net/sunrpc/xprtrdma/fmr_ops.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (c) 2015, 2017 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_mr *mr)
  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         mr->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
  59                                        sizeof(u64), GFP_KERNEL);
  60         if (!mr->fmr.fm_physaddrs)
  61                 goto out_free;
  62
  63         mr->mr_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
  64                             sizeof(*mr->mr_sg), GFP_KERNEL);
  65         if (!mr->mr_sg)
  66                 goto out_free;
  67
  68         sg_init_table(mr->mr_sg, RPCRDMA_MAX_FMR_SGES);
  69
  70         mr->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
  71                                      &fmr_attr);
  72         if (IS_ERR(mr->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(mr->fmr.fm_mr));
  80
  81 out_free:
  82         kfree(mr->mr_sg);
  83         kfree(mr->fmr.fm_physaddrs);
  84         return -ENOMEM;
  85 }
  86
  87 static int
  88 __fmr_unmap(struct rpcrdma_mr *mr)
  89 {
  90         LIST_HEAD(l);
  91         int rc;
  92
  93         list_add(&mr->fmr.fm_mr->list, &l);
  94         rc = ib_unmap_fmr(&l);
  95         list_del(&mr->fmr.fm_mr->list);
  96         return rc;
  97 }
  98
  99 static void
 100 fmr_op_release_mr(struct rpcrdma_mr *mr)
 101 {
 102         LIST_HEAD(unmap_list);
 103         int rc;
 104
 105         /* Ensure MW is not on any rl_registered list */
 106         if (!list_empty(&mr->mr_list))
 107                 list_del(&mr->mr_list);
 108
 109         kfree(mr->fmr.fm_physaddrs);
 110         kfree(mr->mr_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(mr);
 116         if (rc)
 117                 pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
 118                        mr, rc);
 119
 120         rc = ib_dealloc_fmr(mr->fmr.fm_mr);
 121         if (rc)
 122                 pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
 123                        mr, rc);
 124
 125         kfree(mr);
 126 }
 127
 128 /* Reset of a single FMR.
 129  */
 130 static void
 131 fmr_op_recover_mr(struct rpcrdma_mr *mr)
 132 {
 133         struct rpcrdma_xprt *r_xprt = mr->mr_xprt;
 134         int rc;
 135
 136         /* ORDER: invalidate first */
 137         rc = __fmr_unmap(mr);
 138         if (rc)
 139                 goto out_release;
 140
 141         /* ORDER: then DMA unmap */
 142         rpcrdma_mr_unmap_and_put(mr);
 143
 144         r_xprt->rx_stats.mrs_recovered++;
 145         return;
 146
 147 out_release:
 148         pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mr);
 149         r_xprt->rx_stats.mrs_orphaned++;
 150
 151         trace_xprtrdma_dma_unmap(mr);
 152         ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
 153                         mr->mr_sg, mr->mr_nents, mr->mr_dir);
 154
 155         spin_lock(&r_xprt->rx_buf.rb_mrlock);
 156         list_del(&mr->mr_all);
 157         spin_unlock(&r_xprt->rx_buf.rb_mrlock);
 158
 159         fmr_op_release_mr(mr);
 160 }
 161
 162 static int
 163 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
 164             struct rpcrdma_create_data_internal *cdata)
 165 {
 166         ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
 167                                 RPCRDMA_MAX_FMR_SGES);
 168         return 0;
 169 }
 170
 171 /* FMR mode conveys up to 64 pages of payload per chunk segment.
 172  */
 173 static size_t
 174 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
 175 {
 176         return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
 177                      RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
 178 }
 179
 180 /* Use the ib_map_phys_fmr() verb to register a memory region
 181  * for remote access via RDMA READ or RDMA WRITE.
