Nikolay Aleksandrov
2022-Jul-06 19:38 UTC
[Bridge] [PATCH V3 net-next 1/4] net: bridge: add fdb flag to extent locked port feature
On 06/07/2022 21:13, Vladimir Oltean wrote:> Hi Nikolay, > > On Wed, May 25, 2022 at 01:18:49PM +0300, Nikolay Aleksandrov wrote: >>>>>>>> Hi Hans, >>>>>>>> So this approach has a fundamental problem, f->dst is changed without any synchronization >>>>>>>> you cannot rely on it and thus you cannot account for these entries properly. We must be very >>>>>>>> careful if we try to add any new synchronization not to affect performance as well. >>>>>>>> More below... >>>>>>>> >>>>>>>>> @@ -319,6 +326,9 @@ static void fdb_delete(struct net_bridge *br, struct net_bridge_fdb_entry *f, >>>>>>>>> if (test_bit(BR_FDB_STATIC, &f->flags)) >>>>>>>>> fdb_del_hw_addr(br, f->key.addr.addr); >>>>>>>>> >>>>>>>>> + if (test_bit(BR_FDB_ENTRY_LOCKED, &f->flags) && !test_bit(BR_FDB_OFFLOADED, &f->flags)) >>>>>>>>> + atomic_dec(&f->dst->locked_entry_cnt); >>>>>>>> >>>>>>>> Sorry but you cannot do this for multiple reasons: >>>>>>>> - f->dst can be NULL >>>>>>>> - f->dst changes without any synchronization >>>>>>>> - there is no synchronization between fdb's flags and its ->dst >>>>>>>> >>>>>>>> Cheers, >>>>>>>> Nik >>>>>>> >>>>>>> Hi Nik, >>>>>>> >>>>>>> if a port is decoupled from the bridge, the locked entries would of >>>>>>> course be invalid, so maybe if adding and removing a port is accounted >>>>>>> for wrt locked entries and the count of locked entries, would that not >>>>>>> work? >>>>>>> >>>>>>> Best, >>>>>>> Hans >>>>>> >>>>>> Hi Hans, >>>>>> Unfortunately you need the correct amount of locked entries per-port if you want >>>>>> to limit their number per-port, instead of globally. So you need a >>>>>> consistent >>>>> >>>>> Hi Nik, >>>>> the used dst is a port structure, so it is per-port and not globally. >>>>> >>>>> Best, >>>>> Hans >>>>> >>>> >>>> Yeah, I know. :) That's why I wrote it, if the limit is not a feature requirement I'd suggest >>>> dropping it altogether, it can be enforced externally (e.g. from user-space) if needed. >>>> >>>> By the way just fyi net-next is closed right now due to merge window. And one more >>>> thing please include a short log of changes between versions when you send a new one. >>>> I had to go look for v2 to find out what changed. >>>> >>> >>> Okay, I will drop the limit in the bridge module, which is an easy thing >>> to do. :) (It is mostly there to ensure against DOS attacks if someone >>> bombards a locked port with random mac addresses.) >>> I have a similar limitation in the driver, which should then probably be >>> dropped too? >>> >> >> That is up to you/driver, I'd try looking for similar problems in other switch drivers >> and check how those were handled. There are people in the CC above that can >> directly answer that. :) > > Not sure whom you're referring to?I meant people who have dealt with hardware resource management in the drivers.> > In fact I was pretty sure that I didn't see any OOM protection in the > source code of the Linux bridge driver itself either, so I wanted to > check that for myself, so I wrote a small "killswitch" program that's > supposed to, well, kill a switch. It took me a while to find a few free > hours to do the test, sorry for that. > > https://github.com/vladimiroltean/killswitch/blob/master/src/killswitch.c > > Sure enough, I can kill a Marvell Armada 3720 device with 1GB of RAM > within 3 minutes of running the test program. >I don't think that is new or surprising, if there isn't anything to control the device resources you'll get there. You don't really need to write any new programs you can easily do it with mausezahn. I have tests that add over 10 million fdbs on devices for a few seconds. The point is it's not the bridge's task to limit memory consumption or to watch for resource management. You can limit new entries from the device driver (in case of swdev learning) or you can use a daemon to watch the number of entries and disable learning. There are many different ways to avoid this. We've discussed it before and I don't mind adding a hard fdb per-port limit in the bridge as long as it's done properly. We've also discussed LRU and similar algorithms for fdb learning and eviction. But any hardcoded limits or limits that can break current default use cases are unacceptable, they must be opt-in.> [ 273.864203] ksoftirqd/0: page allocation failure: order:0, mode:0x40a20(GFP_ATOMIC|__GFP_COMP), nodemask=(null),cpuset=/,mems_allowed=0 > [ 273.876426] CPU: 0 PID: 12 Comm: ksoftirqd/0 Not tainted 5.18.7-rc1-00013-g52b92343db13 #74 > [ 273.884775] Hardware name: CZ.NIC Turris Mox Board (DT) > [ 273.889994] Call trace: > [ 273.892437] dump_backtrace.part.0+0xc8/0xd4 > [ 273.896721] show_stack+0x18/0x70 > [ 273.900039] dump_stack_lvl+0x68/0x84 > [ 273.903703] dump_stack+0x18/0x34 > [ 273.907017] warn_alloc+0x114/0x1a0 > [ 273.910508] __alloc_pages+0xbb0/0xbe0 > [ 273.914257] cache_grow_begin+0x60/0x300 > [ 273.918183] fallback_alloc+0x184/0x220 > [ 273.922017] ____cache_alloc_node+0x174/0x190 > [ 273.926373] kmem_cache_alloc+0x1a4/0x220 > [ 273.930381] fdb_create+0x40/0x430 > [ 273.933784] br_fdb_update+0x198/0x210 > [ 273.937532] br_handle_frame_finish+0x244/0x530 > [ 273.942063] br_handle_frame+0x1c0/0x270 > [ 273.945986] __netif_receive_skb_core.constprop.0+0x29c/0xd30 > [ 273.951734] __netif_receive_skb_list_core+0xe8/0x210 > [ 273.956784] netif_receive_skb_list_internal+0x180/0x29c > [ 273.962091] napi_gro_receive+0x174/0x190 > [ 273.966099] mvneta_rx_swbm+0x6b8/0xb40 > [ 273.969935] mvneta_poll+0x684/0x900 > [ 273.973506] __napi_poll+0x38/0x18c > [ 273.976988] net_rx_action+0xe8/0x280 > [ 273.980643] __do_softirq+0x124/0x2a0 > [ 273.984299] run_ksoftirqd+0x4c/0x60 > [ 273.987871] smpboot_thread_fn+0x23c/0x270 > [ 273.991963] kthread+0x10c/0x110 > [ 273.995188] ret_from_fork+0x10/0x20 > > (followed by lots upon lots of vomiting, followed by ...) > > [ 311.138590] Out of memory and no killable processes... > [ 311.143774] Kernel panic - not syncing: System is deadlocked on memory > [ 311.150295] CPU: 0 PID: 6 Comm: kworker/0:0 Not tainted 5.18.7-rc1-00013-g52b92343db13 #74 > [ 311.158550] Hardware name: CZ.NIC Turris Mox Board (DT) > [ 311.163766] Workqueue: events rht_deferred_worker > [ 311.168477] Call trace: > [ 311.170916] dump_backtrace.part.0+0xc8/0xd4 > [ 311.175188] show_stack+0x18/0x70 > [ 311.178501] dump_stack_lvl+0x68/0x84 > [ 311.182159] dump_stack+0x18/0x34 > [ 311.185466] panic+0x168/0x328 > [ 311.188515] out_of_memory+0x568/0x584 > [ 311.192261] __alloc_pages+0xb04/0xbe0 > [ 311.196006] __alloc_pages_bulk+0x15c/0x604 > [ 311.200185] alloc_pages_bulk_array_mempolicy+0xbc/0x24c > [ 311.205491] __vmalloc_node_range+0x238/0x550 > [ 311.209843] __vmalloc_node_range+0x1c0/0x550 > [ 311.214195] kvmalloc_node+0xe0/0x124 > [ 311.217856] bucket_table_alloc.isra.0+0x40/0x150 > [ 311.222554] rhashtable_rehash_alloc.isra.0+0x20/0x8c > [ 311.227599] rht_deferred_worker+0x7c/0x540 > [ 311.231775] process_one_work+0x1d0/0x320 > [ 311.235779] worker_thread+0x70/0x440 > [ 311.239435] kthread+0x10c/0x110 > [ 311.242661] ret_from_fork+0x10/0x20 > [ 311.246238] SMP: stopping secondary CPUs > [ 311.250161] Kernel Offset: disabled > [ 311.253642] CPU features: 0x000,00020009,00001086 > [ 311.258338] Memory Limit: none > [ 311.261390] ---[ end Kernel panic - not syncing: System is deadlocked on memory ]--- > > That can't be quite alright? Shouldn't we have some sort of protection > in the bridge itself too, not just tell hardware driver writers to deal > with it? Or is it somewhere, but it needs to be enabled/configured?This is expected, if you'd like feel free to add a hard learning limit in the driver and the bridge (again if implemented properly). Nothing can save you if someone has L2 access to the device, they can poison any table if learning is enabled.
Vladimir Oltean
2022-Jul-06 20:21 UTC
[Bridge] [PATCH V3 net-next 1/4] net: bridge: add fdb flag to extent locked port feature
On Wed, Jul 06, 2022 at 10:38:04PM +0300, Nikolay Aleksandrov wrote:> I don't think that is new or surprising, if there isn't anything to control the > device resources you'll get there. You don't really need to write any new programs > you can easily do it with mausezahn. I have tests that add over 10 million fdbs on > devices for a few seconds.Of course it isn't new, but that doesn't make the situation in any way better, quite the opposite...> The point is it's not the bridge's task to limit memory consumption or to watch for resource > management. You can limit new entries from the device driver (in case of swdev learning) or > you can use a daemon to watch the number of entries and disable learning. There are many > different ways to avoid this. We've discussed it before and I don't mind adding a hard fdb > per-port limit in the bridge as long as it's done properly. We've also discussed LRU and similar > algorithms for fdb learning and eviction. But any hardcoded limits or limits that can break > current default use cases are unacceptable, they must be opt-in.I don't think you can really say that it's not the bridge's task to limit memory consumption when what it does is essentially allocate memory from untrusted and unbounded user input, in kernel softirq context. That's in fact the problem, the kernel OOM killer will kick in, but there will be no process to kill. This is why the kernel deadlocks on memory and dies. Maybe where our expectations differ is that I believe that a Linux bridge shouldn't need gazillions of tweaks to not kill the kernel? There are many devices in production using a bridge without such configuration, you can't just make it opt-in. Of course, performance under heavy stress is a separate concern, and maybe user space monitoring would be a better idea for that. I know you changed jobs, but did Cumulus Linux have an application to monitor and limit the FDB entry count? Is there some standard application which does this somewhere, or does everybody roll their own? Anyway, limiting FDB entry count from user space is still theoretically different from not dying. If you need to schedule a task to dispose of the weight while the ship is sinking from softirq context, you may never get to actually schedule that task in time. AFAIK the bridge UAPI doesn't expose a pre-programmed limit, so what needs to be done is for user space to manually delete entries until the count falls below the limit.