On Thu, Jan 30, 2020 at 12:40 PM Magnus Karlsson <magnus.karlsson@xxxxxxxxx> wrote: > > On Thu, Jan 30, 2020 at 10:59 AM Magnus Karlsson > <magnus.karlsson@xxxxxxxxx> wrote: > > > > On Thu, Jan 30, 2020 at 10:37 AM Maxim Mikityanskiy > > <maximmi@xxxxxxxxxxxx> wrote: > > > > > > > -----Original Message----- > > > > From: Magnus Karlsson <magnus.karlsson@xxxxxxxxx> > > > > Sent: 27 January, 2020 17:55 > > > > To: Maxim Mikityanskiy <maximmi@xxxxxxxxxxxx> > > > > Cc: Ryan Goodfellow <rgoodfel@xxxxxxx>; xdp-newbies@xxxxxxxxxxxxxxx; Tariq > > > > Toukan <tariqt@xxxxxxxxxxxx>; Saeed Mahameed <saeedm@xxxxxxxxxxxx>; Moshe > > > > Shemesh <moshe@xxxxxxxxxxxx> > > > > Subject: Re: zero-copy between interfaces > > > > > > > > On Mon, Jan 27, 2020 at 3:01 PM Maxim Mikityanskiy <maximmi@xxxxxxxxxxxx> > > > > wrote: > > > > > > > > > > On 2020-01-22 23:43, Ryan Goodfellow wrote: > > > > > > On Tue, Jan 21, 2020 at 01:40:50PM +0000, Maxim Mikityanskiy wrote: > > > > > >>>> I've posted output from the program in debugging mode here > > > > > >>>> > > > > > >>>> - https://gitlab.com/mergetb/tech/network- > > > > emulation/kernel/snippets/1930375 > > > > > >>>> > > > > > >>>> Yes, you are correct in that forwarding works for a brief period and > > > > then stops. > > > > > >>>> I've noticed that the number of packets that are forwarded is equal > > > > to the size > > > > > >>>> of the producer/consumer descriptor rings. I've posted two ping > > > > traces from a > > > > > >>>> client ping that shows this. > > > > > >>>> > > > > > >>>> - https://gitlab.com/mergetb/tech/network- > > > > emulation/kernel/snippets/1930376 > > > > > >>>> - https://gitlab.com/mergetb/tech/network- > > > > emulation/kernel/snippets/1930377 > > > > > >> > > > > > >> These snippets are not available. > > > > > > > > > > > > Apologies, I had the wrong permissions set. They should be available > > > > now. > > > > > > > > > > > >> > > > > > >>>> > > > > > >>>> I've also noticed that when the forwarding stops, the CPU usage for > > > > the proc > > > > > >>>> running the program is pegged, which is not the norm for this program > > > > as it uses > > > > > >>>> a poll call with a timeout on the xsk fd. > > > > > >> > > > > > >> This information led me to a guess what may be happening. On the RX > > > > > >> side, mlx5e allocates pages in bulks for performance reasons and to > > > > > >> leverage hardware features targeted to performance. In AF_XDP mode, > > > > > >> bulking of frames is also used (on x86, the bulk size is 64 with > > > > > >> striding RQ enabled, and 8 otherwise, however, it's implementation > > > > > >> details that might change later). If you don't put enough frames to XSK > > > > > >> Fill Ring, the driver will be demanding more frames and return from > > > > > >> poll() immediately. Basically, in the application, you should put as > > > > > >> many frames to the Fill Ring as you can. Please check if that could be > > > > > >> the root cause of your issue. > > > > > > > > > > > > The code in this application makes an effort to relenish the fill ring > > > > as fast > > > > > > as possible. The basic loop of the application is to first check if > > > > there are > > > > > > any descriptors to be consumed from the completion queue or any > > > > descriptors that > > > > > > can be added to the fill queue, and only then to move on to moving > > > > packets > > > > > > through the rx and tx rings. > > > > > > > > > > > > https://gitlab.com/mergetb/tech/network-emulation/kernel/blob/v5.5- > > > > moa/samples/bpf/xdpsock_multidev.c#L452-474 > > > > > > > > > > I reproduced your issue and did my investigation, and here is what I > > > > found: > > > > > > > > > > 1. Commit df0ae6f78a45 (that you found during bisect) introduces an > > > > > important behavioral change (which I thought was not that important). > > > > > xskq_nb_avail used to return min(entries, dcnt), but after the change it > > > > > just returns entries, which may be as big as the ring size. > > > > > > > > > > 2. xskq_peek_addr updates q->ring->consumer only when q->cons_tail > > > > > catches up with q->cons_head. So, before that patch and one previous > > > > > patch, cons_head - cons_tail was not more than 16, so the consumer index > > > > > was updated periodically. Now consumer is updated only when the whole > > > > > ring is exhausted. > > > > > > > > > > 3. The application can't replenish the fill ring if the consumer index > > > > > doesn't move. As a consequence, refilling the descriptors by the > > > > > application can't happen in parallel with using them by the driver. It > > > > > should have some performance penalty and possibly even lead to packet > > > > > drops, because the driver uses all the descriptors and only then > > > > > advances the consumer index, and then it has to wait until the > > > > > application refills the ring, busy-looping and losing packets. > > > > > > > > This will happen if user space always fills up the whole ring, which > > > > might or might not happen all depending on the app. > > > > > > Yes, that's right, and as far as I know, it's common to fill as many > > > frames as the application can (there was no reason to do it other way > > > till now). > > > > > > > With that said, it > > > > might provide better performance to update it once in a while. It > > > > might also be the case that we will get better performance with the > > > > new scheme if we only fill half the fill ring. > > > > > > Yes, it may improve performance. However, I don't think it's correct to > > > set such a limitation on the app. > > Actually, a much worse limitation to put on an application is to say > that you have to have a certain amount of buffers on some ring for the > zero-copy feature to work. For example that we need at least 64 > buffers on the fill ring for all the NIC cards to work in zero-copy > mode. That would be a bad thing to have to put in the documentation. > An OS is supposed to abstract away HW differences, and with this > current limitation in your driver, it shines through for sure. What we > would like to put in the documentation is a statement along the lines > of: "for best performance, make sure you have plenty of buffers on the > fill ring so that the NIC can work as efficiently as possible". Not a > statement that it does not work on Mellanox unless you put enough > buffers on the fill ring. So my advice (and wish) is that you fix this > in your driver. With that said, I will still look into what is the > best way to get at least the sample to work for you. But there is no > way to make sure every single app works for you in zero-copy mode, > unless you support arbitrary amount of buffers on the fill ring. I > guess that sooner or later, a customer of yours will get into this > situation one way or the other, so why not fix it now. > > /Magnus > > > I will examine what provides the best performance. On one hand it is > > the number of updates to shared ring state (minimized by current > > scheme) and the ability for the user app to but buffers on the fill > > ring. Stating that putting e.g. half the packets on the fill ring > > provides better performance (for some application) is not a > > limitation. It is performance tuning advise :-). I have now made a set of measurements. First I just made a variation study using the xdpsock app, varying the amount of packets the kernel can withdraw from a consumer ring (fill and Tx) before updating global state. For the 1 core case (app and driver on the same core) the more frequent you do this update, the better. The reason for this is that it costs very little to update the state since the application is not running. And it is beneficial for the app to have a freshly updated state when it starts to execute as it can operate on more packets. For the 2 core case (app on one core, driver on another) it is the complete opposite: the fewer updates to global state, the better. The reason for this is that it costs a lot to update global state as it triggers cache coherency actions between the two cores. What I did then was to compare the current scheme, update only when grabbing new packets, to a new scheme were we also update the global consumer pointer when we are exiting Rx or Tx processing in the NAPI context. On two cores the current scheme gets 0.5 to 1 Mpps more in throughput than also updating the pointer at the end of NAPI. But for 1 core case, the new scheme is the best and generates between 0.2 and 0.3 Mpps more in throughput than the current one. But all in all, the current scheme is more beneficial than the proposed one if we say that both the 1 core and the 2 core case is equally important. Something to note is that the xdpsock application only puts batch size (64) of packets in the fill ring in every iteration, and this might lead to some good pacing for the current scheme and the 2 core case. I.e., we do not get into the case of the fill ring only being full or empty. But I will run this on some real apps to get some more results, and I know that Björn has an optimized xdpsock application that puts many more packets into the fill queue than 64. This optimization might actually make the new proposal (also updating at the end of NAPI) be better and make the current scheme suffer. We will examine this further and get back. /Magnus > > > > I will look into this > > > > and see what I get. > > > > > > > > > 4. As mlx5e allocates frames in batches, the consequences are even more > > > > > severe: it's a deadlock where the driver waits for the application, and > > > > > vice versa. The driver never reaches the point where cons_tail gets > > > > > equal to cons_head. E.g., if cons_tail + 3 == cons_head, and the batch > > > > > size requested by the driver is 8, the driver won't peek anything from > > > > > the fill ring waiting for difference between cons_tail and cons_head to > > > > > increase to be at least 8. On the other hand, the application can't put > > > > > anything to the ring, because it still thinks that the consumer index is > > > > > 0. As cons_tail never reaches cons_head, the consumer index doesn't get > > > > > updated, hence the deadlock. > > > > > > > > Good thing that you detected this. Maybe I should get a Mellanox card > > > > :-). This