Hi,
在 2023/3/27 22:58, Mathias Nyman 写道:
On 27.3.2023 4.11, Hongyu Xie wrote:
It was first found on HUAWEI Kirin 9006C platform with a builtin xhci controller during stress cycle test(stress-ng, glmark2, x11perf, S4...).
phase one: [26788.706878] PM: dpm_run_callback(): platform_pm_thaw+0x0/0x68 returns -12 [26788.706878] PM: Device xhci-hcd.1.auto failed to thaw async: error -12 ... phase two: [28650.583496] [2023:01:19 04:43:29]Unable to handle kernel NULL pointer dereference at virtual address 0000000000000028 ... [28650.583526] user pgtable: 4k pages, 39-bit VAs, pgdp=000000027862a000 [28650.583557] [0000000000000028] pgd=0000000000000000 ... [28650.583587] pc : xhci_suspend+0x154/0x5b0 [28650.583618] lr : xhci_suspend+0x148/0x5b0 [28650.583618] sp : ffffffc01c7ebbd0 [28650.583618] x29: ffffffc01c7ebbd0 x28: ffffffec834d0000 [28650.583618] x27: ffffffc0106a3cc8 x26: ffffffb2c540c848 [28650.583618] x25: 0000000000000000 x24: ffffffec82ee30b0 [28650.583618] x23: ffffffb43b31c2f8 x22: 0000000000000000 [28650.583618] x21: 0000000000000000 x20: ffffffb43b31c000 [28650.583648] x19: ffffffb43b31c2a8 x18: 0000000000000001 [28650.583648] x17: 0000000000000803 x16: 00000000fffffffe [28650.583648] x15: 0000000000001000 x14: ffffffb150b67e00 [28650.583648] x13: 00000000f0000000 x12: 0000000000000001 [28650.583648] x11: 0000000000000000 x10: 0000000000000a80 [28650.583648] x9 : ffffffc01c7eba00 x8 : ffffffb43ad10ae0 [28650.583648] x7 : ffffffb84cd98dc0 x6 : 0000000cceb6a101 [28650.583679] x5 : 00ffffffffffffff x4 : 0000000000000001 [28650.583679] x3 : 0000000000000011 x2 : 0000000000e2cfa8 [28650.583679] x1 : 00000000823535e1 x0 : 0000000000000000
gdb: (gdb) l *(xhci_suspend+0x154) 0xffffffc010b6cd44 is in xhci_suspend (/.../drivers/usb/host/xhci.c:854). 849 { 850 struct xhci_ring *ring; 851 struct xhci_segment *seg; 852 853 ring = xhci->cmd_ring; 854 seg = ring->deq_seg; (gdb) disassemble 0xffffffc010b6cd44 ... 0xffffffc010b6cd40 <+336>: ldr x22, [x19, #160] 0xffffffc010b6cd44 <+340>: ldr x20, [x22, #40] 0xffffffc010b6cd48 <+344>: mov w1, #0x0 // #0
During phase one, platform_pm_thaw called xhci_plat_resume which called xhci_resume. The rest possible calling routine might be xhci_resume->xhci_init->xhci_mem_init, and xhci->cmd_ring was cleaned in xhci_mem_cleanup before xhci_mem_init returned -ENOMEM.
During phase two, systemd was tring to hibernate again and called xhci_suspend, then xhci_clear_command_ring dereferenced xhci->cmd_ring which was already NULL.
Any comments on the questions I had on the first version of the patch?
Sorry, didn't notice your reply in the first version.
xhci_mem_init() failing with -ENOMEM looks like the real problem here.
Are we really running out of memory? does kmemleak say anything?
It looks like running out of memory, since it was running a stress test. But can't go any further without more details. Didn't run with kmemleak open.
Any chance you could look into where exactly xhci_mem_init() fails as xhci_mem_init() always returns -ENOMEM on failure?
Can't reproduce the problem for a very long time. Still don't know where did it fail in xhci_mem_init. But I think you can't blame xhci driver for memory shortage, and you can't fix that.
So if xhci->cmd_ring is NULL, xhci_clear_command_ring just return.
This hides the problem more than solves it. Root cause is still unknown
You were saying "If xhci_mem_init() failed then...it shouldn't be...", and I agree with it. Further more, I think functions that calling xhci_mem_init needs to check xhci_mem_init's return value, but it needs another patch to do this. This patch is saying that xhci_clear_command_ring should check a pointer before using it, because somewhere else might clear cmd_ring, that's all.
Thanks Mathias
Thanks
Hongyu Xie