Message ID | 20210916134748.67712-1-songmuchun@bytedance.com (mailing list archive) |
---|---|
Headers | show |
Series | Use obj_cgroup APIs to charge the LRU pages | expand |
Hi Muchun! On Thu, Sep 16, 2021 at 09:47:35PM +0800, Muchun Song wrote: > This version is rebased over linux 5.15-rc1, because Shakeel has asked me > if I could do that. I rework some code suggested by Roman as well in this > version. I have not removed the Acked-by tags which are from Roman, because > this version is not based on the folio relevant. If Roman wants me to > do this, please let me know, thanks. I'm fine with this, thanks for clarifying. > > Since the following patchsets applied. All the kernel memory are charged > with the new APIs of obj_cgroup. > > [v17,00/19] The new cgroup slab memory controller[1] > [v5,0/7] Use obj_cgroup APIs to charge kmem pages[2] > > But user memory allocations (LRU pages) pinning memcgs for a long time - > it exists at a larger scale and is causing recurring problems in the real > world: page cache doesn't get reclaimed for a long time, or is used by the > second, third, fourth, ... instance of the same job that was restarted into > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory, > and make page reclaim very inefficient. I've an idea: what if we use struct list_lru_memcg as an intermediate object between an individual page and struct mem_cgroup? It could contain a pointer to a memory cgroup structure (not even sure if a reference is needed), and a lru page can contain a pointer to the lruvec instead of memcg/objcg. This approach can probably simplify the locking scheme. But what's more important, it can dramatically reduce the number of css_get()/put() calls. The latter are not particularly cheap after the deletion of a cgroup: they are atomic_dec()'s. As a result, the reclaim efficiency could be much better. The downside: we will need to update page->lruvec_memcg pointers on reparenting pages during the cgroup removal. This is a rough idea, maybe there are significant reasons why it's not possible or will be way worse. But I think it's worth discussing. What do you think? Thanks!
On Fri, Sep 17, 2021 at 9:29 AM Roman Gushchin <guro@fb.com> wrote: > > Hi Muchun! > > On Thu, Sep 16, 2021 at 09:47:35PM +0800, Muchun Song wrote: > > This version is rebased over linux 5.15-rc1, because Shakeel has asked me > > if I could do that. I rework some code suggested by Roman as well in this > > version. I have not removed the Acked-by tags which are from Roman, because > > this version is not based on the folio relevant. If Roman wants me to > > do this, please let me know, thanks. > > I'm fine with this, thanks for clarifying. > > > > > Since the following patchsets applied. All the kernel memory are charged > > with the new APIs of obj_cgroup. > > > > [v17,00/19] The new cgroup slab memory controller[1] > > [v5,0/7] Use obj_cgroup APIs to charge kmem pages[2] > > > > But user memory allocations (LRU pages) pinning memcgs for a long time - > > it exists at a larger scale and is causing recurring problems in the real > > world: page cache doesn't get reclaimed for a long time, or is used by the > > second, third, fourth, ... instance of the same job that was restarted into > > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory, > > and make page reclaim very inefficient. > > I've an idea: what if we use struct list_lru_memcg as an intermediate object > between an individual page and struct mem_cgroup? > > It could contain a pointer to a memory cgroup structure (not even sure if a > reference is needed), and a lru page can contain a pointer to the lruvec instead > of memcg/objcg. Hi Roman, If I understand properly, here you mean the struct page has a pointer to the struct lruvec not struct list_lru_memcg. What's the functionality of the struct list_lru_memcg? Would you mind exposing more details? > > This approach can probably simplify the locking scheme. But what's more > important, it can dramatically reduce the number of css_get()/put() calls. > The latter are not particularly cheap after the deletion of a cgroup: > they are atomic_dec()'s. As a result, the reclaim efficiency could be much > better. The downside: we will need to update page->lruvec_memcg pointers on > reparenting pages during the cgroup removal. Here we need to update page->lruvec_memcg pointers one by one, right? Because the lru lock is per lruvec, the locking scheme still need to be as proposed by this series when the page->lruvec_memcg is changed If I understand properly. It's likely that I don't get your point. Looking forward to your further details. Thanks. > > This is a rough idea, maybe there are significant reasons why it's not possible > or will be way worse. But I think it's worth discussing. What do you think? > > Thanks!
