This patch updates the initialization place for the AER queue, so it’s initialized once, at controller initialization, and not every time controller is enabled.
While the original version works for a non-SR-IOV device, as it’s hard to interact with the controller if it’s not enabled, the multiple reinitialization is not necessarily correct. With the SR/IOV feature enabled a segfault can happen: a VF can have its controller disabled, while a namespace can still be attached to the controller through the parent PF. An event generated in such case ends up on an uninitialized queue. While it’s an interesting question whether a VF should support AER in the first place, I don’t think it must be answered today. Signed-off-by: Łukasz Gieryk <lukasz.gie...@linux.intel.com> --- hw/nvme/ctrl.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/hw/nvme/ctrl.c b/hw/nvme/ctrl.c index 9d0432a2e5..7d41318961 100644 --- a/hw/nvme/ctrl.c +++ b/hw/nvme/ctrl.c @@ -5980,8 +5980,6 @@ static int nvme_start_ctrl(NvmeCtrl *n) nvme_set_timestamp(n, 0ULL); - QTAILQ_INIT(&n->aer_queue); - nvme_select_iocs(n); return 0; @@ -6960,6 +6958,8 @@ static void nvme_init_ctrl(NvmeCtrl *n, PCIDevice *pci_dev) id->cmic |= NVME_CMIC_MULTI_CTRL; } + QTAILQ_INIT(&n->aer_queue); + NVME_CAP_SET_MQES(cap, 0x7ff); NVME_CAP_SET_CQR(cap, 1); NVME_CAP_SET_TO(cap, 0xf); -- 2.25.1