net/sunrpc/xprtrdma/fmr_ops.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2015 Oracle.  All rights reserved.
   4  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
   5  */
   6
   7 /* Lightweight memory registration using Fast Memory Regions (FMR).
   8  * Referred to sometimes as MTHCAFMR mode.
   9  *
  10  * FMR uses synchronous memory registration and deregistration.
  11  * FMR registration is known to be fast, but FMR deregistration
  12  * can take tens of usecs to complete.
  13  */
  14
  15 /* Normal operation
  16  *
  17  * A Memory Region is prepared for RDMA READ or WRITE using the
  18  * ib_map_phys_fmr verb (fmr_op_map). When the RDMA operation is
  19  * finished, the Memory Region is unmapped using the ib_unmap_fmr
  20  * verb (fmr_op_unmap).
  21  */
  22
  23 #include "xprt_rdma.h"
  24
  25 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  26 # define RPCDBG_FACILITY        RPCDBG_TRANS
  27 #endif
  28
  29 /* Maximum scatter/gather per FMR */
  30 #define RPCRDMA_MAX_FMR_SGES    (64)
  31
  32 /* Access mode of externally registered pages */
  33 enum {
  34         RPCRDMA_FMR_ACCESS_FLAGS        = IB_ACCESS_REMOTE_WRITE |
  35                                           IB_ACCESS_REMOTE_READ,
  36 };
  37
  38 bool
  39 fmr_is_supported(struct rpcrdma_ia *ia)
  40 {
  41         if (!ia->ri_device->alloc_fmr) {
  42                 pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
  43                         ia->ri_device->name);
  44                 return false;
  45         }
  46         return true;
  47 }
  48
  49 static int
  50 fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *mw)
  51 {
  52         static struct ib_fmr_attr fmr_attr = {
  53                 .max_pages      = RPCRDMA_MAX_FMR_SGES,
  54                 .max_maps       = 1,
  55                 .page_shift     = PAGE_SHIFT
  56         };
  57
  58         mw->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
  59                                        sizeof(u64), GFP_KERNEL);
  60         if (!mw->fmr.fm_physaddrs)
  61                 goto out_free;
  62
  63         mw->mw_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
  64                             sizeof(*mw->mw_sg), GFP_KERNEL);
  65         if (!mw->mw_sg)
  66                 goto out_free;
  67
  68         sg_init_table(mw->mw_sg, RPCRDMA_MAX_FMR_SGES);
  69
  70         mw->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
  71                                      &fmr_attr);
  72         if (IS_ERR(mw->fmr.fm_mr))
  73                 goto out_fmr_err;
  74
  75         return 0;
  76
  77 out_fmr_err:
  78         dprintk("RPC:       %s: ib_alloc_fmr returned %ld\n", __func__,
  79                 PTR_ERR(mw->fmr.fm_mr));
  80
  81 out_free:
  82         kfree(mw->mw_sg);
  83         kfree(mw->fmr.fm_physaddrs);
  84         return -ENOMEM;
  85 }
  86
  87 static int
  88 __fmr_unmap(struct rpcrdma_mw *mw)
  89 {
  90         LIST_HEAD(l);
  91         int rc;
  92
  93         list_add(&mw->fmr.fm_mr->list, &l);
  94         rc = ib_unmap_fmr(&l);
  95         list_del(&mw->fmr.fm_mr->list);
  96         return rc;
  97 }
  98
  99 static void
 100 fmr_op_release_mr(struct rpcrdma_mw *r)
 101 {
 102         LIST_HEAD(unmap_list);
 103         int rc;
 104
 105         /* Ensure MW is not on any rl_registered list */
 106         if (!list_empty(&r->mw_list))
 107                 list_del(&r->mw_list);
 108
 109         kfree(r->fmr.fm_physaddrs);
 110         kfree(r->mw_sg);
 111
 112         /* In case this one was left mapped, try to unmap it
 113          * to prevent dealloc_fmr from failing with EBUSY
 114          */
 115         rc = __fmr_unmap(r);
 116         if (rc)
 117                 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
 118                        r, rc);
 119
 120         rc = ib_dealloc_fmr(r->fmr.fm_mr);
 121         if (rc)
 122                 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
 123                        r, rc);
 124
 125         kfree(r);
 126 }
 127
 128 /* Reset of a single FMR.
