[SCHED] VCPU should be initialized completely before deletion There is a race in vsched_del_vcpu() - we can kill migration_thread() even if it has not started yet, i.e. migration_thread() function is not called at all. So, migrate_live_tasks() and migrate_dead_tasks() will not be called on this vcpu while migration thread is killed. But there can be some tasks, that have already migrated on thi...This bug can be easily reproduced. On a busy host with many running
tasks user can run:
In this case, after the second vzctl, migration thread on VCPU 2 will be
created and just waked up, but it can be not really started (scheduled)
yet if there are a lot of other more priority tasks running on the host.
If it will not be scheduled before the third vzctl call, there will be
kernel bug in vsche...
[BC] Fix potential beancounter refcount leakOn some error paths we forget to put beancounter.
This patch fixes two such places:
- sys_setluid()
- bc_entry_open()
[SCHED] find_idle_vcpu() mask check fixIn find_idle_vcpu() we skip VCPU's with ID's that is not
set in physical '*cpus' mask. It's incorrect.
We must skip VCPU's that has appropriate VCPU->last_pcpu
[SCHED] Cleanup: use vcpu_last_pcpu macro instead of vcpu->last_pcpuReplace vcpu->last_pcpu by vcpu_last_pcpu(vcpu),
to fix compilation without CONFIG_VSCHED_VCPU
[IOPRIO] elevator switch oops fix When elevator switch happens and UBs persist, putting of async cfqq can happen second time due to non-NULL value in array.http://bugzilla.openvz.org/show_bug.cgi?id=526
[SCHED] find_busiest_queue() should select VCPUs from given vsched onlyIn new scheme, we choose vsched in find_busiest_vsched(),
i.e. before find_busiest_queue(), so when we look
for busiest queue we must consider this vsched VCPU's only.
[SCHED] remove debug hunk from previous balance patch My previous patch for load_balance() contains wrong condition statement, that I forget to remove after debugging.In 028stab025.1 load_balance() will not pull tasks from a busiest VCPU's,
if there are < 2 tasks running on current VCPU. Attached patch removes
this incorrect check and fixes the problem.
[IOPRIO] new cfq queue putting mechanismIt's better to use original cfqq put function from CFQ then rewrite it.
Use elevator_ops structure for exporting it.
[SCHED] Improve vcpu scheduling taking into account cache hotness In original VZ kernel schedule_vcpu() takes next VCPU from vsched->active list, and it doesn't take in to account vcpu->last_pcpu, so VCPU's can jump from PCPU to PCPU too often.Try to skip 'hot' VCPU's, i.e. VCPU's that were running on some
other PCPU recently.
Time slice threshold is tunable via /proc/sys/kernel/vcpu_hot_timeslice
[IOPRIO] Fix cfqq index calculation in async caseField ioprio of task_struct consits of two numbers:
1) value of class (bits 14-16),
2) value of data (bits 0-13).
Value of data is allowed to belong the range [0, 7].
In current implementation of cfq_set_request tsk->ioprio is
used as index of *async_cfqq[8] array.
It is wrong because tsk->ioprio can be >> 8.
This can cause to either corruption or reading insufficient value:
cfq_set_request...
[IOPRIO] Call bc_findcreate_cfq_bc() out of q->queue_loc (fix 2)Fix to fix for call bc_findcreate_cfq_bc() out of q->queue_lock.
iopriv should be initialized in both cases.
[SCHED] Improve idle load balanceIdle balance is called from an idle thread on rebalance_tick().
load_balance() tries to find busiest group in idle_vsched,
where there are no really running tasks.
With this patch, load_balance() will try to find a busiest vsched first,
and in case of success, then find busiest group inside this vsched, and
so on...
[LOCKDEP] Another fix for virtualized filesystems lockdepAs described before, filesystems in our kernels are
no longer static objects and thus lockdep refuses to
work. This was (wrongly) fixed by setting one static
class for all super block's semaphores and locks.
It turned out that different filesystems use different
lock ordering for sb locks and some other ones, e.g.
UDF may take inode->i_mutex under sb->s_lock, while
ext3 takes sb->s_lock under ...
[PROC] mainstream: race between proc_lookup() and sys_delete_module()Fix for the race between proc_lookup() and sys_delete_module():
proc_lookup() can find PDE under proc_subdir_lock,
on 2nd CPU sys_delete_module() removes pde and module,
then first CPU tries to get de and module in proc_get_inode()...
Bum...
