| Message ID | 749be737-bbba-cf4d-0d97-7657e3b1b76b@huawei.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: > Hi, Paul: > > I try to debug this problem and found this solution could work well for both problem scene. > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > index 85c5a88..dbc14a7 100644 > --- a/kernel/rcu/tree_plugin.h > +++ b/kernel/rcu/tree_plugin.h > @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > if (__rcu_reclaim(rdp->rsp->name, list)) > cl++; > c++; > - local_bh_enable(); > + _local_bh_enable(); > cond_resched_rcu_qs(); > list = next; > } > > > The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use > local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, > what do you think about it? Please give me some suggestion. From what I can see, there is absolutely no guarantee that cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee that it will process any pending softirqs -- and that is not part of its job in any case. So I cannot recommend the above patch. On efficient handling of large invalid packets (that is still the issue, right?), I must defer to Dave and Eric. Thanx, Paul > Thanks. > Ding > > On 2016/11/21 9:28, Ding Tianhong wrote: > > > > > > On 2016/11/21 8:13, Paul E. McKenney wrote: > >> On Sat, Nov 19, 2016 at 12:22:09AM -0800, Paul E. McKenney wrote: > >>> On Sat, Nov 19, 2016 at 03:50:32PM +0800, Ding Tianhong wrote: > >>>> > >>>> > >>>> On 2016/11/18 21:01, Paul E. McKenney wrote: > >>>>> On Fri, Nov 18, 2016 at 08:40:09PM +0800, Ding Tianhong wrote: > >>>>>> The commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") > >>>>>> will introduce a new problem that when huge IP abnormal packet arrived, > >>>>>> it may cause OOM and break the kernel, just like this: > >>>>>> > >>>>>> [ 79.441538] mlx4_en: eth5: Leaving promiscuous mode steering mode:2 > >>>>>> [ 100.067032] ksoftirqd/0: page allocation failure: order:0, mode:0x120 > >>>>>> [ 100.067038] CPU: 0 PID: 3 Comm: ksoftirqd/0 Tainted: G OE ----V------- 3.10.0-327.28.3.28.x86_64 #1 > >>>>>> [ 100.067039] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-20161018_184732-HGH1000003483 04/01/2014 > >>>>>> [ 100.067041] 0000000000000120 00000000b080d798 ffff8802afd5b968 ffffffff81638cb9 > >>>>>> [ 100.067045] ffff8802afd5b9f8 ffffffff81171380 0000000000000010 0000000000000000 > >>>>>> [ 100.067048] ffff8802befd8000 00000000ffffffff 0000000000000001 00000000b080d798 > >>>>>> [ 100.067050] Call Trace: > >>>>>> [ 100.067057] [<ffffffff81638cb9>] dump_stack+0x19/0x1b > >>>>>> [ 100.067062] [<ffffffff81171380>] warn_alloc_failed+0x110/0x180 > >>>>>> [ 100.067066] [<ffffffff81175b16>] __alloc_pages_nodemask+0x9b6/0xba0 > >>>>>> [ 100.067070] [<ffffffff8151e400>] ? skb_add_rx_frag+0x90/0xb0 > >>>>>> [ 100.067075] [<ffffffff811b6fba>] alloc_pages_current+0xaa/0x170 > >>>>>> [ 100.067080] [<ffffffffa06b9be0>] mlx4_alloc_pages.isra.24+0x40/0x170 [mlx4_en] > >>>>>> [ 100.067083] [<ffffffffa06b9dec>] mlx4_en_alloc_frags+0xdc/0x220 [mlx4_en] > >>>>>> [ 100.067086] [<ffffffff8152eeb8>] ? __netif_receive_skb+0x18/0x60 > >>>>>> [ 100.067088] [<ffffffff8152ef40>] ? netif_receive_skb+0x40/0xc0 > >>>>>> [ 100.067092] [<ffffffffa06bb521>] mlx4_en_process_rx_cq+0x5f1/0xec0 [mlx4_en] > >>>>>> [ 100.067095] [<ffffffff8131027d>] ? list_del+0xd/0x30 > >>>>>> [ 100.067098] [<ffffffff8152c90f>] ? __napi_complete+0x1f/0x30 > >>>>>> [ 100.067101] [<ffffffffa06bbeef>] mlx4_en_poll_rx_cq+0x9f/0x170 [mlx4_en] > >>>>>> [ 100.067103] [<ffffffff8152f372>] net_rx_action+0x152/0x240 > >>>>>> [ 100.067107] [<ffffffff81084d1f>] __do_softirq+0xef/0x280 > >>>>>> [ 100.067109] [<ffffffff81084ee0>] run_ksoftirqd+0x30/0x50 > >>>>>> [ 100.067114] [<ffffffff810ae93f>] smpboot_thread_fn+0xff/0x1a0 > >>>>>> [ 100.067117] [<ffffffff8163e269>] ? schedule+0x29/0x70 > >>>>>> [ 100.067120] [<ffffffff810ae840>] ? lg_double_unlock+0x90/0x90 > >>>>>> [ 100.067122] [<ffffffff810a5d4f>] kthread+0xcf/0xe0 > >>>>>> [ 100.067124] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 > >>>>>> [ 100.067127] [<ffffffff81649198>] ret_from_fork+0x58/0x90 > >>>>>> [ 100.067129] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 > >>>>>> > >>>>>> ================================cut here===================================== > >>>>>> > >>>>>> The reason is that the huge abnormal IP packet will be received to net stack > >>>>>> and be dropped finally by dst_release, and the dst_release would use the rcuos > >>>>>> callback-offload kthread to free the packet, but the cond_resched_rcu_qs() will > >>>>>> calling do_softirq() to receive more and more IP abnormal packets which will be > >>>>>> throw into the RCU callbacks again later, the number of received packet is much > >>>>>> greater than the number of packets freed, it will exhaust the memory and then OOM, > >>>>>> so don't try to process any pending softirqs in the rcuos callback-offload kthread > >>>>>> is a more effective solution. > >>>>> > >>>>> OK, but we could still have softirqs processed by the grace-period kthread > >>>>> as a result of any number of other events. So this change might reduce > >>>>> the probability of this problem, but it doesn't eliminate it. > >>>>> > >>>>> How huge are