diff options
author | Patrick Williams <iawillia@us.ibm.com> | 2012-07-06 15:40:43 -0500 |
---|---|---|
committer | A. Patrick Williams III <iawillia@us.ibm.com> | 2012-07-11 08:18:41 -0500 |
commit | 14a2721d2c87dc13f1ef66818c41cd0848dd52db (patch) | |
tree | 66ae0220b0443f9a87d87075279ba9a5c08ba412 /src/kernel/cpumgr.C | |
parent | a23283c6facfee055c9c6d43e23a04ca02edc467 (diff) | |
download | talos-hostboot-14a2721d2c87dc13f1ef66818c41cd0848dd52db.tar.gz talos-hostboot-14a2721d2c87dc13f1ef66818c41cd0848dd52db.zip |
Live-lock issues in memory allocator.
* Debug tool for PageManager.
* Support PageMgr allocations of non-2^k size.
* Switch page-allocation to always be in kernel-mode.
While investigating issue 44511, I noticed two problesm with the
memory page allocator (PageManager). First, the allocator did
not support allocations of pages which were not a power of 2,
which would result in pages appearing to "leak". Second, in
situations where a large allocation was requested and there was
not a large chunk available, the allocation would enter a
live-lock condition where coalescing would never occur.
Switched the PageManager so that all allocations happen in
kernel space. This allows us to force memory-release operations
on the syscall path when we are out of memory and also put in
place a task_yield call which will allow coalescing to eventually
occur. Issue 44523 is suppose to fully resolve any of these
live-lock paths.
RTC: 44511
Change-Id: Ifefd5d0996ee6914e291c862fac0c7b76980717f
Reviewed-on: http://gfw160.austin.ibm.com:8080/gerrit/1330
Tested-by: Jenkins Server
Reviewed-by: Daniel M. Crowell <dcrowell@us.ibm.com>
Reviewed-by: A. Patrick Williams III <iawillia@us.ibm.com>
Diffstat (limited to 'src/kernel/cpumgr.C')
-rw-r--r-- | src/kernel/cpumgr.C | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/src/kernel/cpumgr.C b/src/kernel/cpumgr.C index 3bb52b63c..2a5c2edcc 100644 --- a/src/kernel/cpumgr.C +++ b/src/kernel/cpumgr.C @@ -192,11 +192,11 @@ void CpuManager::executePeriodics(cpu_t * i_cpu) if(0 == (i_cpu->periodic_count % CPU_PERIODIC_CHECK_MEMORY)) { uint64_t pcntAvail = PageManager::queryAvail(); - if(pcntAvail < 16) + if(pcntAvail < PageManager::LOWMEM_NORM_LIMIT) { VmmManager::flushPageTable(); ++(i_cpu->periodic_count); // prevent another flush below - if(pcntAvail < 5) + if(pcntAvail < PageManager::LOWMEM_CRIT_LIMIT) { VmmManager::castOutPages(VmmManager::CRITICAL); } |