On Wed, Nov 6, 2024 at 7:57 AM Jesper Dangaard Brouer <hawk@xxxxxxxxxx> wrote: > > > > On 06/11/2024 14.25, Jesper Dangaard Brouer wrote: > > > > On 26/10/2024 09.33, Yunsheng Lin wrote: > >> On 2024/10/25 22:07, Jesper Dangaard Brouer wrote: > >> > >> ... > >> > >>> > >>>>> You and Jesper seems to be mentioning a possible fact that there might > >>>>> be 'hundreds of gigs of memory' needed for inflight pages, it would > >>>>> be nice > >>>>> to provide more info or reasoning above why 'hundreds of gigs of > >>>>> memory' is > >>>>> needed here so that we don't do a over-designed thing to support > >>>>> recording > >>>>> unlimited in-flight pages if the driver unbound stalling turns out > >>>>> impossible > >>>>> and the inflight pages do need to be recorded. > >>>> > >>>> I don't have a concrete example of a use that will blow the limit you > >>>> are setting (but maybe Jesper does), I am simply objecting to the > >>>> arbitrary imposing of any limit at all. It smells a lot of "640k ought > >>>> to be enough for anyone". > >>>> > >>> > >>> As I wrote before. In *production* I'm seeing TCP memory reach 24 GiB > >>> (on machines with 384GiB memory). I have attached a grafana screenshot > >>> to prove what I'm saying. > >>> > >>> As my co-worker Mike Freemon, have explain to me (and more details in > >>> blogposts[1]). It is no coincident that graph have a strange "sealing" > >>> close to 24 GiB (on machines with 384GiB total memory). This is because > >>> TCP network stack goes into a memory "under pressure" state when 6.25% > >>> of total memory is used by TCP-stack. (Detail: The system will stay in > >>> that mode until allocated TCP memory falls below 4.68% of total memory). > >>> > >>> [1] > >>> https://blog.cloudflare.com/unbounded-memory-usage-by-tcp-for-receive-buffers-and-how-we-fixed-it/ > >> > >> Thanks for the info. > > > > Some more info from production servers. > > > > (I'm amazed what we can do with a simple bpftrace script, Cc Viktor) > > > > In below bpftrace script/oneliner I'm extracting the inflight count, for > > all page_pool's in the system, and storing that in a histogram hash. > > > > sudo bpftrace -e ' > > rawtracepoint:page_pool_state_release { @cnt[probe]=count(); > > @cnt_total[probe]=count(); > > $pool=(struct page_pool*)arg0; > > $release_cnt=(uint32)arg2; > > $hold_cnt=$pool->pages_state_hold_cnt; > > $inflight_cnt=(int32)($hold_cnt - $release_cnt); > > @inflight=hist($inflight_cnt); > > } > > interval:s:1 {time("\n%H:%M:%S\n"); > > print(@cnt); clear(@cnt); > > print(@inflight); > > print(@cnt_total); > > }' > > > > The page_pool behavior depend on how NIC driver use it, so I've run this > > on two prod servers with drivers bnxt and mlx5, on a 6.6.51 kernel. > > > > Driver: bnxt_en > > - kernel 6.6.51 > > > > @cnt[rawtracepoint:page_pool_state_release]: 8447 > > @inflight: > > [0] 507 | | > > [1] 275 | | > > [2, 4) 261 | | > > [4, 8) 215 | | > > [8, 16) 259 | | > > [16, 32) 361 | | > > [32, 64) 933 | | > > [64, 128) 1966 | | > > [128, 256) 937052 |@@@@@@@@@ | > > [256, 512) 5178744 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| > > [512, 1K) 73908 | | > > [1K, 2K) 1220128 |@@@@@@@@@@@@ | > > [2K, 4K) 1532724 |@@@@@@@@@@@@@@@ | > > [4K, 8K) 1849062 |@@@@@@@@@@@@@@@@@@ | > > [8K, 16K) 1466424 |@@@@@@@@@@@@@@ | > > [16K, 32K) 858585 |@@@@@@@@ | > > [32K, 64K) 693893 |@@@@@@ | > > [64K, 128K) 170625 |@ | > > > > Driver: mlx5_core > > - Kernel: 6.6.51 > > > > @cnt[rawtracepoint:page_pool_state_release]: 1975 > > @inflight: > > [128, 256) 28293 |@@@@ | > > [256, 512) 184312 |@@@@@@@@@@@@@@@@@@@@@@@@@@@ | > > [512, 1K) 0 | | > > [1K, 2K) 4671 | | > > [2K, 4K) 342571 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| > > [4K, 8K) 180520 |@@@@@@@@@@@@@@@@@@@@@@@@@@@ | > > [8K, 16K) 96483 |@@@@@@@@@@@@@@ | > > [16K, 32K) 25133 |@@@ | > > [32K, 64K) 8274 |@ | > > > > > > The key thing to notice that we have up-to 128,000 pages in flight on > > these random production servers. The NIC have 64 RX queue configured, > > thus also 64 page_pool objects. > > > > I realized that we primarily want to know the maximum in-flight pages. > > So, I modified the bpftrace oneliner to track the max for each page_pool > in the system. > > sudo bpftrace -e ' > rawtracepoint:page_pool_state_release { @cnt[probe]=count(); > @cnt_total[probe]=count(); > $pool=(struct page_pool*)arg0; > $release_cnt=(uint32)arg2; > $hold_cnt=$pool->pages_state_hold_cnt; > $inflight_cnt=(int32)($hold_cnt - $release_cnt); > $cur=@inflight_max[$pool]; > if ($inflight_cnt > $cur) { > @inflight_max[$pool]=$inflight_cnt;} > } > interval:s:1 {time("\n%H:%M:%S\n"); > print(@cnt); clear(@cnt); > print(@inflight_max); > print(@cnt_total); > }' > > I've attached the output from the script. > For unknown reason this system had 199 page_pool objects. > > The 20 top users: > > $ cat out02.inflight-max | grep inflight_max | tail -n 20 > @inflight_max[0xffff88829133d800]: 26473 > @inflight_max[0xffff888293c3e000]: 27042 > @inflight_max[0xffff888293c3b000]: 27709 > @inflight_max[0xffff8881076f2800]: 29400 > @inflight_max[0xffff88818386e000]: 29690 > @inflight_max[0xffff8882190b1800]: 29813 > @inflight_max[0xffff88819ee83800]: 30067 > @inflight_max[0xffff8881076f4800]: 30086 > @inflight_max[0xffff88818386b000]: 31116 > @inflight_max[0xffff88816598f800]: 36970 > @inflight_max[0xffff8882190b7800]: 37336 > @inflight_max[0xffff888293c38800]: 39265 > @inflight_max[0xffff888293c3c800]: 39632 > @inflight_max[0xffff888293c3b800]: 43461 > @inflight_max[0xffff888293c3f000]: 43787 > @inflight_max[0xffff88816598f000]: 44557 > @inflight_max[0xffff888132ce9000]: 45037 > @inflight_max[0xffff888293c3f800]: 51843 > @inflight_max[0xffff888183869800]: 62612 > @inflight_max[0xffff888113d08000]: 73203 > > Adding all values together: > > grep inflight_max out02.inflight-max | awk 'BEGIN {tot=0} {tot+=$2; > printf "total:" tot "\n"}' | tail -n 1 > > total:1707129 > > Worst case we need a data structure holding 1,707,129 pages. > Fortunately, we don't need a single data structure as this will be split > between 199 page_pool's. > > --Jesper Is there any specific reason for why we need to store the pages instead of just scanning the page tables to look for them? We should already know how many we need to look for and free. If we were to just scan the page structs and identify the page pool pages that are pointing to our pool we should be able to go through and clean them up. It won't be the fastest approach, but this should be an exceptional case to handle things like a hot plug removal of a device where we can essentially run this in the background before we free the device. Then it would just be a matter of modifying the pool so that it will drop support for doing DMA unmapping and essentially just become a place for the freed pages to go to die.