these huge IP packets? Is the underlying problem that they > >>>>> are too large to use the memory-allocator fastpaths? > >>>>> > >>>>> Thanx, Paul > >>>>> > >>>> > >>>> I use the 40G mellanox NiC to receive packet, and the testgine could send Mac abnormal packet and > >>>> IP abnormal packet to full speed. > >>>> > >>>> The Mac abnormal packet would be dropped at low level and not be received to net stack, > >>>> but the IP abnormal packet will introduce this problem, every packet will looks as new dst first and > >>>> release later by dst_release because it is meaningless. > >>>> > >>>> dst_release->call_rcu(&dst->rcu_head, dst_destroy_rcu); > >>>> > >>>> so all packet will be freed until the rcuos callback-offload kthread processing, it will be a infinite loop > >>>> if huge packet is coming because the do_softirq will load more and more packet to the rcuos processing kthread, > >>>> so I still could not find a better way to fix this, btw, it is really hard to say the driver use too large memory-allocater > >>>> fastpaths, there is no memory leak and the Ixgbe may meet the same problem too. > >> > >> And following up on my fastpath point -- from what I can see, one > >> big effect of the large invalid packets is that they push processing > >> off of a number of fastpaths. If these packets could be rejected with > >> less per-packet processing, I bet that things would work much better. > >> > >> Thanx, Paul > > > > Yes, and I found the WARN_ON_ONCE(!irqs_disabled()) will be triggered if use _local_bh_enable here, > > so I think we could ask some help from Eric and David how to reject the huge number packets. > > > > Thanks > > Ding > > > >> > >>> The overall effect of these two patches is to move from enabling bh > >>> (and processing recent softirqs) to enabling bh without processing > >>> recent softirqs. Is this really the correct way to solve this problem? > >>> What about this solution is avoiding re-introducing the original > >>> softlockups? Have you talked to the networking guys about this issue? > >>> > >>> Thanx, Paul > >>> > >>>> Thanks. > >>>> Ding > >>>> > >>>> > >>>>>> Fix commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") > >>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> > >>>>>> > >>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> > >>>>>> --- > >>>>>> kernel/rcu/tree_plugin.h | 3 +-- > >>>>>> 1 file changed, 1 insertion(+), 2 deletions(-) > >>>>>> > >>>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > >>>>>> index 85c5a88..760c3b5 100644 > >>>>>> --- a/kernel/rcu/tree_plugin.h > >>>>>> +++ b/kernel/rcu/tree_plugin.h > >>>>>> @@ -2172,8 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > >>>>>> if (__rcu_reclaim(rdp->rsp->name, list)) > >>>>>> cl++; > >>>>>> c++; > >>>>>> - local_bh_enable(); > >>>>>> - cond_resched_rcu_qs(); > >>>>>> + _local_bh_enable(); > >>>>>> list = next; > >>>>>> } > >>>>>> trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); > >>>>>> -- > >>>>>> 1.9.0 > >>>>>> > >>>>>> > >>>>>> > >>>>> > >>>>> > >>>>> . > >>>>> > >>>> > >> > >> > >> . > >> >
On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: > On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: > > Hi, Paul: > > > > I try to debug this problem and found this solution could work well for both problem scene. > > > > > > diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > > index 85c5a88..dbc14a7 100644 > > --- a/kernel/rcu/tree_plugin.h > > +++ b/kernel/rcu/tree_plugin.h > > @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > > if (__rcu_reclaim(rdp->rsp->name, list)) > > cl++; > > c++; > > - local_bh_enable(); > > + _local_bh_enable(); > > cond_resched_rcu_qs(); > > list = next; > > } > > > > > > The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use > > local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, > > what do you think about it? Please give me some suggestion. > > From what I can see, there is absolutely no guarantee that > cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee > that it will process any pending softirqs -- and that is not part of > its job in any case. So I cannot recommend the above patch. > > On efficient handling of large invalid packets (that is still the issue, > right?), I must defer to Dave and Eric. On the perhaps unlikely off-chance that there is a fix for this outside of networking, what symptoms are you seeing without this fix in place? Still RCU CPU stall warnings? Soft lockups? Something else? Thanx, Paul > > Thanks. > > Ding > > > > On 2016/11/21 9:28, Ding Tianhong wrote: > > > > > > > > > On 2016/11/21 8:13, Paul E. McKenney wrote: > > >> On Sat, Nov 19, 2016 at 12:22:09AM -0800, Paul E. McKenney wrote: > > >>> On Sat, Nov 19, 2016 at 03:50:32PM +0800, Ding Tianhong wrote: > > >>>> > > >>>> > > >>>> On 2016/11/18 21:01, Paul E. McKenney wrote: > > >>>>> On Fri, Nov 18, 2016 at 08:40:09PM +0800, Ding Tianhong wrote: > > >>>>>> The commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") > > >>>>>> will introduce a new problem that when huge IP abnormal packet arrived, > > >>>>>> it may cause OOM and break the kernel, just like this: > > >>>>>> > > >>>>>> [ 79.441538] mlx4_en: eth5: