *,int ); +struct kern_ipc_perm *ipc_obtain_object (struct ipc_ids *ids,int id);void kernel_to_ipc64_perm (struct kern_ipc_perm *in,struct ipc64_perm *out);void ipc64_perm_to_ipc_perm (struct ipc64_perm *in,struct ipc_perm *out);
@@ -168,6 +169,7 @@ static inline void ipc_unlock(struct kern_ipc_perm *perm) }struct kern_ipc_perm *ipc_lock_check (struct ipc_ids *ids,int id); +struct kern_ipc_perm *ipc_obtain_object_check (struct ipc_ids *ids,int id);int ipcget (struct ipc_namespace *ns,struct ipc_ids *ids,struct ipc_ops *ops,struct ipc_params *params); void free_ipcs(struct ipc_namespace *ns, struct ipc_ids *ids,
Signature · 1 line
1.7.7.6
PATCH 3/7 ipc: introduce lockless pre_down ipcctl
In reply to Rik van RielFrom: Davidlohr Bueso <[email protected]>
Various forms of ipc use the ipcctl_pre_down() function to retrieve an ipc object and check permissions, mostly for IPC_RMID and IPC_SET commands.
Introduce ipcctl_pre_down_nolock(), a lockless version of this function. The locking version is maintained, yet modified to call the nolock version, without affecting its semantics, thus transparent to all ipc callers.
--- ipc/util.c | 31 ++++++++++++++++++++++++++----- ipc/util.h | 3 +++ 2 files changed, 29 insertions(+), 5 deletions(-)
ipc/util.c +26 -5
@@ -825,11 +825,28 @@ struct kern_ipc_perm *ipcctl_pre_down(struct ipc_namespace *ns, struct ipc64_perm *perm, int extra_perm) {struct kern_ipc_perm *ipcp; + + ipcp =ipcctl_pre_down_nolock (ns, ids, id, cmd, perm, extra_perm); +if (IS_ERR (ipcp)) + goto out; + +spin_lock (&ipcp->lock); +out: +return ipcp; +} + +struct kern_ipc_perm *ipcctl_pre_down_nolock (struct ipc_namespace *ns, +struct ipc_ids *ids,int id,int cmd, +struct ipc64_perm *perm,int extra_perm) +{kuid_t euid; -int err; +int err = -EPERM ; +struct kern_ipc_perm *ipcp;down_write (&ids->rw_mutex); - ipcp =ipc_lock_check (ids, id); +rcu_read_lock (); + + ipcp =ipc_obtain_object_check (ids, id);if (IS_ERR (ipcp)) { err =PTR_ERR (ipcp); goto out_up;
@@ -838,17 +855,21 @@ struct kern_ipc_perm *ipcctl_pre_down(struct ipc_namespace *ns,audit_ipc_obj (ipcp);if (cmd ==IPC_SET )audit_ipc_set_perm (extra_perm, perm->uid, - perm->gid, perm->mode); + perm->gid, perm->mode); euid =current_euid ();if (uid_eq (euid, ipcp->cuid) ||uid_eq (euid, ipcp->uid) ||ns_capable (ns->user_ns,CAP_SYS_ADMIN ))return ipcp; - err = -EPERM ; -ipc_unlock (ipcp); out_up: +/* +* Unsuccessful lookup, unlock and return +* the corresponding error. +*/ +rcu_read_unlock ();up_write (&ids->rw_mutex); +return ERR_PTR (err); }
ipc/util.h +3 -0
@@ -128,6 +128,9 @@ struct kern_ipc_perm *ipc_obtain_object(struct ipc_ids *ids, int id);void kernel_to_ipc64_perm (struct kern_ipc_perm *in,struct ipc64_perm *out);void ipc64_perm_to_ipc_perm (struct ipc64_perm *in,struct ipc_perm *out);int ipc_update_perm (struct ipc64_perm *in,struct kern_ipc_perm *out); +struct kern_ipc_perm *ipcctl_pre_down_nolock (struct ipc_namespace *ns, +struct ipc_ids *ids,int id,int cmd, +struct ipc64_perm *perm,int extra_perm);struct kern_ipc_perm *ipcctl_pre_down (struct ipc_namespace *ns,struct ipc_ids *ids,int id,int cmd,struct ipc64_perm *perm,int extra_perm);
Signature · 1 line
1.7.7.6
PATCH 4/7 ipc,sem: do not hold ipc lock more than necessary
In reply to Rik van RielFrom: Davidlohr Bueso <[email protected]>
I have corrected the two locking bugs in this patch, but left it otherwise untouched - Rik.
---8<---
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%
--- ipc/sem.c | 155 ++++++++++++++++++++++++++++++++++++++++++------------------ ipc/util.h | 5 ++ 2 files changed, 114 insertions(+), 46 deletions(-)
ipc/sem.c +109 -46
@@ -204,13 +204,34 @@ static inline struct sem_array *sem_lock(struct ipc_namespace *ns, int id) return container_of(ipcp, struct sem_array, sem_perm); } +static inline struct sem_array *sem_obtain_object (struct ipc_namespace *ns,int id) +{ +struct kern_ipc_perm *ipcp =ipc_obtain_object (&sem_ids (ns), id); + +if (IS_ERR (ipcp)) +return ERR_CAST (ipcp); + + return container_of(ipcp, struct sem_array, sem_perm); +} +static inline struct sem_array *sem_lock_check (struct ipc_namespace *ns,int id) {struct kern_ipc_perm *ipcp =ipc_lock_check (&sem_ids (ns), id);if (IS_ERR (ipcp)) -return (struct sem_array *)ipcp; +return ERR_CAST (ipcp); + + return container_of(ipcp, struct sem_array, sem_perm); +} + +static inline struct sem_array *sem_obtain_object_check (struct ipc_namespace *ns, +int id) +{ +struct kern_ipc_perm *ipcp =ipc_obtain_object_check (&sem_ids (ns), id); + +if (IS_ERR (ipcp)) +return ERR_CAST (ipcp); return container_of(ipcp, struct sem_array, sem_perm); }
@@ -234,6 +255,16 @@ static inline void sem_putref(struct sem_array *sma)ipc_unlock (&(sma)->sem_perm); } +/* +* Call inside the rcu read section. +*/ +static inline void sem_getref (struct sem_array *sma) +{ +spin_lock (&(sma)->sem_perm.lock); +ipc_rcu_getref (sma); +ipc_unlock (&(sma)->sem_perm); +} +static inline void sem_rmid (struct ipc_namespace *ns,struct sem_array *s) {ipc_rmid (&sem_ids (ns), &s->sem_perm);
@@ -842,18 +873,25 @@ static int semctl_nolock(struct ipc_namespace *ns, int semid,case SEM_STAT : {struct semid64_ds tbuf; -int id; +int id =0 ; + +memset (&tbuf,0 ,sizeof (tbuf));if (cmd ==SEM_STAT ) { - sma =sem_lock (ns, semid); -if (IS_ERR (sma)) -return PTR_ERR (sma); +rcu_read_lock (); + sma =sem_obtain_object (ns, semid); +if (IS_ERR (sma)) { + err =PTR_ERR (sma); + goto out_unlock; + } id = sma->sem_perm.id; }else { - sma =sem_lock_check (ns, semid); -if (IS_ERR (sma)) -return PTR_ERR (sma); - id =0 ; +rcu_read_lock (); + sma =sem_obtain_object_check (ns, semid); +if (IS_ERR (sma)) { + err =PTR_ERR (sma); + goto out_unlock; + } } err = -EACCES ;
@@ -864,13 +902,11 @@ static int semctl_nolock(struct ipc_namespace *ns, int semid,if (err) goto out_unlock; -memset (&tbuf,0 ,sizeof (tbuf)); -kernel_to_ipc64_perm (&sma->sem_perm, &tbuf.sem_perm); tbuf.sem_otime = sma->sem_otime; tbuf.sem_ctime = sma->sem_ctime; tbuf.sem_nsems = sma->sem_nsems; -sem_unlock (sma); +rcu_read_unlock ();if (copy_semid_to_user (arg.buf, &tbuf, version))return -EFAULT ;return id;
@@ -879,7 +915,7 @@ static int semctl_nolock(struct ipc_namespace *ns, int semid,return -EINVAL ; } out_unlock: -sem_unlock (sma); +rcu_read_unlock ();return err; }
@@ -888,27 +924,34 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, {struct sem_array *sma;struct sem * curr; -int err; +int err, nsems;ushort fast_sem_io[SEMMSL_FAST ];ushort * sem_io = fast_sem_io; -int nsems;struct list_head tasks; - sma =sem_lock_check (ns, semid); -if (IS_ERR (sma)) +INIT_LIST_HEAD (&tasks); + +rcu_read_lock (); + sma =sem_obtain_object_check (ns, semid); +if (IS_ERR (sma)) { +rcu_read_unlock ();return PTR_ERR (sma); + } -INIT_LIST_HEAD (&tasks); nsems = sma->sem_nsems; err = -EACCES ;if (ipcperms (ns, &sma->sem_perm, - (cmd ==SETVAL || cmd ==SETALL ) ?S_IWUGO :S_IRUGO )) - goto out_unlock; + (cmd ==SETVAL || cmd ==SETALL ) ?S_IWUGO :S_IRUGO )) { +rcu_read_unlock (); + goto out_wakeup; + } err =security_sem_semctl (sma, cmd); -if (err) - goto out_unlock; +if (err) { +rcu_read_unlock (); + goto out_wakeup; + } err = -EACCES ;switch (cmd) {
@@ -918,7 +961,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,int i;if (nsems >SEMMSL_FAST ) { -sem_getref_and_unlock (sma); +sem_getref (sma); sem_io =ipc_alloc (sizeof (ushort)*nsems);if (sem_io ==NULL ) {
@@ -934,6 +977,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, } } +spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) sem_io[i] = sma->sem_base[i].semval;sem_unlock (sma);
@@ -947,7 +991,8 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,int i;struct sem_undo *un; -sem_getref_and_unlock (sma); +ipc_rcu_getref (sma); +rcu_read_unlock ();if (nsems >SEMMSL_FAST ) { sem_io =ipc_alloc (sizeof (ushort)*nsems);
@@ -997,6 +1042,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,if (semnum <0 || semnum >= nsems) goto out_unlock; +spin_lock (&sma->sem_perm.lock); curr = &sma->sem_base[semnum];switch (cmd) {
@@ -1034,10 +1080,11 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, goto out_unlock; } } + out_unlock:sem_unlock (sma); +out_wakeup:wake_up_sem_queue_do (&tasks); - out_free:if (sem_io != fast_sem_io)ipc_free (sem_io,sizeof (ushort)*nsems);
@@ -1088,29 +1135,35 @@ static int semctl_down(struct ipc_namespace *ns, int semid,return -EFAULT ; } - ipcp =ipcctl_pre_down (ns, &sem_ids (ns), semid, cmd, - &semid64.sem_perm,0 ); + ipcp =ipcctl_pre_down_nolock (ns, &sem_ids (ns), semid, cmd, + &semid64.sem_perm,0 );if (IS_ERR (ipcp))return PTR_ERR (ipcp); sma = container_of(ipcp, struct sem_array, sem_perm); err =security_sem_semctl (sma, cmd); -if (err) +if (err) { +rcu_read_unlock (); goto out_unlock; + }switch (cmd){case IPC_RMID : +ipc_lock_object (&sma->sem_perm);freeary (ns, ipcp); goto out_up;case IPC_SET : +ipc_lock_object (&sma->sem_perm); err =ipc_update_perm (&semid64.sem_perm, ipcp);if (err) goto out_unlock; sma->sem_ctime =get_seconds ();break ;default : +rcu_read_unlock (); err = -EINVAL ; + goto out_up; } out_unlock:
@@ -1248,16 +1301,18 @@ static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)spin_unlock (&ulp->lock);if (likely (un!=NULL )) goto out; -rcu_read_unlock ();/* no undo structure around - allocate one. */ /* step 1: figure out the size of the semaphore array */ - sma =sem_lock_check (ns, semid); -if (IS_ERR (sma)) + sma =sem_obtain_object_check (ns, semid); +if (IS_ERR (sma)) { +rcu_read_unlock ();return ERR_CAST (sma); + } nsems = sma->sem_nsems; -sem_getref_and_unlock (sma); +ipc_rcu_getref (sma); +rcu_read_unlock ();/* step 2: allocate new undo structure */ new =kzalloc (sizeof (struct sem_undo ) +sizeof (short )*nsems,GFP_KERNEL );
@@ -1392,7 +1447,8 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,INIT_LIST_HEAD (&tasks); - sma =sem_lock_check (ns, semid); +rcu_read_lock (); + sma =sem_obtain_object_check (ns, semid);if (IS_ERR (sma)) {if (un)rcu_read_unlock ();
@@ -1400,6 +1456,24 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, goto out_free; } + error = -EFBIG ; +if (max >= sma->sem_nsems) { +rcu_read_unlock (); + goto out_wakeup; + } + + error = -EACCES ; +if (ipcperms (ns, &sma->sem_perm, alter ?S_IWUGO :S_IRUGO )) { +rcu_read_unlock (); + goto out_wakeup; + } + + error =security_sem_semop (sma, sops, nsops, alter); +if (error) { +rcu_read_unlock (); + goto out_wakeup; + } +/* * semid identifiers are not unique - find_alloc_undo may have * allocated an undo structure, it was invalidated by an RMID
@@ -1408,6 +1482,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,* "un" itself is guaranteed by rcu. */ error = -EIDRM ; +ipc_lock_object (&sma->sem_perm);if (un) {if (un->semid ==-1 ) {rcu_read_unlock ();
@@ -1425,18 +1500,6 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, } } - error = -EFBIG ; -if (max >= sma->sem_nsems) - goto out_unlock_free; - - error = -EACCES ; -if (ipcperms (ns, &sma->sem_perm, alter ?S_IWUGO :S_IRUGO )) - goto out_unlock_free; - - error =security_sem_semop (sma, sops, nsops, alter); -if (error) - goto out_unlock_free; - error =try_atomic_semop (sma, sops, nsops, un,task_tgid_vnr (current));if (error <=0 ) {if (alter && error ==0 )
@@ -1539,7 +1602,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, out_unlock_free:sem_unlock (sma); - +out_wakeup:wake_up_sem_queue_do (&tasks); out_free:if (sops != fast_sops)
ipc/util.h +5 -0
@@ -171,6 +171,11 @@ static inline void ipc_unlock(struct kern_ipc_perm *perm)rcu_read_unlock (); } +static inline void ipc_lock_object (struct kern_ipc_perm *perm) +{ +spin_lock (&perm->lock); +} +struct kern_ipc_perm *ipc_lock_check (struct ipc_ids *ids,int id);struct kern_ipc_perm *ipc_obtain_object_check (struct ipc_ids *ids,int id); int ipcget(struct ipc_namespace *ns, struct ipc_ids *ids,
Signature · 1 line
1.7.7.6
PATCH 5/7 ipc,sem: open code and rename sem_lock
In reply to Rik van RielRename sem_lock to sem_obtain_lock, so we can introduce a sem_lock function 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.
--- ipc/sem.c | 23 +++++++++++++++++++---- 1 files changed, 19 insertions(+), 4 deletions(-)
ipc/sem.c +19 -4
@@ -194,14 +194,29 @@ void __init sem_init (void)* sem_lock_(check_) routines are called in the paths where the rw_mutex * is not held. */ -static inline struct sem_array *sem_lock (struct ipc_namespace *ns,int id) +static inline struct sem_array *sem_obtain_lock (struct ipc_namespace *ns,int id) { -struct kern_ipc_perm *ipcp =ipc_lock (&sem_ids (ns), id); +struct kern_ipc_perm *ipcp; +rcu_read_lock (); + ipcp =ipc_obtain_object (&sem_ids (ns), id);if (IS_ERR (ipcp)) -return (struct sem_array *)ipcp; + goto err1; + +spin_lock (&ipcp->lock); + +/* ipc_rmid() may have already freed the ID while sem_lock +* was spinning: verify that the structure is still valid +*/ +if (ipcp->deleted) + goto err0; return container_of(ipcp, struct sem_array, sem_perm); +err0: +spin_unlock (&ipcp->lock); +err1: +rcu_read_unlock (); +return ERR_PTR (-EINVAL ); }static inline struct sem_array *sem_obtain_object (struct ipc_namespace *ns,int id)
@@ -1562,7 +1577,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, goto out_free; } - sma =sem_lock (ns, semid); + sma =sem_obtain_lock (ns, semid);/* * Wait until it's guaranteed that no wakeup_sem_queue_do() is ongoing.
Signature · 1 line
1.7.7.6
Re: [PATCH 5/7] ipc,sem: open code and rename sem_lock
Quoted 7 lines · Rik van Riel
Rename
sem_locktosem_obtain_lock, so we can introduce asem_lockfunction later that only locks thesem_arrayand does nothing else.Open code the locking from
ipc_lockinsem_obtain_lock, so we can introduce finer grained locking for thesem_arrayin the next patch.Signed-off-by Rik van Riel <[email protected]>
PATCH 6/7 ipc,sem: have only one list in struct sem_queue
In reply to Rik van RielHaving 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.
--- ipc/sem.c | 65 +++++++++++++++++++++++++++++++----------------------------- 1 files changed, 34 insertions(+), 31 deletions(-)
ipc/sem.c +34 -31
@@ -99,7 +99,6 @@ struct sem {/* One queue for each sleeping process in the system. */ struct sem_queue { -struct list_head simple_list;/* queue of pending operations */ struct list_head list;/* queue of pending operations */ struct task_struct *sleeper;/* this process */ struct sem_undo *undo;/* undo structure */
@@ -517,7 +516,7 @@ static void wake_up_sem_queue_prepare(struct list_head *pt, q->status =IN_WAKEUP ; q->pid = error; -list_add_tail (&q->simple_list, pt); +list_add_tail (&q->list, pt); }/**
@@ -535,7 +534,7 @@ static void wake_up_sem_queue_do(struct list_head *pt)int did_something; did_something = !list_empty (pt); -list_for_each_entry_safe (q, t, pt, simple_list) { +list_for_each_entry_safe (q, t, pt, list) {wake_up_process (q->sleeper);/* q can disappear immediately after writing q->status. */ smp_wmb ();
@@ -548,9 +547,7 @@ static void wake_up_sem_queue_do(struct list_head *pt)static void unlink_queue (struct sem_array *sma,struct sem_queue *q) {list_del (&q->list); -if (q->nsops ==1 ) -list_del (&q->simple_list); -else +if (q->nsops >1 ) sma->complex_count--; }
@@ -603,9 +600,9 @@ static int check_restart(struct sem_array *sma, struct sem_queue *q) }/* * semval is 0. Check if there are wait-for-zero semops. -* They must be the first entries in the per-semaphore simple queue +* They must be the first entries in the per-semaphore queue */ - h = list_first_entry(&curr->sem_pending, struct sem_queue, simple_list); + h = list_first_entry(&curr->sem_pending, struct sem_queue, list);BUG_ON (h->nsops !=1 );BUG_ON (h->sops[0 ].sem_num != q->sops[0 ].sem_num);
@@ -625,8 +622,9 @@ static int check_restart(struct sem_array *sma, struct sem_queue *q)* @pt: list head for the tasks that must be woken up. * * update_queue must be called after a semaphore in a semaphore array -* was modified. If multiple semaphore were modified, then @semnum -* must be set to -1. +* was modified. If multiple semaphores were modified, update_queue must +* be called with semnum = -1, as well as with the number of each modified +* semaphore. * The tasks that must be woken up are added to @pt. The return code * is stored in q->pid. * The function return 1 if at least one semop was completed successfully.
@@ -636,30 +634,19 @@ static int update_queue(struct sem_array *sma, int semnum, struct list_head *pt)struct sem_queue *q;struct list_head *walk;struct list_head *pending_list; -int offset;int semop_completed =0 ; -/* if there are complex operations around, then knowing the semaphore -* that was modified doesn't help us. Assume that multiple semaphores -* were modified. -*/ -if (sma->complex_count) - semnum =-1 ; - -if (semnum ==-1 ) { +if (semnum ==-1 ) pending_list = &sma->sem_pending; - offset = offsetof(struct sem_queue , list); - }else { +else pending_list = &sma->sem_base[semnum].sem_pending; - offset = offsetof(struct sem_queue , simple_list); - } again: walk = pending_list->next;while (walk != pending_list) {int error, restart; - q = (struct sem_queue *)((char *)walk - offset); + q = container_of(walk, struct sem_queue, list); walk = walk->next;/* If we are scanning the single sop, per-semaphore list of
@@ -718,9 +705,18 @@ static void do_smart_update(struct sem_array *sma, struct sembuf *sops, int nsopif (sma->complex_count || sops ==NULL ) {if (update_queue (sma,-1 , pt)) otime =1 ; + } + +if (!sops) { +/* No semops; something special is going on. */ +for (i =0 ; i < sma->sem_nsems; i++) { +if (update_queue (sma, i, pt)) + otime =1 ; + } goto done; } +/* Check the semaphores that were modified. */ for (i =0 ; i < nsops; i++) { if (sops[i].sem_op > 0 || (sops[i].sem_op < 0 &&
@@ -791,6 +787,7 @@ static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)struct sem_queue *q, *tq; struct sem_array *sma = container_of(ipcp, struct sem_array, sem_perm);struct list_head tasks; +int i;/* Free the existing undo structures for this semaphore set. */ assert_spin_locked (&sma->sem_perm.lock);
@@ -809,6 +806,13 @@ static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)unlink_queue (sma, q);wake_up_sem_queue_prepare (&tasks, q, -EIDRM ); } +for (i =0 ; i < sma->sem_nsems; i++) { +struct sem *sem = sma->sem_base + i; +list_for_each_entry_safe (q, tq, &sem->sem_pending, list) { +unlink_queue (sma, q); +wake_up_sem_queue_prepare (&tasks, q, -EIDRM ); + } + }/* Remove the semaphore set from the IDR */ sem_rmid (ns, sma);
@@ -1532,21 +1536,20 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, queue.undo = un; queue.pid =task_tgid_vnr (current); queue.alter = alter; -if (alter) -list_add_tail (&queue.list, &sma->sem_pending); -else -list_add (&queue.list, &sma->sem_pending);if (nsops ==1 ) {struct sem *curr; curr = &sma->sem_base[sops->sem_num];if (alter) -list_add_tail (&queue.simple_list, &curr->sem_pending); +list_add_tail (&queue.list, &curr->sem_pending);else -list_add (&queue.simple_list, &curr->sem_pending); +list_add (&queue.list, &curr->sem_pending); }else { -INIT_LIST_HEAD (&queue.simple_list); +if (alter) +list_add_tail (&queue.list, &sma->sem_pending); +else +list_add (&queue.list, &sma->sem_pending); sma->complex_count++; }
Signature · 1 line
1.7.7.6
Re: [PATCH 6/7] ipc,sem: have only one list in struct sem_queue
Quoted 5 lines · Rik 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 <[email protected]>
PATCH 7/7 ipc,sem: fine grained locking for semtimedop
In reply to Rik van RielIntroduce 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
--- ipc/sem.c | 152 ++++++++++++++++++++++++++++++++++++++++++------------------- 1 files changed, 105 insertions(+), 47 deletions(-)
ipc/sem.c +105 -47
@@ -94,6 +94,7 @@struct sem {int semval;/* current value */ int sempid;/* pid of last operation */ +spinlock_t lock;/* spinlock for fine-grained semtimedop */ struct list_head sem_pending;/* pending single-sop operations */ };
@@ -137,7 +138,6 @@ struct sem_undo_list {#define sem_ids (ns) ((ns)->ids[IPC_SEM_IDS]) -#define sem_unlock (sma) ipc_unlock(&(sma)->sem_perm)#define sem_checkid (sma, semid) ipc_checkid(&sma->sem_perm, semid)static int newary (struct ipc_namespace *,struct ipc_params *);
@@ -190,19 +190,83 @@ void __init sem_init (void) }/* +* If the sem_array contains just one semaphore, or if multiple +* semops are performed in one syscall, or if there are complex +* operations pending, the whole sem_array is locked. +* If one semop is performed on an array with multiple semaphores, +* get a shared lock on the array, and lock the individual semaphore. +* +* Carefully guard against sma->complex_count changing between zero +* and non-zero while we are spinning for the lock. The value of +* sma->complex_count cannot change while we are holding the lock, +* so sem_unlock should be fine. +*/ +static inline int sem_lock (struct sem_array *sma,struct sembuf *sops, +int nsops) +{ +int locknum; +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 (&sma->sem_perm.lock); + goto lock_all; + } + locknum = sops->sem_num; + }else { +int i; +/* Lock the sem_array, and all the semaphore locks */ + lock_all: +spin_lock (&sma->sem_perm.lock); +for (i =0 ; i < sma->sem_nsems; i++) { +struct sem *sem = sma->sem_base + i; +spin_lock (&sem->lock); + } + locknum =-1 ; + } +return locknum; +} + +static inline void sem_unlock (struct sem_array *sma,int locknum) +{ +if (locknum ==-1 ) { +int i; +for (i =0 ; i < sma->sem_nsems; i++) { +struct sem *sem = sma->sem_base + i; +spin_unlock (&sem->lock); + } +spin_unlock (&sma->sem_perm.lock); + }else { +struct sem *sem = sma->sem_base + locknum; +spin_unlock (&sem->lock); + } +rcu_read_unlock (); +} + +/* * sem_lock_(check_) routines are called in the paths where the rw_mutex * is not held. */ -static inline struct sem_array *sem_obtain_lock (struct ipc_namespace *ns,int id) +static inline struct sem_array *sem_obtain_lock (struct ipc_namespace *ns, +int id,struct sembuf *sops,int nsops,int *locknum) {struct kern_ipc_perm *ipcp; +struct sem_array *sma;rcu_read_lock (); ipcp =ipc_obtain_object (&sem_ids (ns), id);if (IS_ERR (ipcp)) goto err1; -spin_lock (&ipcp->lock); + sma = container_of(ipcp, struct sem_array, sem_perm); + *locknum =sem_lock (sma, sops, nsops);/* ipc_rmid() may have already freed the ID while sem_lock * was spinning: verify that the structure is still valid
@@ -210,9 +274,9 @@ static inline struct sem_array *sem_obtain_lock(struct ipc_namespace *ns, int idif (ipcp->deleted) goto err0; - return container_of(ipcp, struct sem_array, sem_perm); +return sma; err0: -spin_unlock (&ipcp->lock); +sem_unlock (sma, *locknum); err1:rcu_read_unlock ();return ERR_PTR (-EINVAL );
@@ -228,17 +292,6 @@ static inline struct sem_array *sem_obtain_object(struct ipc_namespace *ns, int return container_of(ipcp, struct sem_array, sem_perm); } -static inline struct sem_array *sem_lock_check (struct ipc_namespace *ns, -int id) -{ -struct kern_ipc_perm *ipcp =ipc_lock_check (&sem_ids (ns), id); - -if (IS_ERR (ipcp)) -return ERR_CAST (ipcp); - - return container_of(ipcp, struct sem_array, sem_perm); -} -static inline struct sem_array *sem_obtain_object_check (struct ipc_namespace *ns,int id) {
@@ -252,21 +305,21 @@ static inline struct sem_array *sem_obtain_object_check(struct ipc_namespace *nsstatic inline void sem_lock_and_putref (struct sem_array *sma) { -ipc_lock_by_ptr (&sma->sem_perm); +rcu_read_lock (); +sem_lock (sma,NULL ,-1 );ipc_rcu_putref (sma); }static inline void sem_getref_and_unlock (struct sem_array *sma) {ipc_rcu_getref (sma); -ipc_unlock (&(sma)->sem_perm); +sem_unlock (sma,-1 ); }static inline void sem_putref (struct sem_array *sma) { -ipc_lock_by_ptr (&sma->sem_perm); -ipc_rcu_putref (sma); -ipc_unlock (&(sma)->sem_perm); +sem_lock_and_putref (sma); +sem_unlock (sma,-1 ); }/*
@@ -274,9 +327,9 @@ static inline void sem_putref(struct sem_array *sma)*/ static inline void sem_getref (struct sem_array *sma) { -spin_lock (&(sma)->sem_perm.lock); +sem_lock (sma,NULL ,-1 );ipc_rcu_getref (sma); -ipc_unlock (&(sma)->sem_perm); +sem_unlock (sma,-1 ); }static inline void sem_rmid (struct ipc_namespace *ns,struct sem_array *s)
@@ -369,15 +422,18 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params) sma->sem_base = (struct sem *) &sma[1 ]; -for (i =0 ; i < nsems; i++) +for (i =0 ; i < nsems; i++) {INIT_LIST_HEAD (&sma->sem_base[i].sem_pending); +spin_lock_init (&sma->sem_base[i].lock); +spin_lock (&sma->sem_base[i].lock); + } sma->complex_count =0 ;INIT_LIST_HEAD (&sma->sem_pending);INIT_LIST_HEAD (&sma->list_id); sma->sem_nsems = nsems; sma->sem_ctime =get_seconds (); -sem_unlock (sma); +sem_unlock (sma,-1 );return sma->sem_perm.id; }
@@ -816,7 +872,7 @@ static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)/* Remove the semaphore set from the IDR */ sem_rmid (ns, sma); -sem_unlock (sma); +sem_unlock (sma,-1 );wake_up_sem_queue_do (&tasks); ns->used_sems -= sma->sem_nsems;
@@ -990,16 +1046,16 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,sem_lock_and_putref (sma);if (sma->sem_perm.deleted) { -sem_unlock (sma); +sem_unlock (sma,-1 ); err = -EIDRM ; goto out_free; } } -spin_lock (&sma->sem_perm.lock); +sem_lock (sma,NULL ,-1 );for (i =0 ; i < sma->sem_nsems; i++) sem_io[i] = sma->sem_base[i].semval; -sem_unlock (sma); +sem_unlock (sma,-1 ); err =0 ;if (copy_to_user (array, sem_io, nsems*sizeof (ushort))) err = -EFAULT ;
@@ -1036,7 +1092,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, }sem_lock_and_putref (sma);if (sma->sem_perm.deleted) { -sem_unlock (sma); +sem_unlock (sma,-1 ); err = -EIDRM ; goto out_free; }
@@ -1061,7 +1117,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,if (semnum <0 || semnum >= nsems) goto out_unlock; -spin_lock (&sma->sem_perm.lock); +sem_lock (sma,NULL ,-1 ); curr = &sma->sem_base[semnum];switch (cmd) {
@@ -1101,7 +1157,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, } out_unlock: -sem_unlock (sma); +sem_unlock (sma,-1 ); out_wakeup:wake_up_sem_queue_do (&tasks); out_free:
@@ -1169,11 +1225,11 @@ static int semctl_down(struct ipc_namespace *ns, int semid,switch (cmd){case IPC_RMID : -ipc_lock_object (&sma->sem_perm); +sem_lock (sma,NULL ,-1 );freeary (ns, ipcp); goto out_up;case IPC_SET : -ipc_lock_object (&sma->sem_perm); +sem_lock (sma,NULL ,-1 ); err =ipc_update_perm (&semid64.sem_perm, ipcp);if (err) goto out_unlock;
@@ -1186,7 +1242,7 @@ static int semctl_down(struct ipc_namespace *ns, int semid, } out_unlock: -sem_unlock (sma); +sem_unlock (sma,-1 ); out_up:up_write (&sem_ids (ns).rw_mutex);return err;
@@ -1343,7 +1399,7 @@ static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)/* step 3: Acquire the lock on semaphore array */ sem_lock_and_putref (sma);if (sma->sem_perm.deleted) { -sem_unlock (sma); +sem_unlock (sma,-1 );kfree (new); un =ERR_PTR (-EIDRM ); goto out;
@@ -1371,7 +1427,7 @@ static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid) success:spin_unlock (&ulp->lock);rcu_read_lock (); -sem_unlock (sma); +sem_unlock (sma,-1 ); out:return un; }
@@ -1411,7 +1467,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,struct sembuf fast_sops[SEMOPM_FAST ];struct sembuf * sops = fast_sops, *sop;struct sem_undo *un; -int undos =0 , alter =0 , max; +int undos =0 , alter =0 , max, locknum;struct sem_queue queue;unsigned long jiffies_left =0 ;struct ipc_namespace *ns;
@@ -1501,7 +1557,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,* "un" itself is guaranteed by rcu. */ error = -EIDRM ; -ipc_lock_object (&sma->sem_perm); + locknum =sem_lock (sma, sops, nsops);if (un) {if (un->semid ==-1 ) {rcu_read_unlock ();
@@ -1558,7 +1614,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, sleep_again: current->state =TASK_INTERRUPTIBLE ; -sem_unlock (sma); +sem_unlock (sma, locknum);if (timeout) jiffies_left =schedule_timeout (jiffies_left);
@@ -1580,7 +1636,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, goto out_free; } - sma =sem_obtain_lock (ns, semid); + sma =sem_obtain_lock (ns, semid, sops, nsops, &locknum);/* * Wait until it's guaranteed that no wakeup_sem_queue_do() is ongoing.
