| Message ID | 1663325892-9825-1-git-send-email-zhaoyang.huang@unisoc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [RFC] mm: check global free_list if there is ongoing reclaiming when pcp fail | expand |
On Fri, Sep 16, 2022 at 06:58:12PM +0800, zhaoyang.huang wrote: > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > Check the global free list again even if rmqueue_bulk failed for pcp pages when > there is ongoing reclaiming, which could eliminate potential direct reclaim by > chance. > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com> Patch does not apply and may be based on a custom kernel that introduced a problem. There is no description of what problem this is trying to fix. Checking the status of reclaim for a specific zone in this path would be a little unexpected. If allocation pressure is exceeding the ability of reclaim to make progress then the caller likely needs to take action like direct reclaim. If the allocation failure is due to a high-order failure then it may need to enter direct compaction etc.
On Mon, Sep 19, 2022 at 6:22 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > On Fri, Sep 16, 2022 at 06:58:12PM +0800, zhaoyang.huang wrote: > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > Check the global free list again even if rmqueue_bulk failed for pcp pages when > > there is ongoing reclaiming, which could eliminate potential direct reclaim by > > chance. > > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > Patch does not apply and may be based on a custom kernel that introduced > a problem. There is no description of what problem this is trying to > fix. Checking the status of reclaim for a specific zone in this path would > be a little unexpected. If allocation pressure is exceeding the ability > of reclaim to make progress then the caller likely needs to take action > like direct reclaim. If the allocation failure is due to a high-order > failure then it may need to enter direct compaction etc. Agree with the above comment. This is a proposal aiming at avoiding direct reclaiming things with minimum cost, that is to say, about 5 CPU instructions in return with the overhead of function calls which has both of several loops inside and potential throttle sleep by IO congestion etc. > > -- > Mel Gorman > SUSE Labs
On Tue, Sep 20, 2022 at 09:45:35AM +0800, Zhaoyang Huang wrote: > On Mon, Sep 19, 2022 at 6:22 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > > > On Fri, Sep 16, 2022 at 06:58:12PM +0800, zhaoyang.huang wrote: > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > Check the global free list again even if rmqueue_bulk failed for pcp pages when > > > there is ongoing reclaiming, which could eliminate potential direct reclaim by > > > chance. > > > > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > Patch does not apply and may be based on a custom kernel that introduced > > a problem. There is no description of what problem this is trying to > > fix. Checking the status of reclaim for a specific zone in this path would > > be a little unexpected. If allocation pressure is exceeding the ability > > of reclaim to make progress then the caller likely needs to take action > > like direct reclaim. If the allocation failure is due to a high-order > > failure then it may need to enter direct compaction etc. > > Agree with the above comment. This is a proposal aiming at avoiding > direct reclaiming things with minimum cost, that is to say, about 5 > CPU instructions in return with the overhead of function calls which > has both of several loops inside and potential throttle sleep by IO > congestion etc. If the refill fails and kswapd is failing to keep up then actions like direct reclaim or compaction are inevitable. At best, this patch would race to allocate pages in one context that are being freed in parallel by another context. Nak.
On Tue, Sep 20, 2022 at 4:46 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > On Tue, Sep 20, 2022 at 09:45:35AM +0800, Zhaoyang Huang wrote: > > On Mon, Sep 19, 2022 at 6:22 PM Mel Gorman <mgorman@techsingularity.net> wrote: > > > > > > On Fri, Sep 16, 2022 at 06:58:12PM +0800, zhaoyang.huang wrote: > > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > > > Check the global free list again even if rmqueue_bulk failed for pcp pages when > > > > there is ongoing reclaiming, which could eliminate potential direct reclaim by > > > > chance. > > > > > > > > Signed-off-by: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > Patch does not apply and may be based on a custom kernel that introduced > > > a problem. There is no description of what problem this is trying to > > > fix. Checking the status of reclaim for a specific zone in this path would > > > be a little unexpected. If allocation pressure is exceeding the ability > > > of reclaim to make progress then the caller likely needs to take action > > > like direct reclaim. If the allocation failure is due to a high-order > > > failure then it may need to enter direct compaction etc. > > > > Agree with the above comment. This is a proposal aiming at avoiding > > direct reclaiming things with minimum cost, that is to say, about 5 > > CPU instructions in return with the overhead of function calls which > > has both of several loops inside and potential throttle sleep by IO > > congestion etc. > > If the refill fails and kswapd is failing to keep up then actions like > direct reclaim or compaction are inevitable. At best, this patch would > race to allocate pages in one context that are being freed in parallel by > another context. > > Nak. ok, I have noticed that the latest modification has made some changes on this path. thanks for comment > > -- > Mel Gorman > SUSE Labs
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index e008a3d..7e99f7d 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -3729,7 +3729,8 @@ struct page *rmqueue(struct zone *preferred_zone, migratetype != MIGRATE_MOVABLE) { page = rmqueue_pcplist(preferred_zone, zone, order, gfp_flags, migratetype, alloc_flags); - goto out; + if (page || !test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) + goto out; } }