Leaving promiscuous mode steering mode:2 > > >>>>>> [ 100.067032] ksoftirqd/0: page allocation failure: order:0, mode:0x120 > > >>>>>> [ 100.067038] CPU: 0 PID: 3 Comm: ksoftirqd/0 Tainted: G OE ----V------- 3.10.0-327.28.3.28.x86_64 #1 > > >>>>>> [ 100.067039] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-20161018_184732-HGH1000003483 04/01/2014 > > >>>>>> [ 100.067041] 0000000000000120 00000000b080d798 ffff8802afd5b968 ffffffff81638cb9 > > >>>>>> [ 100.067045] ffff8802afd5b9f8 ffffffff81171380 0000000000000010 0000000000000000 > > >>>>>> [ 100.067048] ffff8802befd8000 00000000ffffffff 0000000000000001 00000000b080d798 > > >>>>>> [ 100.067050] Call Trace: > > >>>>>> [ 100.067057] [<ffffffff81638cb9>] dump_stack+0x19/0x1b > > >>>>>> [ 100.067062] [<ffffffff81171380>] warn_alloc_failed+0x110/0x180 > > >>>>>> [ 100.067066] [<ffffffff81175b16>] __alloc_pages_nodemask+0x9b6/0xba0 > > >>>>>> [ 100.067070] [<ffffffff8151e400>] ? skb_add_rx_frag+0x90/0xb0 > > >>>>>> [ 100.067075] [<ffffffff811b6fba>] alloc_pages_current+0xaa/0x170 > > >>>>>> [ 100.067080] [<ffffffffa06b9be0>] mlx4_alloc_pages.isra.24+0x40/0x170 [mlx4_en] > > >>>>>> [ 100.067083] [<ffffffffa06b9dec>] mlx4_en_alloc_frags+0xdc/0x220 [mlx4_en] > > >>>>>> [ 100.067086] [<ffffffff8152eeb8>] ? __netif_receive_skb+0x18/0x60 > > >>>>>> [ 100.067088] [<ffffffff8152ef40>] ? netif_receive_skb+0x40/0xc0 > > >>>>>> [ 100.067092] [<ffffffffa06bb521>] mlx4_en_process_rx_cq+0x5f1/0xec0 [mlx4_en] > > >>>>>> [ 100.067095] [<ffffffff8131027d>] ? list_del+0xd/0x30 > > >>>>>> [ 100.067098] [<ffffffff8152c90f>] ? __napi_complete+0x1f/0x30 > > >>>>>> [ 100.067101] [<ffffffffa06bbeef>] mlx4_en_poll_rx_cq+0x9f/0x170 [mlx4_en] > > >>>>>> [ 100.067103] [<ffffffff8152f372>] net_rx_action+0x152/0x240 > > >>>>>> [ 100.067107] [<ffffffff81084d1f>] __do_softirq+0xef/0x280 > > >>>>>> [ 100.067109] [<ffffffff81084ee0>] run_ksoftirqd+0x30/0x50 > > >>>>>> [ 100.067114] [<ffffffff810ae93f>] smpboot_thread_fn+0xff/0x1a0 > > >>>>>> [ 100.067117] [<ffffffff8163e269>] ? schedule+0x29/0x70 > > >>>>>> [ 100.067120] [<ffffffff810ae840>] ? lg_double_unlock+0x90/0x90 > > >>>>>> [ 100.067122] [<ffffffff810a5d4f>] kthread+0xcf/0xe0 > > >>>>>> [ 100.067124] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 > > >>>>>> [ 100.067127] [<ffffffff81649198>] ret_from_fork+0x58/0x90 > > >>>>>> [ 100.067129] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 > > >>>>>> > > >>>>>> ================================cut here===================================== > > >>>>>> > > >>>>>> The reason is that the huge abnormal IP packet will be received to net stack > > >>>>>> and be dropped finally by dst_release, and the dst_release would use the rcuos > > >>>>>> callback-offload kthread to free the packet, but the cond_resched_rcu_qs() will > > >>>>>> calling do_softirq() to receive more and more IP abnormal packets which will be > > >>>>>> throw into the RCU callbacks again later, the number of received packet is much > > >>>>>> greater than the number of packets freed, it will exhaust the memory and then OOM, > > >>>>>> so don't try to process any pending softirqs in the rcuos callback-offload kthread > > >>>>>> is a more effective solution. > > >>>>> > > >>>>> OK, but we could still have softirqs processed by the grace-period kthread > > >>>>> as a result of any number of other events. So this change might reduce > > >>>>> the probability of this problem, but it doesn't eliminate it. > > >>>>> > > >>>>> How huge are these huge IP packets? Is the underlying problem that they > > >>>>> are too large to use the memory-allocator fastpaths? > > >>>>> > > >>>>> Thanx, Paul > > >>>>> > > >>>> > > >>>> I use the 40G mellanox NiC to receive packet, and the testgine could send Mac abnormal packet and > > >>>> IP abnormal packet to full speed. > > >>>> > > >>>> The Mac abnormal packet would be dropped at low level and not be received to net stack, > > >>>> but the IP abnormal packet will introduce this problem, every packet will looks as new dst first and > > >>>> release later by dst_release because it is meaningless. > > >>>> > > >>>> dst_release->call_rcu(&dst->rcu_head, dst_destroy_rcu); > > >>>> > > >>>> so all packet will be freed until the rcuos callback-offload kthread processing, it will be a infinite loop > > >>>> if huge packet is coming because the do_softirq will load more and more packet to the rcuos processing kthread, > > >>>> so I still could not find a better way to fix this, btw, it is really hard to say the driver use too large memory-allocater > > >>>> fastpaths, there is no memory leak and the Ixgbe may meet the same problem too. > > >> > > >> And following up on my fastpath point -- from what I can see, one > > >> big effect of the large invalid packets is that they push processing > > >> off of a number of fastpaths. If these packets could be rejected with > > >> less per-packet processing, I bet that things would work much better. > > >> > > >> Thanx, Paul > > > > > > Yes, and I found the WARN_ON_ONCE(!irqs_disabled()) will be triggered if use _local_bh_enable here, > > > so I think we could ask some help from Eric and David how to reject the huge number packets. > > > > > > Thanks > > > Ding > > > > > >> > > >>> The overall effect of these two patches is to move from enabling bh > > >>> (and processing recent softirqs) to enabling bh without processing > > >>> recent softirqs. Is this really the correct way to solve this problem? > > >>> What about this solution is avoiding re-introducing the original > > >>> softlockups? Have you talked to the networking guys about this issue? > > >>> > > >>> Thanx, Paul > > >>> > > >>>> Thanks. > > >>>> Ding > > >>>> > > >>>> > > >>>>>> Fix commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") > > >>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> > > >>>>>> > > >>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> > > >>>>>> --- > > >>>>>> kernel/rcu/tree_plugin.h | 3 +-- > > >>>>>> 1 file changed, 1 insertion(+), 2 deletions(-) > > >>>>>> > > >>>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > > >>>>>> index 85c5a88..760c3b5 100644 > > >>>>>> --- a/kernel/rcu/tree_plugin.h > > >>>>>> +++ b/kernel/rcu/tree_plugin.h > > >>>>>> @@ -2172,8 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > > >>>>>> if (__rcu_reclaim(rdp->rsp->name, list)) > > >>>>>> cl++; > > >>>>>> c++; > > >>>>>> - local_bh_enable(); > > >>>>>> - cond_resched_rcu_qs(); > > >>>>>> + _local_bh_enable(); > > >>>>>> list = next; > > >>>>>> } > > >>>>>> trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); > > >>>>>> -- > > >>>>>> 1.9.0 > > >>>>>> > > >>>>>> > > >>>>>> > > >>>>> > > >>>>> > > >>>>> . > > >>>>> > > >>>> > > >> > > >> > > >> . > > >> > >
On 2017/1/4 8:57, Paul E. McKenney wrote: > On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: >> On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: >>> Hi, Paul: >>> >>> I try to debug this problem and found this solution could work well for both problem scene. >>> >>> >>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h >>> index 85c5a88..dbc14a7 100644 >>> --- a/kernel/rcu/tree_plugin.h >>> +++ b/kernel/rcu/tree_plugin.h >>> @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) >>> if (__rcu_reclaim(rdp->rsp->name, list)) >>> cl++; >>> c++; >>> - local_bh_enable(); >>> + _local_bh_enable(); >>> cond_resched_rcu_qs(); >>> list = next; >>> } >>> >>> >>> The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use >>> local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, >>> what do you think about it? Please give me some suggestion. >> >> From what I can see, there is absolutely no guarantee that >> cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee >> that it will process any pending softirqs -- and that is not part of >> its job in any case. So I cannot recommend the above patch. >> >> On efficient handling of large invalid packets (that is still the issue, >> right?), I must defer to Dave and Eric. > > On the perhaps unlikely off-chance that there is a fix for this outside > of networking, what symptoms are you seeing without this fix in place? > Still RCU CPU stall warnings? Soft lockups? Something else? > > Thanx, Paul > Hi Paul: I was still try to test and fix this by another way, but could explain more about this problem. when the huge packets coming, the packets was abnormal and will be freed by dst_release->call_rcu(dst_destroy_rcu), so the rcuos kthread will handle the dst_destroy_rcu to free them, but when the rcuos was looping ,I fould the local_bh_enable() will call do_softirq to receive a certain number of packets which is abnormal and need to be free, but more packets is coming so when cond_resched_rcu_qs run, it will do the ksoftirqd and do softirq again, so rcuos kthread need free more, it looks more and more worse and lead to OOM because many more packets need to be freed. So I think the do_softirq in the local_bh_enable is not need here, the cond_resched_rcu_qs() will handle the do_softirq once, it is enough. and recently I found that the Eric has upstream a new patch named (softirq: Let ksoftirqd do its job) may fix this, and still test it, not get any results yet. Thanks Ding >>> Thanks. >>> Ding >>> >>> On 2016/11/21 9:28, Ding Tianhong wrote: >>>> >>>> >>>> On 2016/11/21 8:13, Paul E. McKenney wrote: >>>>> On Sat, Nov 19, 2016 at 12:22:09AM -0800, Paul E. McKenney wrote: >>>>>> On Sat, Nov 19, 2016 at 03:50:32PM +0800, Ding Tianhong wrote: >>>>>>> >>>>>>> >>>>>>> On 2016/11/18 21:01, Paul E. McKenney wrote: >>>>>>>> On Fri, Nov 18, 2016 at 08:40:09PM +0800, Ding Tianhong wrote: >>>>>>>>> The commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") >>>>>>>>> will introduce a new problem that when huge IP abnormal packet arrived, >>>>>>>>> it may cause OOM and break the kernel, just like this: >>>>>>>>> >>>>>>>>> [ 79.441538] mlx4_en: eth5: Leaving promiscuous mode steering mode:2 >>>>>>>>> [ 100.067032] ksoftirqd/0: page allocation failure: order:0, mode:0x120 >>>>>>>>> [ 100.067038] CPU: 0 PID: 3 Comm: ksoftirqd/0 Tainted: G OE ----V------- 3.10.0-327.28.3.28.x86_64 #1 >>>>>>>>> [ 100.067039] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.9.1-0-gb3ef39f-20161018_184732-HGH1000003483 04/01/2014 >>>>>>>>> [ 100.067041] 0000000000000120 00000000b080d798 ffff8802afd5b968 ffffffff81638cb9 >>>>>>>>> [ 100.067045] ffff8802afd5b9f8 