From: Junxian Huang huangjunxian6@hisilicon.com
driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I9MANQ
----------------------------------------------------------------------
CEQEs are handled in interrupt handler currently. This may cause the CPU core staying in interrupt context too long and lead to soft lockup under heavy load.
Handle CEQEs in tasklet and set an upper limit for the number of CEQE handled by a single call of tasklet.
Fixes: a5073d6054f7 ("RDMA/hns: Add eq support of hip08") Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com --- drivers/infiniband/hw/hns/hns_roce_device.h | 1 + drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 80 ++++++++++++--------- 2 files changed, 48 insertions(+), 33 deletions(-)
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index 958ed5bc5..bdc62dcf4 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -802,6 +802,7 @@ struct hns_roce_eq { int shift; int event_type; int sub_type; + struct tasklet_struct tasklet; };
struct hns_roce_eq_table { diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index e6ecfe013..bd49ffa4e 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -6467,33 +6467,11 @@ static struct hns_roce_ceqe *next_ceqe_sw_v2(struct hns_roce_eq *eq) !!(eq->cons_index & eq->entries)) ? ceqe : NULL; }
-static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_dev *hr_dev, - struct hns_roce_eq *eq) +static irqreturn_t hns_roce_v2_ceq_int(struct hns_roce_eq *eq) { - struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq); - irqreturn_t ceqe_found = IRQ_NONE; - u32 cqn; - - while (ceqe) { - /* Make sure we read CEQ entry after we have checked the - * ownership bit - */ - dma_rmb(); - - cqn = hr_reg_read(ceqe, CEQE_CQN); + tasklet_schedule(&eq->tasklet);
- hns_roce_cq_completion(hr_dev, cqn); - - ++eq->cons_index; - ceqe_found = IRQ_HANDLED; - atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]); - - ceqe = next_ceqe_sw_v2(eq); - } - - update_eq_db(eq); - - return IRQ_RETVAL(ceqe_found); + return IRQ_HANDLED; }
static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr) @@ -6504,7 +6482,7 @@ static irqreturn_t hns_roce_v2_msix_interrupt_eq(int irq, void *eq_ptr)
if (eq->type_flag == HNS_ROCE_CEQ) /* Completion event interrupt */ - int_work = hns_roce_v2_ceq_int(hr_dev, eq); + int_work = hns_roce_v2_ceq_int(eq); else /* Asynchronous event interrupt */ int_work = hns_roce_v2_aeq_int(hr_dev, eq); @@ -6872,6 +6850,34 @@ free_cmd_mbox: return ret; }
+static void hns_roce_ceq_task(struct tasklet_struct *task) +{ + struct hns_roce_eq *eq = from_tasklet(eq, task, tasklet); + struct hns_roce_ceqe *ceqe = next_ceqe_sw_v2(eq); + struct hns_roce_dev *hr_dev = eq->hr_dev; + int ceqe_num = 0; + u32 cqn; + + while (ceqe && ceqe_num < hr_dev->caps.ceqe_depth) { + /* Make sure we read CEQ entry after we have checked the + * ownership bit + */ + dma_rmb(); + + cqn = hr_reg_read(ceqe, CEQE_CQN); + + hns_roce_cq_completion(hr_dev, cqn); + + ++eq->cons_index; + ++ceqe_num; + atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_CEQE_CNT]); + + ceqe = next_ceqe_sw_v2(eq); + } + + update_eq_db(eq); +} + static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num, int comp_num, int aeq_num, int other_num) { @@ -6903,21 +6909,24 @@ static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num, j - other_num - aeq_num);
for (j = 0; j < irq_num; j++) { - if (j < other_num) + if (j < other_num) { ret = request_irq(hr_dev->irq[j], hns_roce_v2_msix_interrupt_abn, 0, hr_dev->irq_names[j], hr_dev); - - else if (j < (other_num + comp_num)) + } else if (j < (other_num + comp_num)) { + tasklet_setup(&eq_table->eq[j - other_num].tasklet, + hns_roce_ceq_task); ret = request_irq(eq_table->eq[j - other_num].irq, hns_roce_v2_msix_interrupt_eq, 0, hr_dev->irq_names[j + aeq_num], &eq_table->eq[j - other_num]); - else + } else { ret = request_irq(eq_table->eq[j - other_num].irq, hns_roce_v2_msix_interrupt_eq, 0, hr_dev->irq_names[j - comp_num], &eq_table->eq[j - other_num]); + } + if (ret) { dev_err(hr_dev->dev, "request irq error!\n"); goto err_request_failed; @@ -6928,11 +6937,13 @@ static int __hns_roce_request_irq(struct hns_roce_dev *hr_dev, int irq_num,
err_request_failed: for (j -= 1; j >= 0; j--) - if (j < other_num) + if (j < other_num) { free_irq(hr_dev->irq[j], hr_dev); - else + } else { free_irq(eq_table->eq[j - other_num].irq, &eq_table->eq[j - other_num]); + tasklet_kill(&eq_table->eq[j - other_num].tasklet); + }
err_kzalloc_failed: for (i -= 1; i >= 0; i--) @@ -6953,8 +6964,11 @@ static void __hns_roce_free_irq(struct hns_roce_dev *hr_dev) for (i = 0; i < hr_dev->caps.num_other_vectors; i++) free_irq(hr_dev->irq[i], hr_dev);
- for (i = 0; i < eq_num; i++) + for (i = 0; i < eq_num; i++) { free_irq(hr_dev->eq_table.eq[i].irq, &hr_dev->eq_table.eq[i]); + if (i < hr_dev->caps.num_comp_vectors) + tasklet_kill(&hr_dev->eq_table.eq[i].tasklet); + }
for (i = 0; i < irq_num; i++) kfree(hr_dev->irq_names[i]);