@@ -1619,7 +1675,7 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,unlink_queue (sma, &queue); out_unlock_free: -sem_unlock (sma); +sem_unlock (sma, locknum); out_wakeup:wake_up_sem_queue_do (&tasks); out_free:
@@ -1693,12 +1749,14 @@ void exit_sem(struct task_struct *tsk) semid =-1 ; else semid = un->semid; -rcu_read_unlock (); -if (semid ==-1 ) +if (semid ==-1 ) { +rcu_read_unlock ();break ; + } - sma =sem_lock_check (tsk->nsproxy->ipc_ns, un->semid); + sma =sem_obtain_object_check (tsk->nsproxy->ipc_ns, un->semid); +sem_lock (sma,NULL ,-1 );/* exit_sem raced with IPC_RMID, nothing to do */ if (IS_ERR (sma))
@@ -1709,7 +1767,7 @@ void exit_sem(struct task_struct *tsk)/* exit_sem raced with IPC_RMID+semget() that created * exactly the same semid. Nothing to do. */ -sem_unlock (sma); +sem_unlock (sma,-1 );continue ; }
@@ -1749,7 +1807,7 @@ void exit_sem(struct task_struct *tsk)/* maybe some queued-up processes were waiting for this */ INIT_LIST_HEAD (&tasks);do_smart_update (sma,NULL ,0 ,1 , &tasks); -sem_unlock (sma); +sem_unlock (sma,-1 );wake_up_sem_queue_do (&tasks);kfree_rcu (un, rcu);
Signature · 1 line
1.7.7.6
Re: [PATCH 7/7] ipc,sem: fine grained locking for semtimedop
Quoted 29 lines · Rik 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_arraylock 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
Quoted 2 lines
Signed-off-by Rik van Riel <[email protected]>Suggested-by Linus Torvalds <[email protected]>
Re: [PATCH 7/7] ipc,sem: fine grained locking for semtimedop
In reply to Rik van RielSorry for the late reply; I've been swamped and am behind on my upstream mail.
Quoted 6 lines · Rik van Riel
ipc/sem.c +6
-0+static inline int sem_lock (struct sem_array *sma,struct sembuf *sops, +int nsops) +{ +int locknum; +if (nsops ==1 && !sma->complex_count) { +struct sem *sem = sma->sem_base + sops->sem_num;
ipc/sem.c void __init sem_init (void) · line 219
Quoted 10 lines
ipc/sem.c +10
-0+ +/* 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 (&sma->sem_perm.lock);
I believe this should be spin_unlock(&sem->lock) instead ?
ipc/sem.c void __init sem_init (void) · line 228
Quoted 9 lines
ipc/sem.c +9
-0+ goto lock_all; + } + locknum = sops->sem_num; + }else { +int i; +/* Lock the sem_array, and all the semaphore locks */ + lock_all: +spin_lock (&sma->sem_perm.lock); +for (i =0 ; i < sma->sem_nsems; i++) {
Do we have to lock every sem from the array instead of just the ones that are being operated on in sops ?
(I'm not sure either way, as I don't fully understand the queueing of complex ops)
If we want to keep the loop as is, then we may at least remove the sops argument to sem_lock() since we only care about nsops.
ipc/sem.c void __init sem_init (void) · line 235
Quoted 7 lines
ipc/sem.c +7
-0+struct sem *sem = sma->sem_base + i; +spin_lock (&sem->lock); + } + locknum =-1 ; + } +return locknum; +}
That's all I have. Very nice test results BTW!
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH 7/7] ipc,sem: fine grained locking for semtimedop
Quoted 13 lines · Michel Lespinasse
Sorry for the late reply; I've been swamped and am behind on my upstream mail.
Quoted 10 lines · Rik van Riel
ipc/sem.c +10
-0+static inline int sem_lock (struct sem_array *sma,struct sembuf *sops, +int nsops) +{ +int locknum; +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); +
Quoted 8 lines
ipc/sem.c void __init sem_init (void) · line 219
Quoted 6 lines
ipc/sem.c +6
-0+/* +* 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 (&sma->sem_perm.lock);I believe this should be
spin_unlock(&sem->lock)instead ?
You are right. Good catch.
I'll send a one-liner fix for this to Andrew Morton, since he already applied the patches in -mm.
Quoted 10 lines
ipc/sem.c void __init sem_init (void) · line 228
Quoted 9 lines
ipc/sem.c +9
-0+ goto lock_all; + } + locknum = sops->sem_num; + }else { +int i; +/* Lock the sem_array, and all the semaphore locks */ + lock_all: +spin_lock (&sma->sem_perm.lock); +for (i =0 ; i < sma->sem_nsems; i++) {
Quoted 4 lines
Do we have to lock every sem from the array instead of just the ones that are being operated on in sops ?
(I'm not sure either way, as I don't fully understand the queueing of complex ops)
We should be able to get away with locking just the ones that are being operated on. However, that does require we remember which ones we locked, so we know which ones to unlock again.
I do not know whether that additional complexity is worthwhile, especially considering that in Davidlohr's profile, the major caller locking everybody appears to be semctl, which passes no semops into the kernel but simply operates on all semaphores at once (SET_ALL).
Quoted 2 lines
If we want to keep the loop as is, then we may at least remove the sops argument to
sem_lock()since we only care about nsops.
We need to know exactly which semaphore to lock when we are locking only one. The sops argument is used to figure out which one.
Signature · 1 line
All rights reversed
PATCH 7/7 fix for sem_lock
In reply to Michel LespinasseQuoted 9 lines
Review
Quoted 7 lines
(quoted diff) +6
-0+ /* + * 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(&sma->sem_perm.lock);I believe this should be
spin_unlock(&sem->lock)instead ?
Michel, thanks for spotting this!
Andrew, could you fold this fix into my patch 7/7 before submitting things for 3.10? Thank you.
--->8---
Fix a typo in sem_lock. Of course we need to unlock the local semaphore lock before jumping to lock_all, in the rare case that somebody started a complex operation while we were spinning on the spinlock.
Can be folded into patch 7/7 before merging
--- ipc/sem.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)
ipc/sem.c +1 -1
@@ -346,7 +346,7 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,* we may need to look at things we did not lock here. */ if (unlikely (sma->complex_count)) { -spin_unlock (&sma->sem_perm.lock); +spin_unlock (&sem->lock); goto lock_all; } locknum = sops->sem_num;
Quoted 5 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The series looks sane to me, and I like how each individual step is pretty small and makes sense.
It *would* be lovely to see this run with the actual Swingbench numbers. The microbenchmark always looked much nicer. Do the additional multi-semaphore scalability patches on top of Davidlohr's patches help with the swingbench issue, or are we still totally swamped by the ipc lock there?
Maybe there were already numbers for that, but the last swingbench numbers I can actually recall was from before the finer-grained locking..
And obviously, getting this tested so that there aren't any more missed wakeups etc would be lovely. I'm assuming the plan is that this all goes through Andrew? Do we have big semop users who could test it on real loads? Considering that I *suspect* the main users are things like Oracle etc, I'd assume that there's some RH lab or partner or similar that is interested in making sure this not only helps, but also that it doesn't break anything ;)
Linus
Quoted 10 lines · Linus Torvalds
It *would* be lovely to see this run with the actual Swingbench numbers. The microbenchmark always looked much nicer. Do the additional multi-semaphore scalability patches on top of Davidlohr's patches help with the swingbench issue, or are we still totally swamped by the ipc lock there?
Maybe there were already numbers for that, but the last swingbench numbers I can actually recall was from before the finer-grained locking..
Ok, and if the spinlock is still a big deal even with the finer granularity, it might be interesting to hear if Michel's fast locks make a difference. I'm guessing that this series might actually make it easier/cleaner to do the semaphore locking using another lock, since the ipc_lock got split up and out...
I think Michel did it for some socket code too. I think that was fairly independent and was for netperf.
Linus
Quoted 7 lines · Linus Torvalds
Quoted 5 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Quoted 8 lines
The series looks sane to me, and I like how each individual step is pretty small and makes sense.
It *would* be lovely to see this run with the actual Swingbench numbers. The microbenchmark always looked much nicer. Do the additional multi-semaphore scalability patches on top of Davidlohr's patches help with the swingbench issue, or are we still totally swamped by the ipc lock there?
Yes, I'm testing this patchset with my swingbench workloads. I should have some numbers by today or tomorrow.
Quoted 4 lines
Maybe there were already numbers for that, but the last swingbench numbers I can actually recall was from before the finer-grained locking..
Right, I couldn't get Oracle to run on the with the previous patches, hopefully the bug(s) are now addressed.
Thanks,
Davidlohr
Quoted 5 lines · On Wed, 20 Mar 2013 15:55:30 -0400 Rik van Riel <[email protected]> wrote:
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Quoted 7 lines
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
These patches conflict pretty badly with Peter's:
ipc-clamp-with-min.patch ipc-separate-msg-allocation-from-userspace-copy.patch ipc-tighten-msg-copy-loops.patch ipc-set-efault-as-default-error-in-load_msg.patch ipc-remove-msg-handling-from-queue-scan.patch ipc-implement-msg_copy-as-a-new-receive-mode.patch ipc-simplify-msg-list-search.patch ipc-refactor-msg-list-search-into-separate-function.patch ipc-refactor-msg-list-search-into-separate-function-fix.patch
(they're at http://ozlabs.org/~akpm/mmots/broken-out/)
We're in a bit of a mess at present.
Last month Peter sent a ten-patch series which fixed an oops (Subject:
"ipc MSG_COPY fixes"). The series did other stuff, so we merged into mainline just two bugfix patches. Then davej hit a trinity oops (Subject: "ipc/testmsg GPF.") and testing confirmed that the remaining eight patches in Peter's series fixed that up. Nobody has yet identified which of the eight does the good deed. So we need to either:
a) revert the original two fixes "ipc: fix potential oops when src msg > 4k w/ MSG_COPY" and "ipc: don't allocate a copy larger than max" or
b) try reverting "ipc: don't allocate a copy larger than max", as "ipc: fix potential oops when src msg > 4k w/ MSG_COPY" looks pretty harmless or
c) work out which of the remaining 8 fixes the new oops and merge that or
d) merge all 8 fixes or
e) something else.
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting 88b9e456b1649722673ff ("ipc: don't allocate a copy larger than max") fixes things up?
Quoted 31 lines · Andrew Morton
Quoted 5 lines · On Wed, 20 Mar 2013 15:55:30 -0400 Rik van Riel <[email protected]> wrote:
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Quoted 7 lines
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
These patches conflict pretty badly with Peter's:
ipc-clamp-with-min.patch ipc-separate-msg-allocation-from-userspace-copy.patch ipc-tighten-msg-copy-loops.patch ipc-set-efault-as-default-error-in-load_msg.patch ipc-remove-msg-handling-from-queue-scan.patch ipc-implement-msg_copy-as-a-new-receive-mode.patch ipc-simplify-msg-list-search.patch ipc-refactor-msg-list-search-into-separate-function.patch ipc-refactor-msg-list-search-into-separate-function-fix.patch
(they're at http://ozlabs.org/~akpm/mmots/broken-out/)
Quoted 7 lines
We're in a bit of a mess at present.
Last month Peter sent a ten-patch series which fixed an oops (Subject:
"ipcMSG_COPYfixes"). The series did other stuff, so we merged into mainline just two bugfix patches. Then davej hit a trinity oops (Subject: "ipc/testmsg GPF.") and testing confirmed that the remaining eight patches in Peter's series fixed that up.
Just to clarify the history.
A while back on linux-next both Dave and I were hitting ipc oopses on trinity, which was fallout from the MSG_COPY implementation.
One of those oopses was the ipc/testmsg oops.
I was trying to push out the new tty ldisc patchset and trinity is a decent tool for exploring race conditions (which the tty layer was full of), so the oopses were in my way.
Because it was nearly impossible to static analyze the existing ipc code, I just started cleaning up. As I did, I found the two by-then-obvious bug fixes. Also, assigning the 'msg' variable in the search loop looked suspicious so I factored that search loop out as well. That was the whole 10-patch "ipc MSG_COPY fixes" series.
I wasn't getting the ipc/testmsg bug anymore and I let Dave know, so all was good.
When Andrew asked me if the whole series needed to go into 3.9, I said I didn't know.
So when just the 2 of 10 patches went in, Dave started to hit the ipc/testmsg bug again on 3.9.
Quoted 12 lines
Nobody has yet
identified which of the eight does the good deed. So we need to either:a) revert the original two fixes "ipc: fix potential oops when src msg > 4k w/
MSG_COPY" and "ipc: don't allocate a copy larger than max" orb) try reverting "ipc: don't allocate a copy larger than max", as "ipc: fix potential oops when src msg > 4k w/
MSG_COPY" looks pretty harmless or
Quoted 5 lines
c) work out which of the remaining 8 fixes the new oops and merge that or
d) merge all 8 fixes or
e) something else.
The fix is in the other 8 patches. But I have to say, the base code was such a mess I have no idea which of the other 8 patches fixes the ipc/testmsg oops or how.
I guarantee reverting those 2 fixes from my series will not make the ipc/testmsg oops go away.
Quoted 4 lines
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?
Quoted 10 lines · Andrew Morton
Quoted 8 lines · On Wed, 20 Mar 2013 15:55:30 -0400 Rik van Riel <[email protected]> wrote:
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
Quoted 6 lines
Quoted 4 lines
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
These patches conflict pretty badly with Peter's:
On one point I'm a little confused: my series has been in linux-next for a while. On what tree is this series based?
Quoted 16 lines · On Thu, 21 Mar 2013 17:50:05 -0400 Peter Hurley <[email protected]>…
Quoted 15 lines · Andrew Morton
Quoted 12 lines · On Wed, 20 Mar 2013 15:55:30 -0400 Rik van Riel <[email protected]> wrote:
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
Quoted 4 lines
Quoted 1 line
These patches conflict pretty badly with Peter's:
On one point I'm a little confused: my series has been in linux-next for a while. On what tree is this series based?
It'll be based on mainline. People often forget to peek into linux-next when preparing patches. In the great majority of cases that's OK. Occasionally, we lose...
Quoted 18 lines · Andrew Morton
Quoted 16 lines · On Thu, 21 Mar 2013 17:50:05 -0400 Peter Hurley <[email protected]>…
Quoted 15 lines · Andrew Morton
Quoted 12 lines · On Wed, 20 Mar 2013 15:55:30 -0400 Rik van Riel <[email protected]> wrote:
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
Quoted 8 lines
Quoted 4 lines
Quoted 1 line
These patches conflict pretty badly with Peter's:
On one point I'm a little confused: my series has been in linux-next for a while. On what tree is this series based?
It'll be based on mainline. People often forget to peek into linux-next when preparing patches. In the great majority of cases that's OK. Occasionally, we lose...
I'll be happy to rebase the series onto linux-next, to make merging easier for you.
Signature · 1 line
All rights reversed.
Quoted 4 lines · Andrew Morton
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?
Ok, with that reverted it's been grinding away for a few hours without incident. Normally I see the oops within a minute or so.
Dave
Quoted 10 lines · On Tue, 26 Mar 2013 15:28:52 -0400 Dave Jones <[email protected]> wrote:
Quoted 4 lines · Andrew Morton
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?Ok, with that reverted it's been grinding away for a few hours without incident. Normally I see the oops within a minute or so.
OK, thanks, I queued a revert:
From: Andrew Morton <[email protected]>
Subject: revert "ipc: don't allocate a copy larger than max"
Revert 88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.
---
ipc/msg.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-)
ipc/msg.c renamed from ipc/msg.c~revert-ipc-dont-allocate-a-copy-larger-than-max +2 -4
@@ -820,17 +820,15 @@ long do_msgrcv(int msqid, void __user *bstruct msg_msg *copy =NULL ;unsigned long copy_number =0 ; - ns = current->nsproxy->ipc_ns; -if (msqid <0 || (long ) bufsz <0 )return -EINVAL ;if (msgflg &MSG_COPY ) { - copy =prepare_copy (buf,min_t (size_t, bufsz, ns->msg_ctlmax), - msgflg, &msgtyp, ©_number); + copy =prepare_copy (buf, bufsz, msgflg, &msgtyp, ©_number);if (IS_ERR (copy))return PTR_ERR (copy); } mode =convert_mode (&msgtyp, msgflg); + ns = current->nsproxy->ipc_ns; msq =msg_lock_check (ns, msqid);if (IS_ERR (msq)) {
_
Quoted 12 lines · Andrew Morton
Quoted 9 lines · On Tue, 26 Mar 2013 15:28:52 -0400 Dave Jones <[email protected]> wrote:
Quoted 4 lines · Andrew Morton
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?Ok, with that reverted it's been grinding away for a few hours without incident. Normally I see the oops within a minute or so.
Quoted 7 lines
OK, thanks, I queued a revert:
From: Andrew Morton <[email protected]>
Subject: revert "ipc: don't allocate a copy larger than max"Revert
88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: l2tp_ppp l2tp_netlink l2tp_core llc2 phonet netrom rose af_key af_rxrpc caif_socket caif can_raw cmtp kernelcapi can_bcm can nfnetlink ipt_ULOG scsi_transport_iscsi af_802154 ax25 atm ipx pppoe pppox x25 nfc irda ppp_generic p8023 slhc p8022 appletalk decnet crc_ccitt rds psnap llc lockd sunrpc ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 xt_conntrack nf_conntrack ip6table_filter ip6_tables snd_hda_codec_realtek btusb snd_hda_intel bluetooth snd_hda_codec raid0 snd_pcm rfkill microcode serio_raw pcspkr snd_page_alloc edac_core snd_timer snd soundcore r8169 mii vhost_net tun macvtap macvlan kvm_amd kvm radeon backlight drm_kms_helper ttm CPU 3 Pid: 1850, comm: trinity-child37 Tainted: G B 3.9.0-rc4+ #7 Gigabyte Technology Co., Ltd. GA-MA78GM-S2H/GA-MA78GM-S2H RIP: 0010:[<ffffffff812c20fb>] [<ffffffff812c20fb>] free_msg+0x2b/0x40 RSP: 0018:ffff8800a1d3bdd0 EFLAGS: 00010202 RAX: 0000000000000000 RBX: 6b6b6b6b6b6b6b6b RCX: 0000000000000000 RDX: 0000000000000000 RSI: 0000000000000000 RDI: ffffffff810b6ced RBP: ffff8800a1d3bde0 R08: 0000000000000000 R09: 0000000000000001 R10: 0000000000000001 R11: 0000000000000001 R12: ffff88009997e620 R13: ffffffff81c7ace0 R14: ffff8800caf359d8 R15: ffffffff81c7b024 FS: 00007f2d7be64740(0000) GS:ffff88012ac00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b CR2: 00007f8bd7bb6000 CR3: 00000000a1f0a000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Process trinity-child37 (pid: 1850, threadinfo ffff8800a1d3a000, task ffff8800a1e62490) Stack: 6b6b6b6b6b6b6b6b ffff8800caf35928 ffff8800a1d3be18 ffffffff812c289f 0000000000000000 ffffffff81c7ace0 ffff8800caf35928 0000000000000000 ffff8800a1d3be28 ffff8800a1d3bec8 ffffffff812c2a93 ffff8800a1d3be40 Call Trace: [<ffffffff812c289f>] freeque+0xcf/0x140 [<ffffffff812c2a93>] msgctl_down.constprop.9+0x183/0x200 [<ffffffff810767cf>] ? up_read+0x1f/0x40 [<ffffffff816c8f94>] ? __do_page_fault+0x214/0x5b0 [<ffffffff810b94be>] ? lock_release_non_nested+0x23e/0x320 [<ffffffff812c2da9>] sys_msgctl+0x139/0x400 [<ffffffff816c5d4d>] ? retint_swapgs+0xe/0x13 [<ffffffff810b6c55>] ? trace_hardirqs_on_caller+0x115/0x1a0 [<ffffffff8134b39e>] ? trace_hardirqs_on_thunk+0x3a/0x3f [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b Code: 66 66 66 66 90 55 48 89 e5 41 54 49 89 fc 53 e8 fc 5e 01 00 49 8b 5c 24 20 4c 89 e7 e8 8f af ed ff 48 85 db 75 05 eb 13 4c 89 e3 <4c> 8b 23 48 89 df e8 7a af ed ff 4d 85 e4 75 ed 5b 41 5c 5d c3
(Taint is from an ext4 double-free I just reported in a separate thread)
decoded..
0: 66 66 66 66 90 data32 data32 data32 xchg %ax,%ax 5: 55 push %rbp 6: 48 89 e5 mov %rsp,%rbp 9: 41 54 push %r12 b: 49 89 fc mov %rdi,%r12 e: 53 push %rbx f: e8 fc 5e 01 00 callq 0x15f10 14: 49 8b 5c 24 20 mov 0x20(%r12),%rbx 19: 4c 89 e7 mov %r12,%rdi 1c: e8 8f af ed ff callq 0xffffffffffedafb0 21: 48 85 db test %rbx,%rbx 24: 75 05 jne 0x2b 26: eb 13 jmp 0x3b 28: 4c 89 e3 mov %r12,%rbx 2b:* 4c 8b 23 mov (%rbx),%r12 <-- trapping instruction 2e: 48 89 df mov %rbx,%rdi 31: e8 7a af ed ff callq 0xffffffffffedafb0 36: 4d 85 e4 test %r12,%r12 39: 75 ed jne 0x28 3b: 5b pop %rbx 3c: 41 5c pop %r12 3e: 5d pop %rbp 3f: c3 retq
Disassembly of free_msg shows..
seg = msg->next;
kfree(msg);
while (seg != NULL) {
struct msg_msgseg *tmp = seg->next;
30b: 4c 8b 23 mov (%rbx),%r12
kfree(seg);
30e: 48 89 df mov %rbx,%rdi
311: e8 00 00 00 00 callq 316 <free_msg+0x36>Looks like seg was already kfree'd.
Dave
Quoted 7 lines · Dave Jones
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOCLooks like seg was already kfree'd.
