Created attachment 170670 [details] GELI: Use a dedicated uma zone for writes to onetime devices The attached patch lets GELI use a dedicated uma zone for writes to onetime devices as they are likely to originate from the vm page daemon. Without the patch the system could deadlock because the vm daemon was waiting for pages to be written to disk, while GELI was waiting for the vm daemon to make room for the buffer GELI needed to actually write the pages: (kgdb) where #0 sched_switch (td=0xfffff800055bf9a0, newtd=0xfffff80002341000, flags=<value optimized out>) at /usr/src/sys/kern/sched_ule.c:1969 #1 0xffffffff80962635 in mi_switch (flags=<value optimized out>, newtd=0x0) at /usr/src/sys/kern/kern_synch.c:455 #2 0xffffffff809aaa3a in sleepq_wait (wchan=0x0, pri=0) at /usr/src/sys/kern/subr_sleepqueue.c:637 #3 0xffffffff80962038 in _sleep (ident=<value optimized out>, lock=<value optimized out>, priority=<value optimized out>, wmesg=0xffffffff80e826ee "vmwait", sbt=0, pr=<value optimized out>, flags=<value optimized out>) at /usr/src/sys/kern/kern_synch.c:229 #4 0xffffffff80c1ac6b in vm_wait () at /usr/src/sys/vm/vm_page.c:2705 #5 0xffffffff80c06a9f in kmem_back (object=0xffffffff8144d6f0, addr=18446741874805047296, size=69632, flags=<value optimized out>) at /usr/src/sys/vm/vm_kern.c:356 #6 0xffffffff80c068d2 in kmem_malloc (vmem=0xffffffff813aa500, size=69632, flags=2) at /usr/src/sys/vm/vm_kern.c:316 #7 0xffffffff80bfd7d6 in uma_large_malloc (size=69632, wait=2) at /usr/src/sys/vm/uma_core.c:1106 #8 0xffffffff8092f614 in malloc (size=<value optimized out>, mtp=0xffffffff81b4d520, flags=0) at /usr/src/sys/kern/kern_malloc.c:513 #9 0xffffffff81b4ab99 in g_eli_crypto_run (wr=0xfffff80002560040, bp=0xfffff80008a86d90) at /usr/src/sys/modules/geom/geom_eli/../../../geom/eli/g_eli_privacy.c:262 #10 0xffffffff81b3e860 in g_eli_worker (arg=0xfffff80002560040) at /usr/src/sys/modules/geom/geom_eli/../../../geom/eli/g_eli.c:565 #11 0xffffffff80910f5c in fork_exit (callout=0xffffffff81b3e0b0 <g_eli_worker>, arg=0xfffff80002560040, frame=0xfffffe005005ec00) at /usr/src/sys/kern/kern_fork.c:1034 #12 0xffffffff80c33f0e in fork_trampoline () at /usr/src/sys/amd64/amd64/exception.S:611 #13 0x0000000000000000 in ?? () (kgdb) p vm_cnt $16 = {v_swtch = 0, v_trap = 0, v_syscall = 0, v_intr = 0, v_soft = 0, v_vm_faults = 0, v_io_faults = 0, v_cow_faults = 0, v_cow_optim = 0, v_zfod = 0, v_ozfod = 0, v_swapin = 0, v_swapout = 0, v_swappgsin = 0, v_swappgsout = 0, v_vnodein = 0, v_vnodeout = 0, v_vnodepgsin = 0, v_vnodepgsout = 0, v_intrans = 0, v_reactivated = 0, v_pdwakeups = 22197, v_pdpages = 0, v_tcached = 0, v_dfree = 0, v_pfree = 0, v_tfree = 0, v_page_size = 4096, v_page_count = 247688, v_free_reserved = 372, v_free_target = 5320, v_free_min = 1609, v_free_count = 2, v_wire_count = 140735, v_active_count = 96194, v_inactive_target = 7980, v_inactive_count = 10756, v_cache_count = 0, v_pageout_free_min = 34, v_interrupt_free_min = 2, v_free_severe = 990, v_forks = 0, v_vforks = 0, v_rforks = 0, v_kthreads = 0, v_forkpages = 0, v_vforkpages = 0, v_rforkpages = 0, v_kthreadpages = 0, v_spare = 0xffffffff8144d5ac} A sysctl is added to optionally use the zone for GELI writes in general, without letting common writes cut into the reserve for onetime writes. This may reduce latency for larger writes and as we need to keep a couple of items in the zone anyway, the impact on the zone size is minor. Initial testing seems to indicate that the sysctl could be safely enabled by default in the future. Currently a single zone with a somewhat humongous item size sufficient for all GELI writes is being used. While this may look a bit wasteful, in practice we don't need a lot of items, so this seem tolerable for now. The best solution would probably be to only use the dedicated uma zone for common writes if the size is above 65356 bytes, the largest zone item size internally used by malloc. Currently the zone isn't used for reads as those are less time critical and usually are small enough for malloc() to succeed right away anyway. Example length distribution when reproducing ElectroBSD with -j4 and 1 GB of RAM: gpt/swap-ada1.eli BIO_WRITE value ------------- Distribution ------------- count < 4096 | 0 4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 4965848 8192 |@@@@@ 943980 12288 |@@ 362668 16384 |@ 161485 20480 |@ 120939 24576 | 87827 28672 | 57402 32768 | 40470 36864 | 42243 40960 | 28543 45056 | 20347 49152 | 15235 53248 | 13450 57344 | 9535 61440 | 9952 65536 |@ 179360 69632 | 0 gpt/swap-ada1.eli BIO_READ value ------------- Distribution ------------- count < 4096 | 0 4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 4645114 8192 | 0 12288 | 0 16384 | 3446 20480 | 0 Note that the GELI overhead is not accounted for here and only the results for the swap device are shown. Zone use: [fk@elektrobier3 ~]$ vmstat -z | egrep 'ITEM|eli' | column -t ITEM SIZE LIMIT USED FREE REQ FAIL SLEEP g_eli: 172032, 0, 0, 14, 8077487, 0, 0 This includes writes to gpt/dpool-ada1.eli and gpt/dpool-ada1.eli. Discussion: While the zone served 8077487 memory requests total, 14 items were sufficient for this and therefore the zone only withheld 172032 * 14 bytes plus zone meta data from the rest of the system. Obtained from: ElectroBSD
Created attachment 182268 [details] GELI: Use a dedicated uma zone for unauthenticated writes to onetime devices This updated patch set falls back to using malloc if the the zone item size is too small or if authentication is enabled. The zone item size can be modified more conveniently through the kern.geom.eli.max_uma_bio_length tunable. The second patch in the set works around #218911 but is only relevant if the zone item size is modified. The previous version of the patch would cause crashes when setting kern.geom.eli.use_uma_for_all_writes=1 and using the device for a ZFS pool with recordsize=1M without adjusting the zone item size to fit. Authenticated writes (even when made to onetime devices) could result in crashes as well.