On 2021/9/18 14:16, Yunsheng Lin wrote:

On 2021/9/18 8:23, Chengchang Tang wrote:

...

Add new implementation for mmap by using the new mmap entry API.

Subsequent new features will use this new implementation. The old mmap which use a hard-corded offset will not be further expanded.

And forward compatibility is also guaranteed in this patch.

Signed-off-by: Chengchang Tang tangchengchang@huawei.com

drivers/infiniband/hw/hns/hns_roce_device.h | 21 ++++ drivers/infiniband/hw/hns/hns_roce_main.c | 144 +++++++++++++++++++++++++++- include/uapi/rdma/hns-abi.h | 21 +++- 3 files changed, 181 insertions(+), 5 deletions(-)

diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h index 9467c39..ca45694 100644 --- a/drivers/infiniband/hw/hns/hns_roce_device.h +++ b/drivers/infiniband/hw/hns/hns_roce_device.h @@ -225,11 +225,23 @@ struct hns_roce_uar { unsigned long logic_idx; };

+struct hns_user_mmap_entry {

• struct rdma_user_mmap_entry rdma_entry;
• u64 address;
• u8 mmap_flag;

+};

+enum hns_roce_mmap_type {

• HNS_ROCE_MMAP_TYPE_DB = 1,

+};

struct hns_roce_ucontext { struct ib_ucontext ibucontext; struct hns_roce_uar uar; struct list_head page_list; struct mutex page_mutex;

• bool mmap_key_support;
• struct rdma_user_mmap_entry *db_mmap_entry;

};

struct hns_roce_pd { @@ -1049,6 +1061,12 @@ static inline struct hns_roce_srq *to_hr_srq(struct ib_srq *ibsrq) return container_of(ibsrq, struct hns_roce_srq, ibsrq); }

+static inline struct hns_user_mmap_entry *to_hns_mmap(

• struct rdma_user_mmap_entry *rdma_entry)
  

+{

• return container_of(rdma_entry, struct hns_user_mmap_entry, rdma_entry);

+}

static inline void hns_roce_write64_k(__le32 val[2], void __iomem *dest) { writeq(*(u64 *)val, dest); @@ -1259,4 +1277,7 @@ int hns_roce_init(struct hns_roce_dev *hr_dev); void hns_roce_exit(struct hns_roce_dev *hr_dev); int hns_roce_fill_res_cq_entry(struct sk_buff *msg, struct ib_cq *ib_cq); +struct rdma_user_mmap_entry *hns_roce_user_mmap_entry_insert(

• struct ib_ucontext *ucontext, u64 address,
  

• size_t length, u8 mmap_flag);
  

#endif /* _HNS_ROCE_DEVICE_H */ diff --git a/drivers/infiniband/hw/hns/hns_roce_main.c b/drivers/infiniband/hw/hns/hns_roce_main.c index 5d39bd0..b8be8b7 100644 --- a/drivers/infiniband/hw/hns/hns_roce_main.c +++ b/drivers/infiniband/hw/hns/hns_roce_main.c @@ -291,6 +291,80 @@ static int hns_roce_modify_device(struct ib_device *ib_dev, int mask, return 0; }

+struct rdma_user_mmap_entry *hns_roce_user_mmap_entry_insert(

• 	struct ib_ucontext *ucontext, u64 address,
  

• 	size_t length, u8 mmap_flag)
  

+{

• struct hns_user_mmap_entry *entry;
• int ret;
• entry = kzalloc(sizeof(*entry), GFP_KERNEL);
• if (!entry)

• return NULL;
  

• entry->address = address;
• entry->mmap_flag = mmap_flag;
• ret = rdma_user_mmap_entry_insert(ucontext, &entry->rdma_entry,

• 			  length);
  

• if (ret) {

• kfree(entry);
  

• return NULL;
  

• }
• return &entry->rdma_entry;

+}

+static void hns_roce_dealloc_uar_entry(struct hns_roce_ucontext *context) +{

It seems there is "context->mmap_key_support" checking in alloc_uar, do we need something like below:

if (!context->mmap_key_support) return;

OK, I will add relevant checks here to ensure the symmetry of the code logic. But I need to explain, the code here will not cause a run-time error, because the validity of this pointer will be checked by rdma_user_mmap_entry_remove().