Just to clarify: is this you testing -mm that has Davidlohr's and Rik's scalability patches? Or mainline?
Linus
Quoted 11 lines · Linus Torvalds
Quoted 7 lines · Dave Jones
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOCLooks like seg was already kfree'd.
Just to clarify: is this you testing -mm that has Davidlohr's and Rik's scalability patches? Or mainline?
mainline. Your current tree.
Dave
Quoted 2 lines · Dave Jones
mainline. Your current tree.
Ok, that's what I thought you were generally testing, just wanted to verify. The subject kind of implied otherwise..
Linus
Quoted 5 lines · Dave Jones
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
Do you have CONFIG_CHECKPOINT_RESTORE enabled? Does it go away if you disable it?
I think I foud at least one bug in the MSG_COPY stuff: it leaks the "copy" allocation if
mode == SEARCH_LESSEQUAL
but maybe I'm misreading it. And that shouldn't cause the problem you see, but it's indicative of how badly tested and thought through the MSG_COPY code is.
Totally UNTESTED leak fix appended. Stanislav?
Linus
ipc/msg.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
ipc/msg.c +3 -3
@@ -870,6 +870,7 @@ long do_msgrcv(int msqid, void __user *buf, size_t bufsz, long msgtyp, msg =copy_msg (msg, copy);if (IS_ERR (msg)) goto out_unlock; + copy =NULL ;break ; } }else
@@ -976,10 +977,9 @@ out_unlock:break ; } } -if (IS_ERR (msg)) { -free_copy (copy); +free_copy (copy); +if (IS_ERR (msg))return PTR_ERR (msg); - } bufsz =msg_handler (buf, msg, bufsz);free_msg (msg);
Quoted 13 lines · Linus Torvalds
Quoted 5 lines · Dave Jones
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOCDo you have
CONFIG_CHECKPOINT_RESTOREenabled? Does it go away if you disable it?I think I foud at least one bug in the
MSG_COPYstuff: it leaks the "copy" allocation if
Quoted 7 lines
mode == SEARCH_LESSEQUALbut maybe I'm misreading it. And that shouldn't cause the problem you see, but it's indicative of how badly tested and thought through the
MSG_COPYcode is.Totally UNTESTED leak fix appended. Stanislav?
I'll give it a shot.
Btw, something that's really bothering me is just how much bogus 'follow-on' spew we have lately. I'm not sure what changed, but it seems to have gotten worse.
Here's an oops I just hit..
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60 PGD 10fd95067 PUD 10f767067 PMD 0 Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: phonet netrom llc2 af_key rose af_rxrpc caif_socket caif can_raw cmtp kernelcapi ipt_ULOG nfnetlink can_bcm can scsi_transport_iscsi af_802154 irda ax25 atm ipx x25 p8023 p8022 appletalk pppoe decnet pppox ppp_generic nfc rds slhc psnap crc_ccitt llc lockd sunrpc ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 xt_conntrack nf_conntrack ip6table_filter ip6_tables raid0 snd_hda_codec_realtek snd_hda_intel btusb snd_hda_codec bluetooth microcode serio_raw snd_pcm edac_core pcspkr snd_page_alloc rfkill snd_timer snd soundcore r8169 mii vhost_net tun macvtap macvlan kvm_amd kvm radeon backlight drm_kms_helper ttm CPU 2 Pid: 958, comm: trinity-child20 Not tainted 3.9.0-rc4+ #7 Gigabyte Technology Co., Ltd. GA-MA78GM-S2H/GA-MA78GM-S2H RIP: 0010:[<ffffffff812c24ca>] [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60 RSP: 0018:ffff880117bb5e88 EFLAGS: 00010246 RAX: ffffffffffffffff RBX: 0000000000000004 RCX: 0000000000000078 RDX: 0000000000000004 RSI: fffffffffffffffe RDI: 000000000000000f RBP: ffff880117bb5e88 R08: 0000000000000004 R09: 0000000000000001 R10: ffff880117bb8000 R11: 0000000000000001 R12: fffffffffffffffe R13: ffff88010fd10308 R14: ffff88010fd10258 R15: ffffffffffffffff FS: 00007fa89c256740(0000) GS:ffff88012aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 000000000000000f CR3: 000000010f76c000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Process trinity-child20 (pid: 958, threadinfo ffff880117bb4000, task ffff880117bb8000) Stack: ffff880117bb5f68 ffffffff812c3746 0000000000000000 ffff880117bb8000 ffff880117bb8000 ffff880117bb8000 ffffffff81c7ace0 ffffffff812c2430 0000000000000004 0000000000000000 000000000000ffff 00000000652a928e Call Trace: [<ffffffff812c3746>] do_msgrcv+0x1a6/0x5f0 [<ffffffff812c2430>] ? msg_security+0x10/0x10 [<ffffffff810b6c55>] ? trace_hardirqs_on_caller+0x115/0x1a0 [<ffffffff8134b39e>] ? trace_hardirqs_on_thunk+0x3a/0x3f [<ffffffff812c3ba5>] sys_msgrcv+0x15/0x20 [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b Code: c3 48 c7 c0 f2 ff ff ff eb e5 0f 1f 80 00 00 00 00 66 66 66 66 90 55 83 fa 02 48 89 e5 74 3a 7e 28 83 fa 03 74 13 83 fa 04 75 0a <48> 39 37 b8 01 00 00 00 7e 02 31 c0 5d c3 48 3b 37 74 f7 b8 01 RIP [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60 RSP <ffff880117bb5e88> CR2: 000000000000000f ---[ end trace 8f0713d2aacb6aa3 ]--- BUG: sleeping function called from invalid context at kernel/rwsem.c:20 in_atomic(): 1, irqs_disabled(): 0, pid: 958, name: trinity-child20 INFO: lockdep is turned off. Pid: 958, comm: trinity-child20 Tainted: G D 3.9.0-rc4+ #7 Call Trace: [<ffffffff8107dba5>] __might_sleep+0x145/0x200 [<ffffffff816c215a>] down_read+0x2a/0xa0 [<ffffffff8105fa14>] exit_signals+0x24/0x130 [<ffffffff8104b80c>] do_exit+0xbc/0xd10 [<ffffffff81048b25>] ? kmsg_dump+0x1b5/0x230 [<ffffffff81048995>] ? kmsg_dump+0x25/0x230 [<ffffffff816c6b6c>] oops_end+0x9c/0xe0 [<ffffffff816b7a40>] no_context+0x268/0x275 [<ffffffff816b7ac5>] __bad_area_nosemaphore+0x78/0x1d1 [<ffffffff816b7c31>] bad_area_nosemaphore+0x13/0x15 [<ffffffff816c90e6>] __do_page_fault+0x366/0x5b0 [<ffffffff810b72b5>] ? __lock_acquire+0x2e5/0x1a00 [<ffffffff810b3348>] ? trace_hardirqs_off_caller+0x28/0xc0 [<ffffffff8134b3dd>] ? trace_hardirqs_off_thunk+0x3a/0x3c [<ffffffff816c933e>] do_page_fault+0xe/0x10 [<ffffffff816c5fa2>] page_fault+0x22/0x30 [<ffffffff812c24ca>] ? testmsg.isra.5+0x1a/0x60 [<ffffffff812d80b6>] ? security_msg_queue_msgrcv+0x16/0x20 [<ffffffff812c3746>] do_msgrcv+0x1a6/0x5f0 [<ffffffff812c2430>] ? msg_security+0x10/0x10 [<ffffffff810b6c55>] ? trace_hardirqs_on_caller+0x115/0x1a0 [<ffffffff8134b39e>] ? trace_hardirqs_on_thunk+0x3a/0x3f [<ffffffff812c3ba5>] sys_msgrcv+0x15/0x20 [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b note: trinity-child20[958] exited with preempt_count 1 BUG: scheduling while atomic: trinity-child20/958/0x10000002 INFO: lockdep is turned off. Modules linked in: phonet netrom llc2 af_key rose af_rxrpc caif_socket caif can_raw cmtp kernelcapi ipt_ULOG nfnetlink can_bcm can scsi_transport_iscsi af_802154 irda ax25 atm ipx x25 p8023 p8022 appletalk pppoe decnet pppox ppp_generic nfc rds slhc psnap crc_ccitt llc lockd sunrpc ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 xt_conntrack nf_conntrack ip6table_filter ip6_tables raid0 snd_hda_codec_realtek snd_hda_intel btusb snd_hda_codec bluetooth microcode serio_raw snd_pcm edac_core pcspkr snd_page_alloc rfkill snd_timer snd soundcore r8169 mii vhost_net tun macvtap macvlan kvm_amd kvm radeon backlight drm_kms_helper ttm Pid: 958, comm: trinity-child20 Tainted: G D 3.9.0-rc4+ #7 Call Trace: [<ffffffff816b8792>] __schedule_bug+0x61/0x70 [<ffffffff816c38af>] __schedule+0x8df/0x950 [<ffffffff81083bb8>] __cond_resched+0x18/0x30 [<ffffffff816c3dea>] _cond_resched+0x3a/0x50 [<ffffffff8117712f>] munlock_vma_pages_range+0xbf/0xe0 [<ffffffff8117b6e7>] exit_mmap+0x57/0x160 [<ffffffff811a9b5e>] ? __khugepaged_exit+0xee/0x130 [<ffffffff8119c485>] ? kmem_cache_free+0x335/0x380 [<ffffffff811a9b5e>] ? __khugepaged_exit+0xee/0x130 [<ffffffff81042047>] mmput+0x77/0x100 [<ffffffff8104b9e1>] do_exit+0x291/0xd10 [<ffffffff81048b25>] ? kmsg_dump+0x1b5/0x230 [<ffffffff81048995>] ? kmsg_dump+0x25/0x230 [<ffffffff816c6b6c>] oops_end+0x9c/0xe0 [<ffffffff816b7a40>] no_context+0x268/0x275 [<ffffffff816b7ac5>] __bad_area_nosemaphore+0x78/0x1d1 [<ffffffff816b7c31>] bad_area_nosemaphore+0x13/0x15 [<ffffffff816c90e6>] __do_page_fault+0x366/0x5b0 [<ffffffff810b72b5>] ? __lock_acquire+0x2e5/0x1a00 [<ffffffff810b3348>] ? trace_hardirqs_off_caller+0x28/0xc0 [<ffffffff8134b3dd>] ? trace_hardirqs_off_thunk+0x3a/0x3c [<ffffffff816c933e>] do_page_fault+0xe/0x10 [<ffffffff816c5fa2>] page_fault+0x22/0x30 [<ffffffff812c24ca>] ? testmsg.isra.5+0x1a/0x60 [<ffffffff812d80b6>] ? security_msg_queue_msgrcv+0x16/0x20 [<ffffffff812c3746>] do_msgrcv+0x1a6/0x5f0 [<ffffffff812c2430>] ? msg_security+0x10/0x10 [<ffffffff810b6c55>] ? trace_hardirqs_on_caller+0x115/0x1a0 [<ffffffff8134b39e>] ? trace_hardirqs_on_thunk+0x3a/0x3f [<ffffffff812c3ba5>] sys_msgrcv+0x15/0x20 [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b
That's a ton of not-very-interesting crap that makes me thankful I have a serial console. Most users aren't so lucky. Is there any way we can silence all that if we've just oopsed ?
Related: is there any value in printing out all the ? symbols on kernels with frame pointers enabled ?
Dave
Quoted 4 lines · Dave Jones
Btw, something that's really bothering me is just how much bogus 'follow-on' spew we have lately. I'm not sure what changed, but it seems to have gotten worse.
.. we have many more sanity checks, and they tend to trigger in the case of us killing somebody holding locks etc.
Quoted 2 lines
Related: is there any value in printing out all the ? symbols on kernels with frame pointers enabled ?
I've found them useful. You can often see what happened just before.
Of course, I'd still prefer gcc not generate unnecessarily big stack frames (which is one of the bigger reasons for old stuff shining through) but as things are, they often give hints about what happened just before.
Linus
Quoted 5 lines · Dave Jones
Here's an oops I just hit..
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60
Btw, looking at the code leading up to this, what the f*ck is wrong with the IPC stuff?
It's using the generic list stuff for most of the lists, but then it open-codes the accesses.
So instead of using
for_each_entry(walk_msg, &msq->q_messages, m_list) {
..
}the ipc/msg.c code does all that by hand, with
tmp = msq->q_messages.next;
while (tmp != &msq->q_messages) {
struct msg_msg *walk_msg; walk_msg = list_entry(tmp, struct msg_msg, m_list);
...
tmp = tmp->next;}
Ugh. The code is near unreadable. And then it has magic memory barriers etc, implying that it doesn't lock the data structures, but no comments about them. See expunge_all() and pipelined_send().
The code seems entirely random, and it's badly set up (annoyance of the day: crazy helper functions in ipc/msgutil.c to make sure that (a) you have to spend more effort looking for them, and (b) they won't get inlined).
Clearly nobody has cared for the crazy IPC message code in a long time.
Linus
Quoted 34 lines · Linus Torvalds
Quoted 5 lines · Dave Jones
Here's an oops I just hit..
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60Btw, looking at the code leading up to this, what the f*ck is wrong with the IPC stuff?
It's using the generic list stuff for most of the lists, but then it open-codes the accesses.
So instead of using
for_each_entry(walk_msg, &msq->q_messages, m_list) { .. }the
ipc/msg.ccode does all that by hand, withtmp = msq->q_messages.next; while (tmp != &msq->q_messages) { struct msg_msg *walk_msg;walk_msg = list_entry(tmp, struct msg_msg, m_list); ... tmp = tmp->next;}
Ugh. The code is near unreadable. And then it has magic memory barriers etc, implying that it doesn't lock the data structures, but no comments about them. See
expunge_all()andpipelined_send().
Quoted 6 lines
The code seems entirely random, and it's badly set up (annoyance of the day: crazy helper functions in
ipc/msgutil.cto make sure that (a) you have to spend more effort looking for them, and (b) they won't get inlined).Clearly nobody has cared for the crazy IPC message code in a long time.
Exactly that's what my patch series does; clean this mess up.
This is what I wrote to Andrew a couple of days ago.
On Tue, 2013-03-26 at 22:33 -0400, Peter Hurley wrote:
I just figured out how the queue is being corrupted and why my series
Quoted 4 lines
fixes it.
With
MSG_COPYset, the queue scan can exit with the local variable
'msg'
Quoted 1 line
pointing to a real msg if the
msg_counternever reaches the
copy_number.
Quoted 27 lines
The remaining execution looks like this:
if (!IS_ERR(msg)) { .... if (msgflg & MSG_COPY) goto out_unlock; ....
out_unlock:msg_unlock(msq); break; } } if (IS_ERR(msg)) .... bufsz = msg_handler(); free_msg(msg); <<---- msg never unlinkedSince the msg should not have been found (because it failed the match criteria), the if (!IS_ERR(msg)) clause should never have executed.
That's why my refactor fixes resolve this; because msg is not inadvertently treated as a found msg.
Quoted 7 lines
But let's be honest; the real bug here is the poor structure of this function that even made this possible. The deepest nesting executes a goto to a label in the middle of an if clause. Yuck! No wonder this thing's fragile.
So my recommendation still stands. The series that fixes this has been getting tested in linux-next for a month. Fixing this some other way
is
Quoted 3 lines
just asking for more trouble.
But why not just revert
MSG_COPYaltogether for 3.9?
Regards,
Peter Hurley
Quoted 82 lines · Peter Hurley
Quoted 34 lines · Linus Torvalds
Quoted 5 lines · Dave Jones
Here's an oops I just hit..
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60Btw, looking at the code leading up to this, what the f*ck is wrong with the IPC stuff?
It's using the generic list stuff for most of the lists, but then it open-codes the accesses.
So instead of using
for_each_entry(walk_msg, &msq->q_messages, m_list) { .. }the
ipc/msg.ccode does all that by hand, withtmp = msq->q_messages.next; while (tmp != &msq->q_messages) { struct msg_msg *walk_msg;walk_msg = list_entry(tmp, struct msg_msg, m_list); ... tmp = tmp->next;}
Ugh. The code is near unreadable. And then it has magic memory barriers etc, implying that it doesn't lock the data structures, but no comments about them. See
expunge_all()andpipelined_send().Quoted 6 lines
The code seems entirely random, and it's badly set up (annoyance of the day: crazy helper functions in
ipc/msgutil.cto make sure that (a) you have to spend more effort looking for them, and (b) they won't get inlined).Clearly nobody has cared for the crazy IPC message code in a long time.
Exactly that's what my patch series does; clean this mess up.
This is what I wrote to Andrew a couple of days ago.
On Tue, 2013-03-26 at 22:33 -0400, Peter Hurley wrote:
I just figured out how the queue is being corrupted and why my seriesQuoted 4 lines
fixes it.
With
MSG_COPYset, the queue scan can exit with the local variable'msg'
Quoted 1 line
pointing to a real msg if the
msg_counternever reaches the
copy_number.Quoted 27 lines
The remaining execution looks like this:
if (!IS_ERR(msg)) { .... if (msgflg & MSG_COPY) goto out_unlock; ....
out_unlock:msg_unlock(msq); break; } } if (IS_ERR(msg)) .... bufsz = msg_handler(); free_msg(msg); <<---- msg never unlinkedSince the msg should not have been found (because it failed the match criteria), the if (!IS_ERR(msg)) clause should never have executed.
That's why my refactor fixes resolve this; because msg is not inadvertently treated as a found msg.
Quoted 11 lines
Quoted 7 lines
But let's be honest; the real bug here is the poor structure of this function that even made this possible. The deepest nesting executes a goto to a label in the middle of an if clause. Yuck! No wonder this thing's fragile.
So my recommendation still stands. The series that fixes this has been getting tested in linux-next for a month. Fixing this some other way
is
Quoted 3 lines
just asking for more trouble.
But why not just revert
MSG_COPYaltogether for 3.9?
If you guys are already looking at this, the conversions between size_t, long and int in the do_msgrcv/load_msg/alloc_msg code are a mess. You could trigger anything from:
[ 33.046572] BUG: unable to handle kernel paging request at ffff88003c2c7000 [ 33.047721] IP: [<ffffffff83dbcb34>] bad_from_user+0x4/0x6 [ 33.048528] PGD 7232067 PUD 7233067 PMD 3ffff067 PTE 800000003c2c7060 [ 33.049506] Oops: 0002 [#1] PREEMPT SMP DEBUG_PAGEALLOC [ 33.050029] Modules linked in: [ 33.050029] CPU 0 [ 33.050029] Pid: 6885, comm: a.out Tainted: G W 3.9.0-rc4-next-20130328-sasha-00017-g1463000 #321 [ 33.050029] RIP: 0010:[<ffffffff83dbcb34>] [<ffffffff83dbcb34>] bad_from_user+0x4/0x6 [ 33.050029] RSP: 0018:ffff88003462be40 EFLAGS: 00010246 [ 33.050029] RAX: 0000000000000000 RBX: 00000000fffffffb RCX: 00000000ff06ae2b [ 33.050029] RDX: 00000000fffffffb RSI: 00007fffed36d2a0 RDI: ffff88003c2c7000 [ 33.050029] RBP: ffff88003462be88 R08: 0000000000000280 R09: 0000000000000000 [ 33.050029] R10: 0000000000000000 R11: 0000000000000000 R12: 00000000fffffffb [ 33.050029] R13: 00007fffed36d2a0 R14: 0000000000000000 R15: 0000000000000000 [ 33.050029] FS: 00007f6990044700(0000) GS:ffff88003dc00000(0000) knlGS:0000000000000000 [ 33.050029] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 33.050029] CR2: ffff88003c2c7000 CR3: 00000000347c8000 CR4: 00000000000406f0 [ 33.050029] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 33.050029] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 [ 33.050029] Process a.out (pid: 6885, threadinfo ffff88003462a000, task ffff880034cb3000) [ 33.050029] Stack: [ 33.050029] ffffffff8192a6a9 ffff88003462be98 ffff88003b331e00 ffff88003ddd01e0 [ 33.050029] 0000000000000000 0000000000000000 0000000000000001 0000000000000000 [ 33.050029] 0000000000000000 ffff88003462bf68 ffffffff8192bb34 0000000000000000 [ 33.050029] Call Trace: [ 33.050029] [<ffffffff8192a6a9>] ? load_msg+0x59/0x100 [ 33.050029] [<ffffffff8192bb34>] do_msgrcv+0x74/0x5b0 [ 33.050029] [<ffffffff81202c85>] ? user_exit+0xb5/0xe0 [ 33.050029] [<ffffffff8192a750>] ? load_msg+0x100/0x100 [ 33.050029] [<ffffffff8117cdcd>] ? trace_hardirqs_on+0xd/0x10 [ 33.050029] [<ffffffff81076ea0>] ? syscall_trace_enter+0x20/0x2e0 [ 33.050029] [<ffffffff8192c080>] SyS_msgrcv+0x10/0x20 [ 33.050029] [<ffffffff83db7e58>] tracesys+0xe1/0xe6 [ 33.050029] Code: e9 1f ee c3 fd b9 f2 ff ff ff e9 28 ee c3 fd b8 f2 ff ff ff e9 2f ee c3 fd ba f2 ff ff ff e9 bf f1 c3 fd 90 90 90 90 89 d1 31 c0 <f3> aa 89 d0 c3 01 ca e9 50 fa c4 fd c1 e1 06 01 ca eb 08 48 8d [ 33.050029] RIP [<ffffffff83dbcb34>] bad_from_user+0x4/0x6 [ 33.050029] RSP <ffff88003462be40> [ 33.050029] CR2: ffff88003c2c7000 [ 33.050029] ---[ end trace 9bba0da8a88b1faa ]---
To:
=============================================================================
[ 1393.475659] BUG kmalloc-4096 (Tainted: G W ): Padding overwritten. 0xffff88004e00f8f8-0xffff88004e00ffff [ 1393.477469] ----------------------------------------------------------------------------- [ 1393.477469] [ 1393.478980] Disabling lock debugging due to kernel taint [ 1393.479730] INFO: Slab 0xffffea0001380200 objects=7 used=7 fp=0x (null) flags=0x1ffc0000004081 [ 1393.480030] Pid: 25258, comm: trinity-child54 Tainted: G B W 3.9.0-rc4-next-20130328-sasha-00017-g1463000 #321 [ 1393.480030] Call Trace: [ 1393.480030] [<ffffffff8125a3ca>] slab_err+0xaa/0xd0 [ 1393.480030] [<ffffffff81179e5e>] ? put_lock_stats.isra.14+0xe/0x40 [ 1393.480030] [<ffffffff8125af14>] slab_pad_check+0x104/0x170 [ 1393.480030] [<ffffffff8125b045>] check_slab+0xc5/0xd0 [ 1393.480030] [<ffffffff83d67748>] free_debug_processing+0x52/0x204 [ 1393.480030] [<ffffffff83dafc5d>] ? _raw_spin_unlock_irqrestore+0x5d/0xb0 [ 1393.480030] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.480030] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.480030] [<ffffffff83d67931>] __slab_free+0x37/0x3f7 [ 1393.480030] [<ffffffff81a2268c>] ? __debug_check_no_obj_freed+0x16c/0x220 [ 1393.480030] [<ffffffff811c8ad7>] ? rcu_irq_exit+0x1c7/0x260 [ 1393.480030] [<ffffffff8125c07d>] ? kfree+0x20d/0x330 [ 1393.480030] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.480030] [<ffffffff8125c137>] kfree+0x2c7/0x330 [ 1393.480030] [<ffffffff8192a583>] free_msg+0x33/0x40 [ 1393.480030] [<ffffffff8192a739>] load_msg+0xe9/0x100 [ 1393.480030] [<ffffffff8192bb34>] do_msgrcv+0x74/0x5b0 [ 1393.480030] [<ffffffff81202c85>] ? user_exit+0xb5/0xe0 [ 1393.480030] [<ffffffff8192a750>] ? load_msg+0x100/0x100 [ 1393.480030] [<ffffffff8117cdcd>] ? trace_hardirqs_on+0xd/0x10 [ 1393.480030] [<ffffffff81076ea0>] ? syscall_trace_enter+0x20/0x2e0 [ 1393.480030] [<ffffffff8192c080>] SyS_msgrcv+0x10/0x20 [ 1393.480030] [<ffffffff83db7e58>] tracesys+0xe1/0xe6 [ 1393.480030] Padding ffff88004e00f8f8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.480030] Padding ffff88004e00f908: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.480030] Padding ffff88004e00f918: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.480030] Padding ffff88004e00f928: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f938: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f948: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f958: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f968: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f978: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f988: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f998: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9a8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9b8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9c8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9d8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9e8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00f9f8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fa98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00faa8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fab8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fac8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fad8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fae8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00faf8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fb98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fba8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fbb8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fbc8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fbd8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fbe8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fbf8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fc98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fca8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fcb8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fcc8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fcd8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fce8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fcf8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fd98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fda8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fdb8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fdc8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fdd8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fde8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fdf8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fe98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fea8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00feb8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fec8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fed8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fee8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fef8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff08: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff18: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff28: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff38: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff48: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff58: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff68: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff78: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff88: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ff98: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ffa8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ffb8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ffc8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ffd8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00ffe8: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.509882] Padding ffff88004e00fff8: 00 00 00 00 00 00 00 00 ........ [ 1393.509882] FIX kmalloc-4096: Restoring 0xffff88004e00f8f8-0xffff88004e00ffff=0x5a [ 1393.509882] [ 1393.689228] ============================================================================= [ 1393.690761] BUG kmalloc-4096 (Tainted: G B W ): Redzone overwritten [ 1393.690761] ----------------------------------------------------------------------------- [ 1393.690761] [ 1393.690761] INFO: 0xffff88004e00f7b0-0xffff88004e00f7b7. First byte 0x0 instead of 0xcc [ 1393.690761] INFO: Slab 0xffffea0001380200 objects=7 used=6 fp=0xffff88004e008000 flags=0x1ffc0000004081 [ 1393.690761] INFO: Object 0xffff88004e00e7b0 @offset=26544 fp=0x (null) [ 1393.690761] [ 1393.690761] Bytes b4 ffff88004e00e7a0: 00 00 00 00 00 00 00 00 5a 5a 5a 5a 5a 5a 5a 5a ........ZZZZZZZZ [ 1393.690761] Object ffff88004e00e7b0: 48 91 00 4e 00 88 ff ff 6b 6b 6b 6b 6b 6b 6b 6b H..N....kkkkkkkk [ 1393.690761] Object ffff88004e00e7c0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e7d0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e7e0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e7f0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e800: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e810: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e820: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e830: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e840: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e850: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e860: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e870: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e880: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e890: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8a0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8b0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8c0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8d0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8e0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e8f0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e900: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e910: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e920: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e930: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e940: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e950: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e960: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e970: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e980: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e990: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9a0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9b0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9c0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9d0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9e0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00e9f0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea00: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea10: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea20: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea30: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea40: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea50: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea60: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea70: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea80: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ea90: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eaa0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eab0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eac0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ead0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eae0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eaf0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb00: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb10: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb20: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb30: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb40: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb50: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb60: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb70: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb80: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eb90: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eba0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ebb0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ebc0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ebd0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ebe0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ebf0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec00: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec10: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec20: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec30: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec40: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec50: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec60: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec70: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec80: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ec90: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00eca0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ecb0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ecc0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ecd0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ece0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b kkkkkkkkkkkkkkkk [ 1393.690761] Object ffff88004e00ecf0: 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 6b 00 00 00 00 00 kkkkkkkkkkk..... [ 1393.690761] Object ffff88004e00ed00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed10: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed20: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed30: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed40: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed50: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed60: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed70: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ed90: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eda0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00edb0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00edc0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00edd0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ede0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00edf0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee10: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee20: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee30: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee40: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee50: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee60: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee70: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ee90: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eea0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eeb0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eec0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eed0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eee0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eef0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef10: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef20: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef30: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef40: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef50: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef60: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef70: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00ef90: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00efa0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00efb0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00efc0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00efd0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00efe0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00eff0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f040: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f050: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f060: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f070: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f080: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f090: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f0f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f100: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f110: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f120: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f130: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f140: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f150: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f160: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f170: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f180: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f190: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f1f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f200: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f210: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f220: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f230: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f240: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f250: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f260: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f270: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f280: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f290: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f2f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f300: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f310: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f320: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f330: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f340: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f350: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f360: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f370: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f380: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f390: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f3f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f400: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f410: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f420: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f430: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f440: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f450: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f460: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f470: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f480: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f490: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f4f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f500: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f510: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f520: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f530: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f540: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f550: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f560: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f570: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f580: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f590: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f5f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f600: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f610: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f620: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f630: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f640: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f650: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f660: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f670: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f680: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f690: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6b0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f6f0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f700: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f710: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f720: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f730: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f740: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f750: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f760: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f770: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f780: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f790: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Object ffff88004e00f7a0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................ [ 1393.690761] Redzone ffff88004e00f7b0: 00 00 00 00 00 00 00 00 ........ [ 1393.690761] Padding ffff88004e00f8f0: 00 00 00 00 00 00 00 00 ........ [ 1393.690761] Pid: 25258, comm: trinity-child54 Tainted: G B W 3.9.0-rc4-next-20130328-sasha-00017-g1463000 #321 [ 1393.690761] Call Trace: [ 1393.690761] [<ffffffff81259268>] ? print_section+0x38/0x40 [ 1393.690761] [<ffffffff812593a1>] print_trailer+0x131/0x140 [ 1393.690761] [<ffffffff812597f4>] check_bytes_and_report+0xc4/0x120 [ 1393.690761] [<ffffffff8125a781>] check_object+0x51/0x240 [ 1393.690761] [<ffffffff83d677bd>] free_debug_processing+0xc7/0x204 [ 1393.690761] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.690761] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.690761] [<ffffffff83d67931>] __slab_free+0x37/0x3f7 [ 1393.690761] [<ffffffff81a2268c>] ? __debug_check_no_obj_freed+0x16c/0x220 [ 1393.690761] [<ffffffff811c8ad7>] ? rcu_irq_exit+0x1c7/0x260 [ 1393.690761] [<ffffffff8125c07d>] ? kfree+0x20d/0x330 [ 1393.690761] [<ffffffff8192a583>] ? free_msg+0x33/0x40 [ 1393.690761] [<ffffffff8125c137>] kfree+0x2c7/0x330 [ 1393.690761] [<ffffffff8192a583>] free_msg+0x33/0x40 [ 1393.690761] [<ffffffff8192a739>] load_msg+0xe9/0x100 [ 1393.690761] [<ffffffff8192bb34>] do_msgrcv+0x74/0x5b0 [ 1393.690761] [<ffffffff81202c85>] ? user_exit+0xb5/0xe0 [ 1393.690761] [<ffffffff8192a750>] ? load_msg+0x100/0x100 [ 1393.690761] [<ffffffff8117cdcd>] ? trace_hardirqs_on+0xd/0x10 [ 1393.690761] [<ffffffff81076ea0>] ? syscall_trace_enter+0x20/0x2e0 [ 1393.690761] [<ffffffff8192c080>] SyS_msgrcv+0x10/0x20 [ 1393.690761] [<ffffffff83db7e58>] tracesys+0xe1/0xe6 [ 1393.690761] FIX kmalloc-4096: Restoring 0xffff88004e00f7b0-0xffff88004e00f7b7=0xcc
By just playing with the 'msgsz' parameter with MSG_COPY set.
