[dpdk-dev] [PATCH 00/41] Memory Hotplug for DPDK

[Olivier Matz]

Hi,

On Tue, Mar 20, 2018 at 01:51:31PM +0000, Burakov, Anatoly wrote:
> On 20-Mar-18 12:42 PM, Olivier Matz wrote:
> > On Tue, Mar 20, 2018 at 10:27:55AM +0000, Burakov, Anatoly wrote:
> > > On 19-Mar-18 5:30 PM, Olivier Matz wrote:
> > > > Hi Anatoly,
> > > > 
> > > > On Sat, Mar 03, 2018 at 01:45:48PM +0000, Anatoly Burakov wrote:
> > > > > This patchset introduces dynamic memory allocation for DPDK (aka memory
> > > > > hotplug). Based upon RFC submitted in December [1].
> > > > > 
> > > > > Dependencies (to be applied in specified order):
> > > > > - IPC bugfixes patchset [2]
> > > > > - IPC improvements patchset [3]
> > > > > - IPC asynchronous request API patch [4]
> > > > > - Function to return number of sockets [5]
> > > > > 
> > > > > Deprecation notices relevant to this patchset:
> > > > > - General outline of memory hotplug changes [6]
> > > > > - EAL NUMA node count changes [7]
> > > > > 
> > > > > The vast majority of changes are in the EAL and malloc, the external API
> > > > > disruption is minimal: a new set of API's are added for contiguous memory
> > > > > allocation for rte_memzone, and a few API additions in rte_memory due to
> > > > > switch to memseg_lists as opposed to memsegs. Every other API change is
> > > > > internal to EAL, and all of the memory allocation/freeing is handled
> > > > > through rte_malloc, with no externally visible API changes.
> > > > > 
> > > > > Quick outline of all changes done as part of this patchset:
> > > > > 
> > > > >    * Malloc heap adjusted to handle holes in address space
> > > > >    * Single memseg list replaced by multiple memseg lists
> > > > >    * VA space for hugepages is preallocated in advance
> > > > >    * Added alloc/free for pages happening as needed on rte_malloc/rte_free
> > > > >    * Added contiguous memory allocation API's for rte_memzone
> > > > >    * Integrated Pawel Wodkowski's patch for registering/unregistering memory
> > > > >      with VFIO [8]
> > > > >    * Callbacks for registering memory allocations
> > > > >    * Multiprocess support done via DPDK IPC introduced in 18.02
> > > > > 
> > > > > The biggest difference is a "memseg" now represents a single page (as opposed to
> > > > > being a big contiguous block of pages). As a consequence, both memzones and
> > > > > malloc elements are no longer guaranteed to be physically contiguous, unless
> > > > > the user asks for it at reserve time. To preserve whatever functionality that
> > > > > was dependent on previous behavior, a legacy memory option is also provided,
> > > > > however it is expected (or perhaps vainly hoped) to be temporary solution.
> > > > > 
> > > > > Why multiple memseg lists instead of one? Since memseg is a single page now,
> > > > > the list of memsegs will get quite big, and we need to locate pages somehow
> > > > > when we allocate and free them. We could of course just walk the list and
> > > > > allocate one contiguous chunk of VA space for memsegs, but this
> > > > > implementation uses separate lists instead in order to speed up many
> > > > > operations with memseg lists.
> > > > > 
> > > > > For v1, the following limitations are present:
> > > > > - FreeBSD does not even compile, let alone run
> > > > > - No 32-bit support
> > > > > - There are some minor quality-of-life improvements planned that aren't
> > > > >     ready yet and will be part of v2
> > > > > - VFIO support is only smoke-tested (but is expected to work), VFIO support
> > > > >     with secondary processes is not tested; work is ongoing to validate VFIO
> > > > >     for all use cases
> > > > > - Dynamic mapping/unmapping memory with VFIO is not supported in sPAPR
> > > > >     IOMMU mode - help from sPAPR maintainers requested
> > > > > 
> > > > > Nevertheless, this patchset should be testable under 64-bit Linux, and
> > > > > should work for all use cases bar those mentioned above.
> > > > > 
> > > > > [1] http://dpdk.org/dev/patchwork/bundle/aburakov/Memory_RFC/
> > > > > [2] http://dpdk.org/dev/patchwork/bundle/aburakov/IPC_Fixes/
> > > > > [3] http://dpdk.org/dev/patchwork/bundle/aburakov/IPC_Improvements/
> > > > > [4] http://dpdk.org/dev/patchwork/bundle/aburakov/IPC_Async_Request/
> > > > > [5] http://dpdk.org/dev/patchwork/bundle/aburakov/Num_Sockets/
> > > > > [6] http://dpdk.org/dev/patchwork/patch/34002/
> > > > > [7] http://dpdk.org/dev/patchwork/patch/33853/
> > > > > [8] http://dpdk.org/dev/patchwork/patch/24484/
> > > > 
> > > > I did a quick pass on your patches (unfortunately, I don't have
> > > > the time to really dive in it).
> > > > 
> > > > I have few questions/comments:
> > > > 
> > > > - This is really a big patchset. Thank you for working on this topic.
> > > >     I'll try to test our application with it as soon as possible.
> > > > 
> > > > - I see from patch 17 that it is possible that rte_malloc() expands
> > > >     the heap by requesting more memory to the OS? Did I understand well?
> > > >     Today, a good property of rte_malloc() compared to malloc() is that
> > > >     it won't interrupt the process (the worst case is a spinlock). This
> > > >     is appreciable on a dataplane core. Will it change?
> > > 
> > > Hi Olivier,
> > > 
> > > Not sure what you mean by "interrupt the process". The new rte_malloc will
> > > _mostly_ work just like the old one. There are now two levels of locks: the
> > > heap lock, and the system allocation lock. If your rte_malloc call requests
> > > amount of memory that can be satisfied by already allocated memory, then
> > > only the heap lock is engaged - or, to put it in other words, things work as
> > > before.
> > > 
> > > When you *don't* have enough memory allocated, previously rte_malloc would
> > > just fail. Now, it instead will lock the second lock and try to allocate
> > > more memory from the system. This requires IPC (to ensure all processes have
> > > allocated/freed the same memory), so this will take way longer (timeout is
> > > set to wait up to 5 seconds, although under normal circumstances it's taking
> > > a lot less - depending on how many processes you have running, but generally
> > > under 100ms), and will block other system allocations (i.e. if another
> > > rte_malloc call on another heap is trying to request more memory from the
> > > system).
> > > 
> > > So, in short - you can't allocate from the same heap in parallel (same as
> > > before), and you can't have parallel system memory allocation requests
> > > (regardless of from which heap it comes from). The latter *only* applies to
> > > system memory allocations - that is, if one heap is allocating system memory
> > > while another heap receives allocation request *and is able to satisfy it
> > > from already allocated memory*, it will not block, because the second lock
> > > is never engaged.
> > 
> > OK. Let's imagine you are using rte_malloc() on a dataplane core, and
> > you run out of memory. Previously, the allocation would just fail. Now,
> > if my understanding is correct, it can block for a long time, which can
> > be a problem on a dataplane core, because it will cause packet losses,
> > especially if it also blocks allocations on other cores during that
> > time. In this case, it could be useful to make the dynamic heap resizing
> > feature optional.
> 
> Why would anyone in their right mind call rte_malloc on fast path? If you're
> referring to mempool allocations/deallocations, then this is a completely
> separate subject, as mempool alloc/free is not handled by rte_malloc but is
> handled by rte_mempool itself - as far as rte_malloc is concerned, that
> memory is already allocated and it will not touch it.
> 
> As for "making heap resizing feature optional", i'm working on functionality
> that would essentially enable that. Specifically, i'm adding API's to set
> allocation limits and a callback which will get triggered once allocator
> tries to allocate beyond said limits, with an option of returning -1 and
> thus preventing this allocation from completing. While this is kind of a
> round-about way of doing it, it would have similar effect.

