Back to home page

OSCL-LXR
 
 

     

0001 /*
0002  * Copyright (c) 2016 Hisilicon Limited.
0003  * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
0004  *
0005  * This software is available to you under a choice of one of two
0006  * licenses.  You may choose to be licensed under the terms of the GNU
0007  * General Public License (GPL) Version 2, available from the file
0008  * COPYING in the main directory of this source tree, or the
0009  * OpenIB.org BSD license below:
0010  *
0011  *     Redistribution and use in source and binary forms, with or
0012  *     without modification, are permitted provided that the following
0013  *     conditions are met:
0014  *
0015  *      - Redistributions of source code must retain the above
0016  *        copyright notice, this list of conditions and the following
0017  *        disclaimer.
0018  *
0019  *      - Redistributions in binary form must reproduce the above
0020  *        copyright notice, this list of conditions and the following
0021  *        disclaimer in the documentation and/or other materials
0022  *        provided with the distribution.
0023  *
0024  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
0025  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
0026  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
0027  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
0028  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
0029  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
0030  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
0031  * SOFTWARE.
0032  */
0033 
0034 #include <linux/vmalloc.h>
0035 #include <rdma/ib_umem.h>
0036 #include "hns_roce_device.h"
0037 #include "hns_roce_cmd.h"
0038 #include "hns_roce_hem.h"
0039 
0040 static u32 hw_index_to_key(int ind)
0041 {
0042     return ((u32)ind >> 24) | ((u32)ind << 8);
0043 }
0044 
0045 unsigned long key_to_hw_index(u32 key)
0046 {
0047     return (key << 24) | (key >> 8);
0048 }
0049 
0050 static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
0051 {
0052     struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
0053     struct ib_device *ibdev = &hr_dev->ib_dev;
0054     int err;
0055     int id;
0056 
0057     /* Allocate a key for mr from mr_table */
0058     id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max,
0059                  GFP_KERNEL);
0060     if (id < 0) {
0061         ibdev_err(ibdev, "failed to alloc id for MR key, id(%d)\n", id);
0062         return -ENOMEM;
0063     }
0064 
0065     mr->key = hw_index_to_key(id); /* MR key */
0066 
0067     err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table,
0068                  (unsigned long)id);
0069     if (err) {
0070         ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err);
0071         goto err_free_bitmap;
0072     }
0073 
0074     return 0;
0075 err_free_bitmap:
0076     ida_free(&mtpt_ida->ida, id);
0077     return err;
0078 }
0079 
0080 static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
0081 {
0082     unsigned long obj = key_to_hw_index(mr->key);
0083 
0084     hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj);
0085     ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj);
0086 }
0087 
0088 static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,
0089             struct ib_udata *udata, u64 start)
0090 {
0091     struct ib_device *ibdev = &hr_dev->ib_dev;
0092     bool is_fast = mr->type == MR_TYPE_FRMR;
0093     struct hns_roce_buf_attr buf_attr = {};
0094     int err;
0095 
0096     mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num;
0097     buf_attr.page_shift = is_fast ? PAGE_SHIFT :
0098                   hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT;
0099     buf_attr.region[0].size = mr->size;
0100     buf_attr.region[0].hopnum = mr->pbl_hop_num;
0101     buf_attr.region_count = 1;
0102     buf_attr.user_access = mr->access;
0103     /* fast MR's buffer is alloced before mapping, not at creation */
0104     buf_attr.mtt_only = is_fast;
0105 
0106     err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,
