From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I8HYSU

--------------------------------------------------------------------------

hr_dev->caps.congest_type stands for the supported algorithm set, but not the specific algorithm type. The correct configuration item should be ilog2(hr_qp->congest_type).

Fixes: 09f1b7cb29b2 ("RDMA/hns: Support congestion control algorithm configuration at QP granularity") Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com --- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index bd2d8e447c19..1cfcefb28d7e 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -5171,6 +5171,7 @@ static int fill_congest_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr, struct hns_roce_congestion_algorithm congest_field; struct ib_device *ibdev = ibqp->device; struct hns_roce_dev *hr_dev = to_hr_dev(ibdev); + struct hns_roce_qp *hr_qp = to_hr_qp(ibqp); u32 dip_idx = 0; int ret;

@@ -5184,7 +5185,7 @@ static int fill_congest_field(struct ib_qp *ibqp, const struct ib_qp_attr *attr,

hr_reg_write(context, QPC_CONGEST_ALGO_TMPL_ID, hr_dev->congest_algo_tmpl_id + - hr_dev->caps.congest_type * HNS_ROCE_CONGEST_SIZE); + ilog2(hr_qp->congest_type) * HNS_ROCE_CONGEST_SIZE); hr_reg_clear(qpc_mask, QPC_CONGEST_ALGO_TMPL_ID); hr_reg_write(&context->ext, QPCEX_CONGEST_ALG_SEL, congest_field.alg_sel);

From: Luoyouming luoyouming@huawei.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I87LV8

--------------------------------------------------------------------------

Due to hardware limitations, UD mode can only configure DCQCN. Therefore, the default configuration in this mode is set to DCQCN, and additional validation is applied to user configuration.

Fixes: 09f1b7cb29b2 ("RDMA/hns: Support congestion control algorithm configuration at QP granularity") Signed-off-by: Luoyouming luoyouming@huawei.com --- drivers/infiniband/hw/hns/hns_roce_qp.c | 34 +++++++++++++++++-------- 1 file changed, 24 insertions(+), 10 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c index e8e5641ec692..52ee28181ae4 100644 --- a/drivers/infiniband/hw/hns/hns_roce_qp.c +++ b/drivers/infiniband/hw/hns/hns_roce_qp.c @@ -1095,7 +1095,10 @@ static inline void default_congest_type(struct hns_roce_dev *hr_dev, { struct hns_roce_caps *caps = &hr_dev->caps;

- hr_qp->congest_type = 1 << caps->default_congest_type; + if (hr_qp->ibqp.qp_type == IB_QPT_UD) + hr_qp->congest_type = HNS_ROCE_CREATE_QP_FLAGS_DCQCN; + else + hr_qp->congest_type = 1 << caps->default_congest_type; }

static int set_congest_type(struct hns_roce_qp *hr_qp, @@ -1103,6 +1106,17 @@ static int set_congest_type(struct hns_roce_qp *hr_qp, { int ret = 0;

+ if (hr_qp->ibqp.qp_type == IB_QPT_UD && + !(ucmd->congest_type_flags & HNS_ROCE_CREATE_QP_FLAGS_DCQCN)) { + struct hns_roce_dev *hr_dev = to_hr_dev(hr_qp->ibqp.device); + + ibdev_err_ratelimited(&hr_dev->ib_dev, + "UD just support DCQCN. unsupported congest type 0x%llx.\n", + ucmd->congest_type_flags); + + return -EINVAL; + } + if (ucmd->congest_type_flags & HNS_ROCE_CREATE_QP_FLAGS_DCQCN) hr_qp->congest_type = HNS_ROCE_CREATE_QP_FLAGS_DCQCN; else if (ucmd->congest_type_flags & HNS_ROCE_CREATE_QP_FLAGS_LDCP) @@ -1117,19 +1131,16 @@ static int set_congest_type(struct hns_roce_qp *hr_qp, return ret; }