...

• rdma_user_mmap_entry_remove(context->db_mmap_entry);

+}

+static int hns_roce_alloc_uar_entry(struct ib_ucontext *uctx) +{

• struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
• u64 address;
• if (!context->mmap_key_support)

• return 0;
  

• address = context->uar.pfn << PAGE_SHIFT;
• context->db_mmap_entry =

• hns_roce_user_mmap_entry_insert(uctx, address, PAGE_SIZE,
  

• 				HNS_ROCE_MMAP_TYPE_DB);
  

• if (!context->db_mmap_entry)

• return -ENOMEM;
  

• return 0;

+}

+static void get_ucontext_config(struct hns_roce_ucontext *context,

• 		struct hns_roce_ib_alloc_ucontext *ucmd)
  

+{

• struct hns_roce_dev *hr_dev = to_hr_dev(context->ibucontext.device);
• if (ucmd->comp & HNS_ROCE_ALLOC_UCTX_COMP_CONFIG && hr_dev->hw_rev !=

• 	HNS_ROCE_HW_VER1)
  

• context->mmap_key_support = !!(ucmd->config &
  

• 			    HNS_ROCE_UCTX_REQ_EN_MMAP_KEY);
  

+}

+static void ucontext_set_resp(struct ib_ucontext *uctx,

• 	      struct hns_roce_ib_alloc_ucontext_resp *resp)
  

+{

• struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
• struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
• resp->qp_tab_size = hr_dev->caps.num_qps;
• resp->cqe_size = hr_dev->caps.cqe_sz;
• resp->srq_tab_size = hr_dev->caps.num_srqs;
• if (context->mmap_key_support) {

• resp->config |= HNS_ROCE_UCTX_RESP_MMAP_KEY_EN;
  

• resp->db_mmap_key =
  

• 	rdma_user_mmap_get_offset(context->db_mmap_entry);
  

• }

+}

static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata) { @@ -298,24 +372,35 @@ static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx, struct hns_roce_ucontext *context = to_hr_ucontext(uctx); struct hns_roce_ib_alloc_ucontext_resp resp = {}; struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);

• struct hns_roce_ib_alloc_ucontext ucmd = {};

  if (!hr_dev->active) return -EAGAIN;

• resp.qp_tab_size = hr_dev->caps.num_qps;
• resp.srq_tab_size = hr_dev->caps.num_srqs;

• if (udata->inlen) {

• ret = ib_copy_from_udata(&ucmd, udata,
  

• 			 min(udata->inlen, sizeof(ucmd)));
  

• if (ret)
  

• 	return ret;
  

• }

• get_ucontext_config(context, &ucmd);

  ret = hns_roce_uar_alloc(hr_dev, &context->uar); if (ret) goto error_fail_uar_alloc;

• ret = hns_roce_alloc_uar_entry(uctx);

• if (ret)

• goto error_fail_uar_entry;
  

• 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(&context->page_list); mutex_init(&context->page_mutex); }

• resp.cqe_size = hr_dev->caps.cqe_sz;

• ucontext_set_resp(uctx, &resp);

  ret = ib_copy_to_udata(udata, &resp, min(udata->outlen, sizeof(resp)));

@@ -325,6 +410,9 @@ static int hns_roce_alloc_ucontext(struct ib_ucontext *uctx, return 0;

error_fail_copy_to_udata:

• hns_roce_dealloc_uar_entry(context);

+error_fail_uar_entry: ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);

error_fail_uar_alloc: @@ -336,10 +424,12 @@ static void hns_roce_dealloc_ucontext(struct ib_ucontext *ibcontext) struct hns_roce_ucontext *context = to_hr_ucontext(ibcontext); struct hns_roce_dev *hr_dev = to_hr_dev(ibcontext->device);

• hns_roce_dealloc_uar_entry(context);
• ida_free(&hr_dev->uar_ida.ida, (int)context->uar.logic_idx);

}

-static int hns_roce_mmap(struct ib_ucontext *context, +static int hns_roce_legacy_mmap(struct ib_ucontext *context, struct vm_area_struct *vma) { struct hns_roce_dev *hr_dev = to_hr_dev(context->device); @@ -371,6 +461,51 @@ static int hns_roce_mmap(struct ib_ucontext *context, } }