Thanks,
Sasha
Quoted 4 lines · Sasha Levin
If you guys are already looking at this, the conversions between
size_t, long and int in the do_msgrcv/load_msg/alloc_msg code are a mess. You could trigger anything from:
Good catch.
Let's just change the "(long)bufsz < 0" into "bufsz > INT_MAX".
I suspect we should change some of the "int" arguments to "size_t" too so that we don't have these kinds of odd "different routines see different values due to subtle casting errors", but in the end we don't really want to ever help people have these kinds of potential overflow issues. We already limit normal read/write/sendmsg etc to INT_MAX (although we tend to *truncate* it to INT_MAX rather than return an error, but I think the simpler patch here is preferable unless somebody complains).
Comments?
Linus
Quoted 2 lines · Sasha Levin
By just playing with the 'msgsz' parameter with
MSG_COPYset.
Hmm. Looking closer, I suspect you're testing without commit 88b9e456b164 ("ipc: don't allocate a copy larger than max"). That should limit the size passed in to prepare_copy -> load_copy to msg_ctlmax.
Now, I think it's possibly still a good idea to limit bufsz to INT_MAX regardless, but as far as I can see that prepare_copy -> load_copy path is the only place that can get confused. Everybody else uses size_t (or "long" in the case of r_maxsize) as far as I can tell.
Linus
Quoted 8 lines · Linus Torvalds
Quoted 2 lines · Sasha Levin
By just playing with the 'msgsz' parameter with
MSG_COPYset.Hmm. Looking closer, I suspect you're testing without commit
88b9e456b164("ipc: don't allocate a copy larger than max"). That should limit the size passed in toprepare_copy->load_copytomsg_ctlmax.
That commit has a revert in the -next trees, do we need a revert of the revert?
commit ff6577a3e714ccae02d4400e989762c19c37b0b3
Author: Andrew Morton <[email protected]>
Date: Wed Mar 27 10:24:02 2013 +1100
revert "ipc: don't allocate a copy larger than max"
Revert 88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.
Thanks,
Sasha
Quoted 5 lines · Sasha Levin
Quoted 4 lines · Linus Torvalds
Quoted 2 lines · Sasha Levin
By just playing with the 'msgsz' parameter with
MSG_COPYset.
Quoted 7 lines
Quoted 4 lines
Hmm. Looking closer, I suspect you're testing without commit
88b9e456b164("ipc: don't allocate a copy larger than max"). That should limit the size passed in toprepare_copy->load_copytomsg_ctlmax.That commit has a revert in the -next trees, do we need a revert of the revert?
Yeah, I told Andrew to drop that, but I think he's travelling.
Dave
Quoted 6 lines · Linus Torvalds
I think I foud at least one bug in the
MSG_COPYstuff: it leaks the "copy" allocation ifmode == SEARCH_LESSEQUALbut maybe I'm misreading it.
Yes, you're misreading it. copy_msg() returns the 'copy' address when copying is successful. So this patch double-frees 'copy'.
Andrew has had the patches that fix this for a month but because they're fairly extensive he didn't want to apply them for 3.9.
Regards,
Peter
Quoted 10 lines · Peter Hurley
Quoted 6 lines · Linus Torvalds
I think I foud at least one bug in the
MSG_COPYstuff: it leaks the "copy" allocation ifmode == SEARCH_LESSEQUALbut maybe I'm misreading it.
Yes, you're misreading it.
copy_msg()returns the 'copy' address when copying is successful. So this patch double-frees 'copy'.
No it doesn't.
The case where "copy_msg()" *is* called and is successful, msg gets set to "copy", my patch sets "copy" to NULL, so no, it doesn't double-free.
That said, it looks like if MSG_COPY is set, we cannot trigger the "mode == SEARCH_LESSEQUAL" case, because it forces msgtyp to 0, which in turn forces mode = SEARCH_ANY.
So the actual leak case seems to not be possible, although that's *really* hard to see from looking at the code. The code is just an unreadable mess.
Linus
29.03.2013 22:43, Linus Torvalds пишет:
Quoted 5 lines · Dave Jones
Quoted 4 lines
Quoted 4 lines
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
Quoted 9 lines
Quoted 1 line
Quoted 1 line
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOCDo you have
CONFIG_CHECKPOINT_RESTOREenabled? Does it go away if you disable it?I think I foud at least one bug in the
MSG_COPYstuff: it leaks the "copy" allocation ifmode == SEARCH_LESSEQUAL
Hello, Linus.
Sorry, but I don't see copy allocation leak.
Dummy message is allocated always in msgflg has MSG_COPY flag being set. Also prepare_copy() use msgtyp as a message number to copy and thus set it to 0.
Quoted 6 lines
but maybe I'm misreading it. And that shouldn't cause the problem you see, but it's indicative of how badly tested and thought through the
MSG_COPYcode is.Totally UNTESTED leak fix appended. Stanislav?
I don't see, how this patch can help. And we should not release it until copy is done in msg_handler, because msg is equal to copy.
Dummy copy message is release either by free_copy() (if msg is error) or free_msg().
But there are two issues here definitely:
1) Poor SELinux support for message
copying. This issue was addressed by Stephen Smalley here:
https://lkml.org/lkml/2013/2/6/663
But look like he didn't sent the patch to Andrew.
2) Copy leak and queue corruption in case of
copy message wasn't found (this was mentioned by Peter in another thread; thanks for catching this, Peter!), because msg will be a valid pointer and all the "message copy" clean up logic doesn't work.
I like Peter's cleanup and fix series. But if it looks like too much changes for this old code, I have another small patch, which should fix the issue:
ipc: set msg back to -EAGAIN if copy wasn't performed
Make sure, that msg pointer is set back to error value in case of MSG_COPY flag is set and desired message to copy wasn't found. This garantees, that msg is either a error pointer or a copy address.
Otherwise last message in queue will be freed without unlinking from queue (which leads to memory corruption) plus dummy allocated copy won't be released.
ipc/msg.c +1 -0
@@ -872,6 +872,7 @@ long do_msgrcv(int msqid, void __user *buf, size_t bufsz, long msgtyp, goto out_unlock;break ; } + msg =ERR_PTR (-EAGAIN ); }else break ; msg_counter++;
Quoted 35 lines
Linus
patch.diff
ipc/msg.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)ipc/msg.c +3
-3@@ -870,6 +870,7 @@ long do_msgrcv(int msqid, void __user *buf, size_t bufsz, long msgtyp, msg =copy_msg (msg, copy);if (IS_ERR (msg)) goto out_unlock; + copy =NULL ;break ; } }else @@ -976,10 +977,9 @@ out_unlock:break ; } } -if (IS_ERR (msg)) { -free_copy (copy); +free_copy (copy); +if (IS_ERR (msg))return PTR_ERR (msg); - } bufsz =msg_handler (buf, msg, bufsz);free_msg (msg);
Signature · 2 lines
Best regards, Stanislav Kinsbursky
Quoted 12 lines · Dave Jones
Now that I have that reverted, I'm not seeing msgrcv traces any more, but I've started seeing this..
general protection fault: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC RIP: 0010:[<ffffffff812c20fb>] [<ffffffff812c20fb>] free_msg+0x2b/0x40 Call Trace: [<ffffffff812c289f>] freeque+0xcf/0x140 [<ffffffff812c2a93>] msgctl_down.constprop.9+0x183/0x200 [<ffffffff812c2da9>] sys_msgctl+0x139/0x400 [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b
Quoted 1 line
Looks like seg was already kfree'd.
Hmm.
I have a suspicion.
The use of ipc_rcu_getref/ipc_rcu_putref() seems a bit iffy.
In particular, the refcount is not an atomic variable, so we absolutely *depend* on the spinlock for it.
However, looking at "freeque()", that's not actually the case. It releases the message queue spinlock *before* it does the ipc_rcu_putref(), and it does that because the thing has become unreachable (it did a msg_rmid(), which will set ->deleted, which in turn will mean that nobody should successfully look it up any more).
HOWEVER.
While the "deleted" flag is serialized, the actual refcount is not. So in *parallel* with the freeque() call, we may have some other user that does something like
...
ipc_rcu_getref(msq);
msg_unlock(msq);
schedule();
ipc_lock_by_ptr(&msq->q_perm);
ipc_rcu_putref(msq);
if (msq->q_perm.deleted) {
err = -EIDRM;
goto out_unlock_free;
}
...which got the lock for the "deleted" test, so that part is all fine, but notice the "ipc_rcu_putref()". It can happen at the same time that freeque() also does its own ipc_rcu_putref().
So now refcount may get buggered, resulting in random early reuse, double free's or leaking of the msq.
There may be some reason I don't see why this cannot happen, but it does look suspicious. I wonder if the refcount should be an atomic value.
The alternative would be to make sure the thing is always locked (and in a rcu-read-safe region) before putref/getref. The only place (apart from the initial allocation, which doesn't matter, because nothing can see if itf that path fails) seems to be that freeque(), but I didn't check everything.
Moving the
msg_unlock(msq);
to the end of freeque() might be the way to go. It's going to cause us to hold the lock for longer, but I'm not sure we care for that path.
Guys, am I missing something? This kind of refcounting problem might well explain the rcu-freeing-time bugs reported with the scalability
patches: long-time race that just got *much* easier to expose with the higher level of parallelism?
Linus
Quoted 7 lines · Linus Torvalds
The alternative would be to make sure the thing is always locked (and in a rcu-read-safe region) before putref/getref. The only place (apart from the initial allocation, which doesn't matter, because nothing can see if itf that path fails) seems to be that
freeque(), but I didn't check everything.
Quoted 6 lines
Moving the
msg_unlock(msq);to the end of
freeque()might be the way to go. It's going to cause us to hold the lock for longer, but I'm not sure we care for that path.
Uhhuh. I think shm_destroy() has the same pattern. And I think that one has locking reasons why it has to do the shm_unlock() before tearing some things down, although I'm not sure..
The good news is that shm doesn't seem to have any users of ipc_rcu_get/putref(), so I don't see anything to race against. So it has the same buggy pattern, but I don't think it can trigger anything.
And ipc/sem.c has the same bug-pattern in freeary(). It does "sem_unlock(sma)" followed by "ipc_rcu_putref(sma);", and it *does* seem to have code that that can race against (sem_lock_and_putref()). The whole "sem_putref()" tries to be careful and gets the lock for the last ref, but getting the lock doesn't help when freeary() does the refcount access without it.
The semaphore case seems to argue fairly strongly for an "atomic_t refcount", because right now it does a lot of "sem_putref()" stuff that just gets the lock for the putref. So not only is that insufficient (if it races against freeary(), but it's also more expensive than just an atomic refcount would have been.
I dunno. I'm still not sure this is triggerable, but it looks bad. But both the semaphore case and the msg cases seem to be solvable by moving the unlock down, and shm seem to have no getref/putref users to race with, so this (whitespace-damaged) patch *may* be sufficient:
ipc/msg.c +0 -0
@@ -284,7 +284,6 @@ static void freeque(struct ipc_namespace *ns,
struct kern_ipc_perm *ipcp)
expunge_all(msq, -EIDRM);
ss_wakeup(&msq->q_senders, 1);
msg_rmid(ns, msq);- msg_unlock(msq);
tmp = msq->q_messages.next;
while (tmp != &msq->q_messages) {
@@ -297,6 +296,7 @@ static void freeque(struct ipc_namespace *ns,
struct kern_ipc_perm *ipcp)
atomic_sub(msq->q_cbytes, &ns->msg_bytes);
security_msg_queue_free(msq);
ipc_rcu_putref(msq);+ msg_unlock(msq);
}
/*
ipc/sem.c +0 -0
@@ -766,12 +766,12 @@ static void freeary(struct ipc_namespace *ns,
struct kern_ipc_perm *ipcp)
/* Remove the semaphore set from the IDR */
sem_rmid(ns, sma);- sem_unlock(sma);
wake_up_sem_queue_do(&tasks);
ns->used_sems -= sma->sem_nsems;
security_sem_free(sma);
ipc_rcu_putref(sma);+ sem_unlock(sma);
}
static unsigned long copy_semid_to_user(void __user *buf, structsemid64_ds *in, int version)
(I didn't check very carefully, it's possible that we end up having some locking problem if we move the unlocks down later, but it *looks* fine)
Anybody?
Linus
Quoted 5 lines · Linus Torvalds
I dunno. I'm still not sure this is triggerable, but it looks bad. But both the semaphore case and the msg cases seem to be solvable by moving the unlock down, and shm seem to have no getref/putref users to race with, so this (whitespace-damaged) patch *may* be sufficient:
Well, the patch doesn't seem to cause any problems, at least neither lockdep nor spinlock sleep debugging complains. I have no idea whether it actually fixes any problems, though.
I do wonder if this might explain the problem Emmanuel saw. A double free of a RCU-freeable object would possibly result in exactly the kind of mess that Emmanuel reported with the semaphore scalability patches.
Emmanuel, can you try the attached patch? I think it applies cleanly on top of the scalability series too without any changes, but I didn't check if the patches perhaps changed some of the naming or something.
Linus
ipc/msg.c | 2 +- ipc/sem.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
ipc/msg.c +1 -1
@@ -284,7 +284,6 @@ static void freeque(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)expunge_all (msq, -EIDRM );ss_wakeup (&msq->q_senders,1 );msg_rmid (ns, msq); -msg_unlock (msq); tmp = msq->q_messages.next;while (tmp != &msq->q_messages) {
@@ -297,6 +296,7 @@ static void freeque(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)atomic_sub (msq->q_cbytes, &ns->msg_bytes);security_msg_queue_free (msq);ipc_rcu_putref (msq); +msg_unlock (msq); }/*
ipc/sem.c +1 -1
@@ -766,12 +766,12 @@ static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)/* Remove the semaphore set from the IDR */ sem_rmid (ns, sma); -sem_unlock (sma);wake_up_sem_queue_do (&tasks); ns->used_sems -= sma->sem_nsems;security_sem_free (sma);ipc_rcu_putref (sma); +sem_unlock (sma); }static unsigned long copy_semid_to_user (void __user *buf,struct semid64_ds *in,int version)
Hi Linus,
Quoted 3 lines · Linus Torvalds
Emmanuel, can you try the attached patch? I think it applies cleanly on top of the scalability series too without any changes, but I didn't check if the patches perhaps changed some of the naming or something.
I had to slightly modify the patch since it wouldn't match the changes introduced by 7-7-ipc-sem-fine-grained-locking-for-semtimedop.patch, hope that was the right thing to do. So, what I tried was: original 7 patches + the one liner + your patch blindly modified by me on the top of 3.9-rc4 and I'm still having twilight zone issues.
Quoted 8 lines · Emmanuel Benisty
Hi Linus,
Quoted 3 lines · Linus Torvalds
Emmanuel, can you try the attached patch? I think it applies cleanly on top of the scalability series too without any changes, but I didn't check if the patches perhaps changed some of the naming or something.
Quoted 5 lines
I had to slightly modify the patch since it wouldn't match the changes introduced by 7-7-ipc-sem-fine-grained-locking-for-semtimedop.patch, hope that was the right thing to do. So, what I tried was: original 7 patches + the one liner + your patch blindly modified by me on the top of 3.9-rc4 and I'm still having twilight zone issues.
Not sure which one liner you refer to, but, if you haven't already done so, please try with these fixes (queued in linux-next):
http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/?id=a9cead0347283f3e72a39e7b76a3cc479b048e51
http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/?id=4db64b89525ac357cba754c3120065adddd9ec31
I've been trying to reproduce your twilight zone problem on five different machines now without any luck. Is there anything you're doing to trigger the issue? Does the machine boot ok and then do weird things, say after X starts, open some program, etc?
Thanks,
Davidlohr
Hi Davidlohr,
Quoted 10 lines · Davidlohr Bueso
Not sure which one liner you refer to, but, if you haven't already done so, please try with these fixes (queued in linux-next):
http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/?id=a9cead0347283f3e72a39e7b76a3cc479b048e51
http://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git/commit/?id=4db64b89525ac357cba754c3120065adddd9ec31I've been trying to reproduce your twilight zone problem on five different machines now without any luck. Is there anything you're doing to trigger the issue? Does the machine boot ok and then do weird things, say after X starts, open some program, etc?
I was missing a9cead0, thanks. What I usually do is starting a standard session, which looks like this:
init-+-5*[agetty]
|-bash---startx---xinit-+-X---2*[{X}]
| `-dwm---sh---sleep
|-bash---chromium-+-chromium
| |-chromium-+-chromium
| | `-2*[{chromium}]
|
|-chromium-sandbo---chromium---chromium---4*[chromium---4*[{chromium}]]
| `-65*[{chromium}]
|-crond
|-dbus-daemon
|-klogd
|-syslogd
|-tmux-+-alsamixer
| |-bash---bash
| |-bash
| |-htop
| `-newsbeuter---{newsbeuter}
|-udevd
|-urxvtd-+-bash---pstree
| `-bash---tmux`-wpa_supplicant
Then I start building a random package and the problems start. They may also happen without compiling but this seems to trigger the bug quite quickly. Anyway, some progress here, I hope: dmesg seems to be willing to reveal some secrets (using some pastebin service since this is pretty big):
https://gist.github.com/anonymous/5275120
Thanks.
Quoted 4 lines · Emmanuel Benisty
Then I start building a random package and the problems start. They may also happen without compiling but this seems to trigger the bug quite quickly.
I suspect it's about preemption, and the build just results in enough scheduling load that you start hitting whatever race there is.
Quoted 5 lines
Anyway, some progress here, I hope: dmesg seems to be willing to reveal some secrets (using some pastebin service since this is pretty big):
That looks like exactly the exit_sem() bug that Davidlohr was talking about, where the
/* exit_sem raced with IPC_RMID, nothing to do */
if (IS_ERR(sma))
continue;should be moved to *before* the
sem_lock(sma, NULL, -1);
call. And apparently the bug I had found is already fixed in -next.
Linus
Quoted 22 lines · Linus Torvalds
Quoted 4 lines · Emmanuel Benisty
Then I start building a random package and the problems start. They may also happen without compiling but this seems to trigger the bug quite quickly.
I suspect it's about preemption, and the build just results in enough scheduling load that you start hitting whatever race there is.
Quoted 5 lines
Anyway, some progress here, I hope: dmesg seems to be willing to reveal some secrets (using some pastebin service since this is pretty big):
That looks like exactly the
exit_sem()bug that Davidlohr was talking about, where the/* exit_sem raced with IPC_RMID, nothing to do */ if (IS_ERR(sma)) continue;
Quoted 5 lines
should be moved to *before* the
sem_lock(sma, NULL, -1);call. And apparently the bug I had found is already fixed in -next.
I just tried the 7 original patches + the 2 one liners from -next + modified Linus' patch (attached) on the top of 3.9-rc4 using PREEMPT_NONE and after moving sem_lock(sma, NULL, -1) as explained above. I was building two packages at the same time, went away for 30 seconds, came back and everything froze as soon as I touched the laptop's touchpad. Maybe a coincidence but anyway... Another shot in the dark, I had this weird message when trying to build gcc:
semop(2): encountered an error: Identifier removed
ipc/msg.c | 2 +- ipc/sem.c | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
ipc/msg.c +1 -1
@@ -284,7 +284,6 @@ static void freeque(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)expunge_all (msq, -EIDRM );ss_wakeup (&msq->q_senders,1 );msg_rmid (ns, msq); -msg_unlock (msq); tmp = msq->q_messages.next;while (tmp != &msq->q_messages) {
@@ -297,6 +296,7 @@ static void freeque(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)atomic_sub (msq->q_cbytes, &ns->msg_bytes);security_msg_queue_free (msq);ipc_rcu_putref (msq); +msg_unlock (msq); }/*
ipc/sem.c +1 -1
@@ -766,12 +766,12 @@ static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)/* Remove the semaphore set from the IDR */ sem_rmid (ns, sma); -sem_unlock (sma,-1 );wake_up_sem_queue_do (&tasks); ns->used_sems -= sma->sem_nsems;security_sem_free (sma);ipc_rcu_putref (sma); +sem_unlock (sma,-1 ); }static unsigned long copy_semid_to_user (void __user *buf,struct semid64_ds *in,int version)
Quoted 3 lines · Emmanuel Benisty
I just tried the 7 original patches + the 2 one liners from -next + modified Linus' patch (attached)
.. that patch looks fine.
Quoted 7 lines
on the top of 3.9-rc4 using
PREEMPT_NONEand after movingsem_lock(sma, NULL, -1)as explained above. I was building two packages at the same time, went away for 30 seconds, came back and everything froze as soon as I touched the laptop's touchpad. Maybe a coincidence but anyway... Another shot in the dark, I had this weird message when trying to build gcc:
semop(2): encountered an error: Identifier removed
This came from the gcc build?
That's just crazy. No normal app uses sysv semaphores. I have an older gcc build environment, and some grepping shows it has some ipc semaphore use in the libstdc++ testsuite, and some libmudflap hooks, but that should be very very minor.
You seem to trigger errors really trivially easily, which is really odd. It's sounding less and less like some subtle race, and more like the error just happens all the time. If you can make even the gcc build generate errors, I don't think they can be some rare blue-moon thing.
I notice that your dmesg says that your kernel is compiled by gcc-4.8.1 prerelease. Is there any chance that you could try to install a known-stable gcc, like 4.7.2 or something. It's entirely possible that it's a kernel bug even if it's triggered by some more aggressive compiler optimization or something, but it would be really good to try to see if this might be gcc-specific.
For example, I wonder if your gcc might miscompile idr_alloc() or something, so that we get the same ID for different ipc objects. That would certainly potentially cause chaos.
Hmm?
Linus
Quoted 5 lines · Linus Torvalds
Quoted 3 lines
Another shot in
the dark, I had this weird message when trying to build gcc:
semop(2): encountered an error: Identifier removedThis came from the gcc build?
yes, very early in the build process, IIRC this line was repeated a few times and the build just stalled.
Quoted 11 lines
That's just crazy. No normal app uses sysv semaphores. I have an older gcc build environment, and some grepping shows it has some ipc semaphore use in the libstdc++ testsuite, and some libmudflap hooks, but that should be very very minor.
I notice that your dmesg says that your kernel is compiled by gcc-4.8.1 prerelease. Is there any chance that you could try to install a known-stable gcc, like 4.7.2 or something. It's entirely possible that it's a kernel bug even if it's triggered by some more aggressive compiler optimization or something, but it would be really good to try to see if this might be gcc-specific.
I could build a kernel on another machine on which 4.7.2 is installed, kernel oops'd as well:
http://i.imgur.com/uk6gmq1.jpg
FWIW, I have a few things disabled in my config so here is the one I used, maybe I'm missing something (but again, everything works perfectly with your tree):
https://gist.github.com/anonymous/5275566
Lastly, just FTR, I tried Andrew's 3.9-rc4-mm1 and got the same issues, unsurprisingly I guess.
Quoted 6 lines · Emmanuel Benisty
On Sat, Mar 30, 2013 at 12:10 PM, Linus Torvalds
Quoted 2 lines
This came from the gcc build?
yes, very early in the build process, IIRC this line was repeated a few times and the build just stalled.
Ok, we're bringing out the crazy hacks now.
The attached patch is just insane, doesn't really even work in general, and only even compiles on 64-bit. But it should work in *practice* to find if somebody adds the same RCU head to the RCU lists twice, and ignore the second time it happens (and give a warning that hopefully pinpoints the backtrace).
