net/sunrpc/xprtrdma/frwr_ops.c

   1 /*
   2  * Copyright (c) 2015 Oracle.  All rights reserved.
   3  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   4  */
   5
   6 /* Lightweight memory registration using Fast Registration Work
   7  * Requests (FRWR). Also referred to sometimes as FRMR mode.
   8  *
   9  * FRWR features ordered asynchronous registration and deregistration
  10  * of arbitrarily sized memory regions. This is the fastest and safest
  11  * but most complex memory registration mode.
  12  */
  13
  14 #include "xprt_rdma.h"
  15
  16 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  17 # define RPCDBG_FACILITY        RPCDBG_TRANS
  18 #endif
  19
  20 static int
  21 __frwr_init(struct rpcrdma_mw *r, struct ib_pd *pd, struct ib_device *device,
  22             unsigned int depth)
  23 {
  24         struct rpcrdma_frmr *f = &r->r.frmr;
  25         int rc;
  26
  27         f->fr_mr = ib_alloc_fast_reg_mr(pd, depth);
  28         if (IS_ERR(f->fr_mr))
  29                 goto out_mr_err;
  30         f->fr_pgl = ib_alloc_fast_reg_page_list(device, depth);
  31         if (IS_ERR(f->fr_pgl))
  32                 goto out_list_err;
  33         return 0;
  34
  35 out_mr_err:
  36         rc = PTR_ERR(f->fr_mr);
  37         dprintk("RPC:       %s: ib_alloc_fast_reg_mr status %i\n",
  38                 __func__, rc);
  39         return rc;
  40
  41 out_list_err:
  42         rc = PTR_ERR(f->fr_pgl);
  43         dprintk("RPC:       %s: ib_alloc_fast_reg_page_list status %i\n",
  44                 __func__, rc);
  45         ib_dereg_mr(f->fr_mr);
  46         return rc;
  47 }
  48
  49 static void
  50 __frwr_release(struct rpcrdma_mw *r)
  51 {
  52         int rc;
  53
  54         rc = ib_dereg_mr(r->r.frmr.fr_mr);
  55         if (rc)
  56                 dprintk("RPC:       %s: ib_dereg_mr status %i\n",
  57                         __func__, rc);
  58         ib_free_fast_reg_page_list(r->r.frmr.fr_pgl);
  59 }
  60
  61 static int
  62 frwr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
  63              struct rpcrdma_create_data_internal *cdata)
  64 {
  65         struct ib_device_attr *devattr = &ia->ri_devattr;
  66         int depth, delta;
  67
  68         ia->ri_max_frmr_depth =
  69                         min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
  70                               devattr->max_fast_reg_page_list_len);
  71         dprintk("RPC:       %s: device's max FR page list len = %u\n",
  72                 __func__, ia->ri_max_frmr_depth);
  73
  74         /* Add room for frmr register and invalidate WRs.
  75          * 1. FRMR reg WR for head
  76          * 2. FRMR invalidate WR for head
  77          * 3. N FRMR reg WRs for pagelist
  78          * 4. N FRMR invalidate WRs for pagelist
  79          * 5. FRMR reg WR for tail
  80          * 6. FRMR invalidate WR for tail
  81          * 7. The RDMA_SEND WR
  82          */
  83         depth = 7;
  84
  85         /* Calculate N if the device max FRMR depth is smaller than
  86          * RPCRDMA_MAX_DATA_SEGS.
  87          */
  88         if (ia->ri_max_frmr_depth < RPCRDMA_MAX_DATA_SEGS) {
  89                 delta = RPCRDMA_MAX_DATA_SEGS - ia->ri_max_frmr_depth;
  90                 do {
  91                         depth += 2; /* FRMR reg + invalidate */
  92                         delta -= ia->ri_max_frmr_depth;
  93                 } while (delta > 0);
  94         }
  95
  96         ep->rep_attr.cap.max_send_wr *= depth;
  97         if (ep->rep_attr.cap.max_send_wr > devattr->max_qp_wr) {
  98                 cdata->max_requests = devattr->max_qp_wr / depth;
  99                 if (!cdata->max_requests)
 100                         return -EINVAL;
 101                 ep->rep_attr.cap.max_send_wr = cdata->max_requests *
 102                                                depth;
 103         }
 104
 105         return 0;
 106 }
 107
 108 /* FRWR mode conveys a list of pages per chunk segment. The
 109  * maximum length of that list is the FRWR page list depth.