-static void set_congest_param(struct hns_roce_dev *hr_dev, +static int set_congest_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, struct hns_roce_ib_create_qp *ucmd) { - int ret; - - if (ucmd->comp_mask & HNS_ROCE_CREATE_QP_MASK_CONGEST_TYPE) { - ret = set_congest_type(hr_qp, ucmd); - if (ret == 0) - return; - } + if (ucmd->comp_mask & HNS_ROCE_CREATE_QP_MASK_CONGEST_TYPE) + return set_congest_type(hr_qp, ucmd);

default_congest_type(hr_dev, hr_qp); + + return 0; }

static void set_qp_notify_param(struct hns_roce_qp *hr_qp, @@ -1234,7 +1245,10 @@ static int set_qp_param(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, ret = set_uqp_create_flag_param(hr_dev, hr_qp, init_attr, ucmd); if (ret) return ret; - set_congest_param(hr_dev, hr_qp, ucmd); + + ret = set_congest_param(hr_dev, hr_qp, ucmd); + if (ret) + return ret; } else { if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK) {

From: wenglianfa wenglianfa@huawei.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I87LFL

--------------------------------------------------------------------------

In the current solution, the pseudo WC enables the user-mode to detect the device error in advance and releases context resources. As a result, there is a high probability that hardware reset and context resource release occur at the same time. During the hardware reset, the MBOX cannot instruct the hardware to stop accessing the memory, but the corresponding resources are released during the reset. The hardware is unaware that the driver has freed resources. Therefore, the remaining tasks of the hardware access invalid memory, and the RAS alarm is reported.

If the driver detects above scenario, the driver will not release the resources.Instead, record it in a linked list. Wait for the roce driver to uninstall before releasing it. In this way, the hardware does not access the invalid memory, and the driver does not cause memory leakage.

Fixes: 306b8c76257b ("RDMA/hns: Do not destroy QP resources in the hw resetting phase") Signed-off-by: wenglianfa wenglianfa@huawei.com --- drivers/infiniband/hw/hns/hns_roce_cq.c | 26 +++++- drivers/infiniband/hw/hns/hns_roce_db.c | 30 +++++-- drivers/infiniband/hw/hns/hns_roce_device.h | 34 +++++++- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 3 + drivers/infiniband/hw/hns/hns_roce_main.c | 8 ++ drivers/infiniband/hw/hns/hns_roce_mr.c | 92 ++++++++++++++++++++- drivers/infiniband/hw/hns/hns_roce_qp.c | 27 +++++- drivers/infiniband/hw/hns/hns_roce_srq.c | 39 +++++++-- 8 files changed, 235 insertions(+), 24 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_cq.c b/drivers/infiniband/hw/hns/hns_roce_cq.c index cd058acfa10f..3110b4fd00ed 100644 --- a/drivers/infiniband/hw/hns/hns_roce_cq.c +++ b/drivers/infiniband/hw/hns/hns_roce_cq.c @@ -182,6 +182,8 @@ static void free_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) if (ret) dev_err(dev, "DESTROY_CQ failed (%d) for CQN %06lx\n", ret, hr_cq->cqn); + if (ret == -EBUSY) + hr_cq->delayed_destroy_flag = true;

xa_erase(&cq_table->array, hr_cq->cqn);

@@ -193,7 +195,11 @@ static void free_cqc(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) complete(&hr_cq->free); wait_for_completion(&hr_cq->free);

- hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); + /* this resource will be freed when the driver is uninstalled, so + * no memory leak will occur. + */ + if (!hr_cq->delayed_destroy_flag) + hns_roce_table_put(hr_dev, &cq_table->table, hr_cq->cqn); }

static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, @@ -203,6 +209,10 @@ static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, struct hns_roce_buf_attr buf_attr = {}; int ret;