0107                   hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,
0108                   udata, start);
0109     if (err)
0110         ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
0111     else
0112         mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count;
0113 
0114     return err;
0115 }
0116 
0117 static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
0118 {
0119     hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr);
0120 }
0121 
0122 static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
0123 {
0124     struct ib_device *ibdev = &hr_dev->ib_dev;
0125     int ret;
0126 
0127     if (mr->enabled) {
0128         ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
0129                           key_to_hw_index(mr->key) &
0130                           (hr_dev->caps.num_mtpts - 1));
0131         if (ret)
0132             ibdev_warn(ibdev, "failed to destroy mpt, ret = %d.\n",
0133                    ret);
0134     }
0135 
0136     free_mr_pbl(hr_dev, mr);
0137     free_mr_key(hr_dev, mr);
0138 }
0139 
0140 static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev,
0141                   struct hns_roce_mr *mr)
0142 {
0143     unsigned long mtpt_idx = key_to_hw_index(mr->key);
0144     struct hns_roce_cmd_mailbox *mailbox;
0145     struct device *dev = hr_dev->dev;
0146     int ret;
0147 
0148     /* Allocate mailbox memory */
0149     mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
0150     if (IS_ERR(mailbox))
0151         return PTR_ERR(mailbox);
0152 
0153     if (mr->type != MR_TYPE_FRMR)
0154         ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr);
0155     else
0156         ret = hr_dev->hw->frmr_write_mtpt(hr_dev, mailbox->buf, mr);
0157     if (ret) {
0158         dev_err(dev, "failed to write mtpt, ret = %d.\n", ret);
0159         goto err_page;
0160     }
0161 
0162     ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
0163                      mtpt_idx & (hr_dev->caps.num_mtpts - 1));
0164     if (ret) {
0165         dev_err(dev, "failed to create mpt, ret = %d.\n", ret);
0166         goto err_page;
0167     }
0168 
0169     mr->enabled = 1;
0170 
0171 err_page:
0172     hns_roce_free_cmd_mailbox(hr_dev, mailbox);
0173 
0174     return ret;
0175 }
0176 
0177 void hns_roce_init_mr_table(struct hns_roce_dev *hr_dev)
0178 {
0179     struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
0180 
0181     ida_init(&mtpt_ida->ida);
0182     mtpt_ida->max = hr_dev->caps.num_mtpts - 1;
0183     mtpt_ida->min = hr_dev->caps.reserved_mrws;
0184 }
0185 
0186 struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc)
0187 {
0188     struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
0189     struct hns_roce_mr *mr;
0190     int ret;
0191 
0192     mr = kzalloc(sizeof(*mr), GFP_KERNEL);
0193     if (mr == NULL)
0194         return  ERR_PTR(-ENOMEM);
0195 
0196     mr->type = MR_TYPE_DMA;
0197     mr->pd = to_hr_pd(pd)->pdn;
0198     mr->access = acc;
0199 
0200     /* Allocate memory region key */
0201     hns_roce_hem_list_init(&mr->pbl_mtr.hem_list);
0202     ret = alloc_mr_key(hr_dev, mr);
0203     if (ret)
0204         goto err_free;
0205 
0206     ret = hns_roce_mr_enable(hr_dev, mr);
0207     if (ret)
0208         goto err_mr;
0209 
0210     mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
0211 
0212     return &mr->ibmr;
0213 err_mr:
0214     free_mr_key(hr_dev, mr);
0215 
0216 err_free:
0217     kfree(mr);
0218     return ERR_PTR(ret);
0219 }
0220 
0221 struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
0222                    u64 virt_addr, int access_flags,
0223                    struct ib_udata *udata)
0224 {
0225     struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
0226     struct hns_roce_mr *mr;
0227     int ret;
0228 
0229     mr = kzalloc(sizeof(*mr), GFP_KERNEL);
0230     if (!mr)
0231         return ERR_PTR(-ENOMEM);
0232 
0233     mr->iova = virt_addr;
0234     mr->size = length;
0235     mr->pd = to_hr_pd(pd)->pdn;
0236     mr->access = access_flags;
0237     mr->type = MR_TYPE_MR;
0238 