[VESTATS] use jiffies instead of cycles for mm statsuse jiffies instead of cycles for mm stats about page allocation latency.
This implementation if very simple but it's strictly not that accurate,
because we can add 10 000 000 (or more) cycles (it's ~ 1 jiffy)
even if actual allocation consumes < 10 000 cycles,
but jiffy has been changed at the moment.
[CPT] Fix IPv6 addresses restoreAll IPv6 addresses based on MAC are created with valid lifetime 0.
We checkpoint them and try to restore, but fail as inet6_addr_add()
returns -EINVAL if valid_lft is zero.
We can use ifaddr flags to find correct values for prefered and
valid life times.
TODO:
Kernel creates automatically local ipv6 address based on MAC address on it
when interface is upped. We can manually remove this addre...
[PATCH] mainstream: fix sys_accept() error path* d_alloc() in sock_attach_fd() fails leaving ->f_dentry NULL
* bail out to out_fd label, which does fput()/__fput() on new file
* but __fput() assumes valid ->f_dentry
[PATCH] revert exec perms race fixBackout exec perms patches:
diff-exec-perm-20060317
diff-exec-perm-jffs2-20060317
This patches are correct, but the race they fix is not exploitable.
They do harm since people can't easily compile 3rd party
(out of mainstream) patches/filesystems, so better drop to make everyone
lifes easier.
[CPT] unlimit dcachesize on restoreRecently we have added adjusting of 3 limits on restore
to not fail because of hitting limits.
Now we have to add another one - dcachesize.
[PATCH] Fix unlocked access to task list from /proc/pid/oom_scoreFailing code was prefetch hidden in
list_for_each_entry(child, &p->children, sibling) {
in badness(). badness() is reachable from two points.
One is proc_oom_score, another is
out_of_memory() => oom_select_bad_process() => badness().
Second path grabs tasklist_lock, while first doesn't.
[CPT] bug in restore net routesWhen netroute section in dump is padded, restore tries
to interpret padding as the next rtnetlink message and deadlocks
interpreting it as message of zero length.
[EXT3] dirindex error handling issues (additional) I've found one more place where we exit with wrong error code. (Andreas Dilger agreed that it should be fixed), but not send it to lkml yet.Signed-Off-By: Monakhov Dmitriy <dmonakhov@openvz.org>
[EXT3] dirindex error handling issues (fix) my previous patch changed error message semantics in do_split(). Initially when ext3_append() failed we just exit without printing error. In fact ext3_append() may fail, it is legal and it's happens qite often (ENOSPC for example). This cause annoying fake error message. So restore this semantic as it was before patch.Signed-off-by: Monakhov Dmitriy <dmonakhov@openvz.org>
Cc: Andreas Dilger <adilger@clusterfs.com>
Cc: Theodore Ts'o <tytso@mit.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
[EXT3] dirindex error handling issues- ext3_dx_find_entry() exit with out setting proper error pointer
- do_split() exit with out setting proper error pointer
it is realy painful because many callers contain folowing code:
de = do_split(handle,dir, &bh, frame, &hinfo, &retval);
if (!(de))
return retval;
<<< WOW retval wasn't changed by do_split(), so caller failed
...
[CPT] Reserve vsyscall section and object IDSupport in CPT for vsyscall is not implemented in 2.6.18 kernel yet.
This patch is needed to synchronize section and object IDs
in 2.6.9 and 2.6.18 kernel to not break 2.6.9 -> 2.6.18 migration.
[LOCKDEP] Fix lockdep for virtualized runqueueslockdep is debugging technology that tracks the
sequence of spin_lock()-s in kernel. Each lock has
a "class" assigned. Rules:
1. class must be a static object as it is registered
within lockdep internals
2. two locks of the same class must be taken with
_nested suffix to explicitly tell lockdep that this
is not a deadlock
Problem is that runqueues and filesystem types are
no longer st...
[LOCKDEP] Fix lockdep for virtualized fstype lockdep is debugging technology that tracks the sequence of spin_lock()-s in kernel. Each lock has a "class" assigned. Rules: 1. class must be a static object as it is registered within lockdep internals 2. two locks of the same class must be taken with _nested suffix to explicitly tell lockdep that this is not a deadlockProblem is that runqueues and filesystem types are
no longer static in OpenVZ kernel.
File system types are resolved that way:
set common class for all super blocks. I haven't
yet found places where two locks are taken for
two super blocks
OpenVZ team