 129  */
 130 static void
 131 fmr_op_recover_mr(struct rpcrdma_mw *mw)
 132 {
 133         struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
 134         int rc;
 135
 136         /* ORDER: invalidate first */
 137         rc = __fmr_unmap(mw);
 138
 139         /* ORDER: then DMA unmap */
 140         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 141                         mw->mw_sg, mw->mw_nents, mw->mw_dir);
 142         if (rc)
 143                 goto out_release;
 144
 145         rpcrdma_put_mw(r_xprt, mw);
 146         r_xprt->rx_stats.mrs_recovered++;
 147         return;
 148
 149 out_release:
 150         pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mw);
 151         r_xprt->rx_stats.mrs_orphaned++;
 152
 153         spin_lock(&r_xprt->rx_buf.rb_mwlock);
 154         list_del(&mw->mw_all);
 155         spin_unlock(&r_xprt->rx_buf.rb_mwlock);
 156
 157         fmr_op_release_mr(mw);
 158 }
 159
 160 static int
 161 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
 162             struct rpcrdma_create_data_internal *cdata)
 163 {
 164         ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
 165                                 RPCRDMA_MAX_FMR_SGES);
 166         return 0;
 167 }
 168
 169 /* FMR mode conveys up to 64 pages of payload per chunk segment.
 170  */
 171 static size_t
 172 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
 173 {
 174         return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
 175                      RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
 176 }
 177
 178 /* Use the ib_map_phys_fmr() verb to register a memory region
 179  * for remote access via RDMA READ or RDMA WRITE.
 180  */
 181 static struct rpcrdma_mr_seg *
 182 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
 183            int nsegs, bool writing, struct rpcrdma_mw **out)
 184 {
 185         struct rpcrdma_mr_seg *seg1 = seg;
 186         int len, pageoff, i, rc;
 187         struct rpcrdma_mw *mw;
 188         u64 *dma_pages;
 189
 190         mw = rpcrdma_get_mw(r_xprt);
 191         if (!mw)
 192                 return ERR_PTR(-ENOBUFS);
 193
 194         pageoff = offset_in_page(seg1->mr_offset);
 195         seg1->mr_offset -= pageoff;     /* start of page */
 196         seg1->mr_len += pageoff;
 197         len = -pageoff;
 198         if (nsegs > RPCRDMA_MAX_FMR_SGES)
 199                 nsegs = RPCRDMA_MAX_FMR_SGES;
 200         for (i = 0; i < nsegs;) {
 201                 if (seg->mr_page)
 202                         sg_set_page(&mw->mw_sg[i],
 203                                     seg->mr_page,
 204                                     seg->mr_len,
 205                                     offset_in_page(seg->mr_offset));
 206                 else
 207                         sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
 208                                    seg->mr_len);
 209                 len += seg->mr_len;
 210                 ++seg;
 211                 ++i;
 212                 /* Check for holes */
 213                 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
 214                     offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
 215                         break;
 216         }
 217         mw->mw_dir = rpcrdma_data_dir(writing);
 218
 219         mw->mw_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
 220                                      mw->mw_sg, i, mw->mw_dir);
 221         if (!mw->mw_nents)
 222                 goto out_dmamap_err;
 223
 224         for (i = 0, dma_pages = mw->fmr.fm_physaddrs; i < mw->mw_nents; i++)
 225                 dma_pages[i] = sg_dma_address(&mw->mw_sg[i]);
 226         rc = ib_map_phys_fmr(mw->fmr.fm_mr, dma_pages, mw->mw_nents,
 227                              dma_pages[0]);
 228         if (rc)
 229                 goto out_maperr;
 230
 231         mw->mw_handle = mw->fmr.fm_mr->rkey;
 232         mw->mw_length = len;
 233         mw->mw_offset = dma_pages[0] + pageoff;
 234
 235         *out = mw;
 236         return seg;
 237
 238 out_dmamap_err:
 239         pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
 240                mw->mw_sg, i);
 241         rpcrdma_put_mw(r_xprt, mw);
 242         return ERR_PTR(-EIO);
 243
 244 out_maperr:
 245         pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
 246                len, (unsigned long long)dma_pages[0],
 247                pageoff, mw->mw_nents, rc);
 248         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 249                         mw->mw_sg, mw->mw_nents, mw->mw_dir);
 250         rpcrdma_put_mw(r_xprt, mw);
 251         return ERR_PTR(-EIO);
 252 }
 253
 254 /* Invalidate all memory regions that were registered for "req".