Calling rte_malloc() in the data path may be required in case the
application needs to allocate an unknown-sized object. I'm not saying
it's a usual or an optimal use case, I just say that it happens.

Waiting for a spinlock is acceptable in datapath, if it is held by
another dataplane core.
Waiting for several hundreds of ms is not an option in that case.

If the feature is going to be optional, it's perfectly fine for me.


> > I have another question about the patchset. Today, it is not really
> > possible for an application to allocate a page. If you want a full page
> > (ex: 2M), you need to allocate 4M because the rte_malloc layer adds a
> > header before the allocated memory. Therefore, if the memory is
> > fragmented a lot with only 2M pages, you cannot allocate them as pages.
> > 
> > It is possible, with your patchset or in the future, to have an access
> > to a page-based allocator? The use-case is to be able for an application
> > to ask for pages in dpdk memory and remap them in a virtually contiguous
> > memory.
> 
> Pages returned from our allocator are already virtually contiguous, there is
> no need to do any remapping. If user specifies proper size and alignment
> (i.e. reserve a memzone with RTE_MEMZONE_2MB and with 2M size and
> alignment), it will essentially cause the allocator to return a memzone
> that's exactly page-size long. Yes, in the background, it will allocate
> another page to store malloc metadata, and yes, memory will become
> fragmented if multiple such allocations will occur. It is not possible
> (neither now nor in the future planned work) to do what you describe unless
> we store malloc data separately from allocated memory (which can be done,
> but is a non-trivial amount of work).
> 
> Malloc stores its metadata right in the hugepage mostly for multiprocess
> purposes - so that the entire heap is always shared between all processes.
> If we want to store malloc metadata separately from allocated memory, a
> replacement mechanism to shared heap metadata will need to be put in place
> (which, again, can be done, but is a non-trivial amount of work - arguably
> for questionable gain).
> 
> That said, use case you have described is already possible - just allocate
> multiple pages from DPDK as a memzone, and overlay your own memory allocator
> over that memory. This will have the same effect.

Yes, that's currently what I'm doing: to get one 2M page, I'm allocating
more 2M with 2M alignement, which actually results in 4M allocation. My
problem today is when the huge pages are already fragmented at dpdk
start (i.e. only isolated pages). So an allocation of > 2M would fail.

So your patchset mostly solves that issue, because rte_malloc() does not
request physically contiguous memory anymore, which means that
physically isolated hugepages are now virtually contiguous, right? So
rte_malloc(4M) will always be succesful until the memory is virtually
fragmented (i.e. after several malloc/free).

Thank you for the clarification.