0239     ret = alloc_mr_key(hr_dev, mr);
0240     if (ret)
0241         goto err_alloc_mr;
0242 
0243     ret = alloc_mr_pbl(hr_dev, mr, udata, start);
0244     if (ret)
0245         goto err_alloc_key;
0246 
0247     ret = hns_roce_mr_enable(hr_dev, mr);
0248     if (ret)
0249         goto err_alloc_pbl;
0250 
0251     mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
0252     mr->ibmr.length = length;
0253 
0254     return &mr->ibmr;
0255 
0256 err_alloc_pbl:
0257     free_mr_pbl(hr_dev, mr);
0258 err_alloc_key:
0259     free_mr_key(hr_dev, mr);
0260 err_alloc_mr:
0261     kfree(mr);
0262     return ERR_PTR(ret);
0263 }
0264 
0265 struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start,
0266                      u64 length, u64 virt_addr,
0267                      int mr_access_flags, struct ib_pd *pd,
0268                      struct ib_udata *udata)
0269 {
0270     struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
0271     struct ib_device *ib_dev = &hr_dev->ib_dev;
0272     struct hns_roce_mr *mr = to_hr_mr(ibmr);
0273     struct hns_roce_cmd_mailbox *mailbox;
0274     unsigned long mtpt_idx;
0275     int ret;
0276 
0277     if (!mr->enabled)
0278         return ERR_PTR(-EINVAL);
0279 
0280     mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
0281     if (IS_ERR(mailbox))
0282         return ERR_CAST(mailbox);
0283 
0284     mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1);
0285 
0286     ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
0287                 mtpt_idx);
0288     if (ret)
0289         goto free_cmd_mbox;
0290 
0291     ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
0292                       mtpt_idx);
0293     if (ret)
0294         ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret);
0295 
0296     mr->enabled = 0;
0297     mr->iova = virt_addr;
0298     mr->size = length;
0299 
0300     if (flags & IB_MR_REREG_PD)
0301         mr->pd = to_hr_pd(pd)->pdn;
0302 
0303     if (flags & IB_MR_REREG_ACCESS)
0304         mr->access = mr_access_flags;
0305 
0306     if (flags & IB_MR_REREG_TRANS) {
0307         free_mr_pbl(hr_dev, mr);
0308         ret = alloc_mr_pbl(hr_dev, mr, udata, start);
0309         if (ret) {
0310             ibdev_err(ib_dev, "failed to alloc mr PBL, ret = %d.\n",
0311                   ret);
0312             goto free_cmd_mbox;
0313         }
0314     }
0315 
0316     ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, mailbox->buf);
0317     if (ret) {
0318         ibdev_err(ib_dev, "failed to write mtpt, ret = %d.\n", ret);
0319         goto free_cmd_mbox;
0320     }
0321 
0322     ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
0323                      mtpt_idx);
0324     if (ret) {
0325         ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret);
0326         goto free_cmd_mbox;
0327     }
0328 
0329     mr->enabled = 1;
0330 
0331 free_cmd_mbox:
0332     hns_roce_free_cmd_mailbox(hr_dev, mailbox);
0333 
0334     if (ret)
0335         return ERR_PTR(ret);
0336     return NULL;
0337 }
0338 
0339 int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
0340 {
0341     struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
0342     struct hns_roce_mr *mr = to_hr_mr(ibmr);
0343 
0344     if (hr_dev->hw->dereg_mr)
0345         hr_dev->hw->dereg_mr(hr_dev);
0346 
0347     hns_roce_mr_free(hr_dev, mr);
0348     kfree(mr);
0349 
0350     return 0;
0351 }
0352 
0353 struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
0354                 u32 max_num_sg)
0355 {
0356     struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
0357     struct device *dev = hr_dev->dev;
0358     struct hns_roce_mr *mr;
0359     int ret;
0360 
0361     if (mr_type != IB_MR_TYPE_MEM_REG)
0362         return ERR_PTR(-EINVAL);
0363 
0364     if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) {
0365         dev_err(dev, "max_num_sg larger than %d\n",
0366             HNS_ROCE_FRMR_MAX_PA);
0367         return ERR_PTR(-EINVAL);
0368     }
0369 
0370     mr = kzalloc(sizeof(*mr), GFP_KERNEL);
0371     if (!mr)
