net/sunrpc/xprtrdma/fmr_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 Memory Regions (FMR).
   7  * Referred to sometimes as MTHCAFMR mode.
   8  *
   9  * FMR uses synchronous memory registration and deregistration.
  10  * FMR registration is known to be fast, but FMR deregistration
  11  * can take tens of usecs to complete.
  12  */
  13
  14 /* Normal operation
  15  *
  16  * A Memory Region is prepared for RDMA READ or WRITE using the
  17  * ib_map_phys_fmr verb (fmr_op_map). When the RDMA operation is
  18  * finished, the Memory Region is unmapped using the ib_unmap_fmr
  19  * verb (fmr_op_unmap).
  20  */
  21
  22 #include "xprt_rdma.h"
  23
  24 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  25 # define RPCDBG_FACILITY        RPCDBG_TRANS
  26 #endif
  27
  28 /* Maximum scatter/gather per FMR */
  29 #define RPCRDMA_MAX_FMR_SGES    (64)
  30
  31 /* Access mode of externally registered pages */
  32 enum {
  33         RPCRDMA_FMR_ACCESS_FLAGS        = IB_ACCESS_REMOTE_WRITE |
  34                                           IB_ACCESS_REMOTE_READ,
  35 };
  36
  37 bool
  38 fmr_is_supported(struct rpcrdma_ia *ia)
  39 {
  40         if (!ia->ri_device->alloc_fmr) {
  41                 pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
  42                         ia->ri_device->name);
  43                 return false;
  44         }
  45         return true;
  46 }
  47
  48 static int
  49 __fmr_init(struct rpcrdma_mw *mw, struct ib_pd *pd)
  50 {
  51         static struct ib_fmr_attr fmr_attr = {
  52                 .max_pages      = RPCRDMA_MAX_FMR_SGES,
  53                 .max_maps       = 1,
  54                 .page_shift     = PAGE_SHIFT
  55         };
  56
  57         mw->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
  58                                        sizeof(u64), GFP_KERNEL);
  59         if (!mw->fmr.fm_physaddrs)
  60                 goto out_free;
  61
  62         mw->mw_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
  63                             sizeof(*mw->mw_sg), GFP_KERNEL);
  64         if (!mw->mw_sg)
  65                 goto out_free;
  66
  67         sg_init_table(mw->mw_sg, RPCRDMA_MAX_FMR_SGES);
  68
  69         mw->fmr.fm_mr = ib_alloc_fmr(pd, RPCRDMA_FMR_ACCESS_FLAGS,
  70                                      &fmr_attr);
  71         if (IS_ERR(mw->fmr.fm_mr))
  72                 goto out_fmr_err;
  73
  74         return 0;
  75
  76 out_fmr_err:
  77         dprintk("RPC:       %s: ib_alloc_fmr returned %ld\n", __func__,
  78                 PTR_ERR(mw->fmr.fm_mr));
  79
  80 out_free:
  81         kfree(mw->mw_sg);
  82         kfree(mw->fmr.fm_physaddrs);
  83         return -ENOMEM;
  84 }
  85
  86 static int
  87 __fmr_unmap(struct rpcrdma_mw *mw)
  88 {
  89         LIST_HEAD(l);
  90         int rc;
  91
  92         list_add(&mw->fmr.fm_mr->list, &l);
  93         rc = ib_unmap_fmr(&l);
  94         list_del_init(&mw->fmr.fm_mr->list);
  95         return rc;
  96 }
  97
  98 static void
  99 __fmr_release(struct rpcrdma_mw *r)
 100 {
 101         LIST_HEAD(unmap_list);
 102         int rc;
 103
 104         kfree(r->fmr.fm_physaddrs);
 105         kfree(r->mw_sg);
 106
 107         /* In case this one was left mapped, try to unmap it
 108          * to prevent dealloc_fmr from failing with EBUSY
 109          */
 110         rc = __fmr_unmap(r);
 111         if (rc)
 112                 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
 113                        r, rc);
 114
 115         rc = ib_dealloc_fmr(r->fmr.fm_mr);
 116         if (rc)
 117                 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
 118                        r, rc);
 119 }
 120
 121 /* Reset of a single FMR.
 122  *
 123  * There's no recovery if this fails. The FMR is abandoned, but
 124  * remains in rb_all. It will be cleaned up when the transport is
 125  * destroyed.