It's ugly. It's broken. It may not work. In other words, I'm not proud of it. But you seem to be the only one able to trigger the issue easily, willing to try crazy crap, so "tag, you're it". Maybe this gives us more information. And maybe it doesn't, and I'm totally wrong about the whole "rcu head added twice" theory.
Linus
include/linux/types.h | 1 + include/net/dst.h | 2 +- kernel/rcu.h | 1 + kernel/rcutree.c | 6 ++++++ 4 files changed, 9 insertions(+), 1 deletion(-)
include/linux/types.h +1 -0
@@ -209,6 +209,7 @@ struct ustat {struct callback_head {struct callback_head *next;void (*func)(struct callback_head *head); +unsigned long magic; };#define rcu_head callback_head
include/net/dst.h +1 -1
@@ -89,7 +89,7 @@ struct dst_entry {* (L1_CACHE_SIZE would be too much) */ #ifdef CONFIG_64BIT -long __pad_to_align_refcnt[2 ]; +long __pad_to_align_refcnt[1 ];#endif /* * __refcnt wants to be on a different cache line from
kernel/rcu.h +1 -0
@@ -98,6 +98,7 @@ static inline bool __rcu_reclaim(char *rn, struct rcu_head *head) {unsigned long offset = (unsigned long )head->func; + head->magic =0 ;if (__is_kfree_rcu_offset (offset)) {RCU_TRACE (trace_rcu_invoke_kfree_callback (rn, head, offset));kfree ((void *)head - offset);
kernel/rcutree.c +6 -0
@@ -2221,6 +2221,9 @@ static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp, } } +/* Unlikely bit-pattern to check double RCU calls! */ +#define RCU_HEAD_MAGIC ((unsigned long)(0xfeeddead1acef8edll)) +/* * Helper function for call_rcu() and friends. The cpu argument will * normally be -1, indicating "currently running CPU". It may specify
@@ -2235,9 +2238,12 @@ __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),struct rcu_data *rdp;WARN_ON_ONCE ((unsigned long )head &0x3 );/* Misaligned rcu_head! */ +if (WARN_ON_ONCE (head->magic ==RCU_HEAD_MAGIC )) +return ;debug_rcu_head_queue (head); head->func = func; head->next =NULL ; + head->magic =RCU_HEAD_MAGIC ;/* * Opportunistically note grace-period endings and beginnings.
Hi Linus,
Quoted 16 lines · Linus Torvalds
Quoted 6 lines · Emmanuel Benisty
On Sat, Mar 30, 2013 at 12:10 PM, Linus Torvalds
Quoted 2 lines
This came from the gcc build?
yes, very early in the build process, IIRC this line was repeated a few times and the build just stalled.
Ok, we're bringing out the crazy hacks now.
The attached patch is just insane, doesn't really even work in general, and only even compiles on 64-bit. But it should work in *practice* to find if somebody adds the same RCU head to the RCU lists twice, and ignore the second time it happens (and give a warning that hopefully pinpoints the backtrace).
Quoted 5 lines
It's ugly. It's broken. It may not work. In other words, I'm not proud of it. But you seem to be the only one able to trigger the issue easily, willing to try crazy crap, so "tag, you're it". Maybe this gives us more information. And maybe it doesn't, and I'm totally wrong about the whole "rcu head added twice" theory.
That's all I could get so far:
https://gist.github.com/anonymous/5279255
Losing wireless is generally the start signal of controlled demolition of the machine.
Quoted 30 lines · Emmanuel Benisty
Quoted 22 lines · Linus Torvalds
Quoted 4 lines · Emmanuel Benisty
Then I start building a random package and the problems start. They may also happen without compiling but this seems to trigger the bug quite quickly.
I suspect it's about preemption, and the build just results in enough scheduling load that you start hitting whatever race there is.
Quoted 5 lines
Anyway, some progress here, I hope: dmesg seems to be willing to reveal some secrets (using some pastebin service since this is pretty big):
That looks like exactly the
exit_sem()bug that Davidlohr was talking about, where the/* exit_sem raced with IPC_RMID, nothing to do */ if (IS_ERR(sma)) continue;Quoted 5 lines
should be moved to *before* the
sem_lock(sma, NULL, -1);call. And apparently the bug I had found is already fixed in -next.
Quoted 8 lines
I just tried the 7 original patches + the 2 one liners from -next + modified Linus' patch (attached) on the top of 3.9-rc4 using
PREEMPT_NONEand after movingsem_lock(sma, NULL, -1)as explained above. I was building two packages at the same time, went away for 30 seconds, came back and everything froze as soon as I touched the laptop's touchpad. Maybe a coincidence but anyway... Another shot in the dark, I had this weird message when trying to build gcc:
semop(2): encountered an error: Identifier removed
*sigh*. I had high hopes for this being the bug triggering your issue, specially after seeing exit_sem() in the trace.
Emmanuel, just to be sure, does your changes reflect the patch below? Specially dropping the rcu read lock before the continue statement (sorry for not mentioning this in the last email).
Anyway, this is still a bug. Andrew, the patch below applies to linux-next, please queue this up if you don't have any objections.
Thanks,
Davidlohr
---8<---
From: Davidlohr Bueso <[email protected]>
Subject: [PATCH] ipc, sem: do not call sem_lock when bogus sma
In exit_sem() we attempt to acquire the sma->sem_perm.lock by calling sem_lock() immediately after obtaining sma. However, if sma isn't valid, then calling sem_lock() will tend to do bad things.
Move the sma error check right after the sem_obtain_object_check() call instead.
--- ipc/sem.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-)
ipc/sem.c +5 -5
@@ -1867,8 +1867,7 @@ void exit_sem(struct task_struct *tsk)struct sem_array *sma;struct sem_undo *un;struct list_head tasks; -int semid; -int i; +int semid, i;rcu_read_lock (); un = list_entry_rcu(ulp->list_proc.next,
@@ -1884,12 +1883,13 @@ void exit_sem(struct task_struct *tsk) } sma =sem_obtain_object_check (tsk->nsproxy->ipc_ns, un->semid); -sem_lock (sma,NULL ,-1 ); -/* exit_sem raced with IPC_RMID, nothing to do */ -if (IS_ERR (sma)) +if (IS_ERR (sma)) { +rcu_read_unlock ();continue ; + } +sem_lock (sma,NULL ,-1 ); un =__lookup_undo (ulp, semid);if (un ==NULL ) {/* exit_sem raced with IPC_RMID+semget() that created
Signature · 1 line
1.7.11.7
Quoted 15 lines · Davidlohr Bueso
ipc/sem.c +1
-2@@ -1867,8 +1867,7 @@ void exit_sem(struct task_struct *tsk)struct sem_array *sma;struct sem_undo *un;struct list_head tasks; -int semid; -int i; +int semid, i;rcu_read_lock (); un = list_entry_rcu(ulp->list_proc.next,
ipc/sem.c void exit_sem(struct task_struct *tsk) · line 1885
Quoted 3 lines
ipc/sem.c +0
-0@@ -1884,12 +1883,13 @@ void exit_sem(struct task_struct *tsk) } sma =sem_obtain_object_check (tsk->nsproxy->ipc_ns, un->semid);
Should we use "semid" here, like Linus suggested, instead of "un->semid"?
Quoted 14 lines
ipc/sem.c +4
-3-sem_lock (sma,NULL ,-1 ); -/* exit_sem raced with IPC_RMID, nothing to do */ -if (IS_ERR (sma)) +if (IS_ERR (sma)) { +rcu_read_unlock ();continue ; + } +sem_lock (sma,NULL ,-1 ); un =__lookup_undo (ulp, semid);if (un ==NULL ) {/* exit_sem raced with IPC_RMID+semget() that created
Signature · 1 line
All rights reversed.
Quoted 2 lines · Rik van Riel
Should we use "semid" here, like Linus suggested, instead of "
un->semid"?
As Davidlohr noted, in linux-next the rcu read-lock is held over the whole thing, so no, un->semid should be stable once "un" has been re-looked-up under the semaphore lock.
In mainline, the problem is that the "sem_lock_check()" is done with "un->semid" *after* we've dropped the RCU read-lock, so "un" at that point is not reliable (it could be free'd at any time underneath us).
That said, I really *really* hate what both mainline and linux-next do with the RCU read lock, and linux-next is arguably worse.
The whole "take the RCU lock in one place, and release it in another" is confusing and bug-prone as hell. And linux-next made it worse: now sem_lock() no longer takes the read-lock (it expects the caller to take it), but sem_unlock() still drops the read-lock. This is all just f*cking crazy.
The rule should be that the rcu read-lock is always and released at the same "level". For example, find_alloc_undo() should just be called with (and unconditionaly return with) the rcu read-lock held, and if it needs to actually do an allocation, it can drop the rcu lock for the duration of the allocation.
This whole "conditional locking" depending on error returns and on whether we have undo's etc is bug-prone and confusing. And when you have totally different locking rules for "sem_lock()" vs "sem_unlock()", you know you're confused.
Linus
Hi Davidlohr,
Quoted 2 lines · Davidlohr Bueso
Specially dropping the rcu read lock before the continue statement (sorry for not mentioning this in the last email).
I was missing this indeed, thanks. Still the same issues however... I'll do some more testing on the same machine but with a totally different environment, within tomorrow hopefully.
Quoted 6 lines · Emmanuel Benisty
I had to slightly modify the patch since it wouldn't match the changes introduced by 7-7-ipc-sem-fine-grained-locking-for-semtimedop.patch, hope that was the right thing to do. So, what I tried was: original 7 patches + the one liner + your patch blindly modified by me on the top of 3.9-rc4 and I'm still having twilight zone issues.
Ok, please send your patch so that I can double-check what you did, but it was simple enough that you probably did the right thing.
Sad. Your case definitely looks like a double rcu-free, as shown by the fact that when you enabled SLUB debugging the oops happened with the use-after-free pattern (it's __rcu_reclaim() doing the "head->func(head);" thing, and "func" is 0x6b6b6b6b6b6b6b6b, so "head" has already been free'd once).
So ipc_rcu_putref() and a refcounting error looked very promising.as a potential explanation.
The 'un' undo structure is also free'd with rcu, but the locking around that seems much more robust. The undo entry is on two lists (sma->list_id, under sma->sem_perm.lock and ulp->list_proc, under ulp->lock). But those locks are actually tested with assert_spin_locked() in all the relevant places, and the code actually looks sane. So I had high hopes for ipc_rcu_putref()...
Hmm. Except for exit_sem() that does odd things. You have preemption enabled, don't you? exit_sem() does a lookup of the first list_proc entry under tcy_read_lock to lookup un->semid, and then it drops the rcu read lock. At which point "un" is no longer reliable, I think. But then it still uses "un->semid", rather than the stable value it looked up under the rcu read lock. Which looks bogus.
So I'd like you to test a few more things:
(a) In exit_sem(), can you change the
sma = sem_lock_check(tsk->nsproxy->ipc_ns, un->semid);
to use just "semid" rather than "un->semid", because I don'tthink "un" is stable here.
(b) does the problem go away if you change disable CONFIG_PREEMPT
(perhaps to PREEMPT_NONE or PREEMPT_VOLUNTARY?)
Linus
Quoted 34 lines · Linus Torvalds
Quoted 6 lines · Emmanuel Benisty
I had to slightly modify the patch since it wouldn't match the changes introduced by 7-7-ipc-sem-fine-grained-locking-for-semtimedop.patch, hope that was the right thing to do. So, what I tried was: original 7 patches + the one liner + your patch blindly modified by me on the top of 3.9-rc4 and I'm still having twilight zone issues.
Ok, please send your patch so that I can double-check what you did, but it was simple enough that you probably did the right thing.
Sad. Your case definitely looks like a double rcu-free, as shown by the fact that when you enabled SLUB debugging the oops happened with the use-after-free pattern (it's
__rcu_reclaim()doing the "head->func(head);" thing, and "func" is0x6b6b6b6b6b6b6b6b, so "head" has already been free'd once).So
ipc_rcu_putref()and a refcounting error looked very promising.as a potential explanation.The 'un' undo structure is also free'd with rcu, but the locking around that seems much more robust. The undo entry is on two lists (
sma->list_id, undersma->sem_perm.lockandulp->list_proc, underulp->lock). But those locks are actually tested withassert_spin_locked()in all the relevant places, and the code actually looks sane. So I had high hopes foripc_rcu_putref()...Hmm. Except for
exit_sem()that does odd things. You have preemption enabled, don't you?exit_sem()does a lookup of the firstlist_procentry undertcy_read_lockto lookupun->semid, and then it drops the rcu read lock. At which point "un" is no longer reliable, I think. But then it still uses "un->semid", rather than the stable value it looked up under the rcu read lock. Which looks bogus.
Quoted 8 lines
So I'd like you to test a few more things:
(a) In
exit_sem(), can you change thesma = sem_lock_check(tsk->nsproxy->ipc_ns, un->semid); to use just "semid" rather than "un->semid", because I don'tthink "un" is stable here.
Well that's not really the case in the new code. We don't drop the rcu read lock until the end of the loop, in sem_unlock(). However, I just noticed that we're checking sma for error after trying to acquire sma->sem_perm.lock:
sma = sem_obtain_object_check(tsk->nsproxy->ipc_ns, un->semid); sem_lock(sma, NULL, -1); /* exit_sem raced with IPC_RMID, nothing to do */ if (IS_ERR(sma)) continue;
The IS_ERR(sma) check should be right after the sem_obtain_object_check() call instead.
Quoted 13 lines · Andrew Morton
Quoted 10 lines · On Tue, 26 Mar 2013 15:28:52 -0400 Dave Jones <[email protected]> wrote:
Quoted 4 lines · Andrew Morton
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?Ok, with that reverted it's been grinding away for a few hours without incident. Normally I see the oops within a minute or so.
Quoted 7 lines
OK, thanks, I queued a revert:
From: Andrew Morton <[email protected]>
Subject: revert "ipc: don't allocate a copy larger than max"Revert
88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.
I owe Peter an apology. I just hit it again with that backed out. Andrew, might as well drop that revert.
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60 PGD 10fd95067 PUD 10f767067 PMD 0 Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC Modules linked in: phonet netrom llc2 af_key rose af_rxrpc caif_socket caif can_raw cmtp kernelcapi ipt_ULOG nfnetlink can_bcm can scsi_transport_iscsi af_802154 irda ax25 atm ipx x25 p8023 p8022 appletalk pppoe decnet pppox ppp_generic nfc rds slhc psnap crc_ccitt llc lockd sunrpc ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 xt_conntrack nf_conntrack ip6table_filter ip6_tables raid0 snd_hda_codec_realtek snd_hda_intel btusb snd_hda_codec bluetooth microcode serio_raw snd_pcm edac_core pcspkr snd_page_alloc rfkill snd_timer snd soundcore r8169 mii vhost_net tun macvtap macvlan kvm_amd kvm radeon backlight drm_kms_helper ttm CPU 2 Pid: 958, comm: trinity-child20 Not tainted 3.9.0-rc4+ #7 Gigabyte Technology Co., Ltd. GA-MA78GM-S2H/GA-MA78GM-S2H RIP: 0010:[<ffffffff812c24ca>] [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60 RSP: 0018:ffff880117bb5e88 EFLAGS: 00010246 RAX: ffffffffffffffff RBX: 0000000000000004 RCX: 0000000000000078 RDX: 0000000000000004 RSI: fffffffffffffffe RDI: 000000000000000f RBP: ffff880117bb5e88 R08: 0000000000000004 R09: 0000000000000001 R10: ffff880117bb8000 R11: 0000000000000001 R12: fffffffffffffffe R13: ffff88010fd10308 R14: ffff88010fd10258 R15: ffffffffffffffff FS: 00007fa89c256740(0000) GS:ffff88012aa00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 000000000000000f CR3: 000000010f76c000 CR4: 00000000000007e0 DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400 Process trinity-child20 (pid: 958, threadinfo ffff880117bb4000, task ffff880117bb8000) Stack: ffff880117bb5f68 ffffffff812c3746 0000000000000000 ffff880117bb8000 ffff880117bb8000 ffff880117bb8000 ffffffff81c7ace0 ffffffff812c2430 0000000000000004 0000000000000000 000000000000ffff 00000000652a928e Call Trace: [<ffffffff812c3746>] do_msgrcv+0x1a6/0x5f0 [<ffffffff812c2430>] ? msg_security+0x10/0x10 [<ffffffff810b6c55>] ? trace_hardirqs_on_caller+0x115/0x1a0 [<ffffffff8134b39e>] ? trace_hardirqs_on_thunk+0x3a/0x3f [<ffffffff812c3ba5>] sys_msgrcv+0x15/0x20 [<ffffffff816cd942>] system_call_fastpath+0x16/0x1b Code: c3 48 c7 c0 f2 ff ff ff eb e5 0f 1f 80 00 00 00 00 66 66 66 66 90 55 83 fa 02 48 89 e5 74 3a 7e 28 83 fa 03 74 13 83 fa 04 75 0a <48> 39 37 b8 01 00 00 00 7e 02 31 c0 5d c3 48 3b 37 74 f7 b8 01 RIP [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60
I think I wasn't seeing that this last week because I had inadvertantly disabled DEBUG_PAGEALLOC
and.. we're back to square one.
Dave
Quoted 27 lines · Dave Jones
Quoted 13 lines · Andrew Morton
Quoted 10 lines · On Tue, 26 Mar 2013 15:28:52 -0400 Dave Jones <[email protected]> wrote:
Quoted 4 lines · Andrew Morton
Whichever way we go, we should get a wiggle on - this has been hanging around for too long. Dave, do you have time to determine whether reverting
88b9e456b1649722673ff("ipc: don't allocate a copy larger than max") fixes things up?Ok, with that reverted it's been grinding away for a few hours without incident. Normally I see the oops within a minute or so.
Quoted 7 lines
OK, thanks, I queued a revert:
From: Andrew Morton <[email protected]>
Subject: revert "ipc: don't allocate a copy larger than max"Revert
88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.I owe Peter an apology. I just hit it again with that backed out. Andrew, might as well drop that revert.
BUG: unable to handle kernel NULL pointer dereference at 000000000000000f IP: [<ffffffff812c24ca>] testmsg.isra.5+0x1a/0x60
Quoted 7 lines
[...snip...]
I think I wasn't seeing that this last week because I had inadvertantly disabled
DEBUG_PAGEALLOCand.. we're back to square one.
Dave
Andrew,
I just realized you're still carrying
commit 4bea54c91bcc5451f237e6b721b0b35eccd01d17
Author: Andrew Morton <[email protected]>
Date: Fri Apr 26 10:55:12 2013 +1000
revert "ipc: don't allocate a copy larger than max"
Revert 88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.
Please drop.
As quoted above, the testing on mainline that attributed the observed oops to the reverted patch was due to a config error.
As Linus pointed out below, this patch fixes real bugs.
Quoted 27 lines · Sasha Levin
Quoted 8 lines · Linus Torvalds
Quoted 2 lines · Sasha Levin
By just playing with the 'msgsz' parameter with
MSG_COPYset.Hmm. Looking closer, I suspect you're testing without commit
88b9e456b164("ipc: don't allocate a copy larger than max"). That should limit the size passed in toprepare_copy->load_copytomsg_ctlmax.That commit has a revert in the -next trees, do we need a revert of the revert?
commit
ff6577a3e714ccae02d4400e989762c19c37b0b3
Author: Andrew Morton <[email protected]>
Date: Wed Mar 27 10:24:02 2013 +1100revert "ipc: don't allocate a copy larger than max"
Revert
88b9e456b164. Dave has confirmed that this was causing oopses during trinity testing.Cc Peter Hurley <[email protected]>Cc Stanislav Kinsbursky <[email protected]>Reported-by Dave Jones <[email protected]>Cc <[email protected]>Signed-off-by Andrew Morton <[email protected]>
Quoted 19 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
On a 24 CPU system, performance numbers with the semop-multi test with N threads and N semaphores, look like this:
Quoted 13 lines
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
After testing these patches with my Oracle Swingbench DSS workload, I can say that there are significant improvements. The ipc lock contention was reduced drastically, specially with higher amounts of benchmark users. As a result, the overall %sys time went down as well. Furthermore, throughput (in transactions per second) was increased.
TPS:
100 users: 1257.21 (vanilla) 2805.06 (v3 patchset) 400 users: 1437.57 (vanilla) 2664.67 (v3 patchset) 800 users: 1236.89 (vanilla) 2750.73 (v3 patchset)
ipc lock contention:
100 users: 8,74% (vanilla) 3.17% (v3 patchset) 400 users: 21,86% (vanilla) 5.23% (v3 patchset) 800 users 84,35% (vanilla) 7.39% (v3 patchset)
As seen with perf, the ipc lock isn't even the main source of contention anymore. Also, no matter how many benchmark users, the lock's user is mostly semctl_main() .
100 users:
3.17% oracle [kernel.kallsyms] [k] _raw_spin_lock
|
--- _raw_spin_lock
|
|--50.53%-- sem_lock
| |
| |--82.60%-- semctl_main
| --17.40%-- sys_semtimedop400 users:
5.23% oracle [kernel.kallsyms] [k] _raw_spin_lock
|
--- _raw_spin_lock
|
|--75.81%-- sem_lock
| |
| |--94.09%-- semctl_main
| --5.91%-- sys_semtimedop800 users:
7.39% oracle [kernel.kallsyms] [k] _raw_spin_lock
|
--- _raw_spin_lock
|
|--81.71%-- sem_lock
| |
| |--64.98%-- semctl_main
| --35.02%-- sys_semtimedopThanks,
Davidlohr
Quoted 5 lines · Davidlohr Bueso
ipc lock contention:
100 users: 8,74% (vanilla) 3.17% (v3 patchset) 400 users: 21,86% (vanilla) 5.23% (v3 patchset) 800 users 84,35% (vanilla) 7.39% (v3 patchset)
Ok, I'd call that pretty much "solved". Sure, it's still visible, but for being a benchmark that apparently does little else than pound on those sysv semaphores, I think we can consider it pretty much fine. I'm going to assume that anybody who actually then does any real work (ie a database) is never going to see even close to this bad contention.
Good job, Rik. I'm assuming we'll be merging this during the 3.10 merge window, and hopefully the merge conflicts will be sorted out too. Rik, Peter, can you look at each others patches and see if you can get that sorted out for Andrew?
Linus
Quoted 14 lines · Linus Torvalds
Quoted 5 lines · Davidlohr Bueso
ipc lock contention:
100 users: 8,74% (vanilla) 3.17% (v3 patchset) 400 users: 21,86% (vanilla) 5.23% (v3 patchset) 800 users 84,35% (vanilla) 7.39% (v3 patchset)Ok, I'd call that pretty much "solved". Sure, it's still visible, but for being a benchmark that apparently does little else than pound on those sysv semaphores, I think we can consider it pretty much fine. I'm going to assume that anybody who actually then does any real work (ie a database) is never going to see even close to this bad contention.
Quoted 4 lines
Good job, Rik. I'm assuming we'll be merging this during the 3.10 merge window, and hopefully the merge conflicts will be sorted out too. Rik, Peter, can you look at each others patches and see if you can get that sorted out for Andrew?
Will do.
I will rebase this series on top of what is in linux-next.
Signature · 1 line
All rights reversed.
Quoted 19 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
On a 24 CPU system, performance numbers with the semop-multi test with N threads and N semaphores, look like this:
Quoted 12 lines
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
Hi Rik,
I plugged this set into enterprise -rt kernel, and beat on four boxen. I ran into no trouble while giving boxen a generic drubbing fwtw.
Some semop-multi -rt numbers for an abby-normal 8 node box, and a mundane 4 node box below.
32 cores+SMT, 3.0.68-rt92
numactl --hardware
available: 4 nodes (0-3)
node 0 cpus: 0 1 2 3 4 5 6 7 32 33 34 35 36 37 38 39 node 0 size: 32733 MB
node 0 free: 29910 MB
node 1 cpus: 8 9 10 11 12 13 14 15 40 41 42 43 44 45 46 47 node 1 size: 32768 MB
node 1 free: 30396 MB
node 2 cpus: 16 17 18 19 20 21 22 23 48 49 50 51 52 53 54 55 node 2 size: 32768 MB
node 2 free: 30568 MB
node 3 cpus: 24 25 26 27 28 29 30 31 56 57 58 59 60 61 62 63 node 3 size: 32767 MB
node 3 free: 28706 MB
node distances:
node 0 1 2 3
0: 10 21 21 21 1: 21 10 21 21 2: 21 21 10 21 3: 21 21 21 10
SCHED_OTHER
-v3 set +v3 set
threads
10 438485 1744599 20 376411 1580813 30 396639 1546841 40 449062 2152861 50 477594 2431344 60 446453 1874476 70 578705 2047884 80 607928 2144898 90 662136 2171074 100 622889 2295879 200 709668 2867273 300 661902 3008695 400 641758 3273250 500 614117 3403775
SCHED_FIFO
-v3 set +v3 set
threads
10 158656 914343
20 99784 1133775 30 84245 1099604 40 89725 1756577 50 85697 1607893 60 84033 1467272 70 86833 1979405 80 91451 1922250 90 92484 1990699 100 90691 2067705 200 103692 2308652 300 101161 2376921 400 103722 2417580 500 108838 2443349
64 core box (poor thing, smi free zone though), 3.0.68-rt92
numactl --hardware
available: 1 nodes (0)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 node 0 size: 8181 MB
node 0 free: 6309 MB
node distances:
node 0
0: 10
SCHED_OTHER
-v3 set +v3 set
threads
10 677534 2304417 20 451507 1873474 30 356876 1542688 40 329585 1500392 50 415761 1318676 60 403482 1380644 70 394089 1185165 80 407408 1191834 90 445764 1249156 100 430823 1245573 200 425470 1421686 300 427092 1480379 400 497900 1516599 500 421927 1551309
SCHED_FIFO
10 323560 1882543 20 226257 1806862 30 187851 1263163 40 205337 881331 50 196806 766666 60 193218 612709 70 209432 1241797 80 240445 1269146 90 219865 1482649 100 227970 1473038 200 201354 1719977 300 183728 1823074 400 175051 1808792 500 243708 1849803
Hi Rik,
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
I was trying your patchset and my machine died while building a package. I could reproduce the bug the (only) two times I tried. There's a poor quality picture here: http://i.imgur.com/MuYuyQC.jpg
Thanks.
-- Emmanuel
Quoted 4 lines · Emmanuel Benisty
I was trying your patchset and my machine died while building a package. I could reproduce the bug the (only) two times I tried. There's a poor quality picture here: http://i.imgur.com/MuYuyQC.jpg
Hmm. The original oops may well have scrolled off the window, what remains is just indicating something is corrupted in slab. Which isn't likely to give much hints..
Can you try to reproduce this with CONFIG_SLUB_DEBUG=y and CONFIG_SLUB_DEBUG_ON=y, and see if there are any earlier messages?
Linus
Hi Linus,
Quoted 6 lines · Linus Torvalds
Quoted 4 lines · Emmanuel Benisty
I was trying your patchset and my machine died while building a package. I could reproduce the bug the (only) two times I tried. There's a poor quality picture here: http://i.imgur.com/MuYuyQC.jpg
Quoted 6 lines
Hmm. The original oops may well have scrolled off the window, what remains is just indicating something is corrupted in slab. Which isn't likely to give much hints..
Can you try to reproduce this with
CONFIG_SLUB_DEBUG=yandCONFIG_SLUB_DEBUG_ON=y, and see if there are any earlier messages?
I could reproduce it but could you please let me know what would be the right tools I should use to catch the original oops? This is what I got but I doubt it will be helpful:
http://i.imgur.com/Mewi1hC.jpg
Thanks.