 182  */
 183 static struct rpcrdma_mr_seg *
 184 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
 185            int nsegs, bool writing, struct rpcrdma_mr **out)
 186 {
 187         struct rpcrdma_mr_seg *seg1 = seg;
 188         int len, pageoff, i, rc;
 189         struct rpcrdma_mr *mr;
 190         u64 *dma_pages;
 191
 192         mr = rpcrdma_mr_get(r_xprt);
 193         if (!mr)
 194                 return ERR_PTR(-ENOBUFS);
 195
 196         pageoff = offset_in_page(seg1->mr_offset);
 197         seg1->mr_offset -= pageoff;     /* start of page */
 198         seg1->mr_len += pageoff;
 199         len = -pageoff;
 200         if (nsegs > RPCRDMA_MAX_FMR_SGES)
 201                 nsegs = RPCRDMA_MAX_FMR_SGES;
 202         for (i = 0; i < nsegs;) {
 203                 if (seg->mr_page)
 204                         sg_set_page(&mr->mr_sg[i],
 205                                     seg->mr_page,
 206                                     seg->mr_len,
 207                                     offset_in_page(seg->mr_offset));
 208                 else
 209                         sg_set_buf(&mr->mr_sg[i], seg->mr_offset,
 210                                    seg->mr_len);
 211                 len += seg->mr_len;
 212                 ++seg;
 213                 ++i;
 214                 /* Check for holes */
 215                 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
 216                     offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
 217                         break;
 218         }
 219         mr->mr_dir = rpcrdma_data_dir(writing);
 220
 221         mr->mr_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
 222                                      mr->mr_sg, i, mr->mr_dir);
 223         if (!mr->mr_nents)
 224                 goto out_dmamap_err;
 225
 226         for (i = 0, dma_pages = mr->fmr.fm_physaddrs; i < mr->mr_nents; i++)
 227                 dma_pages[i] = sg_dma_address(&mr->mr_sg[i]);
 228         rc = ib_map_phys_fmr(mr->fmr.fm_mr, dma_pages, mr->mr_nents,
 229                              dma_pages[0]);
 230         if (rc)
 231                 goto out_maperr;
 232
 233         mr->mr_handle = mr->fmr.fm_mr->rkey;
 234         mr->mr_length = len;
 235         mr->mr_offset = dma_pages[0] + pageoff;
 236
 237         *out = mr;
 238         return seg;
 239
 240 out_dmamap_err:
 241         pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
 242                mr->mr_sg, i);
 243         rpcrdma_mr_put(mr);
 244         return ERR_PTR(-EIO);
 245
 246 out_maperr:
 247         pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
 248                len, (unsigned long long)dma_pages[0],
 249                pageoff, mr->mr_nents, rc);
 250         rpcrdma_mr_unmap_and_put(mr);
 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 @mrs 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 *mrs)
 264 {
 265         struct rpcrdma_mr *mr;
 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(mr, mrs, mr_list) {
 275                 dprintk("RPC:       %s: unmapping fmr %p\n",
 276                         __func__, &mr->fmr);
 277                 trace_xprtrdma_localinv(mr);
 278                 list_add_tail(&mr->fmr.fm_mr->list, &unmap_list);
 279         }
 280         r_xprt->rx_stats.local_inv_needed++;
 281         rc = ib_unmap_fmr(&unmap_list);
 282         if (rc)
 283                 goto out_reset;
 284
 285         /* ORDER: Now DMA unmap all of the req's MRs, and return
 286          * them to the free MW list.
 287          */
 288         while (!list_empty(mrs)) {
 289                 mr = rpcrdma_mr_pop(mrs);
 290                 list_del(&mr->fmr.fm_mr->list);
 291                 rpcrdma_mr_unmap_and_put(mr);
 292         }
 293
 294         return;
 295
 296 out_reset:
 297         pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
 298
 299         while (!list_empty(mrs)) {
 300                 mr = rpcrdma_mr_pop(mrs);
 301                 list_del(&mr->fmr.fm_mr->list);
 302                 fmr_op_recover_mr(mr);
 303         }
 304 }
 305
 306 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
 307         .ro_map                         = fmr_op_map,
 308         .ro_unmap_sync                  = fmr_op_unmap_sync,
 309         .ro_recover_mr                  = fmr_op_recover_mr,
 310         .ro_open                        = fmr_op_open,
 311         .ro_maxpages                    = fmr_op_maxpages,
 312         .ro_init_mr                     = fmr_op_init_mr,
 313         .ro_release_mr                  = fmr_op_release_mr,
 314         .ro_displayname                 = "fmr",
 315         .ro_send_w_inv_ok               = 0,
 316 };