There seems to be a race between creating a zvol with volmode=dev. We create the zvol with the default mode, then destroy it and re-create it with dev mode. It seems that if I/O occurs in that window, it takes the geom code path and walks through a NULL pointer. This seems to be introduced by the OpenZFS merge. I've used this workflow often for over a year, mostly on head, and this appeared only recently. I first saw it on r366500+84ccaf49083c-c272054. I hit it reliably on my workstation with the following command. I also hit it on a VM, though it takes more tries to hit the window. # zfs create -s -V 20G -o primarycache=none -o volmode=dev head_root/testvol zvol_create_minor_impl:1250[1]: Creating ZVOL head_root/testvol... zvol_create_minor_impl:1371[1]: ZVOL head_root/testvol created. Fatal trap 12: page fault while in kernel mode cpuid = 7; apic id = 07 fault virtual address = 0x110 fault code = supervisor read data, page not present instruction pointer = 0x20:0xffffffff82167fca stack pointer = 0x28:0xfffffe000edcdb30 frame pointer = 0x28:0xfffffe000edcdb70 code segment = base rx0, limit 0xfffff, type 0x1b = DPL 0, pres 1, long 1, def32 0, gran 1 processor eflags = interrupt enabled, resume, IOPL = 0 current process = 13 (g_down) trap number = 12 db> acttrace Tracing command zfskern pid 21 tid 100478 td 0xfffffe00610c9800 (CPU 6) cpustop_handler() at cpustop_handler+0x28/frame 0xfffffe0011880e00 ipi_nmi_handler() at ipi_nmi_handler+0x39/frame 0xfffffe0011880e10 trap() at trap+0x56/frame 0xfffffe0011880f20 nmi_calltrap() at nmi_calltrap+0x8/frame 0xfffffe0011880f20 --- trap 0x13, rip = 0xffffffff80c25fb2, rsp = 0xfffffe006168c820, rbp = 0xfffffe006168c830 --- lock_delay() at lock_delay+0x42/frame 0xfffffe006168c830 _mtx_lock_spin_cookie() at _mtx_lock_spin_cookie+0xc1/frame 0xfffffe006168c8a0 __mtx_lock_spin_flags() at __mtx_lock_spin_flags+0xd5/frame 0xfffffe006168c8e0 cnputs() at cnputs+0x58/frame 0xfffffe006168c910 vprintf() at vprintf+0xcd/frame 0xfffffe006168c9e0 printf() at printf+0x43/frame 0xfffffe006168ca40 zvol_free() at zvol_free+0x53/frame 0xfffffe006168ca80 zvol_task_cb() at zvol_task_cb+0x271/frame 0xfffffe006168cae0 taskq_run() at taskq_run+0x1f/frame 0xfffffe006168cb00 taskqueue_run_locked() at taskqueue_run_locked+0xaa/frame 0xfffffe006168cb80 taskqueue_thread_loop() at taskqueue_thread_loop+0x94/frame 0xfffffe006168cbb0 fork_exit() at fork_exit+0x80/frame 0xfffffe006168cbf0 fork_trampoline() at fork_trampoline+0xe/frame 0xfffffe006168cbf0 --- trap 0, rip = 0, rsp = 0, rbp = 0 --- Tracing command geom pid 13 tid 100049 td 0xfffffe0011862700 (CPU 7) kdb_enter() at kdb_enter+0x37/frame 0xfffffe000edcd7e0 vpanic() at vpanic+0x19e/frame 0xfffffe000edcd830 panic() at panic+0x43/frame 0xfffffe000edcd890 trap_fatal() at trap_fatal+0x387/frame 0xfffffe000edcd8f0 trap_pfault() at trap_pfault+0x97/frame 0xfffffe000edcd950 trap() at trap+0x2ab/frame 0xfffffe000edcda60 calltrap() at calltrap+0x8/frame 0xfffffe000edcda60 --- trap 0xc, rip = 0xffffffff82167fca, rsp = 0xfffffe000edcdb30, rbp = 0xfffffe000edcdb70 --- zvol_geom_bio_start() at zvol_geom_bio_start+0x2a/frame 0xfffffe000edcdb70 g_io_schedule_down() at g_io_schedule_down+0x134/frame 0xfffffe000edcdba0 g_down_procbody() at g_down_procbody+0x5c/frame 0xfffffe000edcdbb0 fork_exit() at fork_exit+0x80/frame 0xfffffe000edcdbf0 fork_trampoline() at fork_trampoline+0xe/frame 0xfffffe000edcdbf0 --- trap 0, rip = 0, rsp = 0, rbp = 0 --- (The other CPUs were idle.)
FWIW on current head I don't see this behavior. What I see instead is that we create the zvol with default mode, but don't destroy and recreate it. So it ends up with the wrong mode. Only if I subsequently do "zfs set volmode=dev <volname>" does the zvol get the correct mode. I opened bug 251828 for that.
I still see this behavior on main commit f7cd7fe51c4 (25 December). Remember, it's a race, so if you get lucky, you won't see it.
^Triage: to submitter: is this aging PR still relevant?