Quoted 5 lines · Emmanuel Benisty
I could reproduce it but could you please let me know what would be the right tools I should use to catch the original oops? This is what I got but I doubt it will be helpful:
http://i.imgur.com/Mewi1hC.jpg
In this case, I think the best thing to do would be to comment out all of drm_warn_on_modeset_not_all_locked(), because those warnings make the original problem (that probably caused the lock problem in the first place that it is warning about) scroll away.
That said, you should also take the oneliner fix that Rik posted to patch 7 (subject line: "[PATCH 7/7 part3] fix for sem_lock") and apply that, just to make sure that you aren't possibly hitting a bug with the shared-memory locking introduced by that (unusual) case.
Linus
Quoted 12 lines · Linus Torvalds
Quoted 5 lines · Emmanuel Benisty
I could reproduce it but could you please let me know what would be the right tools I should use to catch the original oops? This is what I got but I doubt it will be helpful:
http://i.imgur.com/Mewi1hC.jpgIn this case, I think the best thing to do would be to comment out all of
drm_warn_on_modeset_not_all_locked(), because those warnings make the original problem (that probably caused the lock problem in the first place that it is warning about) scroll away.
Quoted 4 lines
That said, you should also take the oneliner fix that Rik posted to patch 7 (subject line: "[PATCH 7/7 part3] fix for
sem_lock") and apply that, just to make sure that you aren't possibly hitting a bug with the shared-memory locking introduced by that (unusual) case.
Thanks Linus. I hope I got this right, here's the result (3.9-rc4, 7+1 patches): http://i.imgur.com/BebGZxV.jpg
Quoted 3 lines · Emmanuel Benisty
Thanks Linus. I hope I got this right, here's the result (3.9-rc4, 7+1 patches): http://i.imgur.com/BebGZxV.jpg
Ok, that's *slightly* more informative, but not much. At least now we see the actual page fault information, and see what the bad dereference was.
It seems to be a branch through the rcu list "->func" pointer in the rcu callbacks, and the ->func pointer has been corrupted. Instead of being a valid kernel pointer (or a "kfree_rcu_offset" marker, which is a small number between 0-4096), it has the odd value "0x0000006400000064". Two words of decimal "100", in other words.
That's not one of the usual "use-after-free" patters or anything like that, so I don't see what it would be. So I have to admit to not really having any better clue about what is going on. Sometimes the corruption pattern give a hint of what it was that overwrote it, but not here..
And you never see this problem without Rik's patches? Could you bisect *which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?
Another thing to try might be to enable SLUB debugging (ie make sure that all of
CONFIG_SLUB_DEBUG=y
CONFIG_SLUB=y
CONFIG_SLUB_DEBUG_ON=y
are set in your kernel config), which might help pin things down a bit. Sometimes that makes any allocation problems show up earlier in the path, so that it's more obvious who screwed up.
Linus
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?
With the first four patches only, I got some X server freeze (just tried once).
Quoted 7 lines
Another thing to try might be to enable SLUB debugging (ie make sure that all of
CONFIG_SLUB_DEBUG=y
CONFIG_SLUB=y
CONFIG_SLUB_DEBUG_ON=yare set in your kernel config)
My bad, I forgot to enable this in my former build, sorry. Here's what I get now:
http://i.imgur.com/G6H8KgD.jpg
Quoted 4 lines · Emmanuel Benisty
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
Quoted 8 lines
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?With the first four patches only, I got some X server freeze (just tried once).
Could you try booting with panic=1 so the kernel panics on the first oops?
Maybe that way (if we are lucky) we will be able to capture the first oops, and maybe get an idea of what causes the problem.
Signature · 1 line
All rights reversed.
Quoted 12 lines · Rik van Riel
Quoted 11 lines · Emmanuel Benisty
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?
Quoted 5 lines
Quoted 2 lines
With the first four patches only, I got some X server freeze (just tried once).
Could you try booting with panic=1 so the kernel panics on the first oops?
Sorry that should be "oops=panic"
Quoted 2 lines
Maybe that way (if we are lucky) we will be able to capture the first oops, and maybe get an idea of what causes the problem.
Signature · 1 line
All rights reversed.
Quoted 4 lines · Rik van Riel
Quoted 4 lines
Quoted 2 lines
With the first four patches only, I got some X server freeze (just tried once).
Quoted 9 lines
Quoted 2 lines
Could you try booting with panic=1 so the kernel panics on the first oops?
Sorry that should be "oops=panic"
Quoted 2 lines
Maybe that way (if we are lucky) we will be able to capture the first oops, and maybe get an idea of what causes the problem.
Sorry Rik, I get all kind of weird behaviors (wireless dies, compiling gets stuck and is impossible to kill, can't kill X) with the 4 patches+oops=panic but no trace. Here after is 7+1 patches with oops=panic boot: http://i.imgur.com/1jep1qx.jpg
Quoted 12 lines · Emmanuel Benisty
Quoted 11 lines · Rik van Riel
Quoted 6 lines
Quoted 2 lines
With the first four patches only, I got some X server freeze (just tried once).
Could you try booting with panic=1 so the kernel panics on the first oops?
Sorry that should be "oops=panic"
Quoted 7 lines
Quoted 2 lines
Quoted 2 lines
Maybe that way (if we are lucky) we will be able to capture the first oops, and maybe get an idea of what causes the problem.
Sorry Rik, I get all kind of weird behaviors (wireless dies, compiling gets stuck and is impossible to kill, can't kill X) with the 4 patches+oops=panic but no trace. Here after is 7+1 patches with oops=panic boot: http://i.imgur.com/1jep1qx.jpg
This may be a stupid question, but you re-compile and re-install the kernel modules every time you changed the kernel?
The behaviour you report with just the first four patches is so random, it sounds almost like a mismatched data structure between compiles...
Signature · 1 line
All rights reversed.
Quoted 27 lines · Rik van Riel
Quoted 16 lines · Emmanuel Benisty
Quoted 14 lines · Rik van Riel
Quoted 8 lines
Quoted 3 lines
With the first four patches only, I got some X server freeze (just tried once).
Could you try booting with panic=1 so the kernel panics on the first oops?
Sorry that should be "oops=panic"
Quoted 8 lines
Quoted 2 lines
Quoted 2 lines
Maybe that way (if we are lucky) we will be able to capture the first oops, and maybe get an idea of what causes the problem.
Sorry Rik, I get all kind of weird behaviors (wireless dies, compiling gets stuck and is impossible to kill, can't kill X) with the 4 patches+oops=panic but no trace. Here after is 7+1 patches with oops=panic boot: http://i.imgur.com/1jep1qx.jpg
Quoted 6 lines
This may be a stupid question, but you re-compile and re-install the kernel modules every time you changed the kernel?
The behaviour you report with just the first four patches is so random, it sounds almost like a mismatched data structure between compiles...
Yes it's OK. I even started from scratch just in case but I'm still getting the same weird things. Everything works just fine with a build from Linus' tree which I normally use. I'll try the patches on another machine tomorrow.
Quoted 4 lines · Emmanuel Benisty
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
Quoted 8 lines
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?With the first four patches only, I got some X server freeze (just tried once).
Another random question, just to rule something out. What video driver are you using on your system?
Signature · 1 line
All rights reversed.
Quoted 16 lines · Rik van Riel
Quoted 15 lines · Emmanuel Benisty
Quoted 2 lines · Linus Torvalds
And you never see this problem without Rik's patches?
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?
Quoted 6 lines
Quoted 2 lines
With the first four patches only, I got some X server freeze (just tried once).
Another random question, just to rule something out. What video driver are you using on your system?
intel on i3 330m
Quoted 4 lines · Emmanuel Benisty
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
Quoted 8 lines
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?With the first four patches only, I got some X server freeze (just tried once).
Going over the code again, I found a potential recursive spinlock scenario. Andrew, if you have no objections, please queue this up.
Thanks.
---8<---
From: Davidlohr Bueso <[email protected]>
Subject: [PATCH] ipc, sem: prevent possible deadlock
In semctl_main(), when cmd == GETALL, we're locking sma->sem_perm.lock (through sem_lock_and_putref), yet after the conditional, we lock it again. Unlock sma right after exiting the conditional.
--- ipc/sem.c | 1 + 1 file changed, 1 insertion(+)
ipc/sem.c +1 -0
@@ -1243,6 +1243,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, err = -EIDRM ; goto out_free; } +sem_unlock (sma,-1 ); }sem_lock (sma,NULL ,-1 );
Signature · 1 line
1.7.11.7
Rik van Riel
Reviewed-byQuoted 9 lines · Davidlohr Bueso
From: Davidlohr Bueso <[email protected]>
Subject: [PATCH] ipc, sem: prevent possible deadlockIn
semctl_main(), when cmd == GETALL, we're lockingsma->sem_perm.lock(throughsem_lock_and_putref), yet after the conditional, we lock it again. Unlock sma right after exiting the conditional.Signed-off-by Davidlohr Bueso <[email protected]>
Signature · 1 line
All rights reversed.
Quoted 39 lines · On Tue, 26 Mar 2013 10:59:27 -0700 Davidlohr Bueso <[email protected]>…
Quoted 4 lines · Emmanuel Benisty
Quoted 1 line · Linus Torvalds
And you never see this problem without Rik's patches?
Quoted 8 lines
No, never.
Quoted 4 lines
Could you bisect
*which* patch it starts with? Are the first four ones ok (the moving of the locking around, but without the fine-grained ones), for example?With the first four patches only, I got some X server freeze (just tried once).
Going over the code again, I found a potential recursive spinlock scenario. Andrew, if you have no objections, please queue this up.
Thanks.
---8<---
From: Davidlohr Bueso <[email protected]>
Subject: [PATCH] ipc, sem: prevent possible deadlockIn
semctl_main(), when cmd == GETALL, we're lockingsma->sem_perm.lock(throughsem_lock_and_putref), yet after the conditional, we lock it again. Unlock sma right after exiting the conditional.Signed-off-by Davidlohr Bueso <[email protected]>--- ipc/sem.c | 1 + 1 file changed, 1 insertion(+)ipc/sem.c +0
-0
ipc/sem.c static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, · line 1249
Quoted 7 lines
ipc/sem.c +1
-0@@ -1243,6 +1243,7 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum, err = -EIDRM ; goto out_free; } +sem_unlock (sma,-1 ); }sem_lock (sma,NULL ,-1 );
Looks right.
Do we need the locking at all? What does it actually do?
sem_lock_and_putref(sma);
if (sma->sem_perm.deleted) {
sem_unlock(sma, -1);
err = -EIDRM;
goto out_free;
}
sem_unlock(sma, -1);We're taking the lock, testing an int and then dropping the lock. What's the point in that?
Quoted 10 lines · On Tue, 26 Mar 2013 11:35:33 -0700 Andrew Morton <[email protected]>…
Do we need the locking at all? What does it actually do?
sem_lock_and_putref(sma); if (sma->sem_perm.deleted) { sem_unlock(sma, -1); err = -EIDRM; goto out_free; } sem_unlock(sma, -1);
Quoted 2 lines
We're taking the lock, testing an int and then dropping the lock. What's the point in that?
Rikpoke.
The new semctl_main() is now taking a lock, testing sma->sem_perm.deleted then dropping that lock. It looks wrong. What is that lock testing against? What prevents .deleted from changing value 1ns after we dropped that lock?
Quoted 5 lines · Emmanuel Benisty
I could reproduce it but could you please let me know what would be the right tools I should use to catch the original oops? This is what I got but I doubt it will be helpful:
http://i.imgur.com/Mewi1hC.jpg
You could use either netconsole or blockconsole. Netconsole requires a second machine to capture the information, blockconsole requires a usb key or something similar to write to. In both cases you will get the entire console output in a file, often including very helpful messages leading up to the crash.
Blockconsole currently lives here:
https://git.kernel.org/cgit/linux/kernel/git/joern/bcon2.git/
Jörn
--
Unless something dramatically changes, by 2015 we'll be largely wondering what all the fuss surrounding Linux was really about. -- Rob Enderle
Quoted 2 lines · Jörn Engel
Blockconsole currently lives here:
https://git.kernel.org/cgit/linux/kernel/git/joern/bcon2.git/
Tja, if only that were upstream...
Signature · 5 lines
Regards/Gruss,
Boris.
Sent from a fat crate under my desk. Formatting is fine.
--Quoted 5 lines · Borislav Petkov
Quoted 2 lines · Jörn Engel
Blockconsole currently lives here:
https://git.kernel.org/cgit/linux/kernel/git/joern/bcon2.git/Tja, if only that were upstream...
Linus has a pull request. If he prefers to change code until the relevant bits fit into 80x25, that is his choice. ;)
Jörn
--
The only good bug is a dead bug.
-- Starship Troopers
Quoted 19 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
The first four patches were written by Davidlohr Buesso, and reduce the hold time of the semaphore lock.
The last three patches change the sysv semaphore code locking to be more fine grained, providing a performance boost when multiple semaphores in a semaphore array are being manipulated simultaneously.
On a 24 CPU system, performance numbers with the semop-multi test with N threads and N semaphores, look like this:
Quoted 13 lines
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
Some results with semop-multi on my 4 core laptop:
vanilla v3 patchset
threads
10 5094473 10289146 20 5079946 10187923 30 5041258 10660635 40 4942786 10876009 50 5076437 10759434 60 5139024 10797032 70 5103811 10698323 80 5094850 9959675 90 5085774 10054844 100 4939547 9798291
Quoted 7 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
I'm getting the following false positives from lockdep:
[ 80.492995] =============================================
[ 80.494052] [ INFO: possible recursive locking detected ]
[ 80.494878] 3.9.0-rc4-next-20130325-sasha-00044-gcb6ef58 #315 Tainted: G W
[ 80.496228] ---------------------------------------------
[ 80.497171] trinity-child9/7210 is trying to acquire lock:
[ 80.497934] (&(&sma->sem_base[i].lock)->rlock){+.+...}, at: [<ffffffff8192da37>] newary+0x1c7/0x2a0
[ 80.499202]
[ 80.499202] but task is already holding lock:
[ 80.500031] (&(&sma->sem_base[i].lock)->rlock){+.+...}, at: [<ffffffff8192da37>] newary+0x1c7/0x2a0
[ 80.500031]
[ 80.500031] other info that might help us debug this:
[ 80.500031] Possible unsafe locking scenario:
[ 80.500031]
[ 80.500031] CPU0
[ 80.500031] ----
[ 80.500031] lock(&(&sma->sem_base[i].lock)->rlock);
[ 80.500031] lock(&(&sma->sem_base[i].lock)->rlock);
[ 80.500031]
[ 80.500031] *** DEADLOCK ***
[ 80.500031]
[ 80.500031] May be due to missing lock nesting notation
[ 80.500031]
[ 80.500031] 4 locks held by trinity-child9/7210:
[ 80.500031] #0: (&ids->rw_mutex){+++++.}, at: [<ffffffff8192a422>] ipcget+0x72/0x340
[ 80.500031] #1: (rcu_read_lock){.+.+..}, at: [<ffffffff81929b65>] ipc_addid+0x35/0x230
[ 80.500031] #2: (&(&new->lock)->rlock){+.+...}, at: [<ffffffff81929c00>] ipc_addid+0xd0/0x230
[ 80.500031] #3: (&(&sma->sem_base[i].lock)->rlock){+.+...}, at: [<ffffffff8192da37>] newary+0x1c7/0x2a0
[ 80.500031]
[ 80.500031] stack backtrace:
[ 80.500031] Pid: 7210, comm: trinity-child9 Tainted: G W 3.9.0-rc4-next-20130325-sasha-00044-gcb6ef58 #315
[ 80.500031] Call Trace:
[ 80.500031] [<ffffffff8117f65e>] __lock_acquire+0xc6e/0x1e50
[ 80.500031] [<ffffffff819ff225>] ? idr_get_empty_slot+0x255/0x3c0
[ 80.500031] [<ffffffff8117ca1e>] ? mark_held_locks+0x12e/0x150
[ 80.500031] [<ffffffff811810ba>] lock_acquire+0x1aa/0x240
[ 80.500031] [<ffffffff8192da37>] ? newary+0x1c7/0x2a0
[ 80.500031] [<ffffffff83d8768b>] _raw_spin_lock+0x3b/0x70
[ 80.500031] [<ffffffff8192da37>] ? newary+0x1c7/0x2a0
[ 80.500031] [<ffffffff8192da37>] newary+0x1c7/0x2a0
[ 80.500031] [<ffffffff8192a422>] ? ipcget+0x72/0x340
[ 80.500031] [<ffffffff8192a5d6>] ipcget+0x226/0x340
[ 80.500031] [<ffffffff81930215>] SyS_semget+0x65/0x80
[ 80.500031] [<ffffffff8192d870>] ? semctl_down.constprop.8+0x320/0x320
[ 80.500031] [<ffffffff8192c770>] ? wake_up_sem_queue_do+0xa0/0xa0
[ 80.500031] [<ffffffff8192c690>] ? SyS_msgrcv+0x20/0x20
[ 80.500031] [<ffffffff83d90898>] tracesys+0xe1/0xe6The code is:
for (i = 0; i < nsems; i++) {
INIT_LIST_HEAD(&sma->sem_base[i].sem_pending);
spin_lock_init(&sma->sem_base[i].lock);
spin_lock(&sma->sem_base[i].lock); <---- here
}Thanks,
Sasha
PATCH 1/1 ipc,sem: fix lockdep false positive
Quoted 4 lines · On Mon, 25 Mar 2013 16:21:22 -0400 Sasha Levin <[email protected]> wrote:
Quoted 3 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---
Quoted 8 lines
Quoted 4 lines
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
I'm getting the following false positives from lockdep:
Does this patch fix it?
Andrew, this looks like another one for the queue... ---8<---
Subject: [PATCH -mm -next] ipc,sem: fix lockdep false positive
When locking all the semaphores inside a sem_array, the kernel ends up locking a large number of locks with identical lockdep status. This trips up lockdep. Annotate the code to prevent such warnings.
--- ipc/sem.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-)
ipc/sem.c +2 -2
@@ -357,7 +357,7 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); +spin_lock_nested (&sem->lock,SINGLE_DEPTH_NESTING ); } locknum =-1 ; }
@@ -558,7 +558,7 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params)for (i =0 ; i < nsems; i++) {INIT_LIST_HEAD (&sma->sem_base[i].sem_pending);spin_lock_init (&sma->sem_base[i].lock); -spin_lock (&sma->sem_base[i].lock); +spin_lock_nested (&sma->sem_base[i].lock,SINGLE_DEPTH_NESTING ); } sma->complex_count =0 ;
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 12 lines · Rik van Riel
Quoted 9 lines · On Mon, 25 Mar 2013 16:21:22 -0400 Sasha Levin <[email protected]> wrote:
Quoted 7 lines · Rik van Riel
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Quoted 5 lines
Quoted 3 lines
Hi Rik,
I'm getting the following false positives from lockdep:
Does this patch fix it?
I'll be surprised if it does, because we don't actually have single depth nesting here...
Adding Peter & Ingo for advice about how to proceed
(the one solution I know would involve using arch_spin_lock() directly to bypass the lockdep checks, but there's got to be a better way...)
Quoted 36 lines
Andrew, this looks like another one for the queue... ---8<---
Subject: [PATCH -mm -next] ipc,sem: fix lockdep false positiveWhen locking all the semaphores inside a
sem_array, the kernel ends up locking a large number of locks with identical lockdep status. This trips up lockdep. Annotate the code to prevent such warnings.Signed-off-by Rik van Riel <[email protected]>--- ipc/sem.c | 4 ++-- 1 files changed, 2 insertions(+), 2 deletions(-)ipc/sem.c +2
-2@@ -357,7 +357,7 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); +spin_lock_nested (&sem->lock,SINGLE_DEPTH_NESTING ); } locknum =-1 ; }@@ -558,7 +558,7 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params)for (i =0 ; i < nsems; i++) {INIT_LIST_HEAD (&sma->sem_base[i].sem_pending);spin_lock_init (&sma->sem_base[i].lock); -spin_lock (&sma->sem_base[i].lock); +spin_lock_nested (&sma->sem_base[i].lock,SINGLE_DEPTH_NESTING ); } sma->complex_count =0 ;
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 5 lines · Michel Lespinasse
I'll be surprised if it does, because we don't actually have single depth nesting here...
Adding Peter & Ingo for advice about how to proceed
(the one solution I know would involve usingarch_spin_lock()directly to bypass the lockdep checks, but there's got to be a better way...)
Maybe spin_lock_nest_lock() can help too. I'm not sure, the feature is undocumented.
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 10 lines · Michel Lespinasse
Quoted 5 lines · Michel Lespinasse
I'll be surprised if it does, because we don't actually have single depth nesting here...
Adding Peter & Ingo for advice about how to proceed
(the one solution I know would involve usingarch_spin_lock()directly to bypass the lockdep checks, but there's got to be a better way...)Maybe
spin_lock_nest_lock()can help too. I'm not sure, the feature is undocumented.
I think we should name the locks properly (using 'key') and initialize their lockdep_map using lockdep_init_map instead of letting spin_lock pass the "&sma->sem_base[i].lock" as name.
Thanks,
Sasha
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
In reply to Michel LespinasseQuoted 15 lines · Michel Lespinasse
Quoted 14 lines · Rik van Riel
Quoted 14 lines
Quoted 11 lines · On Mon, 25 Mar 2013 16:21:22 -0400 Sasha Levin <[email protected]> wrote:
Quoted 11 lines · Rik van Riel
Quoted 7 lines
Quoted 7 lines
Include lkml in the CC: this time... *sigh*
---8<---This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line
Quoted 1 line
I'm getting the following false positives from lockdep:
Does this patch fix it?
I'll be surprised if it does, because we don't actually have single depth nesting here...
Adding Peter & Ingo for advice about how to proceed
(the one solution I know would involve usingarch_spin_lock()directly to bypass the lockdep checks, but there's got to be a better way...)
Yeah, it did not.
Thanks,
Sasha
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
In reply to Michel LespinasseQuoted 6 lines · Michel Lespinasse
depth nesting here...
Adding Peter & Ingo for advice about how to proceedReview static inline int sem_lock(struct sem_array
Quoted 2 lines
+++
b/ipc/sem.c(quoted diff) +0
-0@@ -357,7 +357,7 @@ static inline int sem_lock(struct sem_array*sma, struct sembuf *sops,
Quoted 9 lines
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
Quoted 5 lines
ipc/sem.c +1
-1spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); + spin_lock_nested(&sem->lock,
SINGLE_DEPTH_NESTING);Quoted 3 lines
ipc/sem.c +0
-0} locknum =-1 ; }
Right, so as walken said, this isn't going to work right.
I need a little more information as I've not really paid much attention to this stuff. Firstly, is there a limit to sem_nsems or is this a random user specified number? Secondly do we care about lock order at all, or is array order the only order that counts?
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 17 lines · Peter Zijlstra
Quoted 6 lines · Michel Lespinasse
depth nesting here...
Adding Peter & Ingo for advice about how to proceedReview static inline int sem_lock(struct sem_array
Quoted 2 lines
+++
b/ipc/sem.c(quoted diff) +0
-0@@ -357,7 +357,7 @@ static inline int sem_lock(struct sem_array*sma, struct sembuf *sops,
Quoted 9 lines
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
Quoted 5 lines
ipc/sem.c +1
-1spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); + spin_lock_nested(&sem->lock,
SINGLE_DEPTH_NESTING);Quoted 3 lines
ipc/sem.c +0
-0} locknum =-1 ; }
Quoted 6 lines
Right, so as walken said, this isn't going to work right.
I need a little more information as I've not really paid much attention to this stuff. Firstly, is there a limit to
sem_nsemsor is this a random user specified number? Secondly do we care about lock order at all, or is array order the only order that counts?
sem_nsems is user provided as the array size in some semget system call. It's the size of an ipc semaphore array.
complex semop operations take the array's lock plus every semaphore locks; simple semop operations (operating on a single semaphore) only take that one semaphore's lock.
AFAIK no operations take more than one semaphore lock except for the case where they are all taken (in which case the array's lock is also taken). Ideally it'd be best to have lockdep verify that this usage pattern is respected, though.
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 2 lines · Michel Lespinasse
sem_nsemsis user provided as the array size in some semget system call. It's the size of an ipc semaphore array.
So we're basically adding a random (big) number to preempt_count (obviously while preemption is disabled), seems rather costly and undesirable.
Quoted 3 lines
complex semop operations take the array's lock plus every semaphore locks; simple semop operations (operating on a single semaphore) only take that one semaphore's lock.
Right, standard global/local lock like stuff. Is there a way we can add a r/o test to the 'local' lock operation and avoid doing the above?
Maybe something like:
void sma_lock(struct sem_array *sma) /* global */
{
int i;
sma->global_locked = 1;
smp_wmb(); /* can we merge with the LOCK ? */
spin_lock(&sma->global_lock);
/* wait for all local locks to go away */
for (i = 0; i < sma->sem_nsems; i++)
spin_unlock_wait(&sem->sem_base[i]->lock);
}void sma_lock_one(struct sem_array *sma, int nr) /* local */
{
smp_rmb(); /* pairs with wmb in sma_lock() */
if (unlikely(sma->global_locked)) { /* wait for global lock */
while (sma->global_locked)
spin_unlock_wait(&sma->global_lock);
}
spin_lock(&sma->sem_base[nr]->lock);
}This still has the problem of a non-preemptible section of O(sem_nsems) (with the avg wait-time on the local lock). Could we make the global lock a sleeping lock?
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 9 lines · Peter Zijlstra
Quoted 2 lines · Michel Lespinasse
sem_nsemsis user provided as the array size in some semget system call. It's the size of an ipc semaphore array.So we're basically adding a random (big) number to
preempt_count(obviously while preemption is disabled), seems rather costly and undesirable.
Quoted 6 lines
Quoted 3 lines
complex semop operations take the array's lock plus every semaphore locks; simple semop operations (operating on a single semaphore) only take that one semaphore's lock.
Right, standard global/local lock like stuff. Is there a way we can add a r/o test to the 'local' lock operation and avoid doing the above?
That makes me wonder, how did mm_take_all_locks used to work before we turned the anon_vma lock into a mutex?
The code used to use spin_lock_nest_lock, but still has the potential to overflow the preempt counter. How did that ever work right?
Quoted 16 lines
Maybe something like:
void
sma_lock(struct sem_array *sma)/* global */{ int i; sma->global_locked = 1; smp_wmb(); /* can we merge with the LOCK ? */ spin_lock(&sma->global_lock); /* wait for all local locks to go away */ for (i = 0; i < sma->sem_nsems; i++) spin_unlock_wait(&sem->sem_base[i]->lock); }void
sma_lock_one(struct sem_array *sma, int nr)/* local */
Quoted 8 lines
{
smp_rmb(); /* pairs with wmb in sma_lock() */
if (unlikely(sma->global_locked)) { /* wait for global lock */
while (sma->global_locked)
spin_unlock_wait(&sma->global_lock);
}
spin_lock(&sma->sem_base[nr]->lock);
}That is essentially a read-only version of the global rwlock that I originally proposed, where the global lock takes the lock for write and the single version takes the global lock for read, and then one of the semaphore spinlocks.
I could certainly implement and test the above, unless Linus thinks it's too ugly to live :)
Quoted 3 lines
This still has the problem of a non-preemptible section of
O(sem_nsems)(with the avg wait-time on the local lock). Could we make the global lock a sleeping lock?
