Kernel OOPS while creating a NVMe Namespace

[Keith Busch]

On Mon, Jun 10, 2024 at 12:57:02PM +0300, Sagi Grimberg wrote:
> On 10/06/2024 10:51, Venkat Rao Bagalkote wrote:
> > 
> > [  140.209777] BUG: Unable to handle kernel data access at 0x18d7003065646fee
> > [  140.209792] Faulting instruction address: 0xc00000000023b45c
> > [  140.209798] Oops: Kernel access of bad area, sig: 11 [#1]
> > [  140.209802] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=8192 NUMA pSeries
> > [  140.209864] CPU: 2 PID: 129 Comm: kworker/u65:3 Kdump: loaded Not tainted 6.10.0-rc3 #2
> > [  140.209870] Hardware name: IBM,9009-42A POWER9 (raw) 0x4e0202 0xf000005 of:IBM,FW950.A0 (VL950_141) hv:phyp pSeries
> > [  140.209876] Workqueue: nvme-wq nvme_scan_work [nvme_core]
> > [  140.209889] NIP:  c00000000023b45c LR: c008000006a96b20 CTR: c00000000023b42c
> > [  140.209894] REGS: c0000000506078a0 TRAP: 0380   Not tainted (6.10.0-rc3)
> > [  140.209899] MSR:  800000000280b033 <SF,VEC,VSX,EE,FP,ME,IR,DR,RI,LE>  CR: 24000244  XER: 00000000
> > [  140.209975] NIP [c00000000023b45c] synchronize_srcu+0x30/0x1c0
> > [  140.209984] LR [c008000006a96b20] nvme_ns_remove+0x80/0x2d8 [nvme_core]
> > [  140.209994] Call Trace:
> > [  140.209997] [c000000050607b90] [c008000006a96b20] nvme_ns_remove+0x80/0x2d8 [nvme_core]
> > [  140.210008] [c000000050607bd0] [c008000006a972b4] nvme_remove_invalid_namespaces+0x144/0x1ac [nvme_core]
> > [  140.210020] [c000000050607c60] [c008000006a9dbd4] nvme_scan_ns_list+0x19c/0x370 [nvme_core]
> > [  140.210032] [c000000050607d70] [c008000006a9dfc8] nvme_scan_work+0xc8/0x278 [nvme_core]
> > [  140.210043] [c000000050607e40] [c00000000019414c] process_one_work+0x20c/0x4f4
> > [  140.210051] [c000000050607ef0] [c0000000001950cc] worker_thread+0x378/0x544
> > [  140.210058] [c000000050607f90] [c0000000001a164c] kthread+0x138/0x140
> > [  140.210065] [c000000050607fe0] [c00000000000df98] start_kernel_thread+0x14/0x18
> > [  140.210072] Code: 3c4c0134 384282d4 7c0802a6 60000000 7c0802a6 fbc1fff0 fba1ffe8 fbe1fff8 7c7e1b78 f8010010 f821ffb1 e9230010 e9290080> 7c2004ac 71290003 41820008
> > [  140.210093] ---[ end trace 0000000000000000 ]---
> > 
> > Issue is introduced by the patch:
> > be647e2c76b27f409cdd520f66c95be888b553a3.
> 
> Exactly this was the concern when introducing a behavior change in a
> sensitive area of the code
> to silence lockdep...

No risk, no reward. :)

If we really can't figure this out, we can always revert and revisit for
the next merge.

> I'm assuming that the bad dereference is:
>         synchronize_srcu(&ns->ctrl->srcu);

That would have to be it based on the report. Not sure what the problem
could be, though, the ns->ctrl must have been valid or we would have
failed earlier, and the srcu struct can't be released while the
controller is still in use by any namespace.

Anyway, I tested this path quite a bit, but I'll revisit with dynamic
attachments instead and see if that helps.

> btw, looking at the code again, I'm assuming that synchronizing srcu in
> every ns remove
> slows down batch removal of many namespaces greatly...

I may need to test this out, but I thought srcu sync was pretty quick if
there were no active readers, which there shouldn't be here outside
unusual cases.