> On Jun 10, 2019, at 8:07 PM, Tom Talpey <tom@xxxxxxxxxx> wrote: > > On 6/10/2019 6:13 PM, Tom Talpey wrote: >> On 6/10/2019 5:57 PM, Chuck Lever wrote: >>> >>> >>>> On Jun 10, 2019, at 3:14 PM, Tom Talpey <tom@xxxxxxxxxx> wrote: >>>> >>>> On 6/10/2019 1:50 PM, Chuck Lever wrote: >>>>> Hi Tom- >>>>>> On Jun 10, 2019, at 10:50 AM, Tom Talpey <tom@xxxxxxxxxx> wrote: >>>>>> >>>>>> On 6/5/2019 1:25 PM, Chuck Lever wrote: >>>>>>> Hi Tom- >>>>>>>> On Jun 5, 2019, at 12:43 PM, Tom Talpey <tom@xxxxxxxxxx> wrote: >>>>>>>> >>>>>>>> On 6/5/2019 8:15 AM, Chuck Lever wrote: >>>>>>>>> The DRC is not working at all after an RPC/RDMA transport reconnect. >>>>>>>>> The problem is that the new connection uses a different source port, >>>>>>>>> which defeats DRC hash. >>>>>>>>> >>>>>>>>> An NFS/RDMA client's source port is meaningless for RDMA transports. >>>>>>>>> The transport layer typically sets the source port value on the >>>>>>>>> connection to a random ephemeral port. The server already ignores it >>>>>>>>> for the "secure port" check. See commit 16e4d93f6de7 ("NFSD: Ignore >>>>>>>>> client's source port on RDMA transports"). >>>>>>>> >>>>>>>> Where does the entropy come from, then, for the server to not >>>>>>>> match other requests from other mount points on this same client? >>>>>>> The first ~200 bytes of each RPC Call message. >>>>>>> [ Note that this has some fun ramifications for calls with small >>>>>>> RPC headers that use Read chunks. ] >>>>>> >>>>>> Ok, good to know. I forgot that the Linux server implemented this. >>>>>> I have some concerns abot it, honestly, and it's important to remember >>>>>> that it's not the same on all servers. But for the problem you're >>>>>> fixing, it's ok I guess and certainly better than today. Still, the >>>>>> errors are goingto be completely silent, and can lead to data being >>>>>> corrupted. Well, welcome to the world of NFSv3. >>>>> I don't see another option. >>>>> Some regard this checksum as more robust than using the client's >>>>> IP source port. After all, the same argument can be made that >>>>> the server cannot depend on clients to reuse their source port. >>>>> That is simply a convention that many clients adopted before >>>>> servers used a stronger DRC hash mechanism. >>>>>>>> And since RDMA is capable of >>>>>>>> such high IOPS, the likelihood seems rather high. >>>>>>> Only when the server's durable storage is slow enough to cause >>>>>>> some RPC requests to have extremely high latency. >>>>>>> And, most clients use an atomic counter for their XIDs, so they >>>>>>> are also likely to wrap that counter over some long-pending RPC >>>>>>> request. >>>>>>> The only real answer here is NFSv4 sessions. >>>>>>>> Missing the cache >>>>>>>> might actually be safer than hitting, in this case. >>>>>>> Remember that _any_ retransmit on RPC/RDMA requires a fresh >>>>>>> connection, that includes NFSv3, to reset credit accounting >>>>>>> due to the lost half of the RPC Call/Reply pair. >>>>>>> I can very quickly reproduce bad (non-deterministic) behavior >>>>>>> by running a software build on an NFSv3 on RDMA mount point >>>>>>> with disconnect injection. If the DRC issue is addressed, the >>>>>>> software build runs to completion. >>>>>> >>>>>> Ok, good. But I have a better test. >>>>>> >>>>>> In the Connectathon suite, there's a "Special" test called "nfsidem". >>>>>> I wrote this test in, like, 1989 so I remember it :-) >>>>>> >>>>>> This test performs all the non-idempotent NFv3 operations in a loop, >>>>>> and each loop element depends on the previous one, so if there's >>>>>> any failure, the test imemdiately bombs. >>>>>> >>>>>> Nobody seems to understand it, usually when it gets run people will >>>>>> run it without injecting errors, and it "passes" so they decide >>>>>> everything is ok. >>>>>> >>>>>> So my suggestion is to run your flakeway packet-drop harness while >>>>>> running nfsidem in a huge loop (nfsidem 10000). The test is slow, >>>>>> owing to the expensive operations it performs, so you'll need to >>>>>> run it for a long time. >>>>>> >>>>>> You'll almost definitely get a failure or two, since the NFSv3 >>>>>> protocol is flawed by design. But you can compare the behaviors, >>>>>> and even compute a likelihood. I'd love to see some actual numbers. >>>>> I configured the client to disconnect after 23711 RPCs have completed. >>>>> (I can re-run these with more frequent disconnects if you think that >>>>> would be useful). >>>>> Here's a run with the DRC modification: >>>>> [cel@manet ~]$ sudo mount -o vers=3,proto=rdma,sec=sys klimt.ib:/export/tmp /mnt >>>>> [cel@manet ~]$ (cd /mnt; ~/src/cthon04/special/nfsidem 100000) >>>>> testing 100000 idempotencies in directory "./TEST" >>>>> [cel@manet ~]$ sudo umount /mnt >>>>> Here's a run with the stock v5.1 Linux server: >>>>> [cel@manet ~]$ sudo mount -o vers=3,proto=rdma,sec=sys klimt.ib:/export/tmp /mnt >>>>> [cel@manet ~]$ (cd /mnt; ~/src/cthon04/special/nfsidem 100000) >>>>> testing 100000 idempotencies in directory "./TEST" >>>>> [cel@manet ~]$ >>>>> This test reported no errors in either case. We can see that the >>>>> disconnects did trigger retransmits: >>>>> RPC statistics: >>>>> 1888819 RPC requests sent, 1888581 RPC replies received (0 XIDs not found) >>>>> average backlog queue length: 119 >>>> >>>> Ok, well, that's 1.2% error rate, which IMO could be cranked up much >>>> higher for testing purposes. I'd also be sure the server was serving >>>> other workloads during the same time, putting at least some pressure >>>> on the DRC. The op rate of a single nfsidem test is pretty low so I >>>> doubt it's ever evicting anything. >>>> >>>> Ideally, it would be best to >>>> 1) increase the error probability >>>> 2) run several concurrent nfsidem tests, on different connections >>>> 3) apply some other load to the server, e.g. several cthon basics >>>> >>>> The idea being to actually get the needle off of zero and measure some >>>> kind of difference. Otherwise it really isn't giving any information >>>> apart from a slight didn't-break-it confidence. Honestly, I'm surprised >>>> you couldn't evince a failure from stock. On paper, these results don't >>>> actually tell us the patch is doing anything. >>> >>> I boosted the disconnect injection rate to once every 11353 RPCs, >>> and mounted a second share with "nosharecache", running nfsidem on >>> both mounts. Both mounts are subject to disconnect injection. >>> >>> With the current v5.2-rc Linux server, both nfsidem jobs fail within >>> 30 seconds. >>> >>> With my DRC fix applied, both jobs run to completion with no errors. >>> It takes more than an hour. >> Great, that definitely proves it. It's relatively low risk, and >> should go in. >> I think it's worth thinking through the best way to address this >> across multiple transports, especially RDMA, where the port and >> address behaviors may differ from TCP/UDP. Perhaps tossing the whole >> idea of using the 5-tuple should simply be set aside, after all >> these years. Even though NFSv4.1 already has! >> Thanks for humoring me with the extra testing, by the way :-) >> Tom. >>> Here are the op metrics for one of the mounts during a run that >>> completes successfully: >>> >>> RPC statistics: >>> 4091380 RPC requests sent, 4088143 RPC replies received (0 XIDs not found) >>> average backlog queue length: 1800 >>> >>> ACCESS: >>> 300395 ops (7%) 301 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 132 avg bytes received per op: 120 >>> backlog wait: 4.289199 RTT: 0.019821 total execute time: 4.315092 (milliseconds) >>> REMOVE: >>> 300390 ops (7%) 168 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 136 avg bytes received per op: 144 >>> backlog wait: 2.368664 RTT: 0.070106 total execute time: 2.445148 (milliseconds) >>> MKDIR: >>> 200262 ops (4%) 193 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 158 avg bytes received per op: 271 >>> backlog wait: 4.207034 RTT: 0.075421 total execute time: 4.289101 (milliseconds) >>> RMDIR: >>> 200262 ops (4%) 100 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 130 avg bytes received per op: 144 >>> backlog wait: 2.050749 RTT: 0.071676 total execute time: 2.128801 (milliseconds) >>> LOOKUP: >>> 194664 ops (4%) 233 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 136 avg bytes received per op: 176 >>> backlog wait: 5.365984 RTT: 0.020615 total execute time: 5.392769 (milliseconds) >>> SETATTR: >>> 100130 ops (2%) 86 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 160 avg bytes received per op: 144 >>> backlog wait: 3.520603 RTT: 0.066863 total execute time: 3.594327 (milliseconds) >>> WRITE: >>> 100130 ops (2%) 82 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 180 avg bytes received per op: 136 >>> backlog wait: 3.331249 RTT: 0.118316 total execute time: 3.459563 (milliseconds) >>> CREATE: >>> 100130 ops (2%) 95 