Not without breaking your scheme above :)
I suppose making things into a sleeping lock should be possible, but that is another major change in this code. I would rather do things in smaller steps...
Signature · 1 line
All rights reversed.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 5 lines · Rik van Riel
That makes me wonder, how did
mm_take_all_locksused to work before we turned theanon_vmalock into a mutex?The code used to use
spin_lock_nest_lock, but still has the potential to overflow the preempt counter. How did that ever work right?
It did trigger a bunch of warnings, but early on it was understood that KVM would have 'few' vmas when starting and registering the mmu_notifier thing.. then KVM bloated into insanity.
But aside from the warnings, if you overflow the regular preempt_count bits, nothing really bad happens because you start poking at softirq nesting, then hardirq etc.. all of those also disable preemption.
You'll get a few 'unexpected' side-effects for things like serving_softirq()/in_irq() or whatever those functions are called, but other than that things mostly work.
I don't particularly like overflowing preempt count, but its mostly harmless (up to a point). The much worse offender in my book is the duration of the preempt_disable section thus created.
Esp with everything in user control, you can basically create an arbitrary long non-preempt section with the semops.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
In reply to Rik van RielQuoted 15 lines · Rik van Riel
Quoted 15 lines
Maybe something like:
void
sma_lock(struct sem_array *sma)/* global */{ int i; sma->global_locked = 1; smp_wmb(); /* can we merge with the LOCK ? */ spin_lock(&sma->global_lock); /* wait for all local locks to go away */ for (i = 0; i < sma->sem_nsems; i++) spin_unlock_wait(&sem->sem_base[i]->lock); }
Quoted 9 lines
Quoted 9 lines
void
sma_lock_one(struct sem_array *sma, int nr)/* local */{ smp_rmb(); /* pairs with wmb in sma_lock() */ if (unlikely(sma->global_locked)) { /* wait for global lock */ while (sma->global_locked) spin_unlock_wait(&sma->global_lock); } spin_lock(&sma->sem_base[nr]->lock); }
I since realized there's an ordering problem with ->global_locked, we need to use spin_is_locked() or somesuch.
Two competing sma_lock() operations will screw over the separate variable.
Quoted 7 lines
Quoted 2 lines
This still has the problem of a non-preemptible section of
O(sem_nsems)Quoted 2 lines
(with the avg wait-time on the local lock). Could we make the global lock a sleeping lock?
Not without breaking your scheme above :)
How would making sma->global_lock a mutex wreck anything?
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 33 lines · Peter Zijlstra
Quoted 15 lines · Rik van Riel
Quoted 15 lines
Maybe something like:
void
sma_lock(struct sem_array *sma)/* global */{ int i; sma->global_locked = 1; smp_wmb(); /* can we merge with the LOCK ? */ spin_lock(&sma->global_lock); /* wait for all local locks to go away */ for (i = 0; i < sma->sem_nsems; i++) spin_unlock_wait(&sem->sem_base[i]->lock); }Quoted 9 lines
Quoted 9 lines
void
sma_lock_one(struct sem_array *sma, int nr)/* local */{ smp_rmb(); /* pairs with wmb in sma_lock() */ if (unlikely(sma->global_locked)) { /* wait for global lock */ while (sma->global_locked) spin_unlock_wait(&sma->global_lock); } spin_lock(&sma->sem_base[nr]->lock); }I since realized there's an ordering problem with ->global_locked, we need to use
spin_is_locked()or somesuch.Two competing
sma_lock()operations will screw over the separate variable.Quoted 1 line
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
This still has the problem of a non-preemptible section of
O(sem_nsems)Quoted 2 lines
(with the avg wait-time on the local lock). Could we make the global lock a sleeping lock?
Not without breaking your scheme above :)
How would making
sma->global_locka mutex wreck anything?
I don't remember the details (rik probably will), but rcu is also already involved, so there is a non trivial chance that it would...
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
Quoted 6 lines · Michel Lespinasse
Quoted 3 lines
Quoted 1 line
Not without breaking your scheme above :)
How would making
sma->global_locka mutex wreck anything?I don't remember the details (rik probably will), but rcu is also already involved, so there is a non trivial chance that it would...
Ah, that might be true, the callsite might not be able to deal with it, but the proposed global/local lock thing doesn't care -- which is how I read Rik's comment above.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
In reply to Peter ZijlstraQuoted 26 lines · Peter Zijlstra
Quoted 15 lines · Rik van Riel
Quoted 15 lines
Maybe something like:
void
sma_lock(struct sem_array *sma)/* global */{ int i; sma->global_locked = 1; smp_wmb(); /* can we merge with the LOCK ? */ spin_lock(&sma->global_lock); /* wait for all local locks to go away */ for (i = 0; i < sma->sem_nsems; i++) spin_unlock_wait(&sem->sem_base[i]->lock); }Quoted 9 lines
Quoted 9 lines
void
sma_lock_one(struct sem_array *sma, int nr)/* local */{ smp_rmb(); /* pairs with wmb in sma_lock() */ if (unlikely(sma->global_locked)) { /* wait for global lock */ while (sma->global_locked) spin_unlock_wait(&sma->global_lock); } spin_lock(&sma->sem_base[nr]->lock); }
Quoted 5 lines
I since realized there's an ordering problem with ->global_locked, we need to use
spin_is_locked()or somesuch.Two competing
sma_lock()operations will screw over the separate variable.
There may be another problem with your idea.
If there are two single locks coming in for the same semaphore, the first one holds the lock, while the second one is spinning on the lock.
The global lock is spinning on spin_unlock_wait, which may end up finishing after the first single lock holder unlocks, right before the second single lock holder grabs the lock.
At that point, you have both a process that thinks it holds the global lock, and a process that thinks it holds a single lock.
To prevent against this, the single lock probably needs to test whether the global lock is locked, after it acquires the local lock.
If the global lock is locked, it needs to unlock its single lock, and then do a spin_unlock_wait on the global lock, before trying again from the start.
Would that work?
PATCH v2 1/1 ipc,sem: fix lockdep false positive
In reply to Peter ZijlstraQuoted 5 lines · On Wed, 27 Mar 2013 09:42:30 +0100 Peter Zijlstra <[email protected]> wrote:
I since realized there's an ordering problem with ->global_locked, we need to use
spin_is_locked()or somesuch.Two competing
sma_lock()operations will screw over the separate variable.
I created a worse version of the stress test, one that will have the first thread always do two semops at a time (creating a complex operation), and have all the other threads do one at a time. All the threads issue down and then up, so there is a lot of sleeping and waking back up with this test.
The patch below should fix the last lockdep issue (the other issue being fixed by the RCU patch), and is running the various semop tests just fine on a 48 CPU system.
The locking is a little more complicated than I would have liked, but a few hundred million semaphore operations suggest there are probably no race conditions left...
---8<---
Subject: [PATCH -mm -next] ipc,sem: change locking scheme to make lockdep happy
Unfortunately the locking scheme originally proposed has false positives with lockdep. This can be fixed by changing the code to only ever take one lock, and making sure that other relevant locks are not locked, before entering a critical section.
For the "global lock" case, this is done by taking the sem_array lock, and then (potentially) waiting for all the semaphore's spinlocks to be unlocked.
For the "local lock" case, we wait on the sem_array's lock to be free, before taking the semaphore local lock. To prevent races, we need to check again after we have taken the local lock.
--- ipc/sem.c | 55 ++++++++++++++++++++++++++++++++++++++++--------------- 1 files changed, 40 insertions(+), 15 deletions(-)
ipc/sem.c +40 -15
@@ -320,24 +320,39 @@ void __init sem_init (void) }/* -* If the sem_array contains just one semaphore, or if multiple -* semops are performed in one syscall, or if there are complex -* operations pending, the whole sem_array is locked. -* If one semop is performed on an array with multiple semaphores, -* get a shared lock on the array, and lock the individual semaphore. +* If the request contains only one semaphore operation, and there are +* no complex transactions pending, lock only the semaphore involved. +* Otherwise, lock the entire semaphore array, since we either have +* multiple semaphores in our own semops, or we need to look at +* semaphores from other pending complex operations. * * Carefully guard against sma->complex_count changing between zero * and non-zero while we are spinning for the lock. The value of * sma->complex_count cannot change while we are holding the lock, * so sem_unlock should be fine. +* +* The global lock path checks that all the local locks have been released, +* checking each local lock once. This means that the local lock paths +* cannot start their critical sections while the global lock is held. */ 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; +/* +* Another process is holding the global lock on the +* sem_array. Wait for that process to release the lock, +* before acquiring our lock. +*/ +if (unlikely (spin_is_locked (&sma->sem_perm.lock))) { +spin_unlock_wait (&sma->sem_perm.lock); + goto again; + } +/* Lock just the semaphore we are interested in. */ spin_lock (&sem->lock);
@@ -347,17 +362,33 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,*/ if (unlikely (sma->complex_count)) {spin_unlock (&sem->lock); - goto lock_all; + goto lock_array; + } + +/* +* 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); + goto again; } + locknum = sops->sem_num; }else {int i; -/* Lock the sem_array, and all the semaphore locks */ - lock_all: +/* +* Lock the semaphore array, and wait for all of the +* individual semaphore locks to go away. The code +* above ensures no new single-lock holders will enter +* their critical section while the array lock is held. +*/ + lock_array:spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); +spin_unlock_wait (&sem->lock); } locknum =-1 ; }
@@ -367,11 +398,6 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,static inline void sem_unlock (struct sem_array *sma,int locknum) {if (locknum ==-1 ) { -int i; -for (i =0 ; i < sma->sem_nsems; i++) { -struct sem *sem = sma->sem_base + i; -spin_unlock (&sem->lock); - }spin_unlock (&sma->sem_perm.lock); }else {struct sem *sem = sma->sem_base + locknum;
@@ -558,7 +584,6 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params)for (i =0 ; i < nsems; i++) {INIT_LIST_HEAD (&sma->sem_base[i].sem_pending);spin_lock_init (&sma->sem_base[i].lock); -spin_lock (&sma->sem_base[i].lock); } sma->complex_count =0 ;
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
Quoted 11 lines · Rik van Riel
Subject: [PATCH -mm -next] ipc,sem: change locking scheme to make lockdep happy
Unfortunately the locking scheme originally proposed has false positives with lockdep. This can be fixed by changing the code to only ever take one lock, and making sure that other relevant locks are not locked, before entering a critical section.
For the "global lock" case, this is done by taking the
sem_arraylock, and then (potentially) waiting for all the semaphore's spinlocks to be unlocked.
Quoted 7 lines
For the "local lock" case, we wait on the sem_array's lock to be free, before taking the semaphore local lock. To prevent races, we need to check again after we have taken the local lock.
Suggested-by Peter Zijlstra <[email protected]>Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>
TL;DR: The locking algorithm is not familiar for me, but it seems sound. There are some implementation details I don't like. More below...
Quoted 40 lines
--- ipc/sem.c | 55 ++++++++++++++++++++++++++++++++++++++++--------------- 1 files changed, 40 insertions(+), 15 deletions(-)ipc/sem.c +10
-5@@ -320,24 +320,39 @@ void __init sem_init (void) }/* -* If the sem_array contains just one semaphore, or if multiple -* semops are performed in one syscall, or if there are complex -* operations pending, the whole sem_array is locked. -* If one semop is performed on an array with multiple semaphores, -* get a shared lock on the array, and lock the individual semaphore. +* If the request contains only one semaphore operation, and there are +* no complex transactions pending, lock only the semaphore involved. +* Otherwise, lock the entire semaphore array, since we either have +* multiple semaphores in our own semops, or we need to look at +* semaphores from other pending complex operations. * * Carefully guard against sma->complex_count changing between zero * and non-zero while we are spinning for the lock. The value of * sma->complex_count cannot change while we are holding the lock, * so sem_unlock should be fine. +* +* The global lock path checks that all the local locks have been released, +* checking each local lock once. This means that the local lock paths +* cannot start their critical sections while the global lock is held. */ 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;
ipc/sem.c void __init sem_init (void) · line 355
Quoted 10 lines
ipc/sem.c +10
-0+/* +* Another process is holding the global lock on the +* sem_array. Wait for that process to release the lock, +* before acquiring our lock. +*/ +if (unlikely (spin_is_locked (&sma->sem_perm.lock))) { +spin_unlock_wait (&sma->sem_perm.lock); + goto again; + } +
So, there are a few things I don't like about spin_unlock_wait():
1- From a lock ordering point of view, it is strictly equivalent to taking the lock and then releasing it - and yet, lockdep won't catch any deadlocks that involve spin_unlock_wait. (Not your fault here, this should be fixed as a separate change in lockdep. I manually looked at the lock ordering here and found it safe).
2- With the current ticket lock implementation, a stream of lockers can starve spin_unlock_wait() forever. Once again, not your fault and I suspect this could be fixed - I expect spin_unlock_wait() callers actually only want to know that the lock has been passed on, not that it actually got to an unlocked state.
3- Regarding your actual use here - I find it confusing to call spin_unlock_wait() before holding any other lock. The pattern I expect to see is that people take one lock, then see that the other lock they want is already taken, so they release the first lock and wait on the second. So, I'd suggest we remove the sem_perm.lock checks here and deal with this in a retry path later down.
Quoted 9 lines
ipc/sem.c +1
-1/* Lock just the semaphore we are interested in. */ spin_lock (&sem->lock);@@ -347,17 +362,33 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,*/ if (unlikely (sma->complex_count)) {spin_unlock (&sem->lock); - goto lock_all; + goto lock_array;
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops, · line 375
Quoted 10 lines
ipc/sem.c +10
-0+ } + +/* +* 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); + goto again;
This is IMO where the spin_unlock_wait(&sma->sem_perm.lock) would belong - right before the goto again.
Also - I think there is a risk that an endless stream of complex semops could starve a simple semop here, as it would always find the sem_perm.lock to be locked ??? One easy way to guarantee progress would be to goto lock_array instead; however there is then the issue that a complex semop could force an endless stream of following simple semops to take the lock_array path. I'm not sure which of these problems is preferable to have...
Quoted 12 lines
ipc/sem.c +6
-2} + locknum = sops->sem_num; }else {int i; -/* Lock the sem_array, and all the semaphore locks */ - lock_all: +/* +* Lock the semaphore array, and wait for all of the +* individual semaphore locks to go away. The code +* above ensures no new single-lock holders will enter +* their critical section while the array lock is held.
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops, · line 394
Quoted 10 lines
ipc/sem.c +3
-1+*/ + lock_array:spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); +spin_unlock_wait (&sem->lock); } locknum =-1 ; }
Subtle, but it'll work (modulo the starvation issue I mentioned).
Cheers,
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
Quoted 7 lines · Michel Lespinasse
So, there are a few things I don't like about
spin_unlock_wait():1- From a lock ordering point of view, it is strictly equivalent to taking the lock and then releasing it - and yet, lockdep won't catch any deadlocks that involve
spin_unlock_wait. (Not your fault here, this should be fixed as a separate change in lockdep. I manually looked at the lock ordering here and found it safe).
Ooh, I never noticed that, but indeed this shouldn't be hard to cure.
Quoted 5 lines
2- With the current ticket lock implementation, a stream of lockers can starve
spin_unlock_wait()forever. Once again, not your fault and I suspect this could be fixed - I expectspin_unlock_wait()callers actually only want to know that the lock has been passed on, not that it actually got to an unlocked state.
I suppose the question is do we want to fix it or have both semantics and use lock+unlock where appropriate.
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
Quoted 11 lines · Peter Zijlstra
Quoted 7 lines · Michel Lespinasse
So, there are a few things I don't like about
spin_unlock_wait():1- From a lock ordering point of view, it is strictly equivalent to taking the lock and then releasing it - and yet, lockdep won't catch any deadlocks that involve
spin_unlock_wait. (Not your fault here, this should be fixed as a separate change in lockdep. I manually looked at the lock ordering here and found it safe).Ooh, I never noticed that, but indeed this shouldn't be hard to cure.
Quoted 8 lines
Quoted 5 lines
2- With the current ticket lock implementation, a stream of lockers can starve
spin_unlock_wait()forever. Once again, not your fault and I suspect this could be fixed - I expectspin_unlock_wait()callers actually only want to know that the lock has been passed on, not that it actually got to an unlocked state.I suppose the question is do we want to fix it or have both semantics and use lock+unlock where appropriate.
We'd have to look at the users to be sure, but I strongly expect they don't need to get in line waiting - it's sufficient to just wait for the head of the queue to move (or for the queue to be empty).
There are actually very few users - Just drivers/ata/libata-eh.c for the spin_unlock_wait() function, and a couple more (kernel/task_work.c and kernel/exit.c) for the raw_spin_unlock_wait variant. Guess I'm not the only one to dislike that function :)
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
In reply to Michel LespinasseQuoted 110 lines · Michel Lespinasse
Quoted 11 lines · Rik van Riel
Subject: [PATCH -mm -next] ipc,sem: change locking scheme to make lockdep happy
Unfortunately the locking scheme originally proposed has false positives with lockdep. This can be fixed by changing the code to only ever take one lock, and making sure that other relevant locks are not locked, before entering a critical section.
For the "global lock" case, this is done by taking the
sem_arraylock, and then (potentially) waiting for all the semaphore's spinlocks to be unlocked.Quoted 7 lines
For the "local lock" case, we wait on the sem_array's lock to be free, before taking the semaphore local lock. To prevent races, we need to check again after we have taken the local lock.
Suggested-by Peter Zijlstra <[email protected]>Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>TL;DR: The locking algorithm is not familiar for me, but it seems sound. There are some implementation details I don't like. More below...
Quoted 40 lines
--- ipc/sem.c | 55 ++++++++++++++++++++++++++++++++++++++++--------------- 1 files changed, 40 insertions(+), 15 deletions(-)ipc/sem.c +10
-5@@ -320,24 +320,39 @@ void __init sem_init (void) }/* -* If the sem_array contains just one semaphore, or if multiple -* semops are performed in one syscall, or if there are complex -* operations pending, the whole sem_array is locked. -* If one semop is performed on an array with multiple semaphores, -* get a shared lock on the array, and lock the individual semaphore. +* If the request contains only one semaphore operation, and there are +* no complex transactions pending, lock only the semaphore involved. +* Otherwise, lock the entire semaphore array, since we either have +* multiple semaphores in our own semops, or we need to look at +* semaphores from other pending complex operations. * * Carefully guard against sma->complex_count changing between zero * and non-zero while we are spinning for the lock. The value of * sma->complex_count cannot change while we are holding the lock, * so sem_unlock should be fine. +* +* The global lock path checks that all the local locks have been released, +* checking each local lock once. This means that the local lock paths +* cannot start their critical sections while the global lock is held. */ 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;ipc/sem.c void __init sem_init (void) · line 355
Quoted 10 lines
ipc/sem.c +10
-0+/* +* Another process is holding the global lock on the +* sem_array. Wait for that process to release the lock, +* before acquiring our lock. +*/ +if (unlikely (spin_is_locked (&sma->sem_perm.lock))) { +spin_unlock_wait (&sma->sem_perm.lock); + goto again; + } +So, there are a few things I don't like about
spin_unlock_wait():1- From a lock ordering point of view, it is strictly equivalent to taking the lock and then releasing it - and yet, lockdep won't catch any deadlocks that involve
spin_unlock_wait. (Not your fault here, this should be fixed as a separate change in lockdep. I manually looked at the lock ordering here and found it safe).2- With the current ticket lock implementation, a stream of lockers can starve
spin_unlock_wait()forever. Once again, not your fault and I suspect this could be fixed - I expectspin_unlock_wait()callers actually only want to know that the lock has been passed on, not that it actually got to an unlocked state.3- Regarding your actual use here - I find it confusing to call
spin_unlock_wait()before holding any other lock. The pattern I expect to see is that people take one lock, then see that the other lock they want is already taken, so they release the first lock and wait on the second. So, I'd suggest we remove the sem_perm.lock checks here and deal with this in a retry path later down.Quoted 14 lines
ipc/sem.c +6
-1/* Lock just the semaphore we are interested in. */ spin_lock (&sem->lock);@@ -347,17 +362,33 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,*/ if (unlikely (sma->complex_count)) {spin_unlock (&sem->lock); - goto lock_all; + goto lock_array; + } + +/* +* Another process is holding the global lock on the +* sem_array; we cannot enter our critical section,
Quoted 8 lines
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops, · line 375
Quoted 5 lines
ipc/sem.c +5
-0+* but have to wait for the global lock to be released. +*/ +if (unlikely (spin_is_locked (&sma->sem_perm.lock))) { +spin_unlock (&sem->lock); + goto again;This is IMO where the
spin_unlock_wait(&sma->sem_perm.lock)would belong - right before the goto again.
That is where I had it initially. I may have gotten too clever and worked on keeping more accesses read-only. If you want, I can move it back here and re-submit the patch :)
Quoted 7 lines
Also - I think there is a risk that an endless stream of complex semops could starve a simple semop here, as it would always find the sem_perm.lock to be locked ??? One easy way to guarantee progress would be to goto
lock_arrayinstead; however there is then the issue that a complex semop could force an endless stream of following simple semops to take thelock_arraypath. I'm not sure which of these problems is preferable to have...
If starvation turns out to be an issue, there is an even simpler solution:
if (unlikely(spin_is_locked(&sma->sem_perm.lock))) {
spin_unlock(&sem->lock);
spin_lock(&sma->sem_perm.lock);
spin_lock(&sem->lock);
spin_unlock(&sma->sem_perm.lock);
}Followed by unconditionally doing the critical section for holding a single semaphore's lock, because we know that a subsequent taker of sma->sem_perm.lock will either grab a different semaphore's spinlock, or wait on the same semaphore's spinlock, or wait for us to unlock our spinlock.
Signature · 1 line
All rights reversed.
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
Quoted 17 lines · Rik van Riel
Quoted 16 lines · Michel Lespinasse
Quoted 14 lines · Rik van Riel
if (unlikely(sma->complex_count)) { spin_unlock(&sem->lock);ipc/sem.c +11
-1- goto lock_all; + goto lock_array; + } + +/* +* 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); + goto again;
Quoted 6 lines
Quoted 2 lines
This is IMO where the
spin_unlock_wait(&sma->sem_perm.lock)would belong - right before the goto again.That is where I had it initially. I may have gotten too clever and worked on keeping more accesses read-only. If you want, I can move it back here and re-submit the patch :)
I think I would prefer that - I feel having it upper serves little purpose, as you still need to check it here, so we might as well be optimistic and check it only here. Or maybe I've missed the benefit of having it earlier - I just don't see it.
Quoted 11 lines
Quoted 7 lines
Also - I think there is a risk that an endless stream of complex semops could starve a simple semop here, as it would always find the sem_perm.lock to be locked ??? One easy way to guarantee progress would be to goto
lock_arrayinstead; however there is then the issue that a complex semop could force an endless stream of following simple semops to take thelock_arraypath. I'm not sure which of these problems is preferable to have...If starvation turns out to be an issue, there is an even simpler solution:
Quoted 6 lines
if (unlikely(spin_is_locked(&sma->sem_perm.lock))) {
spin_unlock(&sem->lock);
spin_lock(&sma->sem_perm.lock);
spin_lock(&sem->lock);
spin_unlock(&sma->sem_perm.lock);
}This is kinda nice - certainly nicer than falling back to the lock_array case.
It still makes it possible (though less likely) that each simple semop taking this path might make the next one take it too, though. So, I'm not sure how to balance that against the starvation possibility. I'll leave it up to you then :)
Cheers,
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH v2 -mm -next] ipc,sem: fix lockdep false positive
Quoted 8 lines · Michel Lespinasse
Quoted 3 lines
That is where I had it initially. I may have gotten too clever and worked on keeping more accesses read-only. If you want, I can move it back here and re-submit the patch :)
I think I would prefer that - I feel having it upper serves little purpose, as you still need to check it here, so we might as well be optimistic and check it only here. Or maybe I've missed the benefit of having it earlier - I just don't see it.
OK, will do.
Quoted 4 lines
It still makes it possible (though less likely) that each simple semop taking this path might make the next one take it too, though. So, I'm not sure how to balance that against the starvation possibility. I'll leave it up to you then :)
Lets optimistically assume that userland is using the semaphores for synchronization. That rather limits the potential for starvation.
Signature · 1 line
All rights reversed.
PATCH v3 1/1 ipc,sem: fix lockdep false positive
In reply to Michel LespinasseQuoted 2 lines · On Thu, 28 Mar 2013 19:50:47 -0700 Michel Lespinasse <[email protected]> wrote:
This is IMO where the
spin_unlock_wait(&sma->sem_perm.lock)would belong - right before the goto again.
Here is the slightly more optimistic (and probably more readable) version of the patch:
---8<---
Unfortunately the locking scheme originally proposed has false positives with lockdep. This can be fixed by changing the code to only ever take one lock, and making sure that other relevant locks are not locked, before entering a critical section.
For the "global lock" case, this is done by taking the sem_array lock, and then (potentially) waiting for all the semaphore's spinlocks to be unlocked.
For the "local lock" case, we wait on the sem_array's lock to be free, before taking the semaphore local lock. To prevent races, we need to check again after we have taken the local lock.
--- ipc/sem.c | 46 +++++++++++++++++++++++++++++++--------------- 1 files changed, 31 insertions(+), 15 deletions(-)
ipc/sem.c +31 -15
@@ -320,21 +320,26 @@ void __init sem_init (void) }/* -* If the sem_array contains just one semaphore, or if multiple -* semops are performed in one syscall, or if there are complex -* operations pending, the whole sem_array is locked. -* If one semop is performed on an array with multiple semaphores, -* get a shared lock on the array, and lock the individual semaphore. +* If the request contains only one semaphore operation, and there are +* no complex transactions pending, lock only the semaphore involved. +* Otherwise, lock the entire semaphore array, since we either have +* multiple semaphores in our own semops, or we need to look at +* semaphores from other pending complex operations. * * Carefully guard against sma->complex_count changing between zero * and non-zero while we are spinning for the lock. The value of * sma->complex_count cannot change while we are holding the lock, * so sem_unlock should be fine. +* +* The global lock path checks that all the local locks have been released, +* checking each local lock once. This means that the local lock paths +* cannot start their critical sections while the global lock is held. */ 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;
@@ -347,17 +352,34 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,*/ if (unlikely (sma->complex_count)) {spin_unlock (&sem->lock); - goto lock_all; + goto lock_array; + } + +/* +* 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; } + locknum = sops->sem_num; }else {int i; -/* Lock the sem_array, and all the semaphore locks */ - lock_all: +/* +* Lock the semaphore array, and wait for all of the +* individual semaphore locks to go away. The code +* above ensures no new single-lock holders will enter +* their critical section while the array lock is held. +*/ + lock_array:spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); +spin_unlock_wait (&sem->lock); } locknum =-1 ; }
@@ -367,11 +389,6 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,static inline void sem_unlock (struct sem_array *sma,int locknum) {if (locknum ==-1 ) { -int i; -for (i =0 ; i < sma->sem_nsems; i++) { -struct sem *sem = sma->sem_base + i; -spin_unlock (&sem->lock); - }spin_unlock (&sma->sem_perm.lock); }else {struct sem *sem = sma->sem_base + locknum;
@@ -558,7 +575,6 @@ static int newary(struct ipc_namespace *ns, struct ipc_params *params)for (i =0 ; i < nsems; i++) {INIT_LIST_HEAD (&sma->sem_base[i].sem_pending);spin_lock_init (&sma->sem_base[i].lock); -spin_lock (&sma->sem_base[i].lock); } sma->complex_count =0 ;
Re: [PATCH v3 -mm -next] ipc,sem: fix lockdep false positive
Quoted 19 lines · Rik van Riel
Quoted 2 lines · On Thu, 28 Mar 2013 19:50:47 -0700 Michel Lespinasse <[email protected]> wrote:
This is IMO where the
spin_unlock_wait(&sma->sem_perm.lock)would belong - right before the goto again.Here is the slightly more optimistic (and probably more readable) version of the patch:
---8<---
Unfortunately the locking scheme originally proposed has false positives with lockdep. This can be fixed by changing the code to only ever take one lock, and making sure that other relevant locks are not locked, before entering a critical section.For the "global lock" case, this is done by taking the
sem_arraylock, and then (potentially) waiting for all the semaphore's spinlocks to be unlocked.