+ hr_cq->mtr_node = kvmalloc(sizeof(*hr_cq->mtr_node), GFP_KERNEL); + if (!hr_cq->mtr_node) + return -ENOMEM; + buf_attr.page_shift = hr_dev->caps.cqe_buf_pg_sz + PAGE_SHIFT; buf_attr.region[0].size = hr_cq->cq_depth * hr_cq->cqe_size; buf_attr.region[0].hopnum = hr_dev->caps.cqe_hop_num; @@ -211,15 +221,22 @@ static int alloc_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, ret = hns_roce_mtr_create(hr_dev, &hr_cq->mtr, &buf_attr, hr_dev->caps.cqe_ba_pg_sz + PAGE_SHIFT, udata, addr); - if (ret) + if (ret) { ibdev_err(ibdev, "Failed to alloc CQ mtr, ret = %d\n", ret); + kvfree(hr_cq->mtr_node); + }

return ret; }

static void free_cq_buf(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq) { - hns_roce_mtr_destroy(hr_dev, &hr_cq->mtr); + if (hr_cq->delayed_destroy_flag) { + hns_roce_add_unfree_mtr(hr_cq->mtr_node, hr_dev, &hr_cq->mtr); + } else { + hns_roce_mtr_destroy(hr_dev, &hr_cq->mtr); + kvfree(hr_cq->mtr_node); + } }

static int alloc_cq_db(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, @@ -270,7 +287,8 @@ static void free_cq_db(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq, uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext, ibucontext); - hns_roce_db_unmap_user(uctx, &hr_cq->db); + hns_roce_db_unmap_user(uctx, &hr_cq->db, + hr_cq->delayed_destroy_flag); } else { hns_roce_free_db(hr_dev, &hr_cq->db); } diff --git a/drivers/infiniband/hw/hns/hns_roce_db.c b/drivers/infiniband/hw/hns/hns_roce_db.c index 0f5fe892e897..78121524a794 100644 --- a/drivers/infiniband/hw/hns/hns_roce_db.c +++ b/drivers/infiniband/hw/hns/hns_roce_db.c @@ -24,7 +24,7 @@ int hns_roce_db_map_user(struct hns_roce_ucontext *context, unsigned long virt, page = kmalloc(sizeof(*page), GFP_KERNEL); if (!page) { ret = -ENOMEM; - goto out; + goto err_out; }

refcount_set(&page->refcount, 1); @@ -33,8 +33,12 @@ int hns_roce_db_map_user(struct hns_roce_ucontext *context, unsigned long virt, PAGE_SIZE, 0); if (IS_ERR(page->umem)) { ret = PTR_ERR(page->umem); - kfree(page); - goto out; + goto err_page; + } + page->umem_node = kvmalloc(sizeof(*page->umem_node), GFP_KERNEL); + if (!page->umem_node) { + ret = -ENOMEM; + goto err_umem; }

list_add(&page->list, &context->page_list); @@ -45,22 +49,36 @@ int hns_roce_db_map_user(struct hns_roce_ucontext *context, unsigned long virt, db->virt_addr = sg_virt(page->umem->sg_head.sgl) + offset; db->u.user_page = page; refcount_inc(&page->refcount); + mutex_unlock(&context->page_mutex); + return 0;

-out: +err_umem: + ib_umem_release(page->umem); +err_page: + kvfree(page); +err_out: mutex_unlock(&context->page_mutex);

return ret; }

void hns_roce_db_unmap_user(struct hns_roce_ucontext *context, - struct hns_roce_db *db) + struct hns_roce_db *db, + bool delayed_unmap_flag) { + struct hns_roce_dev *hr_dev = to_hr_dev(context->ibucontext.device); + mutex_lock(&context->page_mutex);

refcount_dec(&db->u.user_page->refcount); if (refcount_dec_if_one(&db->u.user_page->refcount)) { list_del(&db->u.user_page->list); - ib_umem_release(db->u.user_page->umem); + if (delayed_unmap_flag) { + hns_roce_add_unfree_umem(db->u.user_page, hr_dev); + } else { + ib_umem_release(db->u.user_page->umem); + kvfree(db->u.user_page->umem_node); + } kfree(db->u.user_page); }

diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index 743453accfd7..a9ff1eb81506 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -409,6 +409,8 @@ struct hns_roce_mr { struct hns_roce_mtr pbl_mtr; u32 npages; dma_addr_t *page_list; + bool delayed_destroy_flag; + struct hns_roce_mtr_node *mtr_node; };

struct hns_roce_mr_table { @@ -475,11 +477,17 @@ struct hns_roce_db_pgdir { dma_addr_t db_dma; };

