On Wed, Dec 14, 2022 at 3:22 AM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Tue, Dec 13, 2022 at 04:52:09PM +0100, Vlastimil Babka wrote:
> > On 12/13/22 15:56, Hyeonggon Yoo wrote:
> > > On Tue, Dec 13, 2022 at 07:52:42PM +0800, Yafang Shao wrote:
> > >> On Tue, Dec 13, 2022 at 1:54 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
> > >> >
> > >> > On 12/12/22 01:37, Yafang Shao wrote:
> > >> > > Currently there's no way to get BPF memory usage, while we can only
> > >> > > estimate the usage by bpftool or memcg, both of which are not reliable.
> > >> > >
> > >> > > - bpftool
> > >> > > `bpftool {map,prog} show` can show us the memlock of each map and
> > >> > > prog, but the memlock is vary from the real memory size. The memlock
> > >> > > of a bpf object is approximately
> > >> > > `round_up(key_size + value_size, 8) * max_entries`,
> > >> > > so 1) it can't apply to the non-preallocated bpf map which may
> > >> > > increase or decrease the real memory size dynamically. 2) the element
> > >> > > size of some bpf map is not `key_size + value_size`, for example the
> > >> > > element size of htab is
> > >> > > `sizeof(struct htab_elem) + round_up(key_size, 8) + round_up(value_size, 8)`
> > >> > > That said the differece between these two values may be very great if
> > >> > > the key_size and value_size is small. For example in my verifaction,
> > >> > > the size of memlock and real memory of a preallocated hash map are,
> > >> > >
> > >> > > $ grep BPF /proc/meminfo
> > >> > > BPF: 1026048 B <<< the size of preallocated memalloc pool
> > >> > >
> > >> > > (create hash map)
> > >> > >
> > >> > > $ bpftool map show
> > >> > > 3: hash name count_map flags 0x0
> > >> > > key 4B value 4B max_entries 1048576 memlock 8388608B
> > >> > >
> > >> > > $ grep BPF /proc/meminfo
> > >> > > BPF: 84919344 B
> > >> > >
> > >> > > So the real memory size is $((84919344 - 1026048)) which is 83893296
> > >> > > bytes while the memlock is only 8388608 bytes.
> > >> > >
> > >> > > - memcg
> > >> > > With memcg we only know that the BPF memory usage is less than
> > >> > > memory.usage_in_bytes (or memory.current in v2). Furthermore, we only
> > >> > > know that the BPF memory usage is less than $MemTotal if the BPF
> > >> > > object is charged into root memcg :)
> > >> > >
> > >> > > So we need a way to get the BPF memory usage especially there will be
> > >> > > more and more bpf programs running on the production environment. The
> > >> > > memory usage of BPF memory is not trivial, which deserves a new item in
> > >> > > /proc/meminfo.
> > >> > >
> > >> > > This patchset introduce a solution to calculate the BPF memory usage.
> > >> > > This solution is similar to how memory is charged into memcg, so it is
> > >> > > easy to understand. It counts three types of memory usage -
> > >> > > - page
> > >> > > via kmalloc, vmalloc, kmem_cache_alloc or alloc pages directly and
> > >> > > their families.
> > >> > > When a page is allocated, we will count its size and mark the head
> > >> > > page, and then check the head page at page freeing.
> > >> > > - slab
> > >> > > via kmalloc, kmem_cache_alloc and their families.
> > >> > > When a slab object is allocated, we will mark this object in this
> > >> > > slab and check it at slab object freeing. That said we need extra memory
> > >> > > to store the information of each object in a slab.
> > >> > > - percpu
> > >> > > via alloc_percpu and its family.
> > >> > > When a percpu area is allocated, we will mark this area in this
> > >> > > percpu chunk and check it at percpu area freeing. That said we need
> > >> > > extra memory to store the information of each area in a percpu chunk.
> > >> > >
> > >> > > So we only need to annotate the allcation to add the BPF memory size,
> > >> > > and the sub of the BPF memory size will be handled automatically at
> > >> > > freeing. We can annotate it in irq, softirq or process context. To avoid
> > >> > > counting the nested allcations, for example the percpu backing allocator,
> > >> > > we reuse the __GFP_ACCOUNT to filter them out. __GFP_ACCOUNT also make
> > >> > > the count consistent with memcg accounting.
> > >> >
> > >> > So you can't easily annotate the freeing places as well, to avoid the whole
> > >> > tracking infrastructure?
> > >>
> > >> The trouble is kfree_rcu(). for example,
> > >> old_item = active_vm_item_set(ACTIVE_VM_BPF);
> > >> kfree_rcu();
> > >> active_vm_item_set(old_item);
> > >> If we want to pass the ACTIVE_VM_BPF into the deferred rcu context, we
> > >> will change lots of code in the RCU subsystem. I'm not sure if it is
> > >> worth it.
> > >
> > > (+Cc rcu folks)
> > >
> > > IMO adding new kfree_rcu() varient for BPF that accounts BPF memory
> > > usage would be much less churn :)
> >
> > Alternatively, just account the bpf memory as freed already when calling
> > kfree_rcu()? I think the amount of memory "in flight" to be freed by rcu is
> > a separate issue (if it's actually an issue) and not something each
> > kfree_rcu() user should think about separately?
>
> If the in-flight memory really does need to be accounted for, then one
> straightforward approach is to use call_rcu() and do the first part of
> the needed accounting at the call_rcu() callsite and the rest of the
> accounting when the callback is invoked. Or, if memory must be freed
> quickly even on ChromeOS and Android, use call_rcu_hurry() instead
> of call_rcu().
>
Right, call_rcu() can make it work.
But I'm not sure if all kfree_rcu() in kernel/bpf can be replaced by call_rcu().
Alexei, any comment on it ?
$ grep -r "kfree_rcu" kernel/bpf/
kernel/bpf/local_storage.c: kfree_rcu(new, rcu);
kernel/bpf/lpm_trie.c: kfree_rcu(node, rcu);
kernel/bpf/lpm_trie.c: kfree_rcu(parent, rcu);
kernel/bpf/lpm_trie.c: kfree_rcu(node, rcu);
kernel/bpf/lpm_trie.c: kfree_rcu(node, rcu);
kernel/bpf/bpf_inode_storage.c: kfree_rcu(local_storage, rcu);
kernel/bpf/bpf_task_storage.c: kfree_rcu(local_storage, rcu);
kernel/bpf/trampoline.c: kfree_rcu(im, rcu);
kernel/bpf/core.c: kfree_rcu(progs, rcu);
kernel/bpf/core.c: * no need to call kfree_rcu(), just call
kfree() directly.
kernel/bpf/core.c: kfree_rcu(progs, rcu);
kernel/bpf/bpf_local_storage.c: * kfree(), else do kfree_rcu().
kernel/bpf/bpf_local_storage.c: kfree_rcu(local_storage, rcu);
kernel/bpf/bpf_local_storage.c: kfree_rcu(selem, rcu);
kernel/bpf/bpf_local_storage.c: kfree_rcu(selem, rcu);
kernel/bpf/bpf_local_storage.c: kfree_rcu(local_storage, rcu);
--
Regards
Yafang