On Fri, Sep 17, 2021 at 06:49:21PM +0800, Muchun Song wrote: > On Fri, Sep 17, 2021 at 9:29 AM Roman Gushchin <guro@fb.com> wrote: > > > > Hi Muchun! > > > > On Thu, Sep 16, 2021 at 09:47:35PM +0800, Muchun Song wrote: > > > This version is rebased over linux 5.15-rc1, because Shakeel has asked me > > > if I could do that. I rework some code suggested by Roman as well in this > > > version. I have not removed the Acked-by tags which are from Roman, because > > > this version is not based on the folio relevant. If Roman wants me to > > > do this, please let me know, thanks. > > > > I'm fine with this, thanks for clarifying. > > > > > > > > Since the following patchsets applied. All the kernel memory are charged > > > with the new APIs of obj_cgroup. > > > > > > [v17,00/19] The new cgroup slab memory controller[1] > > > [v5,0/7] Use obj_cgroup APIs to charge kmem pages[2] > > > > > > But user memory allocations (LRU pages) pinning memcgs for a long time - > > > it exists at a larger scale and is causing recurring problems in the real > > > world: page cache doesn't get reclaimed for a long time, or is used by the > > > second, third, fourth, ... instance of the same job that was restarted into > > > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory, > > > and make page reclaim very inefficient. > > > > I've an idea: what if we use struct list_lru_memcg as an intermediate object > > between an individual page and struct mem_cgroup? > > > > It could contain a pointer to a memory cgroup structure (not even sure if a > > reference is needed), and a lru page can contain a pointer to the lruvec instead > > of memcg/objcg. lruvec_memcg I mean. > > Hi Roman, > > If I understand properly, here you mean the struct page has a pointer > to the struct lruvec not struct list_lru_memcg. What's the functionality > of the struct list_lru_memcg? Would you mind exposing more details? So the basic idea is simple: a lru page charged to a memcg is associated with a per-memcg lruvec (list_lru_memcg), which is associated with a memory cgroup. And after your patches there is a second link of associations: page to objcg to memcg: 1) page->objcg->memcg 2) page->list_lru_memcg->memcg (those are not necessarily direct pointers, but generally speaking, relations). My gut feeling is that if we can merge them into just 2) and use list_lru_memcg as an intermediate object between pages and memory cgroups, the whole thing can be more efficient and beautiful. Yes, on reparenting we'd need to scan over all pages in the lru list, but hopefully we can do it from a worker context. And it's not such a big deal as with slab objects, where we simple had no list of all objects. Again, I'm not 100% sure if it's possible and worth it, so it shouldn't block your patchset if everybody else like it. Thanks
On Sat, Sep 18, 2021 at 8:13 AM Roman Gushchin <guro@fb.com> wrote: > > On Fri, Sep 17, 2021 at 06:49:21PM +0800, Muchun Song wrote: > > On Fri, Sep 17, 2021 at 9:29 AM Roman Gushchin <guro@fb.com> wrote: > > > > > > Hi Muchun! > > > > > > On Thu, Sep 16, 2021 at 09:47:35PM +0800, Muchun Song wrote: > > > > This version is rebased over linux 5.15-rc1, because Shakeel has asked me > > > > if I could do that. I rework some code suggested by Roman as well in this > > > > version. I have not removed the Acked-by tags which are from Roman, because > > > > this version is not based on the folio relevant. If Roman wants me to > > > > do this, please let me know, thanks. > > > > > > I'm fine with this, thanks for clarifying. > > > > > > > > > > > Since the following patchsets applied. All the kernel memory are charged > > > > with the new APIs of obj_cgroup. > > > > > > > > [v17,00/19] The new cgroup slab memory controller[1] > > > > [v5,0/7] Use obj_cgroup APIs to charge kmem pages[2] > > > > > > > > But user memory allocations (LRU pages) pinning memcgs for a long time - > > > > it exists at a larger scale and is causing recurring problems in the real > > > > world: page cache doesn't get reclaimed for a long time, or is used by the > > > > second, third, fourth, ... instance of the same job that was restarted into > > > > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory, > > > > and make page reclaim very inefficient. > > > > > > I've an idea: what if we use struct list_lru_memcg as an intermediate object > > > between an individual page and struct mem_cgroup? > > > > > > It could contain a pointer to a memory cgroup structure (not even sure if a > > > reference is needed), and a lru page can contain a pointer to the lruvec instead > > > of memcg/objcg. > > lruvec_memcg I mean. Thanks for your clarification. > > > > > Hi Roman, > > > > If I understand properly, here you mean the struct page has a pointer > > to the struct lruvec not struct list_lru_memcg. What's the functionality > > of the struct list_lru_memcg? Would you mind exposing more details? > > So the basic idea is simple: a lru page charged to a memcg is associated with > a per-memcg lruvec (list_lru_memcg), which is associated with a memory cgroup. > And after your patches there is a second link of associations: page to objcg > to memcg: > > 1) page->objcg->memcg > 2) page->list_lru_memcg->memcg > > (those are not necessarily direct pointers, but generally speaking, relations). > > My gut feeling is that if we can merge them into just 2) and use list_lru_memcg > as an intermediate object between pages and memory cgroups, the whole thing can > be more efficient and beautiful. > > Yes, on reparenting we'd need to scan over all pages in the lru list, but > hopefully we can do it from a worker context. And it's not such a big deal as > with slab objects, where we simple had no list of all objects. struct list_lru_memcg seems to be redundant, it just contains a pointer to struct mem_cgroup. We need to update each page->lruvec_memcg, why not update page->memcg_data directly to its parent memcg? The update of page->lruvec_memcg should be under both child and parent's lruvec lock, right? I suppose scanning over all pages may be a problem if there are many pages. Thanks. > > Again, I'm not 100% sure if it's possible and worth it, so it shouldn't block > your patchset if everybody else like it. > > Thanks