0372         return ERR_PTR(-ENOMEM);
0373 
0374     mr->type = MR_TYPE_FRMR;
0375     mr->pd = to_hr_pd(pd)->pdn;
0376     mr->size = max_num_sg * (1 << PAGE_SHIFT);
0377 
0378     /* Allocate memory region key */
0379     ret = alloc_mr_key(hr_dev, mr);
0380     if (ret)
0381         goto err_free;
0382 
0383     ret = alloc_mr_pbl(hr_dev, mr, NULL, 0);
0384     if (ret)
0385         goto err_key;
0386 
0387     ret = hns_roce_mr_enable(hr_dev, mr);
0388     if (ret)
0389         goto err_pbl;
0390 
0391     mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
0392     mr->ibmr.length = mr->size;
0393 
0394     return &mr->ibmr;
0395 
0396 err_key:
0397     free_mr_key(hr_dev, mr);
0398 err_pbl:
0399     free_mr_pbl(hr_dev, mr);
0400 err_free:
0401     kfree(mr);
0402     return ERR_PTR(ret);
0403 }
0404 
0405 static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr)
0406 {
0407     struct hns_roce_mr *mr = to_hr_mr(ibmr);
0408 
0409     if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) {
0410         mr->page_list[mr->npages++] = addr;
0411         return 0;
0412     }
0413 
0414     return -ENOBUFS;
0415 }
0416 
0417 int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
0418                unsigned int *sg_offset)
0419 {
0420     struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
0421     struct ib_device *ibdev = &hr_dev->ib_dev;
0422     struct hns_roce_mr *mr = to_hr_mr(ibmr);
0423     struct hns_roce_mtr *mtr = &mr->pbl_mtr;
0424     int ret = 0;
0425 
0426     mr->npages = 0;
0427     mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count,
0428                  sizeof(dma_addr_t), GFP_KERNEL);
0429     if (!mr->page_list)
0430         return ret;
0431 
0432     ret = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, hns_roce_set_page);
0433     if (ret < 1) {
0434         ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n",
0435               mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, ret);
0436         goto err_page_list;
0437     }
0438 
0439     mtr->hem_cfg.region[0].offset = 0;
0440     mtr->hem_cfg.region[0].count = mr->npages;
0441     mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num;
0442     mtr->hem_cfg.region_count = 1;
0443     ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages);
0444     if (ret) {
0445         ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret);
0446         ret = 0;
0447     } else {
0448         mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size);
0449         ret = mr->npages;
0450     }
0451 
0452 err_page_list:
0453     kvfree(mr->page_list);
0454     mr->page_list = NULL;
0455 
0456     return ret;
0457 }
0458 
0459 static void hns_roce_mw_free(struct hns_roce_dev *hr_dev,
0460                  struct hns_roce_mw *mw)
0461 {
0462     struct device *dev = hr_dev->dev;
0463     int ret;
0464 
0465     if (mw->enabled) {
0466         ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
0467                           key_to_hw_index(mw->rkey) &
0468                           (hr_dev->caps.num_mtpts - 1));
0469         if (ret)
0470             dev_warn(dev, "MW DESTROY_MPT failed (%d)\n", ret);
0471 
0472         hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table,
0473                    key_to_hw_index(mw->rkey));
0474     }
0475 
0476     ida_free(&hr_dev->mr_table.mtpt_ida.ida,
0477          (int)key_to_hw_index(mw->rkey));
0478 }
0479 
0480 static int hns_roce_mw_enable(struct hns_roce_dev *hr_dev,
0481                   struct hns_roce_mw *mw)
0482 {
0483     struct hns_roce_mr_table *mr_table = &hr_dev->mr_table;
0484     struct hns_roce_cmd_mailbox *mailbox;
0485     struct device *dev = hr_dev->dev;
0486     unsigned long mtpt_idx = key_to_hw_index(mw->rkey);
0487     int ret;
0488 
0489     /* prepare HEM entry memory */
0490     ret = hns_roce_table_get(hr_dev, &mr_table->mtpt_table, mtpt_idx);
0491     if (ret)
0492         return ret;
0493 
0494     mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
0495     if (IS_ERR(mailbox)) {
0496         ret = PTR_ERR(mailbox);