+struct hns_roce_umem_node { + struct ib_umem *umem; + struct list_head list; +}; + struct hns_roce_user_db_page { struct list_head list; struct ib_umem *umem; unsigned long user_virt; refcount_t refcount; + struct hns_roce_umem_node *umem_node; };

struct hns_roce_db { @@ -531,7 +539,9 @@ struct hns_roce_cq { int is_armed; /* cq is armed */ struct list_head node; /* all armed cqs are on a list */ u8 poe_channel; + bool delayed_destroy_flag; struct hns_roce_notify_conf write_notify; + struct hns_roce_mtr_node *mtr_node; };

struct hns_roce_idx_que { @@ -540,6 +550,7 @@ struct hns_roce_idx_que { unsigned long *bitmap; u32 head; u32 tail; + struct hns_roce_mtr_node *mtr_node; };

struct hns_roce_srq { @@ -565,6 +576,8 @@ struct hns_roce_srq { void (*event)(struct hns_roce_srq *srq, enum hns_roce_event event); struct hns_roce_db rdb; u32 cap_flags; + bool delayed_destroy_flag; + struct hns_roce_mtr_node *mtr_node; };

struct hns_roce_uar_table { @@ -740,6 +753,8 @@ struct hns_roce_qp { u32 config; u8 tc_mode; u8 priority; + bool delayed_destroy_flag; + struct hns_roce_mtr_node *mtr_node; };

struct hns_roce_ib_iboe { @@ -1100,6 +1115,11 @@ struct hns_roce_port { struct hns_roce_scc_param *scc_param; };

+struct hns_roce_mtr_node { + struct hns_roce_mtr mtr; + struct list_head list; +}; + struct hns_roce_dev { struct ib_device ib_dev; struct pci_dev *pci_dev; @@ -1183,6 +1203,10 @@ struct hns_roce_dev {

struct rdma_notify_mem *notify_tbl; size_t notify_num; + struct list_head mtr_unfree_list; /* list of unfree mtr on this dev */ + spinlock_t mtr_unfree_list_lock; /* protect mtr_unfree_list */ + struct list_head umem_unfree_list; /* list of unfree umem on this dev */ + spinlock_t umem_unfree_list_lock; /* protect umem_unfree_list */ };

static inline struct hns_roce_dev *to_hr_dev(struct ib_device *ib_dev) @@ -1464,7 +1488,8 @@ int hns_roce_destroy_cq(struct ib_cq *ib_cq, struct ib_udata *udata); int hns_roce_db_map_user(struct hns_roce_ucontext *context, unsigned long virt, struct hns_roce_db *db); void hns_roce_db_unmap_user(struct hns_roce_ucontext *context, - struct hns_roce_db *db); + struct hns_roce_db *db, + bool delayed_unmap_flag); int hns_roce_alloc_db(struct hns_roce_dev *hr_dev, struct hns_roce_db *db, int order); void hns_roce_free_db(struct hns_roce_dev *hr_dev, struct hns_roce_db *db); @@ -1484,6 +1509,13 @@ int hns_roce_fill_res_qp_entry(struct sk_buff *msg, struct ib_qp *ib_qp); int hns_roce_fill_res_qp_entry_raw(struct sk_buff *msg, struct ib_qp *ib_qp); int hns_roce_fill_res_mr_entry(struct sk_buff *msg, struct ib_mr *ib_mr); int hns_roce_fill_res_mr_entry_raw(struct sk_buff *msg, struct ib_mr *ib_mr); +void hns_roce_add_unfree_umem(struct hns_roce_user_db_page *user_page, + struct hns_roce_dev *hr_dev); +void hns_roce_free_unfree_umem(struct hns_roce_dev *hr_dev); +void hns_roce_add_unfree_mtr(struct hns_roce_mtr_node *pos, + struct hns_roce_dev *hr_dev, + struct hns_roce_mtr *mtr); +void hns_roce_free_unfree_mtr(struct hns_roce_dev *hr_dev); struct hns_user_mmap_entry * hns_roce_user_mmap_entry_insert(struct ib_ucontext *ucontext, u64 address, size_t length, diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index c84d44707a35..447e12d87396 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -6038,6 +6038,9 @@ int hns_roce_v2_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata) "failed to destroy QP, QPN = 0x%06lx, ret = %d.\n", hr_qp->qpn, ret);

