On Tue, Dec 13, 2022 at 11:52 PM Vlastimil Babka <vbabka@xxxxxxx> 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?
>
Not sure if it is a problem for other users as well. But I can explain
why it is an issue for BPF accounting.
In BPF accounting, we need to store something into the task_struct and
use it later. So if the 'storer' and the 'user' are different tasks,
the information will be lost.
> >>
> >> > I thought there was a patchset for a whole
> >> > bfp-specific memory allocator, where accounting would be implemented
> >> > naturally, I would imagine.
> >> >
> >>
> >> I posted a patchset[1] which annotates both allocating and freeing
> >> several months ago.
> >> But unfortunately after more investigation and verification I found
> >> the deferred freeing context is a problem, which can't be resolved
> >> easily.
> >> That's why I finally decided to annotate allocating only.
> >>
> >> [1]. https://lore.kernel.org/linux-mm/20220921170002.29557-1-laoar.shao@xxxxxxxxx/
> >>
> >> > > To store the information of a slab or a page, we need to create a new
> >> > > member in struct page, but we can do it in page extension which can
> >> > > avoid changing the size of struct page. So a new page extension
> >> > > active_vm is introduced. Each page and each slab which is allocated as
> >> > > BPF memory will have a struct active_vm. The reason it is named as
> >> > > active_vm is that we can extend it to other areas easily, for example in
> >> > > the future we may use it to count other memory usage.
> >> > >
> >> > > The new page extension active_vm can be disabled via CONFIG_ACTIVE_VM at
> >> > > compile time or kernel parameter `active_vm=` at runtime.
> >> >
> >> > The issue with page_ext is the extra memory usage, so it was rather intended
> >> > for debugging features that can be always compiled in, but only enabled at
> >> > runtime when debugging is needed. The overhead is only paid when enabled.
> >> > That's at least the case of page_owner and page_table_check. The 32bit
> >> > page_idle is rather an oddity that could have instead stayed 64-bit only.
> >> >
> >>
> >> Right, it seems currently page_ext is for debugging purposes only.
> >>
> >> > But this is proposing a page_ext functionality supposed to be enabled at all
> >> > times in production, with the goal of improved accounting. Not an on-demand
> >> > debugging. I'm afraid the costs will outweight the benefits.
> >> >
> >>
> >> The memory overhead of this new page extension is (8/4096), which is
> >> 0.2% of total memory. Not too big to be acceptable.
> >
> > It's generally unacceptable to increase sizeof(struct page)
> > (nor enabling page_ext by default, and that's the why page_ext is for
> > debugging purposes only)
> >
> >> If the user really
> >> thinks this overhead is not accepted, he can set "active_vm=off" to
> >> disable it.
> >
> > I'd say many people won't welcome adding 0.2% of total memory by default
> > to get BPF memory usage.
>
> Agreed.
>
> >> To reduce the memory overhead further, I have a bold idea.
> >> Actually we don't need to allocate such a page extension for every
> >> page, while we only need to allocate it if the user needs to access
> >> it. That said it seems that we can allocate some kind of page
> >> extensions dynamically rather than preallocate at booting, but I
> >> haven't investigated it deeply to check if it can work. What do you
> >> think?
>
> There's lots of benefits (simplicity) of page_ext being allocated as it is
> today.
These benefits also lead it to debugging purposes only :)
If we can make it run on production env, it will be more useful.
> What you're suggesting will be better solved (in few years :) by
> Matthew's bold ideas about shrinking the current struct page and allocating
> usecase-specific descriptors.
>
So the memory overhead won't be a problem in the future, right ?
--
Regards
Yafang