0497         goto err_table;
0498     }
0499 
0500     ret = hr_dev->hw->mw_write_mtpt(mailbox->buf, mw);
0501     if (ret) {
0502         dev_err(dev, "MW write mtpt fail!\n");
0503         goto err_page;
0504     }
0505 
0506     ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
0507                      mtpt_idx & (hr_dev->caps.num_mtpts - 1));
0508     if (ret) {
0509         dev_err(dev, "MW CREATE_MPT failed (%d)\n", ret);
0510         goto err_page;
0511     }
0512 
0513     mw->enabled = 1;
0514 
0515     hns_roce_free_cmd_mailbox(hr_dev, mailbox);
0516 
0517     return 0;
0518 
0519 err_page:
0520     hns_roce_free_cmd_mailbox(hr_dev, mailbox);
0521 
0522 err_table:
0523     hns_roce_table_put(hr_dev, &mr_table->mtpt_table, mtpt_idx);
0524 
0525     return ret;
0526 }
0527 
0528 int hns_roce_alloc_mw(struct ib_mw *ibmw, struct ib_udata *udata)
0529 {
0530     struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device);
0531     struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
0532     struct ib_device *ibdev = &hr_dev->ib_dev;
0533     struct hns_roce_mw *mw = to_hr_mw(ibmw);
0534     int ret;
0535     int id;
0536 
0537     /* Allocate a key for mw from mr_table */
0538     id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max,
0539                  GFP_KERNEL);
0540     if (id < 0) {
0541         ibdev_err(ibdev, "failed to alloc id for MW key, id(%d)\n", id);
0542         return -ENOMEM;
0543     }
0544 
0545     mw->rkey = hw_index_to_key(id);
0546 
0547     ibmw->rkey = mw->rkey;
0548     mw->pdn = to_hr_pd(ibmw->pd)->pdn;
0549     mw->pbl_hop_num = hr_dev->caps.pbl_hop_num;
0550     mw->pbl_ba_pg_sz = hr_dev->caps.pbl_ba_pg_sz;
0551     mw->pbl_buf_pg_sz = hr_dev->caps.pbl_buf_pg_sz;
0552 
0553     ret = hns_roce_mw_enable(hr_dev, mw);
0554     if (ret)
0555         goto err_mw;
0556 
0557     return 0;
0558 
0559 err_mw:
0560     hns_roce_mw_free(hr_dev, mw);
0561     return ret;
0562 }
0563 
0564 int hns_roce_dealloc_mw(struct ib_mw *ibmw)
0565 {
0566     struct hns_roce_dev *hr_dev = to_hr_dev(ibmw->device);
0567     struct hns_roce_mw *mw = to_hr_mw(ibmw);
0568 
0569     hns_roce_mw_free(hr_dev, mw);
0570     return 0;
0571 }
0572 
0573 static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0574               struct hns_roce_buf_region *region, dma_addr_t *pages,
0575               int max_count)
0576 {
0577     int count, npage;
0578     int offset, end;
0579     __le64 *mtts;
0580     u64 addr;
0581     int i;
0582 
0583     offset = region->offset;
0584     end = offset + region->count;
0585     npage = 0;
0586     while (offset < end && npage < max_count) {
0587         count = 0;
0588         mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
0589                           offset, &count, NULL);
0590         if (!mtts)
0591             return -ENOBUFS;
0592 
0593         for (i = 0; i < count && npage < max_count; i++) {
0594             addr = pages[npage];
0595 
0596             mtts[i] = cpu_to_le64(addr);
0597             npage++;
0598         }
0599         offset += count;
0600     }
0601 
0602     return npage;
0603 }
0604 
0605 static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr)
0606 {
0607     int i;
0608 
0609     for (i = 0; i < attr->region_count; i++)
0610         if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 &&
0611             attr->region[i].hopnum > 0)
0612             return true;
0613 
0614     /* because the mtr only one root base address, when hopnum is 0 means
0615      * root base address equals the first buffer address, thus all alloced
0616      * memory must in a continuous space accessed by direct mode.
0617      */
0618     return false;
0619 }
0620 
0621 static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr)
0622 {
0623     size_t size = 0;
0624     int i;
0625 
0626     for (i = 0; i < attr->region_count; i++)
0627         size += attr->region[i].size;
0628 
0629     return size;
0630 }
0631 
0632 /*
0633  * check the given pages in continuous address space
0634  * Returns 0 on success, or the error page num.