is different from how we implemented Intel's drivers that > > > > just work on any batch size. If it gets 3 packets back, it will use > > > > those. How do you deal with the case when the application just puts a > > > > single buffer in the fill ring and wants to receive a single packet? > > > > > > mlx5e will wait until the full batch is available. As AF_XDP is intended > > > for high-performance apps, this scenario is less expected. We prefer to > > > leverage our performance features. > > > > That you cannot support all applications on top of AF_XDP with your > > zero-copy support seems broken to me. But I give you that you might > > support all the ones you care about with your current batching > > support. Like you say, most apps will put plenty of buffers on the > > fill ring, so this should not be a problem. Can you not implement some > > slow path for these cases? You must have one for the skb case. > > > > > > Why does the Mellanox driver need a specific batch size? This is just > > > > for my understanding so we can find a good solution. > > > > > > The main reason is our performance feature called striding RQ. Skipping > > > all irrelevant details, a batch of 64 pages is posted to the NIC with a > > > single request, and the NIC fills them progressively. This feature is > > > turned on by default on modern NICs, and it's really good for > > > performance. > > > > > > It might be possible to post a smaller batch though, I'm not sure about > > > it, it needs to be checked, but anyway it's not something that we will > > > likely do, because it's a complication of the data path, and if we know > > > more frames are coming, it's much better for the performance to wait for > > > them, rather than to post several incomplete batches. > > > > > > > > So, in my vision, the decision to remove RX_BATCH_SIZE and periodic > > > > > updates of the consumer index was wrong. It totally breaks mlx5e, that > > > > > does batching, and it will affect the performance of any driver, because > > > > > the application can't refill the ring until it gets completely empty and > > > > > the driver starts waiting for frames. I suggest that periodic updates of > > > > > the consumer index should be readded to xskq_cons_peek_addr. > > > > > > > > The reason I wanted to remove RX_BATCH_SIZE is that application > > > > developers complained about it giving rise to counter intuitive > > > > behavior in user space. I will try to dig out the complaints and the > > > > explanations Björn and I had to send which it seemed that users really > > > > should not have to care about. It should just work. > > > > > > I think the counter that doesn't update till the very last moment and > > > then advances by the ring size will also be something to complain about > > > (and I am the first one to complain :D). Such bursts are not good in any > > > case. > > > > Do you have any performance data that shows this scheme is bad for > > performance? The good thing about this scheme is that global state is > > updated less frequently. And the bad thing is what you mentioned. But > > which one has the greatest effect, is the question. > > > > > > I also do not like > > > > to have arbitrary constants like this. Why 16? > > > > > > I believe any batching mechanism has a constant that look arbitrary. > > > This constant should be the golden mean: if it's too small, there is > > > little effect from batching; if it's too big, it gets too bursty. > > > > > > Basically, after your patch it just changed from 16 to the ring size. > > > Maybe we can tie that constant to ring size? Make it ring_size / > > > another_arbitrary_constant? :) > > > > Agreed, I thought about this too. Something tied to ring_size might be > > a good compromise. Will examine this. But I want to base this on > > performance data not idle speculation, so need to run some experiments > > first. > > > > /Magnus > > > > > > Would much prefer not > > > > having to deal with this, unless of course it horribly breaks > > > > something or gives rise to worse performance. Might still be the case > > > > here, but if not, I would like to remove it. > > > > > > > > Thanks: Magnus > > > > > > > > > Magnus, what do you think of the suggestion above? > > > > > > > > > > Thanks, > > > > > Max > > > > > > > > > > >> > > > > > >> I tracked this issue in our internal bug tracker in case we need to > > > > > >> perform actual debugging of mlx5e. I'm looking forward to your feedback > > > > > >> on my assumption above. > > > > > >> > > > > > >>>> The hardware I am using is a Mellanox ConnectX4 2x100G card (MCX416A- > > > > CCAT) > > > > > >>>> running the mlx5 driver. > > > > > >> > > > > > >> This one should run without striding RQ, please verify it with ethtool > > > > > >> --show-priv-flags (the flag name is rx_striding_rq). > > > > > > > > > > > > I do not remember changing this option, so whatever the default is, is > > > > what it > > > > > > was running with. I am traveling this week and do not have access to > > > > these > > > > > > systems, but will ensure that this flag is set properly when I get back. > > > > > > > > > > >
|