ffffffff81171380 0000000000000010 0000000000000000 >>>>>>>>> [ 100.067048] ffff8802befd8000 00000000ffffffff 0000000000000001 00000000b080d798 >>>>>>>>> [ 100.067050] Call Trace: >>>>>>>>> [ 100.067057] [<ffffffff81638cb9>] dump_stack+0x19/0x1b >>>>>>>>> [ 100.067062] [<ffffffff81171380>] warn_alloc_failed+0x110/0x180 >>>>>>>>> [ 100.067066] [<ffffffff81175b16>] __alloc_pages_nodemask+0x9b6/0xba0 >>>>>>>>> [ 100.067070] [<ffffffff8151e400>] ? skb_add_rx_frag+0x90/0xb0 >>>>>>>>> [ 100.067075] [<ffffffff811b6fba>] alloc_pages_current+0xaa/0x170 >>>>>>>>> [ 100.067080] [<ffffffffa06b9be0>] mlx4_alloc_pages.isra.24+0x40/0x170 [mlx4_en] >>>>>>>>> [ 100.067083] [<ffffffffa06b9dec>] mlx4_en_alloc_frags+0xdc/0x220 [mlx4_en] >>>>>>>>> [ 100.067086] [<ffffffff8152eeb8>] ? __netif_receive_skb+0x18/0x60 >>>>>>>>> [ 100.067088] [<ffffffff8152ef40>] ? netif_receive_skb+0x40/0xc0 >>>>>>>>> [ 100.067092] [<ffffffffa06bb521>] mlx4_en_process_rx_cq+0x5f1/0xec0 [mlx4_en] >>>>>>>>> [ 100.067095] [<ffffffff8131027d>] ? list_del+0xd/0x30 >>>>>>>>> [ 100.067098] [<ffffffff8152c90f>] ? __napi_complete+0x1f/0x30 >>>>>>>>> [ 100.067101] [<ffffffffa06bbeef>] mlx4_en_poll_rx_cq+0x9f/0x170 [mlx4_en] >>>>>>>>> [ 100.067103] [<ffffffff8152f372>] net_rx_action+0x152/0x240 >>>>>>>>> [ 100.067107] [<ffffffff81084d1f>] __do_softirq+0xef/0x280 >>>>>>>>> [ 100.067109] [<ffffffff81084ee0>] run_ksoftirqd+0x30/0x50 >>>>>>>>> [ 100.067114] [<ffffffff810ae93f>] smpboot_thread_fn+0xff/0x1a0 >>>>>>>>> [ 100.067117] [<ffffffff8163e269>] ? schedule+0x29/0x70 >>>>>>>>> [ 100.067120] [<ffffffff810ae840>] ? lg_double_unlock+0x90/0x90 >>>>>>>>> [ 100.067122] [<ffffffff810a5d4f>] kthread+0xcf/0xe0 >>>>>>>>> [ 100.067124] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 >>>>>>>>> [ 100.067127] [<ffffffff81649198>] ret_from_fork+0x58/0x90 >>>>>>>>> [ 100.067129] [<ffffffff810a5c80>] ? kthread_create_on_node+0x140/0x140 >>>>>>>>> >>>>>>>>> ================================cut here===================================== >>>>>>>>> >>>>>>>>> The reason is that the huge abnormal IP packet will be received to net stack >>>>>>>>> and be dropped finally by dst_release, and the dst_release would use the rcuos >>>>>>>>> callback-offload kthread to free the packet, but the cond_resched_rcu_qs() will >>>>>>>>> calling do_softirq() to receive more and more IP abnormal packets which will be >>>>>>>>> throw into the RCU callbacks again later, the number of received packet is much >>>>>>>>> greater than the number of packets freed, it will exhaust the memory and then OOM, >>>>>>>>> so don't try to process any pending softirqs in the rcuos callback-offload kthread >>>>>>>>> is a more effective solution. >>>>>>>> >>>>>>>> OK, but we could still have softirqs processed by the grace-period kthread >>>>>>>> as a result of any number of other events. So this change might reduce >>>>>>>> the probability of this problem, but it doesn't eliminate it. >>>>>>>> >>>>>>>> How huge are these huge IP packets? Is the underlying problem that they >>>>>>>> are too large to use the memory-allocator fastpaths? >>>>>>>> >>>>>>>> Thanx, Paul >>>>>>>> >>>>>>> >>>>>>> I use the 40G mellanox NiC to receive packet, and the testgine could send Mac abnormal packet and >>>>>>> IP abnormal packet to full speed. >>>>>>> >>>>>>> The Mac abnormal packet would be dropped at low level and not be received to net stack, >>>>>>> but the IP abnormal packet will introduce this problem, every packet will looks as new dst first and >>>>>>> release later by dst_release because it is meaningless. >>>>>>> >>>>>>> dst_release->call_rcu(&dst->rcu_head, dst_destroy_rcu); >>>>>>> >>>>>>> so all packet will be freed until the rcuos callback-offload kthread processing, it will be a infinite loop >>>>>>> if huge packet is coming because the do_softirq will load more and more packet to the rcuos processing kthread, >>>>>>> so I still could not find a better way to fix this, btw, it is really hard to say the driver use too large memory-allocater >>>>>>> fastpaths, there is no memory leak and the Ixgbe may meet the same problem too. >>>>> >>>>> And following up on my fastpath point -- from what I can see, one >>>>> big effect of the large invalid packets is that they push processing >>>>> off of a number of fastpaths. If these packets could be rejected with >>>>> less per-packet processing, I bet that things would work much better. >>>>> >>>>> Thanx, Paul >>>> >>>> Yes, and I found the WARN_ON_ONCE(!irqs_disabled()) will be triggered if use _local_bh_enable here, >>>> so I think we could ask some help from Eric and David how to reject the huge number packets. >>>> >>>> Thanks >>>> Ding >>>> >>>>> >>>>>> The overall effect of these two patches is to move from enabling bh >>>>>> (and processing recent softirqs) to enabling bh without processing >>>>>> recent softirqs. Is this really the correct way to solve this problem? >>>>>> What about this solution is avoiding re-introducing the original >>>>>> softlockups? Have you talked to the networking guys about this issue? >>>>>> >>>>>> Thanx, Paul >>>>>> >>>>>>> Thanks. >>>>>>> Ding >>>>>>> >>>>>>> >>>>>>>>> Fix commit bedc196915 ("rcu: Fix soft lockup for