0635  */
0636 static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count,
0637                      unsigned int page_shift)
0638 {
0639     size_t page_size = 1 << page_shift;
0640     int i;
0641 
0642     for (i = 1; i < page_count; i++)
0643         if (pages[i] - pages[i - 1] != page_size)
0644             return i;
0645 
0646     return 0;
0647 }
0648 
0649 static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
0650 {
0651     /* release user buffers */
0652     if (mtr->umem) {
0653         ib_umem_release(mtr->umem);
0654         mtr->umem = NULL;
0655     }
0656 
0657     /* release kernel buffers */
0658     if (mtr->kmem) {
0659         hns_roce_buf_free(hr_dev, mtr->kmem);
0660         mtr->kmem = NULL;
0661     }
0662 }
0663 
0664 static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0665               struct hns_roce_buf_attr *buf_attr,
0666               struct ib_udata *udata, unsigned long user_addr)
0667 {
0668     struct ib_device *ibdev = &hr_dev->ib_dev;
0669     size_t total_size;
0670 
0671     total_size = mtr_bufs_size(buf_attr);
0672 
0673     if (udata) {
0674         mtr->kmem = NULL;
0675         mtr->umem = ib_umem_get(ibdev, user_addr, total_size,
0676                     buf_attr->user_access);
0677         if (IS_ERR_OR_NULL(mtr->umem)) {
0678             ibdev_err(ibdev, "failed to get umem, ret = %ld.\n",
0679                   PTR_ERR(mtr->umem));
0680             return -ENOMEM;
0681         }
0682     } else {
0683         mtr->umem = NULL;
0684         mtr->kmem = hns_roce_buf_alloc(hr_dev, total_size,
0685                            buf_attr->page_shift,
0686                            mtr->hem_cfg.is_direct ?
0687                            HNS_ROCE_BUF_DIRECT : 0);
0688         if (IS_ERR(mtr->kmem)) {
0689             ibdev_err(ibdev, "failed to alloc kmem, ret = %ld.\n",
0690                   PTR_ERR(mtr->kmem));
0691             return PTR_ERR(mtr->kmem);
0692         }
0693     }
0694 
0695     return 0;
0696 }
0697 
0698 static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0699             int page_count, unsigned int page_shift)
0700 {
0701     struct ib_device *ibdev = &hr_dev->ib_dev;
0702     dma_addr_t *pages;
0703     int npage;
0704     int ret;
0705 
0706     /* alloc a tmp array to store buffer's dma address */
0707     pages = kvcalloc(page_count, sizeof(dma_addr_t), GFP_KERNEL);
0708     if (!pages)
0709         return -ENOMEM;
0710 
0711     if (mtr->umem)
0712         npage = hns_roce_get_umem_bufs(hr_dev, pages, page_count,
0713                            mtr->umem, page_shift);
0714     else
0715         npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count,
0716                            mtr->kmem, page_shift);
0717 
0718     if (npage != page_count) {
0719         ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage,
0720               page_count);
0721         ret = -ENOBUFS;
0722         goto err_alloc_list;
0723     }
0724 
0725     if (mtr->hem_cfg.is_direct && npage > 1) {
0726         ret = mtr_check_direct_pages(pages, npage, page_shift);
0727         if (ret) {
0728             ibdev_err(ibdev, "failed to check %s page: %d / %d.\n",
0729                   mtr->umem ? "umtr" : "kmtr", ret, npage);
0730             ret = -ENOBUFS;
0731             goto err_alloc_list;
0732         }
0733     }
0734 
0735     ret = hns_roce_mtr_map(hr_dev, mtr, pages, page_count);
0736     if (ret)
0737         ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret);
0738 
0739 err_alloc_list:
0740     kvfree(pages);
0741 
0742     return ret;
0743 }
0744 
0745 int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0746              dma_addr_t *pages, unsigned int page_cnt)
0747 {
0748     struct ib_device *ibdev = &hr_dev->ib_dev;
0749     struct hns_roce_buf_region *r;
0750     unsigned int i, mapped_cnt;
0751     int ret = 0;
0752 
0753     /*
0754      * Only use the first page address as root ba when hopnum is 0, this
0755      * is because the addresses of all pages are consecutive in this case.