 126  */
 127 static void
 128 fmr_op_recover_mr(struct rpcrdma_mw *mw)
 129 {
 130         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
 131         int rc;
 132
 133         /* ORDER: invalidate first */
 134         rc = __fmr_unmap(mw);
 135
 136         /* ORDER: then DMA unmap */
 137         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 138                         mw->mw_sg, mw->mw_nents, mw->mw_dir);
 139         if (rc) {
 140                 pr_err("rpcrdma: FMR reset status %d, %p orphaned\n",
 141                        rc, mw);
 142                 r_xprt->rx_stats.mrs_orphaned++;
 143                 return;
 144         }
 145
 146         rpcrdma_put_mw(r_xprt, mw);
 147         r_xprt->rx_stats.mrs_recovered++;
 148 }
 149
 150 static int
 151 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
 152             struct rpcrdma_create_data_internal *cdata)
 153 {
 154         rpcrdma_set_max_header_sizes(ia, cdata, max_t(unsigned int, 1,
 155                                                       RPCRDMA_MAX_DATA_SEGS /
 156                                                       RPCRDMA_MAX_FMR_SGES));
 157         return 0;
 158 }
 159
 160 /* FMR mode conveys up to 64 pages of payload per chunk segment.
 161  */
 162 static size_t
 163 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
 164 {
 165         return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
 166                      RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
 167 }
 168
 169 static int
 170 fmr_op_init(struct rpcrdma_xprt *r_xprt)
 171 {
 172         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
 173         struct ib_pd *pd = r_xprt->rx_ia.ri_pd;
 174         struct rpcrdma_mw *r;
 175         int i, rc;
 176
 177         spin_lock_init(&buf->rb_mwlock);
 178         INIT_LIST_HEAD(&buf->rb_mws);
 179         INIT_LIST_HEAD(&buf->rb_all);
 180
 181         i = max_t(int, RPCRDMA_MAX_DATA_SEGS / RPCRDMA_MAX_FMR_SGES, 1);
 182         i += 2;                         /* head + tail */
 183         i *= buf->rb_max_requests;      /* one set for each RPC slot */
 184         dprintk("RPC:       %s: initalizing %d FMRs\n", __func__, i);
 185
 186         while (i--) {
 187                 r = kzalloc(sizeof(*r), GFP_KERNEL);
 188                 if (!r)
 189                         return -ENOMEM;
 190
 191                 rc = __fmr_init(r, pd);
 192                 if (rc) {
 193                         kfree(r);
 194                         return rc;
 195                 }
 196
 197                 r->mw_xprt = r_xprt;
 198                 list_add(&r->mw_list, &buf->rb_mws);
 199                 list_add(&r->mw_all, &buf->rb_all);
 200         }
 201         return 0;
 202 }
 203
 204 /* Use the ib_map_phys_fmr() verb to register a memory region
 205  * for remote access via RDMA READ or RDMA WRITE.
 206  */
 207 static int
 208 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
 209            int nsegs, bool writing)
 210 {
 211         struct rpcrdma_mr_seg *seg1 = seg;
 212         int len, pageoff, i, rc;
 213         struct rpcrdma_mw *mw;
 214         u64 *dma_pages;
 215
 216         mw = seg1->rl_mw;
 217         seg1->rl_mw = NULL;
 218         if (mw)
 219                 rpcrdma_defer_mr_recovery(mw);
 220         mw = rpcrdma_get_mw(r_xprt);
 221         if (!mw)
 222                 return -ENOBUFS;
 223
 224         pageoff = offset_in_page(seg1->mr_offset);
 225         seg1->mr_offset -= pageoff;     /* start of page */
 226         seg1->mr_len += pageoff;
 227         len = -pageoff;
 228         if (nsegs > RPCRDMA_MAX_FMR_SGES)
 229                 nsegs = RPCRDMA_MAX_FMR_SGES;
 230         for (i = 0; i < nsegs;) {
 231                 if (seg->mr_page)
 232                         sg_set_page(&mw->mw_sg[i],
 233                                     seg->mr_page,
 234                                     seg->mr_len,
 235                                     offset_in_page(seg->mr_offset));
 236                 else
 237                         sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
 238                                    seg->mr_len);
 239                 len += seg->mr_len;
 240                 ++seg;
 241                 ++i;
 242                 /* Check for holes */
 243                 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
 244                     offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
 245                         break;
 246         }
 247         mw->mw_nents = i;
 248         mw->mw_dir = rpcrdma_data_dir(writing);
 249         if (i == 0)
 250                 goto out_dmamap_err;
 251
 252         if (!ib_dma_map_sg(r_xprt->rx_ia.ri_device,
 253                            mw->mw_sg, mw->mw_nents, mw->mw_dir))
 254                 goto out_dmamap_err;
 255
 256         for (i = 0, dma_pages = mw->fmr.fm_physaddrs; i < mw->mw_nents; i++)
 257                 dma_pages[i] = sg_dma_address(&mw->mw_sg[i]);
 258         rc = ib_map_phys_fmr(mw->fmr.fm_mr, dma_pages, mw->mw_nents,
 259                              dma_pages[0]);
 260         if (rc)
 261                 goto out_maperr;
 262
 263         seg1->rl_mw = mw;
 264         seg1->mr_rkey = mw->fmr.fm_mr->rkey;
 265         seg1->mr_base = dma_pages[0] + pageoff;
 266         seg1->mr_nsegs = mw->mw_nents;
 267         seg1->mr_len = len;
 268         return mw->mw_nents;
 269
 270 out_dmamap_err:
 271         pr_err("rpcrdma: failed to dma map sg %p sg_nents %u\n",
 272                mw->mw_sg, mw->mw_nents);
 273         rpcrdma_defer_mr_recovery(mw);
 274         return -EIO;
 275
 276 out_maperr:
 277         pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
 278                len, (unsigned long long)dma_pages[0],
 279                pageoff, mw->mw_nents, rc);
 280         rpcrdma_defer_mr_recovery(mw);
 281         return -EIO;
 282 }
 283
 284 /* Invalidate all memory regions that were registered for "req".