+ if (ret == -EBUSY) + hr_qp->delayed_destroy_flag = true; + hns_roce_qp_destroy(hr_dev, hr_qp, udata);

return 0; diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index 861c63449519..4eb18fb31726 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -1309,6 +1309,12 @@ static int hns_roce_setup_hca(struct hns_roce_dev *hr_dev) INIT_LIST_HEAD(&hr_dev->uctx_list); spin_lock_init(&hr_dev->uctx_list_lock);

+ INIT_LIST_HEAD(&hr_dev->mtr_unfree_list); + spin_lock_init(&hr_dev->mtr_unfree_list_lock); + + INIT_LIST_HEAD(&hr_dev->umem_unfree_list); + spin_lock_init(&hr_dev->umem_unfree_list_lock); + if (hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_CQ_RECORD_DB || hr_dev->caps.flags & HNS_ROCE_CAP_FLAG_QP_RECORD_DB) { INIT_LIST_HEAD(&hr_dev->pgdir_list); @@ -1555,6 +1561,8 @@ void hns_roce_exit(struct hns_roce_dev *hr_dev, bool bond_cleanup) if (hr_dev->hw->hw_exit) hr_dev->hw->hw_exit(hr_dev); hns_roce_teardown_hca(hr_dev); + hns_roce_free_unfree_umem(hr_dev); + hns_roce_free_unfree_mtr(hr_dev); hns_roce_cleanup_hem(hr_dev);

if (hr_dev->cmd_mod) diff --git a/drivers/infiniband/hw/hns/hns_roce_mr.c b/drivers/infiniband/hw/hns/hns_roce_mr.c index 0b5108d79219..9c0281744c48 100644 --- a/drivers/infiniband/hw/hns/hns_roce_mr.c +++ b/drivers/infiniband/hw/hns/hns_roce_mr.c @@ -83,7 +83,11 @@ static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) { unsigned long obj = key_to_hw_index(mr->key);

- hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj); + /* this resource will be freed when the driver is uninstalled, so + * no memory leak will occur. + */ + if (!mr->delayed_destroy_flag) + hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj); ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj); }

@@ -95,6 +99,10 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, struct hns_roce_buf_attr buf_attr = {}; int err;

+ mr->mtr_node = kvmalloc(sizeof(*mr->mtr_node), GFP_KERNEL); + if (!mr->mtr_node) + return -ENOMEM; + mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num; buf_attr.page_shift = is_fast ? PAGE_SHIFT : hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT; @@ -113,6 +121,7 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr, udata, start); if (err) { ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err); + kvfree(mr->mtr_node); return err; }

@@ -124,7 +133,12 @@ static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,

static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) { - hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr); + if (mr->delayed_destroy_flag && mr->type != MR_TYPE_DMA) { + hns_roce_add_unfree_mtr(mr->mtr_node, hr_dev, &mr->pbl_mtr); + } else { + hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr); + kvfree(mr->mtr_node); + } }

static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr) @@ -139,6 +153,8 @@ static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr if (ret) ibdev_warn(ibdev, "failed to destroy mpt, ret = %d.\n", ret); + if (ret == -EBUSY) + mr->delayed_destroy_flag = true; }