0756      */
0757     if (mtr->hem_cfg.is_direct) {
0758         mtr->hem_cfg.root_ba = pages[0];
0759         return 0;
0760     }
0761 
0762     for (i = 0, mapped_cnt = 0; i < mtr->hem_cfg.region_count &&
0763          mapped_cnt < page_cnt; i++) {
0764         r = &mtr->hem_cfg.region[i];
0765         /* if hopnum is 0, no need to map pages in this region */
0766         if (!r->hopnum) {
0767             mapped_cnt += r->count;
0768             continue;
0769         }
0770 
0771         if (r->offset + r->count > page_cnt) {
0772             ret = -EINVAL;
0773             ibdev_err(ibdev,
0774                   "failed to check mtr%u count %u + %u > %u.\n",
0775                   i, r->offset, r->count, page_cnt);
0776             return ret;
0777         }
0778 
0779         ret = mtr_map_region(hr_dev, mtr, r, &pages[r->offset],
0780                      page_cnt - mapped_cnt);
0781         if (ret < 0) {
0782             ibdev_err(ibdev,
0783                   "failed to map mtr%u offset %u, ret = %d.\n",
0784                   i, r->offset, ret);
0785             return ret;
0786         }
0787         mapped_cnt += ret;
0788         ret = 0;
0789     }
0790 
0791     if (mapped_cnt < page_cnt) {
0792         ret = -ENOBUFS;
0793         ibdev_err(ibdev, "failed to map mtr pages count: %u < %u.\n",
0794               mapped_cnt, page_cnt);
0795     }
0796 
0797     return ret;
0798 }
0799 
0800 int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0801               u32 offset, u64 *mtt_buf, int mtt_max, u64 *base_addr)
0802 {
0803     struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
0804     int mtt_count, left;
0805     u32 start_index;
0806     int total = 0;
0807     __le64 *mtts;
0808     u32 npage;
0809     u64 addr;
0810 
0811     if (!mtt_buf || mtt_max < 1)
0812         goto done;
0813 
0814     /* no mtt memory in direct mode, so just return the buffer address */
0815     if (cfg->is_direct) {
0816         start_index = offset >> HNS_HW_PAGE_SHIFT;
0817         for (mtt_count = 0; mtt_count < cfg->region_count &&
0818              total < mtt_max; mtt_count++) {
0819             npage = cfg->region[mtt_count].offset;
0820             if (npage < start_index)
0821                 continue;
0822 
0823             addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT);
0824             mtt_buf[total] = addr;
0825 
0826             total++;
0827         }
0828 
0829         goto done;
0830     }
0831 
0832     start_index = offset >> cfg->buf_pg_shift;
0833     left = mtt_max;
0834     while (left > 0) {
0835         mtt_count = 0;
0836         mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
0837                           start_index + total,
0838                           &mtt_count, NULL);
0839         if (!mtts || !mtt_count)
0840             goto done;
0841 
0842         npage = min(mtt_count, left);
0843         left -= npage;
0844         for (mtt_count = 0; mtt_count < npage; mtt_count++)
0845             mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]);
0846     }
0847 
0848 done:
0849     if (base_addr)
0850         *base_addr = cfg->root_ba;
0851 
0852     return total;
0853 }
0854 
0855 static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev,
0856                 struct hns_roce_buf_attr *attr,
0857                 struct hns_roce_hem_cfg *cfg,
0858                 unsigned int *buf_page_shift, u64 unalinged_size)
0859 {
0860     struct hns_roce_buf_region *r;
0861     u64 first_region_padding;
0862     int page_cnt, region_cnt;
0863     unsigned int page_shift;
0864     size_t buf_size;
0865 
0866     /* If mtt is disabled, all pages must be within a continuous range */
0867     cfg->is_direct = !mtr_has_mtt(attr);
0868     buf_size = mtr_bufs_size(attr);
0869     if (cfg->is_direct) {
0870         /* When HEM buffer uses 0-level addressing, the page size is
0871          * equal to the whole buffer size, and we split the buffer into
0872          * small pages which is used to check whether the adjacent
0873          * units are in the continuous space and its size is fixed to
0874          * 4K based on hns ROCEE's requirement.
0875          */
0876         page_shift = HNS_HW_PAGE_SHIFT;
0877 
0878         /* The ROCEE requires the page size to be 4K * 2 ^ N. */
0879         cfg->buf_pg_count = 1;
0880         cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT +
0881             order_base_2(DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE));
0882         first_region_padding = 0;
0883     } else {
0884         page_shift = attr->page_shift;
0885         cfg->buf_pg_count = DIV_ROUND_UP(buf_size + unalinged_size,
0886                          1 << page_shift);
0887         cfg->buf_pg_shift = page_shift;
0888         first_region_padding = unalinged_size;
0889     }
0890 
0891     /* Convert buffer size to page index and page count for each region and
0892      * the buffer's offset needs to be appended to the first region.