 255  *
 256  * Sleeps until it is safe for the host CPU to access the
 257  * previously mapped memory regions.
 258  *
 259  * Caller ensures that @mws is not empty before the call. This
 260  * function empties the list.
 261  */
 262 static void
 263 fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mws)
 264 {
 265         struct rpcrdma_mw *mw;
 266         LIST_HEAD(unmap_list);
 267         int rc;
 268
 269         /* ORDER: Invalidate all of the req's MRs first
 270          *
 271          * ib_unmap_fmr() is slow, so use a single call instead
 272          * of one call per mapped FMR.
 273          */
 274         list_for_each_entry(mw, mws, mw_list) {
 275                 dprintk("RPC:       %s: unmapping fmr %p\n",
 276                         __func__, &mw->fmr);
 277                 list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
 278         }
 279         r_xprt->rx_stats.local_inv_needed++;
 280         rc = ib_unmap_fmr(&unmap_list);
 281         if (rc)
 282                 goto out_reset;
 283
 284         /* ORDER: Now DMA unmap all of the req's MRs, and return
 285          * them to the free MW list.
 286          */
 287         while (!list_empty(mws)) {
 288                 mw = rpcrdma_pop_mw(mws);
 289                 dprintk("RPC:       %s: DMA unmapping fmr %p\n",
 290                         __func__, &mw->fmr);
 291                 list_del(&mw->fmr.fm_mr->list);
 292                 ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 293                                 mw->mw_sg, mw->mw_nents, mw->mw_dir);
 294                 rpcrdma_put_mw(r_xprt, mw);
 295         }
 296
 297         return;
 298
 299 out_reset:
 300         pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
 301
 302         while (!list_empty(mws)) {
 303                 mw = rpcrdma_pop_mw(mws);
 304                 list_del(&mw->fmr.fm_mr->list);
 305                 fmr_op_recover_mr(mw);
 306         }
 307 }
 308
 309 /* Use a slow, safe mechanism to invalidate all memory regions
 310  * that were registered for "req".
 311  */
 312 static void
 313 fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
 314                   bool sync)
 315 {
 316         struct rpcrdma_mw *mw;
 317
 318         while (!list_empty(&req->rl_registered)) {
 319                 mw = rpcrdma_pop_mw(&req->rl_registered);
 320                 if (sync)
 321                         fmr_op_recover_mr(mw);
 322                 else
 323                         rpcrdma_defer_mr_recovery(mw);
 324         }
 325 }
 326
 327 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
 328         .ro_map                         = fmr_op_map,
 329         .ro_unmap_sync                  = fmr_op_unmap_sync,
 330         .ro_unmap_safe                  = fmr_op_unmap_safe,
 331         .ro_recover_mr                  = fmr_op_recover_mr,
 332         .ro_open                        = fmr_op_open,
 333         .ro_maxpages                    = fmr_op_maxpages,
 334         .ro_init_mr                     = fmr_op_init_mr,
 335         .ro_release_mr                  = fmr_op_release_mr,
 336         .ro_displayname                 = "fmr",
 337         .ro_send_w_inv_ok               = 0,
 338 };