free_mr_pbl(hr_dev, mr); @@ -1217,3 +1233,75 @@ void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr) /* free buffers */ mtr_free_bufs(hr_dev, mtr); } + +static void hns_roce_copy_mtr(struct hns_roce_mtr *new_mtr, struct hns_roce_mtr *old_mtr) +{ + struct list_head *new_head, *old_head; + int i, j; + + memcpy(new_mtr, old_mtr, sizeof(*old_mtr)); + + for (i = 0; i < HNS_ROCE_MAX_BT_REGION; i++) + for (j = 0; j < HNS_ROCE_MAX_BT_LEVEL; j++) { + new_head = &new_mtr->hem_list.mid_bt[i][j]; + old_head = &old_mtr->hem_list.mid_bt[i][j]; + list_replace(old_head, new_head); + } + + new_head = &new_mtr->hem_list.root_bt; + old_head = &old_mtr->hem_list.root_bt; + list_replace(old_head, new_head); + + new_head = &new_mtr->hem_list.btm_bt; + old_head = &old_mtr->hem_list.btm_bt; + list_replace(old_head, new_head); +} + +void hns_roce_add_unfree_mtr(struct hns_roce_mtr_node *pos, + struct hns_roce_dev *hr_dev, + struct hns_roce_mtr *mtr) +{ + hns_roce_copy_mtr(&pos->mtr, mtr); + + spin_lock(&hr_dev->mtr_unfree_list_lock); + list_add_tail(&pos->list, &hr_dev->mtr_unfree_list); + spin_unlock(&hr_dev->mtr_unfree_list_lock); +} + +void hns_roce_free_unfree_mtr(struct hns_roce_dev *hr_dev) +{ + struct hns_roce_mtr_node *pos, *next; + + spin_lock(&hr_dev->mtr_unfree_list_lock); + list_for_each_entry_safe(pos, next, &hr_dev->mtr_unfree_list, list) { + list_del(&pos->list); + hns_roce_mtr_destroy(hr_dev, &pos->mtr); + kvfree(pos); + } + spin_unlock(&hr_dev->mtr_unfree_list_lock); +} + +void hns_roce_add_unfree_umem(struct hns_roce_user_db_page *user_page, + struct hns_roce_dev *hr_dev) +{ + struct hns_roce_umem_node *pos = user_page->umem_node; + + pos->umem = user_page->umem; + + spin_lock(&hr_dev->umem_unfree_list_lock); + list_add_tail(&pos->list, &hr_dev->umem_unfree_list); + spin_unlock(&hr_dev->umem_unfree_list_lock); +} + +void hns_roce_free_unfree_umem(struct hns_roce_dev *hr_dev) +{ + struct hns_roce_umem_node *pos, *next; + + spin_lock(&hr_dev->umem_unfree_list_lock); + list_for_each_entry_safe(pos, next, &hr_dev->umem_unfree_list, list) { + list_del(&pos->list); + ib_umem_release(pos->umem); + kvfree(pos); + } + spin_unlock(&hr_dev->mtr_unfree_list_lock); +} diff --git a/drivers/infiniband/hw/hns/hns_roce_qp.c b/drivers/infiniband/hw/hns/hns_roce_qp.c index 52ee28181ae4..d3251fa82242 100644 --- a/drivers/infiniband/hw/hns/hns_roce_qp.c +++ b/drivers/infiniband/hw/hns/hns_roce_qp.c @@ -387,6 +387,12 @@ static void free_qpc(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp) { struct hns_roce_qp_table *qp_table = &hr_dev->qp_table;

+ if (hr_qp->delayed_destroy_flag) + return; + + /* this resource will be freed when the driver is uninstalled, so + * no memory leak will occur. + */ if (hr_dev->caps.trrl_entry_sz) hns_roce_table_put(hr_dev, &qp_table->trrl_table, hr_qp->qpn); hns_roce_table_put(hr_dev, &qp_table->irrl_table, hr_qp->qpn); @@ -782,12 +788,17 @@ static int alloc_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, struct ib_device *ibdev = &hr_dev->ib_dev; int ret;

+ hr_qp->mtr_node = kvmalloc(sizeof(*hr_qp->mtr_node), GFP_KERNEL); + if (!hr_qp->mtr_node) + return -ENOMEM; + if (dca_en) { /* DCA must be enabled after the buffer attr is configured. */ ret = hns_roce_enable_dca(hr_dev, hr_qp, udata); if (ret) { ibdev_err(ibdev, "failed to enable DCA, ret = %d.\n", ret); + kvfree(hr_qp->mtr_node); return ret; }