rcu_nocb_kthread") >>>>>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> >>>>>>>>> >>>>>>>>> Signed-off-by: Ding Tianhong <dingtianhong@huawei.com> >>>>>>>>> --- >>>>>>>>> kernel/rcu/tree_plugin.h | 3 +-- >>>>>>>>> 1 file changed, 1 insertion(+), 2 deletions(-) >>>>>>>>> >>>>>>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h >>>>>>>>> index 85c5a88..760c3b5 100644 >>>>>>>>> --- a/kernel/rcu/tree_plugin.h >>>>>>>>> +++ b/kernel/rcu/tree_plugin.h >>>>>>>>> @@ -2172,8 +2172,7 @@ static int rcu_nocb_kthread(void *arg) >>>>>>>>> if (__rcu_reclaim(rdp->rsp->name, list)) >>>>>>>>> cl++; >>>>>>>>> c++; >>>>>>>>> - local_bh_enable(); >>>>>>>>> - cond_resched_rcu_qs(); >>>>>>>>> + _local_bh_enable(); >>>>>>>>> list = next; >>>>>>>>> } >>>>>>>>> trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); >>>>>>>>> -- >>>>>>>>> 1.9.0 >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> . >>>>>>>> >>>>>>> >>>>> >>>>> >>>>> . >>>>> >>> > > > . >
On Wed, Jan 04, 2017 at 03:02:30PM +0800, Ding Tianhong wrote: > > > On 2017/1/4 8:57, Paul E. McKenney wrote: > > On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: > >> On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: > >>> Hi, Paul: > >>> > >>> I try to debug this problem and found this solution could work well for both problem scene. > >>> > >>> > >>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > >>> index 85c5a88..dbc14a7 100644 > >>> --- a/kernel/rcu/tree_plugin.h > >>> +++ b/kernel/rcu/tree_plugin.h > >>> @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > >>> if (__rcu_reclaim(rdp->rsp->name, list)) > >>> cl++; > >>> c++; > >>> - local_bh_enable(); > >>> + _local_bh_enable(); > >>> cond_resched_rcu_qs(); > >>> list = next; > >>> } > >>> > >>> > >>> The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use > >>> local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, > >>> what do you think about it? Please give me some suggestion. > >> > >> From what I can see, there is absolutely no guarantee that > >> cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee > >> that it will process any pending softirqs -- and that is not part of > >> its job in any case. So I cannot recommend the above patch. > >> > >> On efficient handling of large invalid packets (that is still the issue, > >> right?), I must defer to Dave and Eric. > > > > On the perhaps unlikely off-chance that there is a fix for this outside > > of networking, what symptoms are you seeing without this fix in place? > > Still RCU CPU stall warnings? Soft lockups? Something else? > > > > Thanx, Paul > > > > Hi Paul: > > I was still try to test and fix this by another way, but could explain more about this problem. > > when the huge packets coming, the packets was abnormal and will be freed by dst_release->call_rcu(dst_destroy_rcu), > so the rcuos kthread will handle the dst_destroy_rcu to free them, but when the rcuos was looping ,I fould the local_bh_enable() will > call do_softirq to receive a certain number of packets which is abnormal and need to be free, but more packets is coming so when cond_resched_rcu_qs run, > it will do the ksoftirqd and do softirq again, so rcuos kthread need free more, it looks more and more worse and lead to OOM because many more packets need to > be freed. > So I think the do_softirq in the local_bh_enable is not need here, the cond_resched_rcu_qs() will handle the do_softirq once, it is enough. > > and recently I found that the Eric has upstream a new patch named (softirq: Let ksoftirqd do its job) may fix this, and still test it, not get any results yet. OK, I don't see any reasonable way that the RCU callback-offload tasks (rcuos) can figure out whether or not they should let softirqs happen -- unconditionally suppressing them might help your workload, but would break workloads needing low networking latency, of which there are many. So please let me know now things go with Eric's patch. Thanx, Paul
On 2017/1/4 21:48, Paul E. McKenney wrote: > On Wed, Jan 04, 2017 at 03:02:30PM +0800, Ding Tianhong wrote: >> >> >> On 2017/1/4 8:57, Paul E. McKenney wrote: >>> On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: >>>> On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: >>>>> Hi, Paul: >>>>> >>>>> I try to debug this problem and found this solution could work well for both problem scene. >>>>> >>>>> >>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h >>>>> index 85c5a88..dbc14a7 100644 >>>>> --- a/kernel/rcu/tree_plugin.h >>>>> +++ b/kernel/rcu/tree_plugin.h >>>>> @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) >>>>> if (__rcu_reclaim(rdp->rsp->name, list)) >>>>> cl++; >>>>> c++; >>>>> - local_bh_enable(); >>>>> + _local_bh_enable(); >>>>> cond_resched_rcu_qs(); >>>>> list = next; >>>>> } >>>>> >>>>> >>>>> The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use >>>>> local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, >>>>> what do you think about it? Please give me some suggestion. >>>> >>>> From what I can see, there is absolutely no guarantee that >>>> cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee >>>> that it will process any pending softirqs -- and that is not part of >>>> its job in any case. So