0893      */
0894     for (page_cnt = 0, region_cnt = 0; region_cnt < attr->region_count &&
0895          region_cnt < ARRAY_SIZE(cfg->region); region_cnt++) {
0896         r = &cfg->region[region_cnt];
0897         r->offset = page_cnt;
0898         buf_size = hr_hw_page_align(attr->region[region_cnt].size +
0899                         first_region_padding);
0900         r->count = DIV_ROUND_UP(buf_size, 1 << page_shift);
0901         first_region_padding = 0;
0902         page_cnt += r->count;
0903         r->hopnum = to_hr_hem_hopnum(attr->region[region_cnt].hopnum,
0904                          r->count);
0905     }
0906 
0907     cfg->region_count = region_cnt;
0908     *buf_page_shift = page_shift;
0909 
0910     return page_cnt;
0911 }
0912 
0913 static int mtr_alloc_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0914              unsigned int ba_page_shift)
0915 {
0916     struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
0917     int ret;
0918 
0919     hns_roce_hem_list_init(&mtr->hem_list);
0920     if (!cfg->is_direct) {
0921         ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list,
0922                         cfg->region, cfg->region_count,
0923                         ba_page_shift);
0924         if (ret)
0925             return ret;
0926         cfg->root_ba = mtr->hem_list.root_ba;
0927         cfg->ba_pg_shift = ba_page_shift;
0928     } else {
0929         cfg->ba_pg_shift = cfg->buf_pg_shift;
0930     }
0931 
0932     return 0;
0933 }
0934 
0935 static void mtr_free_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
0936 {
0937     hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
0938 }
0939 
0940 /**
0941  * hns_roce_mtr_create - Create hns memory translate region.
0942  *
0943  * @hr_dev: RoCE device struct pointer
0944  * @mtr: memory translate region
0945  * @buf_attr: buffer attribute for creating mtr
0946  * @ba_page_shift: page shift for multi-hop base address table
0947  * @udata: user space context, if it's NULL, means kernel space
0948  * @user_addr: userspace virtual address to start at
0949  */
0950 int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
0951             struct hns_roce_buf_attr *buf_attr,
0952             unsigned int ba_page_shift, struct ib_udata *udata,
0953             unsigned long user_addr)
0954 {
0955     struct ib_device *ibdev = &hr_dev->ib_dev;
0956     unsigned int buf_page_shift = 0;
0957     int buf_page_cnt;
0958     int ret;
0959 
0960     buf_page_cnt = mtr_init_buf_cfg(hr_dev, buf_attr, &mtr->hem_cfg,
0961                     &buf_page_shift,
0962                     udata ? user_addr & ~PAGE_MASK : 0);
0963     if (buf_page_cnt < 1 || buf_page_shift < HNS_HW_PAGE_SHIFT) {
0964         ibdev_err(ibdev, "failed to init mtr cfg, count %d shift %u.\n",
0965               buf_page_cnt, buf_page_shift);
0966         return -EINVAL;
0967     }
0968 
0969     ret = mtr_alloc_mtt(hr_dev, mtr, ba_page_shift);
0970     if (ret) {
0971         ibdev_err(ibdev, "failed to alloc mtr mtt, ret = %d.\n", ret);
0972         return ret;
0973     }
0974 
0975     /* The caller has its own buffer list and invokes the hns_roce_mtr_map()
0976      * to finish the MTT configuration.
0977      */
0978     if (buf_attr->mtt_only) {
0979         mtr->umem = NULL;
0980         mtr->kmem = NULL;
0981         return 0;
0982     }
0983 
0984     ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, udata, user_addr);
0985     if (ret) {
0986         ibdev_err(ibdev, "failed to alloc mtr bufs, ret = %d.\n", ret);
0987         goto err_alloc_mtt;
0988     }
0989 
0990     /* Write buffer's dma address to MTT */
0991     ret = mtr_map_bufs(hr_dev, mtr, buf_page_cnt, buf_page_shift);
0992     if (ret)
0993         ibdev_err(ibdev, "failed to map mtr bufs, ret = %d.\n", ret);
0994     else
0995         return 0;
0996 
0997     mtr_free_bufs(hr_dev, mtr);
0998 err_alloc_mtt:
0999     mtr_free_mtt(hr_dev, mtr);
1000     return ret;
1001 }
1002 
1003 void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
1004 {
1005     /* release multi-hop addressing resource */
1006     hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
1007 
1008     /* free buffers */
1009     mtr_free_bufs(hr_dev, mtr);
1010 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!