Quoted 7 lines
For the "local lock" case, we wait on the sem_array's lock to be free, before taking the semaphore local lock. To prevent races, we need to check again after we have taken the local lock.
Suggested-by Peter Zijlstra <[email protected]>Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>
Looks good.
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Re: [PATCH -mm -next] ipc,sem: fix lockdep false positive
In reply to Peter ZijlstraQuoted 17 lines · Peter Zijlstra
Quoted 6 lines · Michel Lespinasse
depth nesting here...
Adding Peter & Ingo for advice about how to proceedReview static inline int sem_lock(struct sem_array
Quoted 2 lines
+++
b/ipc/sem.c(quoted diff) +0
-0@@ -357,7 +357,7 @@ static inline int sem_lock(struct sem_array*sma, struct sembuf *sops,
Quoted 9 lines
ipc/sem.c static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
Quoted 5 lines
ipc/sem.c +1
-1spin_lock (&sma->sem_perm.lock);for (i =0 ; i < sma->sem_nsems; i++) {struct sem *sem = sma->sem_base + i; -spin_lock (&sem->lock); + spin_lock_nested(&sem->lock,
SINGLE_DEPTH_NESTING);Quoted 3 lines
ipc/sem.c +0
-0} locknum =-1 ; }
Quoted 6 lines
Right, so as walken said, this isn't going to work right.
I need a little more information as I've not really paid much attention to this stuff. Firstly, is there a limit to
sem_nsemsor is this a random user specified number? Secondly do we care about lock order at all, or is array order the only order that counts?
It is a user specified number, and we either lock only one of the semaphore locks, or we lock all of them (in array order, after taking the semaphore array lock).
Signature · 1 line
All rights reversed.
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Another issue that came up is:
[ 96.347341] ================================================
[ 96.348085] [ BUG: lock held when returning to user space! ]
[ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W
[ 96.360300] ------------------------------------------------
[ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held!
[ 96.362019] 1 lock held by trinity-child9/7583:
[ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it seems that this behaviour is actually intentional?
rcu_read_lock();
sma = sem_obtain_object_check(ns, semid);
if (IS_ERR(sma)) {
if (un)
rcu_read_unlock();
error = PTR_ERR(sma);
goto out_free;
}When I've looked at that it seems that not releasing the read lock was (very) intentional.
After that, the only code path that would release the lock starts with:
if (un) {
...So we won't release the lock at all if un is NULL?
Thanks,
Sasha
Quoted 32 lines · Sasha Levin
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Another issue that came up is:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it seems that this behaviour is actually intentional?
rcu_read_lock(); sma = sem_obtain_object_check(ns, semid); if (IS_ERR(sma)) { if (un) rcu_read_unlock(); error = PTR_ERR(sma); goto out_free; }
Quoted 2 lines
When I've looked at that it seems that not releasing the read lock was (very) intentional.
This logic was from the original code, which I also found to be quite confusing.
Quoted 8 lines
After that, the only code path that would release the lock starts with:
if (un) { ...So we won't release the lock at all if un is NULL?
Not necessarily, we do release everything at the end of the function:
out_unlock_free:
sem_unlock(sma, locknum);
Thanks,
Davidlohr
Quoted 33 lines · Davidlohr Bueso
Quoted 32 lines · Sasha Levin
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Another issue that came up is:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it seems that this behaviour is actually intentional?
rcu_read_lock(); sma = sem_obtain_object_check(ns, semid); if (IS_ERR(sma)) { if (un) rcu_read_unlock(); error = PTR_ERR(sma); goto out_free; }
Quoted 5 lines
Quoted 2 lines
When I've looked at that it seems that not releasing the read lock was (very) intentional.
This logic was from the original code, which I also found to be quite confusing.
I wasn't getting this warning with the old code, so there was probably something else that triggers this now.
Quoted 6 lines
Quoted 6 lines
After that, the only code path that would release the lock starts with:
if (un) { ...
Quoted 7 lines
Quoted 2 lines
So we won't release the lock at all if un is NULL?
Not necessarily, we do release everything at the end of the function:
out_unlock_free:sem_unlock(sma, locknum);
Ow, there's a rcu_read_unlock() in sem_unlock()? This complicates things even more I suspect. If un is non-NULL we'll be unlocking rcu lock twice?
if (un->semid == -1) {
rcu_read_unlock();
goto out_unlock_free;
}
[...]
out_unlock_free:
sem_unlock(sma, locknum);Thanks,
Sasha
Quoted 52 lines · Sasha Levin
Quoted 33 lines · Davidlohr Bueso
Quoted 32 lines · Sasha Levin
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Another issue that came up is:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it seems that this behaviour is actually intentional?
rcu_read_lock(); sma = sem_obtain_object_check(ns, semid); if (IS_ERR(sma)) { if (un) rcu_read_unlock(); error = PTR_ERR(sma); goto out_free; }Quoted 5 lines
Quoted 2 lines
When I've looked at that it seems that not releasing the read lock was (very) intentional.
This logic was from the original code, which I also found to be quite confusing.
I wasn't getting this warning with the old code, so there was probably something else that triggers this now.
Quoted 9 lines
Quoted 8 lines
After that, the only code path that would release the lock starts with:
if (un) { ...So we won't release the lock at all if un is NULL?
Quoted 7 lines
Quoted 4 lines
Not necessarily, we do release everything at the end of the function:
out_unlock_free:sem_unlock(sma, locknum);Ow, there's a
rcu_read_unlock()insem_unlock()? This complicates things even more I suspect. If un is non-NULL we'll be unlocking rcu lock twice?
It is uglier than you think...
On success, find_alloc_undo will call rcu_read_lock, so we have the rcu_read_lock held twice :(
Some of the ipc code is quite ugly, but making too many large changes at once is just asking for trouble. I suspect we're going to have to untangle this one bit at a time...
Signature · 1 line
All rights reversed.
PATCH 1/1 ipc,sem: untangle RCU locking with find_alloc_undo
In reply to Sasha LevinQuoted 6 lines · On Tue, 26 Mar 2013 14:07:14 -0400 Sasha Levin <[email protected]> wrote:
Quoted 3 lines
Not necessarily, we do release everything at the end of the function:
out_unlock_free: sem_unlock(sma, locknum);Ow, there's a
rcu_read_unlock()insem_unlock()? This complicates things even more I suspect. If un is non-NULL we'll be unlocking rcu lock twice?
Sasha, this patch should resolve the RCU tangle, by making sure we only ever take the rcu_read_lock once in semtimedop.
---8<---
The ipc semaphore code has a nasty RCU locking tangle, with both find_alloc_undo and semtimedop taking the rcu_read_lock(). The code can be cleaned up somewhat by only taking the rcu_read_lock once.
There are no other callers to find_alloc_undo.
This should also solve the trinity issue reported by Sasha Levin.
--- ipc/sem.c | 31 +++++++++---------------------- 1 files changed, 9 insertions(+), 22 deletions(-)
ipc/sem.c +9 -22
@@ -1646,22 +1646,23 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, alter =1 ; } +INIT_LIST_HEAD (&tasks); +if (undos) { +/* On success, find_alloc_undo takes the rcu_read_lock */ un =find_alloc_undo (ns, semid);if (IS_ERR (un)) { error =PTR_ERR (un); goto out_free; } - }else + }else { un =NULL ; +rcu_read_lock (); + } -INIT_LIST_HEAD (&tasks); - -rcu_read_lock (); sma =sem_obtain_object_check (ns, semid);if (IS_ERR (sma)) { -if (un) -rcu_read_unlock (); +rcu_read_unlock (); error =PTR_ERR (sma); goto out_free; }
@@ -1693,22 +1694,8 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,*/ error = -EIDRM ; locknum =sem_lock (sma, sops, nsops); -if (un) { -if (un->semid ==-1 ) { -rcu_read_unlock (); - goto out_unlock_free; - }else { -/* -* rcu lock can be released, "un" cannot disappear: -* - sem_lock is acquired, thus IPC_RMID is -* impossible. -* - exit_sem is impossible, it always operates on -* current (or a dead task). -*/ - -rcu_read_unlock (); - } - } +if (un && un->semid ==-1 ) + goto out_unlock_free; error =try_atomic_semop (sma, sops, nsops, un,task_tgid_vnr (current));if (error <=0 ) {
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 15 lines · Rik van Riel
Quoted 6 lines · On Tue, 26 Mar 2013 14:07:14 -0400 Sasha Levin <[email protected]> wrote:
Quoted 3 lines
Not necessarily, we do release everything at the end of the function:
out_unlock_free: sem_unlock(sma, locknum);Ow, there's a
rcu_read_unlock()insem_unlock()? This complicates things even more I suspect. If un is non-NULL we'll be unlocking rcu lock twice?Sasha, this patch should resolve the RCU tangle, by making sure we only ever take the
rcu_read_lockonce in semtimedop.---8<---
Quoted 8 lines
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.There are no other callers to
find_alloc_undo.This should also solve the trinity issue reported by Sasha Levin.
I take it this is on top of the patchlet Sasha submitted?
(I hit rcu stall banging on patch set in rt with 60 synchronized core executive model if I let it run long enough, fwtw)
Quoted 70 lines
Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>--- ipc/sem.c | 31 +++++++++---------------------- 1 files changed, 9 insertions(+), 22 deletions(-)ipc/sem.c +9
-23@@ -1646,22 +1646,23 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, alter =1 ; } +INIT_LIST_HEAD (&tasks); +if (undos) { +/* On success, find_alloc_undo takes the rcu_read_lock */ un =find_alloc_undo (ns, semid);if (IS_ERR (un)) { error =PTR_ERR (un); goto out_free; } - }else + }else { un =NULL ; +rcu_read_lock (); + } -INIT_LIST_HEAD (&tasks); - -rcu_read_lock (); sma =sem_obtain_object_check (ns, semid);if (IS_ERR (sma)) { -if (un) -rcu_read_unlock (); +rcu_read_unlock (); error =PTR_ERR (sma); goto out_free; }@@ -1693,22 +1694,8 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,*/ error = -EIDRM ; locknum =sem_lock (sma, sops, nsops); -if (un) { -if (un->semid ==-1 ) { -rcu_read_unlock (); - goto out_unlock_free; - }else { -/* -* rcu lock can be released, "un" cannot disappear: -* - sem_lock is acquired, thus IPC_RMID is -* impossible. -* - exit_sem is impossible, it always operates on -* current (or a dead task). -*/ - -rcu_read_unlock (); - } - } +if (un && un->semid ==-1 ) + goto out_unlock_free; error =try_atomic_semop (sma, sops, nsops, un,task_tgid_vnr (current));if (error <=0 ) { --To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 7 lines · Mike Galbraith
Quoted 5 lines · Rik van Riel
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.
Quoted 5 lines
Quoted 3 lines
There are no other callers to
find_alloc_undo.This should also solve the trinity issue reported by Sasha Levin.
I take it this is on top of the patchlet Sasha submitted?
Indeed, and all the other fixes that got submitted :)
Quoted 2 lines
(I hit rcu stall banging on patch set in rt with 60 synchronized core executive model if I let it run long enough, fwtw)
What are you using to trigger an rcu stall?
Quoted 2 lines
Quoted 2 lines
Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>
Signature · 1 line
All rights reversed.
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 12 lines · Rik van Riel
Quoted 11 lines · Mike Galbraith
Quoted 8 lines · Rik van Riel
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.There are no other callers to
find_alloc_undo.This should also solve the trinity issue reported by Sasha Levin.
Quoted 8 lines
Quoted 1 line
I take it this is on top of the patchlet Sasha submitted?
Indeed, and all the other fixes that got submitted :)
Quoted 2 lines
(I hit rcu stall banging on patch set in rt with 60 synchronized core executive model if I let it run long enough, fwtw)
What are you using to trigger an rcu stall?
Running a model of a userspace task scheduler. That was a fix or so ago now though. I'll try the set again on that box.
-Mike
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
In reply to Rik van RielQuoted 8 lines · Rik van Riel
Quoted 7 lines · Mike Galbraith
Quoted 5 lines · Rik van Riel
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
There are no other callers to
find_alloc_undo.This should also solve the trinity issue reported by Sasha Levin.
I take it this is on top of the patchlet Sasha submitted?
Indeed, and all the other fixes that got submitted :)
I plugged it into my 64 core rt box and beat on it again, no stalls or any other troubles noted.
-Mike
Quoted 35 lines · Sasha Levin
Quoted 4 lines · Rik van Riel
This series makes the sysv semaphore code more scalable, by reducing the time the semaphore lock is held, and making the locking more scalable for semaphore arrays with multiple semaphores.
Hi Rik,
Another issue that came up is:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
I'm a bit confused by what's going on in semtimedop with regards to rcu read lock, it seems that this behaviour is actually intentional?
rcu_read_lock(); sma = sem_obtain_object_check(ns, semid); if (IS_ERR(sma)) { if (un) rcu_read_unlock(); error = PTR_ERR(sma); goto out_free; }When I've looked at that it seems that not releasing the read lock was (very) intentional.
Quoted 6 lines
After that, the only code path that would release the lock starts with:
if (un) { ...So we won't release the lock at all if un is NULL?
Intentions notwithstanding, it is absolutely required to exit any and all RCU read-side critical sections prior to going into user mode.
I suggest removing the "if (un)".
Thanx, Paul
PATCH 1/1 ipc,sem: untangle RCU locking with find_alloc_undo
In reply to Sasha LevinQuoted 9 lines · On Tue, 26 Mar 2013 13:33:07 -0400 Sasha Levin <[email protected]> wrote:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
Sasha, this patch untangles the RCU locking with find_alloc_undo, and should fix the above issue. As a side benefit, this makes the code a little cleaner.
Next up: implement locking in a way that does not trigger any lockdep warnings...
---8<---
Subject: ipc,sem: untangle RCU locking with find_alloc_undo
The ipc semaphore code has a nasty RCU locking tangle, with both find_alloc_undo and semtimedop taking the rcu_read_lock(). The code can be cleaned up somewhat by only taking the rcu_read_lock once.
The only caller of find_alloc_undo is in semtimedop.
This should solve the trinity issue reported by Sasha Levin.
--- ipc/sem.c | 31 +++++++++---------------------- 1 files changed, 9 insertions(+), 22 deletions(-)
ipc/sem.c +9 -22
@@ -1646,22 +1646,23 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, alter =1 ; } +INIT_LIST_HEAD (&tasks); +if (undos) { +/* On success, find_alloc_undo takes the rcu_read_lock */ un =find_alloc_undo (ns, semid);if (IS_ERR (un)) { error =PTR_ERR (un); goto out_free; } - }else + }else { un =NULL ; +rcu_read_lock (); + } -INIT_LIST_HEAD (&tasks); - -rcu_read_lock (); sma =sem_obtain_object_check (ns, semid);if (IS_ERR (sma)) { -if (un) -rcu_read_unlock (); +rcu_read_unlock (); error =PTR_ERR (sma); goto out_free; }
@@ -1693,22 +1694,8 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,*/ error = -EIDRM ; locknum =sem_lock (sma, sops, nsops); -if (un) { -if (un->semid ==-1 ) { -rcu_read_unlock (); - goto out_unlock_free; - }else { -/* -* rcu lock can be released, "un" cannot disappear: -* - sem_lock is acquired, thus IPC_RMID is -* impossible. -* - exit_sem is impossible, it always operates on -* current (or a dead task). -*/ - -rcu_read_unlock (); - } - } +if (un && un->semid ==-1 ) + goto out_unlock_free; error =try_atomic_semop (sma, sops, nsops, un,task_tgid_vnr (current));if (error <=0 ) {
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 20 lines · Rik van Riel
Quoted 9 lines · On Tue, 26 Mar 2013 13:33:07 -0400 Sasha Levin <[email protected]> wrote:
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0It seems that we can leave semtimedop without releasing the rcu read lock.
Sasha, this patch untangles the RCU locking with
find_alloc_undo, and should fix the above issue. As a side benefit, this makes the code a little cleaner.Next up: implement locking in a way that does not trigger any lockdep warnings...
Quoted 8 lines
---8<---
Subject: ipc,sem: untangle RCU locking with
find_alloc_undoThe ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.
indeed!
Quoted 16 lines
The only caller of
find_alloc_undois in semtimedop.This should solve the trinity issue reported by Sasha Levin.
Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>--- ipc/sem.c | 31 +++++++++---------------------- 1 files changed, 9 insertions(+), 22 deletions(-)ipc/sem.c +0
-0
ipc/sem.c SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, · line 1653
Quoted 8 lines
ipc/sem.c +3
-0@@ -1646,22 +1646,23 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, alter =1 ; } +INIT_LIST_HEAD (&tasks); +if (undos) { +/* On success, find_alloc_undo takes the rcu_read_lock */ un =find_alloc_undo (ns, semid);
find_alloc_undo() has some nested rcu_read_locks of its own. We can simplify that as well. Will look into it, but don't want to introduce any more changes until we address all the issues with the patchset, and know it to behave.
Quoted 33 lines
ipc/sem.c +4
-14if (IS_ERR (un)) { error =PTR_ERR (un); goto out_free; } - }else + }else { un =NULL ; +rcu_read_lock (); + } -INIT_LIST_HEAD (&tasks); - -rcu_read_lock (); sma =sem_obtain_object_check (ns, semid);if (IS_ERR (sma)) { -if (un) -rcu_read_unlock (); +rcu_read_unlock (); error =PTR_ERR (sma); goto out_free; }@@ -1693,22 +1694,8 @@ SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops,*/ error = -EIDRM ; locknum =sem_lock (sma, sops, nsops); -if (un) { -if (un->semid ==-1 ) { -rcu_read_unlock (); - goto out_unlock_free; - }else { -/* -* rcu lock can be released, "un" cannot disappear: -* - sem_lock is acquired, thus IPC_RMID is
ipc/sem.c SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, · line 1698
Quoted 10 lines
ipc/sem.c +2
-8-* impossible. -* - exit_sem is impossible, it always operates on -* current (or a dead task). -*/ - -rcu_read_unlock (); - } - } +if (un && un->semid ==-1 ) + goto out_unlock_free;
Yeah, I was tempted in doing something much like this, but didn't want to change any existing logic. Hopefully we can get away with this and it fixes Sasha's issue.
ipc/sem.c SYSCALL_DEFINE4(semtimedop, int, semid, struct sembuf __user *, tsops, · line 1701
Quoted 3 lines
ipc/sem.c +0
-0error =try_atomic_semop (sma, sops, nsops, un,task_tgid_vnr (current));if (error <=0 ) {
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
In reply to Rik van RielQuoted 11 lines · Rik van Riel
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.The only caller of
find_alloc_undois in semtimedop.This should solve the trinity issue reported by Sasha Levin.
Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>
Looks good^Wbetter. I have nothing specific to say other than I've been staring at it for 10 minutes :)
Signature · 2 lines
Michel "Walken" Lespinasse A program is never fully debugged until the last user dies.
Sasha Levin
Tested-byRe: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 10 lines · Michel Lespinasse
Quoted 9 lines · Rik van Riel
The ipc semaphore code has a nasty RCU locking tangle, with both
find_alloc_undoand semtimedop taking thercu_read_lock(). The code can be cleaned up somewhat by only taking thercu_read_lockonce.The only caller of
find_alloc_undois in semtimedop.This should solve the trinity issue reported by Sasha Levin.
Quoted 7 lines
Quoted 2 lines
Reported-by Sasha Levin <[email protected]>Signed-off-by Rik van Riel <[email protected]>Looks good^Wbetter. I have nothing specific to say other than I've been staring at it for 10 minutes :)
Reviewed-by Michel Lespinasse <[email protected]>
And the warnings are gone:
Thanks,
Sasha
Re: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
In reply to Rik van RielQuoted 11 lines · Rik van Riel
Quoted 8 lines · On Tue, 26 Mar 2013 13:33:07 -0400 Sasha Levin <[email protected]> wrote:
Quoted 8 lines
[ 96.347341] ================================================ [ 96.348085] [ BUG: lock held when returning to user space! ] [ 96.348834] 3.9.0-rc4-next-20130326-sasha-00011-gbcb2313 #318 Tainted: G W [ 96.360300] ------------------------------------------------ [ 96.361084] trinity-child9/7583 is leaving the kernel with locks still held! [ 96.362019] 1 lock held by trinity-child9/7583: [ 96.362610] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192eafb>] SYSC_semtimedop+0x1fb/0xec0
Quoted 7 lines
Quoted 1 line
Quoted 1 line
It seems that we can leave semtimedop without releasing the rcu read lock.
Sasha, this patch untangles the RCU locking with
find_alloc_undo, and should fix the above issue. As a side benefit, this makes the code a little cleaner.Next up: implement locking in a way that does not trigger any lockdep warnings...
The following is mostly unrelated to this patch but close enough:
semctl_main() has a snippet that looks like this:
err = -EINVAL;
if(semnum < 0 || semnum >= nsems)
goto out_unlock;
sem_lock(sma, NULL, -1);Which means we'll try unlocking the sma without trying to lock it first. It makes lockdep unhappy:
[ 95.528492] =====================================
[ 95.529251] [ BUG: bad unlock balance detected! ]
[ 95.529897] 3.9.0-rc4-next-20130328-sasha-00014-g91a3267 #319 Tainted: G W
[ 95.530190] -------------------------------------
[ 95.530190] trinity-child14/9123 is trying to release lock (&(&new->lock)->rlock) at:
[ 95.530190] [<ffffffff8192f8e4>] semctl_main+0xe54/0xf00
[ 95.530190] but there are no more locks to release!
[ 95.530190]
[ 95.530190] other info that might help us debug this:
[ 95.530190] 1 lock held by trinity-child14/9123:
[ 95.530190] #0: (rcu_read_lock){.+.+..}, at: [<ffffffff8192ea90>] semctl_main+0x0/0xf00
[ 95.530190]
[ 95.530190] stack backtrace:
[ 95.530190] Pid: 9123, comm: trinity-child14 Tainted: G W 3.9.0-rc4-next-20130328-sasha-00014-g91a3267 #319
[ 95.530190] Call Trace:
[ 95.530190] [<ffffffff8192f8e4>] ? semctl_main+0xe54/0xf00
[ 95.530190] [<ffffffff8117b7e6>] print_unlock_imbalance_bug+0xf6/0x110
[ 95.530190] [<ffffffff8192f8e4>] ? semctl_main+0xe54/0xf00
[ 95.530190] [<ffffffff81180a35>] lock_release_non_nested+0xd5/0x320
[ 95.530190] [<ffffffff8122e3ab>] ? __do_fault+0x42b/0x530
[ 95.530190] [<ffffffff81179da2>] ? get_lock_stats+0x22/0x70
[ 95.530190] [<ffffffff81179e5e>] ? put_lock_stats.isra.14+0xe/0x40
[ 95.530190] [<ffffffff8192f8e4>] ? semctl_main+0xe54/0xf00
[ 95.530190] [<ffffffff81180f1e>] lock_release+0x29e/0x3b0
[ 95.530190] [<ffffffff819451f4>] ? security_ipc_permission+0x14/0x20
[ 95.530190] [<ffffffff83daf33e>] _raw_spin_unlock+0x1e/0x60
[ 95.530190] [<ffffffff8192f8e4>] semctl_main+0xe54/0xf00
[ 95.530190] [<ffffffff8192ea90>] ? SYSC_semtimedop+0xe30/0xe30
[ 95.530190] [<ffffffff8109d188>] ? kvm_clock_read+0x38/0x70
[ 95.530190] [<ffffffff8114feb5>] ? sched_clock_local+0x25/0xa0
[ 95.530190] [<ffffffff811500e8>] ? sched_clock_cpu+0xf8/0x110
[ 95.530190] [<ffffffff83db3814>] ? __do_page_fault+0x514/0x5e0
[ 95.530190] [<ffffffff81179da2>] ? get_lock_stats+0x22/0x70
[ 95.530190] [<ffffffff81179e5e>] ? put_lock_stats.isra.14+0xe/0x40
[ 95.530190] [<ffffffff83db3814>] ? __do_page_fault+0x514/0x5e0
[ 95.530190] [<ffffffff8113dc0e>] ? up_read+0x1e/0x40
[ 95.530190] [<ffffffff83db3814>] ? __do_page_fault+0x514/0x5e0
[ 95.530190] [<ffffffff811c6500>] ? rcu_eqs_exit_common+0x60/0x260
[ 95.530190] [<ffffffff81202b9d>] ? user_enter+0xfd/0x130
[ 95.530190] [<ffffffff81202c85>] ? user_exit+0xb5/0xe0
[ 95.530190] [<ffffffff8192faf9>] SyS_semctl+0x69/0x430
[ 95.530190] [<ffffffff81076ea0>] ? syscall_trace_enter+0x20/0x2e0
[ 95.530190] [<ffffffff83db7d18>] tracesys+0xe1/0xe6I'm thinking that the solution is as simple as:
ipc/sem.c +4 -2
@@ -1333,8 +1333,10 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,/* GETVAL, GETPID, GETNCTN, GETZCNT: fall-through */ } err = -EINVAL ; -if (semnum <0 || semnum >= nsems) - goto out_unlock; +if (semnum <0 || semnum >= nsems) { +rcu_read_unlock (); + goto out_wakeup; + }sem_lock (sma,NULL ,-1 ); curr = &sma->sem_base[semnum];
But I'm not 100% sure if I don't mess up anything else.
Thanks,
Sasha
Rik van Riel
Reviewed-byRe: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 1 line · Sasha Levin
I'm thinking that the solution is as simple as:
Your patch is absolutely correct. All it needs now is your signed-off-by, so Andrew can merge it into -mm :)
Quoted 18 lines
ipc/sem.c +4
-2@@ -1333,8 +1333,10 @@ static int semctl_main(struct ipc_namespace *ns, int semid, int semnum,/* GETVAL, GETPID, GETNCTN, GETZCNT: fall-through */ } err = -EINVAL ; -if (semnum <0 || semnum >= nsems) - goto out_unlock; +if (semnum <0 || semnum >= nsems) { +rcu_read_unlock (); + goto out_wakeup; + }sem_lock (sma,NULL ,-1 ); curr = &sma->sem_base[semnum];
Quoted 1 line
But I'm not 100% sure if I don't mess up anything else.
I checked the surrounding code, it all looks fine.
Signature · 1 line
All rights reversed.
Davidlohr Bueso
Reviewed-byRe: [PATCH -mm -next] ipc,sem: untangle RCU locking with find_alloc_undo
Quoted 8 lines · Rik van Riel
Quoted 1 line · Sasha Levin
I'm thinking that the solution is as simple as:
Your patch is absolutely correct. All it needs now is your signed-off-by, so Andrew can merge it into -mm :)
Reviewed-by Rik van Riel <[email protected]>