I cannot recommend the above patch. >>>> >>>> On efficient handling of large invalid packets (that is still the issue, >>>> right?), I must defer to Dave and Eric. >>> >>> On the perhaps unlikely off-chance that there is a fix for this outside >>> of networking, what symptoms are you seeing without this fix in place? >>> Still RCU CPU stall warnings? Soft lockups? Something else? >>> >>> Thanx, Paul >>> >> >> Hi Paul: >> >> I was still try to test and fix this by another way, but could explain more about this problem. >> >> when the huge packets coming, the packets was abnormal and will be freed by dst_release->call_rcu(dst_destroy_rcu), >> so the rcuos kthread will handle the dst_destroy_rcu to free them, but when the rcuos was looping ,I fould the local_bh_enable() will >> call do_softirq to receive a certain number of packets which is abnormal and need to be free, but more packets is coming so when cond_resched_rcu_qs run, >> it will do the ksoftirqd and do softirq again, so rcuos kthread need free more, it looks more and more worse and lead to OOM because many more packets need to >> be freed. >> So I think the do_softirq in the local_bh_enable is not need here, the cond_resched_rcu_qs() will handle the do_softirq once, it is enough. >> >> and recently I found that the Eric has upstream a new patch named (softirq: Let ksoftirqd do its job) may fix this, and still test it, not get any results yet. > > OK, I don't see any reasonable way that the RCU callback-offload tasks > (rcuos) can figure out whether or not they should let softirqs happen -- > unconditionally suppressing them might help your workload, but would > break workloads needing low networking latency, of which there are many. > > So please let me know now things go with Eric's patch. > Hi Paul: Good news, the Eric's patch could fix this problem, it means that if the softirqd kthread is running, we should not take too much time in the softirq process, this behavior equivalent that we remove the do_softirq in the local_bh_enable(), but this solution looks more perfect, we need to inform the lts kernel maintainer to applied this patch which is not looks like a bugfix. Thanks Ding > Thanx, Paul > > > . >
On Tue, Jan 10, 2017 at 11:20:40AM +0800, Ding Tianhong wrote: > > > On 2017/1/4 21:48, Paul E. McKenney wrote: > > On Wed, Jan 04, 2017 at 03:02:30PM +0800, Ding Tianhong wrote: > >> > >> > >> On 2017/1/4 8:57, Paul E. McKenney wrote: > >>> On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: > >>>> On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: > >>>>> Hi, Paul: > >>>>> > >>>>> I try to debug this problem and found this solution could work well for both problem scene. > >>>>> > >>>>> > >>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h > >>>>> index 85c5a88..dbc14a7 100644 > >>>>> --- a/kernel/rcu/tree_plugin.h > >>>>> +++ b/kernel/rcu/tree_plugin.h > >>>>> @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) > >>>>> if (__rcu_reclaim(rdp->rsp->name, list)) > >>>>> cl++; > >>>>> c++; > >>>>> - local_bh_enable(); > >>>>> + _local_bh_enable(); > >>>>> cond_resched_rcu_qs(); > >>>>> list = next; > >>>>> } > >>>>> > >>>>> > >>>>> The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use > >>>>> local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, > >>>>> what do you think about it? Please give me some suggestion. > >>>> > >>>> From what I can see, there is absolutely no guarantee that > >>>> cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee > >>>> that it will process any pending softirqs -- and that is not part of > >>>> its job in any case. So I cannot recommend the above patch. > >>>> > >>>> On efficient handling of large invalid packets (that is still the issue, > >>>> right?), I must defer to Dave and Eric. > >>> > >>> On the perhaps unlikely off-chance that there is a fix for this outside > >>> of networking, what symptoms are you seeing without this fix in place? > >>> Still RCU CPU stall warnings? Soft lockups? Something else? > >>> > >>> Thanx, Paul > >>> > >> > >> Hi Paul: > >> > >> I was still try to test and fix this by another way, but could explain more about this problem. > >> > >> when the huge packets coming, the packets was abnormal and will be freed by dst_release->call_rcu(dst_destroy_rcu), > >> so the rcuos kthread will handle the dst_destroy_rcu to free them, but when the rcuos was looping ,I fould the local_bh_enable() will > >> call do_softirq to receive a certain number of packets which is abnormal and need to be free, but more packets is coming so when cond_resched_rcu_qs run, > >> it will do the ksoftirqd and do softirq again, so rcuos kthread need free more, it looks more and more worse and lead to OOM because many more packets need to > >> be freed. > >> So I think the do_softirq in the local_bh_enable is not need here, the cond_resched_rcu_qs() will handle the do_softirq once, it is enough. > >> > >> and recently I found that the Eric has upstream a new patch named (softirq: Let ksoftirqd do its job) may fix this, and still test it, not get any results yet. > > > > OK, I don't see any reasonable way that the RCU callback-offload tasks > > (rcuos) can figure out whether or not they should let softirqs happen -- > > unconditionally suppressing them might help your workload, but would > > break workloads needing low networking latency, of which there are many. > > > > So please let me know now things go with Eric's patch. > > > Hi Paul: > > Good news, the Eric's patch could fix this problem, it means that if the softirqd kthread is running, we should not take too much > time in the softirq process, this behavior equivalent that we remove the do_softirq in the local_bh_enable(), but this solution looks more > perfect, we need to inform the lts kernel maintainer to applied this patch which is not looks like a bugfix. Here is hoping! ;-) Thanx, Paul