 110  */
 111 static size_t
 112 frwr_op_maxpages(struct rpcrdma_xprt *r_xprt)
 113 {
 114         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 115
 116         return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
 117                      rpcrdma_max_segments(r_xprt) * ia->ri_max_frmr_depth);
 118 }
 119
 120 /* If FAST_REG or LOCAL_INV failed, indicate the frmr needs to be reset. */
 121 static void
 122 frwr_sendcompletion(struct ib_wc *wc)
 123 {
 124         struct rpcrdma_mw *r;
 125
 126         if (likely(wc->status == IB_WC_SUCCESS))
 127                 return;
 128
 129         /* WARNING: Only wr_id and status are reliable at this point */
 130         r = (struct rpcrdma_mw *)(unsigned long)wc->wr_id;
 131         dprintk("RPC:       %s: frmr %p (stale), status %d\n",
 132                 __func__, r, wc->status);
 133         r->r.frmr.fr_state = FRMR_IS_STALE;
 134 }
 135
 136 static int
 137 frwr_op_init(struct rpcrdma_xprt *r_xprt)
 138 {
 139         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 140         struct ib_device *device = r_xprt->rx_ia.ri_id->device;
 141         unsigned int depth = r_xprt->rx_ia.ri_max_frmr_depth;
 142         struct ib_pd *pd = r_xprt->rx_ia.ri_pd;
 143         int i;
 144
 145         INIT_LIST_HEAD(&buf->rb_mws);
 146         INIT_LIST_HEAD(&buf->rb_all);
 147
 148         i = (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS;
 149         dprintk("RPC:       %s: initializing %d FRMRs\n", __func__, i);
 150
 151         while (i--) {
 152                 struct rpcrdma_mw *r;
 153                 int rc;
 154
 155                 r = kzalloc(sizeof(*r), GFP_KERNEL);
 156                 if (!r)
 157                         return -ENOMEM;
 158
 159                 rc = __frwr_init(r, pd, device, depth);
 160                 if (rc) {
 161                         kfree(r);
 162                         return rc;
 163                 }
 164
 165                 list_add(&r->mw_list, &buf->rb_mws);
 166                 list_add(&r->mw_all, &buf->rb_all);
 167                 r->mw_sendcompletion = frwr_sendcompletion;
 168         }
 169
 170         return 0;
 171 }
 172
 173 /* Post a FAST_REG Work Request to register a memory region
 174  * for remote access via RDMA READ or RDMA WRITE.
 175  */
 176 static int
 177 frwr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
 178             int nsegs, bool writing)
 179 {
 180         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 181         struct ib_device *device = ia->ri_id->device;
 182         enum dma_data_direction direction = rpcrdma_data_dir(writing);
 183         struct rpcrdma_mr_seg *seg1 = seg;
 184         struct rpcrdma_mw *mw = seg1->rl_mw;
 185         struct rpcrdma_frmr *frmr = &mw->r.frmr;
 186         struct ib_mr *mr = frmr->fr_mr;
 187         struct ib_send_wr fastreg_wr, *bad_wr;
 188         u8 key;
 189         int len, pageoff;
 190         int i, rc;
 191         int seg_len;
 192         u64 pa;
 193         int page_no;
 194
 195         pageoff = offset_in_page(seg1->mr_offset);
 196         seg1->mr_offset -= pageoff;     /* start of page */
 197         seg1->mr_len += pageoff;
 198         len = -pageoff;
 199         if (nsegs > ia->ri_max_frmr_depth)
 200                 nsegs = ia->ri_max_frmr_depth;
 201         for (page_no = i = 0; i < nsegs;) {
 202                 rpcrdma_map_one(device, seg, direction);
 203                 pa = seg->mr_dma;
 204                 for (seg_len = seg->mr_len; seg_len > 0; seg_len -= PAGE_SIZE) {
 205                         frmr->fr_pgl->page_list[page_no++] = pa;
 206                         pa += PAGE_SIZE;
 207                 }
 208                 len += seg->mr_len;
 209                 ++seg;
 210                 ++i;
 211                 /* Check for holes */
 212                 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
 213                     offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
 214                         break;
 215         }
 216         dprintk("RPC:       %s: Using frmr %p to map %d segments (%d bytes)\n",
 217                 __func__, mw, i, len);
 218
 219         frmr->fr_state = FRMR_IS_VALID;
 220
 221         memset(&fastreg_wr, 0, sizeof(fastreg_wr));
 222         fastreg_wr.wr_id = (unsigned long)(void *)mw;
 223         fastreg_wr.opcode = IB_WR_FAST_REG_MR;
 224         fastreg_wr.wr.fast_reg.iova_start = seg1->mr_dma + pageoff;
 225         fastreg_wr.wr.fast_reg.page_list = frmr->fr_pgl;
 226         fastreg_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
 227         fastreg_wr.wr.fast_reg.page_list_len = page_no;
 228         fastreg_wr.wr.fast_reg.length = len;
 229         fastreg_wr.wr.fast_reg.access_flags = writing ?
