On 7/31/19 11:07 AM, Rik van Riel wrote: > On Wed, 2019-07-31 at 10:15 -0400, Waiman Long wrote: >> On 7/31/19 9:48 AM, Rik van Riel wrote: >>> On Tue, 2019-07-30 at 17:01 -0400, Waiman Long wrote: >>>> On 7/29/19 8:26 PM, Rik van Riel wrote: >>>>> On Mon, 2019-07-29 at 17:42 -0400, Waiman Long wrote: >>>>> >>>>>> What I have found is that a long running process on a mostly >>>>>> idle >>>>>> system >>>>>> with many CPUs is likely to cycle through a lot of the CPUs >>>>>> during >>>>>> its >>>>>> lifetime and leave behind its mm in the active_mm of those >>>>>> CPUs. My >>>>>> 2-socket test system have 96 logical CPUs. After running the >>>>>> test >>>>>> program for a minute or so, it leaves behind its mm in about >>>>>> half >>>>>> of >>>>>> the >>>>>> CPUs with a mm_count of 45 after exit. So the dying mm will >>>>>> stay >>>>>> until >>>>>> all those 45 CPUs get new user tasks to run. >>>>> OK. On what kernel are you seeing this? >>>>> >>>>> On current upstream, the code in native_flush_tlb_others() >>>>> will send a TLB flush to every CPU in mm_cpumask() if page >>>>> table pages have been freed. >>>>> >>>>> That should cause the lazy TLB CPUs to switch to init_mm >>>>> when the exit->zap_page_range path gets to the point where >>>>> it frees page tables. >>>>> >>>> I was using the latest upstream 5.3-rc2 kernel. It may be the >>>> case >>>> that >>>> the mm has been switched, but the mm_count field of the active_mm >>>> of >>>> the >>>> kthread is not being decremented until a user task runs on a CPU. >>> Is that something we could fix from the TLB flushing >>> code? >>> >>> When switching to init_mm, drop the refcount on the >>> lazy mm? >>> >>> That way that overhead is not added to the context >>> switching code. >> I have thought about that. That will require changing the active_mm >> of >> the current task to point to init_mm, for example. Since TLB flush is >> done in interrupt context, proper coordination between interrupt and >> process context will require some atomic instruction which will >> defect >> the purpose. > Would it be possible to work around that by scheduling > a work item that drops the active_mm? > > After all, a work item runs in a kernel thread, so by > the time the work item is run, either the kernel will > still be running the mm you want to get rid of as > active_mm, or it will have already gotten rid of it > earlier. Yes, that may work. Thanks, Longman