+static int hns_roce_mmap(struct ib_ucontext *uctx, struct vm_area_struct *vma) +{

• struct hns_roce_ucontext *context = to_hr_ucontext(uctx);
• struct hns_roce_dev *hr_dev = to_hr_dev(uctx->device);
• struct ib_device *ibdev = &hr_dev->ib_dev;
• struct rdma_user_mmap_entry *rdma_entry;
• struct hns_user_mmap_entry *entry;
• phys_addr_t pfn;
• pgprot_t prot;
• int ret;
• if (!context->mmap_key_support)

• return hns_roce_legacy_mmap(uctx, vma);
  

• rdma_entry = rdma_user_mmap_entry_get_pgoff(uctx, vma->vm_pgoff);
• if (!rdma_entry) {

• ibdev_err(ibdev, "pgoff[%lu] is invalid entry.\n",
  

• 		vma->vm_pgoff);
  

• return -EINVAL;
  

• }
• entry = to_hns_mmap(rdma_entry);
• pfn = entry->address >> PAGE_SHIFT;
• prot = vma->vm_page_prot;
• switch (entry->mmap_flag) {
• case HNS_ROCE_MMAP_TYPE_DB:

• ret = rdma_user_mmap_io(uctx, vma, pfn,
  

• 			rdma_entry->npages * PAGE_SIZE,
  

• 			pgprot_noncached(prot), rdma_entry);
  

• break;
  

• default:

• ret =  -EINVAL;
  

• }
• rdma_user_mmap_entry_put(rdma_entry);
• return ret;

+}

+static void hns_roce_free_mmap(struct rdma_user_mmap_entry *rdma_entry) +{

• struct hns_user_mmap_entry *entry = to_hns_mmap(rdma_entry);
• kfree(entry);

When will ib stack call mmap_free ops? It seems if the entry is freed here, context->db_mmap_entry may still point to that pointer? would it cause a double free problem?

There is a refcnt for every entry. When refcnt is decremented to 0, it will be released by IB framework through calling this ops.

These objects are managed by the IB framework, drivers should not release them directly. So, there should not be a problem of double free here.

...

+}

static int hns_roce_port_immutable(struct ib_device *ib_dev, u32 port_num, struct ib_port_immutable *immutable) { @@ -444,6 +579,7 @@ static const struct ib_device_ops hns_roce_dev_ops = { .get_link_layer = hns_roce_get_link_layer, .get_port_immutable = hns_roce_port_immutable, .mmap = hns_roce_mmap,

• .mmap_free = hns_roce_free_mmap, .modify_device = hns_roce_modify_device, .modify_qp = hns_roce_modify_qp, .query_ah = hns_roce_query_ah,

diff --git a/include/uapi/rdma/hns-abi.h b/include/uapi/rdma/hns-abi.h index 42b1776..bb9a02d 100644 --- a/include/uapi/rdma/hns-abi.h +++ b/include/uapi/rdma/hns-abi.h @@ -83,11 +83,30 @@ struct hns_roce_ib_create_qp_resp { __aligned_u64 cap_flags; };

+enum hns_roce_alloc_uctx_comp_flag {

• HNS_ROCE_ALLOC_UCTX_COMP_CONFIG = 1 << 0,

+};

+enum hns_roce_alloc_uctx_resp_config {

• HNS_ROCE_UCTX_RESP_MMAP_KEY_EN = 1 << 0,

+};

+enum hns_roce_alloc_uctx_req_config {

• HNS_ROCE_UCTX_REQ_EN_MMAP_KEY = 1 << 0,

To be coniststent with HNS_ROCE_UCTX_RESP_MMAP_KEY_EN, maybe: HNS_ROCE_UCTX_REQ_MMAP_KEY_EN

OK, I will rename it in the next version.

...

+};

+struct hns_roce_ib_alloc_ucontext {

• __u32 comp;
• __u32 config;

+};

struct hns_roce_ib_alloc_ucontext_resp { __u32 qp_tab_size; __u32 cqe_size; __u32 srq_tab_size;

• __u32 reserved;

• __u8 config;
• __u8 rsv[3];
• __aligned_u64 db_mmap_key;

};

struct hns_roce_ib_alloc_pd_resp {

2.9.5

.

.