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 168 avg bytes received per op: 272 >>> backlog wait: 4.099451 RTT: 0.071437 total execute time: 4.177479 (milliseconds) >>> SYMLINK: >>> 100130 ops (2%) 83 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 188 avg bytes received per op: 271 >>> backlog wait: 3.727704 RTT: 0.073534 total execute time: 3.807700 (milliseconds) >>> RENAME: >>> 100130 ops (2%) 68 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 180 avg bytes received per op: 260 >>> backlog wait: 2.659982 RTT: 0.070518 total execute time: 2.738979 (milliseconds) >>> LINK: >>> 100130 ops (2%) 85 retrans (0%) 0 major timeouts >>> avg bytes sent per op: 172 avg bytes received per op: 232 >>> backlog wait: 3.676680 RTT: 0.066773 total execute time: 3.749785 (milliseconds) >>> GETATTR: >>> 230 ops (0%) 81 retrans (35%) 0 major timeouts >>> avg bytes sent per op: 170 avg bytes received per op: 112 >>> backlog wait: 1584.026087 RTT: 0.043478 total execute time: 1584.082609 (milliseconds) > > By the way, that's a *nasty* tail latency on at least one getattr. > What's up with that?? The average round-trip with the server is 43 microseconds. The long total execution time is due to the backlog wait. Likely that some of the GETATTRs are waiting behind other operations. ftrace capture would confirm it. > Tom. > >>> READDIRPLUS: >>> 10 ops (0%) >>> avg bytes sent per op: 149 avg bytes received per op: 1726 >>> backlog wait: 0.000000 RTT: 0.300000 total execute time: 0.400000 (milliseconds) >>> FSINFO: >>> 2 ops (0%) >>> avg bytes sent per op: 112 avg bytes received per op: 80 >>> backlog wait: 0.000000 RTT: 0.000000 total execute time: 0.000000 (milliseconds) >>> NULL: >>> 1 ops (0%) >>> avg bytes sent per op: 40 avg bytes received per op: 24 >>> backlog wait: 2.000000 RTT: 0.000000 total execute time: 3.000000 (milliseconds) >>> READLINK: >>> 1 ops (0%) >>> avg bytes sent per op: 128 avg bytes received per op: 1140 >>> backlog wait: 0.000000 RTT: 0.000000 total execute time: 0.000000 (milliseconds) >>> PATHCONF: >>> 1 ops (0%) >>> avg bytes sent per op: 112 avg bytes received per op: 56 >>> backlog wait: 0.000000 RTT: 0.000000 total execute time: 0.000000 (milliseconds) >>> >>> >>>> Tom. >>>> >>>>> ACCESS: >>>>> 300001 ops (15%) 44 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 132 avg bytes received per op: 120 >>>>> backlog wait: 0.591118 RTT: 0.017463 total execute time: 0.614795 (milliseconds) >>>>> REMOVE: >>>>> 300000 ops (15%) 40 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 136 avg bytes received per op: 144 >>>>> backlog wait: 0.531667 RTT: 0.018973 total execute time: 0.556927 (milliseconds) >>>>> MKDIR: >>>>> 200000 ops (10%) 26 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 158 avg bytes received per op: 272 >>>>> backlog wait: 0.518940 RTT: 0.019755 total execute time: 0.545230 (milliseconds) >>>>> RMDIR: >>>>> 200000 ops (10%) 24 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 130 avg bytes received per op: 144 >>>>> backlog wait: 0.512320 RTT: 0.018580 total execute time: 0.537095 (milliseconds) >>>>> LOOKUP: >>>>> 188533 ops (9%) 21 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 136 avg bytes received per op: 174 >>>>> backlog wait: 0.455925 RTT: 0.017721 total execute time: 0.480011 (milliseconds) >>>>> SETATTR: >>>>> 100000 ops (5%) 11 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 160 avg bytes received per op: 144 >>>>> backlog wait: 0.371960 RTT: 0.019470 total execute time: 0.398330 (milliseconds) >>>>> WRITE: >>>>> 100000 ops (5%) 9 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 180 avg bytes received per op: 136 >>>>> backlog wait: 0.399190 RTT: 0.022860 total execute time: 0.436610 (milliseconds) >>>>> CREATE: >>>>> 100000 ops (5%) 9 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 168 avg bytes received per op: 272 >>>>> backlog wait: 0.365290 RTT: 0.019560 total execute time: 0.391140 (milliseconds) >>>>> SYMLINK: >>>>> 100000 ops (5%) 18 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 188 avg bytes received per op: 272 >>>>> backlog wait: 0.750470 RTT: 0.020150 total execute time: 0.786410 (milliseconds) >>>>> RENAME: >>>>> 100000 ops (5%) 14 retrans (0%) 0 major timeouts >>>>> avg bytes sent per op: 180 avg bytes received per op: 260 >>>>> backlog wait: 0.461650 RTT: 0.020710 total execute time: 0.489670 (milliseconds) >>>>> -- >>>>> Chuck Lever >>> >>> -- >>> Chuck Lever -- Chuck Lever