 285  *
 286  * Sleeps until it is safe for the host CPU to access the
 287  * previously mapped memory regions.
 288  */
 289 static void
 290 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
 291 {
 292         struct rpcrdma_mr_seg *seg;
 293         unsigned int i, nchunks;
 294         struct rpcrdma_mw *mw;
 295         LIST_HEAD(unmap_list);
 296         int rc;
 297
 298         dprintk("RPC:       %s: req %p\n", __func__, req);
 299
 300         /* ORDER: Invalidate all of the req's MRs first
 301          *
 302          * ib_unmap_fmr() is slow, so use a single call instead
 303          * of one call per mapped FMR.
 304          */
 305         for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
 306                 seg = &req->rl_segments[i];
 307                 mw = seg->rl_mw;
 308
 309                 list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
 310
 311                 i += seg->mr_nsegs;
 312         }
 313         rc = ib_unmap_fmr(&unmap_list);
 314         if (rc)
 315                 goto out_reset;
 316
 317         /* ORDER: Now DMA unmap all of the req's MRs, and return
 318          * them to the free MW list.
 319          */
 320         for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
 321                 seg = &req->rl_segments[i];
 322                 mw = seg->rl_mw;
 323
 324                 list_del_init(&mw->fmr.fm_mr->list);
 325                 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 326                                 mw->mw_sg, mw->mw_nents, mw->mw_dir);
 327                 rpcrdma_put_mw(r_xprt, mw);
 328
 329                 i += seg->mr_nsegs;
 330                 seg->mr_nsegs = 0;
 331                 seg->rl_mw = NULL;
 332         }
 333
 334         req->rl_nchunks = 0;
 335         return;
 336
 337 out_reset:
 338         pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
 339
 340         for (i = 0, nchunks = req->rl_nchunks; nchunks; nchunks--) {
 341                 seg = &req->rl_segments[i];
 342                 mw = seg->rl_mw;
 343
 344                 list_del_init(&mw->fmr.fm_mr->list);
 345                 fmr_op_recover_mr(mw);
 346
 347                 i += seg->mr_nsegs;
 348         }
 349 }
 350
 351 /* Use a slow, safe mechanism to invalidate all memory regions
 352  * that were registered for "req".
 353  */
 354 static void
 355 fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 356                   bool sync)
 357 {
 358         struct rpcrdma_mr_seg *seg;
 359         struct rpcrdma_mw *mw;
 360         unsigned int i;
 361
 362         for (i = 0; req->rl_nchunks; req->rl_nchunks--) {
 363                 seg = &req->rl_segments[i];
 364                 mw = seg->rl_mw;
 365
 366                 if (sync)
 367                         fmr_op_recover_mr(mw);
 368                 else
 369                         rpcrdma_defer_mr_recovery(mw);
 370
 371                 i += seg->mr_nsegs;
 372                 seg->mr_nsegs = 0;
 373                 seg->rl_mw = NULL;
 374         }
 375 }
 376
 377 static void
 378 fmr_op_destroy(struct rpcrdma_buffer *buf)
 379 {
 380         struct rpcrdma_mw *r;
 381
 382         while (!list_empty(&buf->rb_all)) {
 383                 r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
 384                 list_del(&r->mw_all);
 385                 __fmr_release(r);
 386                 kfree(r);
 387         }
 388 }
 389
 390 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
 391         .ro_map                         = fmr_op_map,
 392         .ro_unmap_sync                  = fmr_op_unmap_sync,
 393         .ro_unmap_safe                  = fmr_op_unmap_safe,
 394         .ro_recover_mr                  = fmr_op_recover_mr,
 395         .ro_open                        = fmr_op_open,
 396         .ro_maxpages                    = fmr_op_maxpages,
 397         .ro_init                        = fmr_op_init,
 398         .ro_destroy                     = fmr_op_destroy,
 399         .ro_displayname                 = "fmr",
 400 };