 230                                 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_LOCAL_WRITE :
 231                                 IB_ACCESS_REMOTE_READ;
 232         key = (u8)(mr->rkey & 0x000000FF);
 233         ib_update_fast_reg_key(mr, ++key);
 234         fastreg_wr.wr.fast_reg.rkey = mr->rkey;
 235
 236         DECR_CQCOUNT(&r_xprt->rx_ep);
 237         rc = ib_post_send(ia->ri_id->qp, &fastreg_wr, &bad_wr);
 238         if (rc)
 239                 goto out_senderr;
 240
 241         seg1->mr_rkey = mr->rkey;
 242         seg1->mr_base = seg1->mr_dma + pageoff;
 243         seg1->mr_nsegs = i;
 244         seg1->mr_len = len;
 245         return i;
 246
 247 out_senderr:
 248         dprintk("RPC:       %s: ib_post_send status %i\n", __func__, rc);
 249         ib_update_fast_reg_key(mr, --key);
 250         frmr->fr_state = FRMR_IS_INVALID;
 251         while (i--)
 252                 rpcrdma_unmap_one(device, --seg);
 253         return rc;
 254 }
 255
 256 /* Post a LOCAL_INV Work Request to prevent further remote access
 257  * via RDMA READ or RDMA WRITE.
 258  */
 259 static int
 260 frwr_op_unmap(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg)
 261 {
 262         struct rpcrdma_mr_seg *seg1 = seg;
 263         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 264         struct ib_send_wr invalidate_wr, *bad_wr;
 265         int rc, nsegs = seg->mr_nsegs;
 266         struct ib_device *device;
 267
 268         seg1->rl_mw->r.frmr.fr_state = FRMR_IS_INVALID;
 269
 270         memset(&invalidate_wr, 0, sizeof(invalidate_wr));
 271         invalidate_wr.wr_id = (unsigned long)(void *)seg1->rl_mw;
 272         invalidate_wr.opcode = IB_WR_LOCAL_INV;
 273         invalidate_wr.ex.invalidate_rkey = seg1->rl_mw->r.frmr.fr_mr->rkey;
 274         DECR_CQCOUNT(&r_xprt->rx_ep);
 275
 276         read_lock(&ia->ri_qplock);
 277         device = ia->ri_id->device;
 278         while (seg1->mr_nsegs--)
 279                 rpcrdma_unmap_one(device, seg++);
 280         rc = ib_post_send(ia->ri_id->qp, &invalidate_wr, &bad_wr);
 281         read_unlock(&ia->ri_qplock);
 282         if (rc)
 283                 goto out_err;
 284         return nsegs;
 285
 286 out_err:
 287         /* Force rpcrdma_buffer_get() to retry */
 288         seg1->rl_mw->r.frmr.fr_state = FRMR_IS_STALE;
 289         dprintk("RPC:       %s: ib_post_send status %i\n", __func__, rc);
 290         return nsegs;
 291 }
 292
 293 /* After a disconnect, a flushed FAST_REG_MR can leave an FRMR in
 294  * an unusable state. Find FRMRs in this state and dereg / reg
 295  * each.  FRMRs that are VALID and attached to an rpcrdma_req are
 296  * also torn down.
 297  *
 298  * This gives all in-use FRMRs a fresh rkey and leaves them INVALID.
 299  *
 300  * This is invoked only in the transport connect worker in order
 301  * to serialize with rpcrdma_register_frmr_external().
 302  */
 303 static void
 304 frwr_op_reset(struct rpcrdma_xprt *r_xprt)
 305 {
 306         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 307         struct ib_device *device = r_xprt->rx_ia.ri_id->device;
 308         unsigned int depth = r_xprt->rx_ia.ri_max_frmr_depth;
 309         struct ib_pd *pd = r_xprt->rx_ia.ri_pd;
 310         struct rpcrdma_mw *r;
 311         int rc;
 312
 313         list_for_each_entry(r, &buf->rb_all, mw_all) {
 314                 if (r->r.frmr.fr_state == FRMR_IS_INVALID)
 315                         continue;
 316
 317                 __frwr_release(r);
 318                 rc = __frwr_init(r, pd, device, depth);
 319                 if (rc) {
 320                         dprintk("RPC:       %s: mw %p left %s\n",
 321                                 __func__, r,
 322                                 (r->r.frmr.fr_state == FRMR_IS_STALE ?
 323                                         "stale" : "valid"));
 324                         continue;
 325                 }
 326
 327                 r->r.frmr.fr_state = FRMR_IS_INVALID;
 328         }
 329 }
 330
 331 static void
 332 frwr_op_destroy(struct rpcrdma_buffer *buf)
 333 {
 334         struct rpcrdma_mw *r;
 335
 336         while (!list_empty(&buf->rb_all)) {
 337                 r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
 338                 list_del(&r->mw_all);
 339                 __frwr_release(r);
 340                 kfree(r);
 341         }
 342 }
 343
 344 const struct rpcrdma_memreg_ops rpcrdma_frwr_memreg_ops = {
 345         .ro_map                         = frwr_op_map,
 346         .ro_unmap                       = frwr_op_unmap,
 347         .ro_open                        = frwr_op_open,
 348         .ro_maxpages                    = frwr_op_maxpages,
 349         .ro_init                        = frwr_op_init,
 350         .ro_reset                       = frwr_op_reset,
 351         .ro_destroy                     = frwr_op_destroy,
 352         .ro_displayname                 = "frwr",
 353 };