@@ -811,6 +822,7 @@ static int alloc_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, ibdev_err(ibdev, "failed to create WQE mtr, ret = %d.\n", ret); if (dca_en) hns_roce_disable_dca(hr_dev, hr_qp, udata); + kvfree(hr_qp->mtr_node); }

return ret; @@ -819,7 +831,12 @@ static int alloc_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, static void free_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp, struct ib_udata *udata) { - hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr); + if (hr_qp->delayed_destroy_flag) { + hns_roce_add_unfree_mtr(hr_qp->mtr_node, hr_dev, &hr_qp->mtr); + } else { + hns_roce_mtr_destroy(hr_dev, &hr_qp->mtr); + kvfree(hr_qp->mtr_node); + }

if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DYNAMIC_CTX_ATTACH) hns_roce_disable_dca(hr_dev, hr_qp, udata); @@ -959,7 +976,7 @@ static int alloc_user_qp_db(struct hns_roce_dev *hr_dev,

err_sdb: if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) - hns_roce_db_unmap_user(uctx, &hr_qp->sdb); + hns_roce_db_unmap_user(uctx, &hr_qp->sdb, false); err_out: return ret; } @@ -1041,9 +1058,11 @@ static void free_qp_db(struct hns_roce_dev *hr_dev, struct hns_roce_qp *hr_qp,

if (udata) { if (hr_qp->en_flags & HNS_ROCE_QP_CAP_RQ_RECORD_DB) - hns_roce_db_unmap_user(uctx, &hr_qp->rdb); + hns_roce_db_unmap_user(uctx, &hr_qp->rdb, + hr_qp->delayed_destroy_flag); if (hr_qp->en_flags & HNS_ROCE_QP_CAP_SQ_RECORD_DB) - hns_roce_db_unmap_user(uctx, &hr_qp->sdb); + hns_roce_db_unmap_user(uctx, &hr_qp->sdb, + hr_qp->delayed_destroy_flag); if (hr_qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE) qp_user_mmap_entry_remove(hr_qp); } else { diff --git a/drivers/infiniband/hw/hns/hns_roce_srq.c b/drivers/infiniband/hw/hns/hns_roce_srq.c index 1e8b7a65519f..0f2f4e897738 100644 --- a/drivers/infiniband/hw/hns/hns_roce_srq.c +++ b/drivers/infiniband/hw/hns/hns_roce_srq.c @@ -164,14 +164,16 @@ static void free_srqc(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq) if (ret) dev_err(hr_dev->dev, "DESTROY_SRQ failed (%d) for SRQN %06lx\n", ret, srq->srqn); + if (ret == -EBUSY) + srq->delayed_destroy_flag = true;

xa_erase(&srq_table->xa, srq->srqn);

if (refcount_dec_and_test(&srq->refcount)) complete(&srq->free); wait_for_completion(&srq->free); - - hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn); + if (!srq->delayed_destroy_flag) + hns_roce_table_put(hr_dev, &srq_table->table, srq->srqn); }

static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq, @@ -182,6 +184,10 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq, struct hns_roce_buf_attr buf_attr = {}; int ret;

+ idx_que->mtr_node = kvmalloc(sizeof(*idx_que->mtr_node), GFP_KERNEL); + if (!idx_que->mtr_node) + return -ENOMEM; + srq->idx_que.entry_shift = ilog2(HNS_ROCE_IDX_QUE_ENTRY_SZ);

buf_attr.page_shift = hr_dev->caps.idx_buf_pg_sz + PAGE_SHIFT; @@ -195,7 +201,7 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq, udata, addr); if (ret) { ibdev_err(ibdev, "Failed to alloc SRQ idx mtr, ret = %d.\n", ret); - return ret; + goto err_kvmalloc; }

if (!udata) { @@ -213,6 +219,8 @@ static int alloc_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq, return 0; err_idx_mtr: hns_roce_mtr_destroy(hr_dev, &idx_que->mtr); +err_kvmalloc: + kvfree(idx_que->mtr_node);

return ret; } @@ -223,7 +231,12 @@ static void free_srq_idx(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq)

bitmap_free(idx_que->bitmap); idx_que->bitmap = NULL; - hns_roce_mtr_destroy(hr_dev, &idx_que->mtr); + if (srq->delayed_destroy_flag) { + hns_roce_add_unfree_mtr(idx_que->mtr_node, hr_dev, &idx_que->mtr); + } else { + hns_roce_mtr_destroy(hr_dev, &idx_que->mtr); + kvfree(idx_que->mtr_node); + } }

static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev, @@ -234,6 +247,10 @@ static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_buf_attr buf_attr = {}; int ret;