Hi David： The Patch "rcu: Fix soft lockup for rcu_nocb_kthread" has been added to several stable tree， it may introduced an issue in certain special scenarios, The Patch "softirq: Let ksoftirqd do its job" could fix this issue, so I hope you could add this patch to stable tree, thanks. Ding On 2017/1/10 13:51, Paul E. McKenney wrote: > On Tue, Jan 10, 2017 at 11:20:40AM +0800, Ding Tianhong wrote: >> >> >> On 2017/1/4 21:48, Paul E. McKenney wrote: >>> On Wed, Jan 04, 2017 at 03:02:30PM +0800, Ding Tianhong wrote: >>>> >>>> >>>> On 2017/1/4 8:57, Paul E. McKenney wrote: >>>>> On Wed, Dec 28, 2016 at 04:13:15PM -0800, Paul E. McKenney wrote: >>>>>> On Wed, Dec 28, 2016 at 01:58:06PM +0800, Ding Tianhong wrote: >>>>>>> Hi, Paul: >>>>>>> >>>>>>> I try to debug this problem and found this solution could work well for both problem scene. >>>>>>> >>>>>>> >>>>>>> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h >>>>>>> index 85c5a88..dbc14a7 100644 >>>>>>> --- a/kernel/rcu/tree_plugin.h >>>>>>> +++ b/kernel/rcu/tree_plugin.h >>>>>>> @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) >>>>>>> if (__rcu_reclaim(rdp->rsp->name, list)) >>>>>>> cl++; >>>>>>> c++; >>>>>>> - local_bh_enable(); >>>>>>> + _local_bh_enable(); >>>>>>> cond_resched_rcu_qs(); >>>>>>> list = next; >>>>>>> } >>>>>>> >>>>>>> >>>>>>> The cond_resched_rcu_qs() would process the softirq if the softirq is pending, so no need to use >>>>>>> local_bh_enable() to process the softirq twice here, and it will avoid OOM when huge packets arrives, >>>>>>> what do you think about it? Please give me some suggestion. >>>>>> >>>>>> From what I can see, there is absolutely no guarantee that >>>>>> cond_resched_rcu_qs() will do local_bh_enable(), and thus no guarantee >>>>>> that it will process any pending softirqs -- and that is not part of >>>>>> its job in any case. So I cannot recommend the above patch. >>>>>> >>>>>> On efficient handling of large invalid packets (that is still the issue, >>>>>> right?), I must defer to Dave and Eric. >>>>> >>>>> On the perhaps unlikely off-chance that there is a fix for this outside >>>>> of networking, what symptoms are you seeing without this fix in place? >>>>> Still RCU CPU stall warnings? Soft lockups? Something else? >>>>> >>>>> Thanx, Paul >>>>> >>>> >>>> Hi Paul: >>>> >>>> I was still try to test and fix this by another way, but could explain more about this problem. >>>> >>>> when the huge packets coming, the packets was abnormal and will be freed by dst_release->call_rcu(dst_destroy_rcu), >>>> so the rcuos kthread will handle the dst_destroy_rcu to free them, but when the rcuos was looping ,I fould the local_bh_enable() will >>>> call do_softirq to receive a certain number of packets which is abnormal and need to be free, but more packets is coming so when cond_resched_rcu_qs run, >>>> it will do the ksoftirqd and do softirq again, so rcuos kthread need free more, it looks more and more worse and lead to OOM because many more packets need to >>>> be freed. >>>> So I think the do_softirq in the local_bh_enable is not need here, the cond_resched_rcu_qs() will handle the do_softirq once, it is enough. >>>> >>>> and recently I found that the Eric has upstream a new patch named (softirq: Let ksoftirqd do its job) may fix this, and still test it, not get any results yet. >>> >>> OK, I don't see any reasonable way that the RCU callback-offload tasks >>> (rcuos) can figure out whether or not they should let softirqs happen -- >>> unconditionally suppressing them might help your workload, but would >>> break workloads needing low networking latency, of which there are many. >>> >>> So please let me know now things go with Eric's patch. >>> >> Hi Paul: >> >> Good news, the Eric's patch could fix this problem, it means that if the softirqd kthread is running, we should not take too much >> time in the softirq process, this behavior equivalent that we remove the do_softirq in the local_bh_enable(), but this solution looks more >> perfect, we need to inform the lts kernel maintainer to applied this patch which is not looks like a bugfix. > > Here is hoping! ;-) > > Thanx, Paul > > > . >
diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h index 85c5a88..dbc14a7 100644 --- a/kernel/rcu/tree_plugin.h +++ b/kernel/rcu/tree_plugin.h @@ -2172,7 +2172,7 @@ static int rcu_nocb_kthread(void *arg) if (__rcu_reclaim(rdp->rsp->name, list)) cl++; c++; - local_bh_enable(); + _local_bh_enable(); cond_resched_rcu_qs(); list = next; }