+ srq->mtr_node = kvmalloc(sizeof(*srq->mtr_node), GFP_KERNEL); + if (!srq->mtr_node) + return -ENOMEM; + srq->wqe_shift = ilog2(roundup_pow_of_two(max(HNS_ROCE_SGE_SIZE, HNS_ROCE_SGE_SIZE * srq->max_gs))); @@ -247,9 +264,11 @@ static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev, ret = hns_roce_mtr_create(hr_dev, &srq->buf_mtr, &buf_attr, hr_dev->caps.srqwqe_ba_pg_sz + PAGE_SHIFT, udata, addr); - if (ret) + if (ret) { ibdev_err(ibdev, "failed to alloc SRQ buf mtr, ret = %d.\n", ret); + kvfree(srq->mtr_node); + }

return ret; } @@ -257,7 +276,12 @@ static int alloc_srq_wqe_buf(struct hns_roce_dev *hr_dev, static void free_srq_wqe_buf(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq) { - hns_roce_mtr_destroy(hr_dev, &srq->buf_mtr); + if (srq->delayed_destroy_flag) { + hns_roce_add_unfree_mtr(srq->mtr_node, hr_dev, &srq->buf_mtr); + } else { + hns_roce_mtr_destroy(hr_dev, &srq->buf_mtr); + kvfree(srq->mtr_node); + } }

static int alloc_srq_wrid(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq) @@ -427,7 +451,8 @@ static void free_srq_db(struct hns_roce_dev *hr_dev, struct hns_roce_srq *srq, uctx = rdma_udata_to_drv_context(udata, struct hns_roce_ucontext, ibucontext); - hns_roce_db_unmap_user(uctx, &srq->rdb); + hns_roce_db_unmap_user(uctx, &srq->rdb, + srq->delayed_destroy_flag); } else { hns_roce_free_db(hr_dev, &srq->rdb); }

From: Junxian Huang huangjunxian6@hisilicon.com

driver inclusion category: bugfix bugzilla: https://gitee.com/openeuler/kernel/issues/I8HZ7W

--------------------------------------------------------------------------

IB-core and rdma-core restrict the sgid_index specified by users, which is uint8_t/u8 data type, to only be within the range of 0-255, so it's meaningless to support excessively large gid_table_len.

On the other hand, ib-core creates as many sysfs gid files as gid_table_len, most of which are not only useless because of the reason above, but also greatly increase the traversal time of the sysfs gid files for applications.

This patch limits the maximum length of gid table to 256.

Signed-off-by: Junxian Huang huangjunxian6@hisilicon.com --- drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c index 447e12d87396..69f969b7773b 100644 --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c @@ -2268,6 +2268,7 @@ static void set_hem_page_size(struct hns_roce_dev *hr_dev) /* Apply all loaded caps before setting to hardware */ static void apply_func_caps(struct hns_roce_dev *hr_dev) { +#define MAX_GID_TBL_LEN 256 struct hns_roce_caps *caps = &hr_dev->caps; struct hns_roce_v2_priv *priv = hr_dev->priv;

@@ -2303,9 +2304,14 @@ static void apply_func_caps(struct hns_roce_dev *hr_dev) caps->gmv_entry_sz = HNS_ROCE_V3_GMV_ENTRY_SZ;

caps->gmv_hop_num = HNS_ROCE_HOP_NUM_0; - caps->gid_table_len[0] = caps->gmv_bt_num * - (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz);

+ /* It's meaningless to support excessively large gid_table_len, + * as the type of sgid_index in kernel struct ib_global_route + * and userspace struct ibv_global_route are u8/uint8_t (0-255). + */ + caps->gid_table_len[0] = min_t(u32, MAX_GID_TBL_LEN, + caps->gmv_bt_num * + (HNS_HW_PAGE_SIZE / caps->gmv_entry_sz)); caps->gmv_entry_num = caps->gmv_bt_num * (PAGE_SIZE / caps->gmv_entry_sz); } else {