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 /* Transport recovery
  23  *
  24  * After a transport reconnect, fmr_op_map re-uses the MR already
  25  * allocated for the RPC, but generates a fresh rkey then maps the
  26  * MR again. This process is synchronous.
  27  */
  28
  29 #include "xprt_rdma.h"
  30
  31 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  32 # define RPCDBG_FACILITY        RPCDBG_TRANS
  33 #endif
  34
  35 /* Maximum scatter/gather per FMR */
  36 #define RPCRDMA_MAX_FMR_SGES    (64)
  37
  38 static int
  39 fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
  40             struct rpcrdma_create_data_internal *cdata)
  41 {
  42         return 0;
  43 }
  44
  45 /* FMR mode conveys up to 64 pages of payload per chunk segment.
  46  */
  47 static size_t
  48 fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
  49 {
  50         return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
  51                      rpcrdma_max_segments(r_xprt) * RPCRDMA_MAX_FMR_SGES);
  52 }
  53
  54 static int
  55 fmr_op_init(struct rpcrdma_xprt *r_xprt)
  56 {
  57         struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
  58         int mr_access_flags = IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_READ;
  59         struct ib_fmr_attr fmr_attr = {
  60                 .max_pages      = RPCRDMA_MAX_FMR_SGES,
  61                 .max_maps       = 1,
  62                 .page_shift     = PAGE_SHIFT
  63         };
  64         struct ib_pd *pd = r_xprt->rx_ia.ri_pd;
  65         struct rpcrdma_mw *r;
  66         int i, rc;
  67
  68         spin_lock_init(&buf->rb_mwlock);
  69         INIT_LIST_HEAD(&buf->rb_mws);
  70         INIT_LIST_HEAD(&buf->rb_all);
  71
  72         i = (buf->rb_max_requests + 1) * RPCRDMA_MAX_SEGS;
  73         dprintk("RPC:       %s: initializing %d FMRs\n", __func__, i);
  74
  75         while (i--) {
  76                 r = kzalloc(sizeof(*r), GFP_KERNEL);
  77                 if (!r)
  78                         return -ENOMEM;
  79
  80                 r->r.fmr = ib_alloc_fmr(pd, mr_access_flags, &fmr_attr);
  81                 if (IS_ERR(r->r.fmr))
  82                         goto out_fmr_err;
  83
  84                 list_add(&r->mw_list, &buf->rb_mws);
  85                 list_add(&r->mw_all, &buf->rb_all);
  86         }
  87         return 0;
  88
  89 out_fmr_err:
  90         rc = PTR_ERR(r->r.fmr);
  91         dprintk("RPC:       %s: ib_alloc_fmr status %i\n", __func__, rc);
  92         kfree(r);
  93         return rc;
  94 }
  95
  96 static int
  97 __fmr_unmap(struct rpcrdma_mw *r)
  98 {
  99         LIST_HEAD(l);
 100
 101         list_add(&r->r.fmr->list, &l);
 102         return ib_unmap_fmr(&l);
 103 }
 104
 105 /* Use the ib_map_phys_fmr() verb to register a memory region
 106  * for remote access via RDMA READ or RDMA WRITE.
 107  */
 108 static int
 109 fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
 110            int nsegs, bool writing)
 111 {
 112         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 113         struct ib_device *device = ia->ri_device;
 114         enum dma_data_direction direction = rpcrdma_data_dir(writing);
 115         struct rpcrdma_mr_seg *seg1 = seg;
 116         u64 physaddrs[RPCRDMA_MAX_DATA_SEGS];
 117         int len, pageoff, i, rc;
 118         struct rpcrdma_mw *mw;
 119
 120         mw = seg1->rl_mw;
 121         seg1->rl_mw = NULL;
 122         if (!mw) {
 123                 mw = rpcrdma_get_mw(r_xprt);
 124                 if (!mw)
 125                         return -ENOMEM;
 126         } else {
 127                 /* this is a retransmit; generate a fresh rkey */
 128                 rc = __fmr_unmap(mw);
 129                 if (rc)
 130                         return rc;
 131         }
 132
 133         pageoff = offset_in_page(seg1->mr_offset);
 134         seg1->mr_offset -= pageoff;     /* start of page */
 135         seg1->mr_len += pageoff;
 136         len = -pageoff;
 137         if (nsegs > RPCRDMA_MAX_FMR_SGES)
 138                 nsegs = RPCRDMA_MAX_FMR_SGES;
 139         for (i = 0; i < nsegs;) {
 140                 rpcrdma_map_one(device, seg, direction);
 141                 physaddrs[i] = seg->mr_dma;
 142                 len += seg->mr_len;
 143                 ++seg;
 144                 ++i;
 145                 /* Check for holes */
 146                 if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
 147                     offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
 148                         break;
 149         }
 150
 151         rc = ib_map_phys_fmr(mw->r.fmr, physaddrs, i, seg1->mr_dma);
 152         if (rc)
 153                 goto out_maperr;
 154
 155         seg1->rl_mw = mw;
 156         seg1->mr_rkey = mw->r.fmr->rkey;
 157         seg1->mr_base = seg1->mr_dma + pageoff;
 158         seg1->mr_nsegs = i;
 159         seg1->mr_len = len;
 160         return i;
 161
 162 out_maperr:
 163         dprintk("RPC:       %s: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
 164                 __func__, len, (unsigned long long)seg1->mr_dma,
 165                 pageoff, i, rc);
 166         while (i--)
 167                 rpcrdma_unmap_one(device, --seg);
 168         return rc;
 169 }
 170
 171 /* Use the ib_unmap_fmr() verb to prevent further remote
 172  * access via RDMA READ or RDMA WRITE.
 173  */
 174 static int
 175 fmr_op_unmap(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg)
 176 {
 177         struct rpcrdma_ia *ia = &r_xprt->rx_ia;
 178         struct rpcrdma_mr_seg *seg1 = seg;
 179         struct rpcrdma_mw *mw = seg1->rl_mw;
 180         int rc, nsegs = seg->mr_nsegs;
 181
 182         dprintk("RPC:       %s: FMR %p\n", __func__, mw);
 183
 184         seg1->rl_mw = NULL;
 185         while (seg1->mr_nsegs--)
 186                 rpcrdma_unmap_one(ia->ri_device, seg++);
 187         rc = __fmr_unmap(mw);
 188         if (rc)
 189                 goto out_err;
 190         rpcrdma_put_mw(r_xprt, mw);
 191         return nsegs;
 192
 193 out_err:
 194         /* The FMR is abandoned, but remains in rb_all. fmr_op_destroy
 195          * will attempt to release it when the transport is destroyed.
 196          */
 197         dprintk("RPC:       %s: ib_unmap_fmr status %i\n", __func__, rc);
 198         return nsegs;
 199 }
 200
 201 static void
 202 fmr_op_destroy(struct rpcrdma_buffer *buf)
 203 {
 204         struct rpcrdma_mw *r;
 205         int rc;
 206
 207         while (!list_empty(&buf->rb_all)) {
 208                 r = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
 209                 list_del(&r->mw_all);
 210                 rc = ib_dealloc_fmr(r->r.fmr);
 211                 if (rc)
 212                         dprintk("RPC:       %s: ib_dealloc_fmr failed %i\n",
 213                                 __func__, rc);
 214                 kfree(r);
 215         }
 216 }
 217
 218 const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
 219         .ro_map                         = fmr_op_map,
 220         .ro_unmap                       = fmr_op_unmap,
 221         .ro_open                        = fmr_op_open,
 222         .ro_maxpages                    = fmr_op_maxpages,
 223         .ro_init                        = fmr_op_init,
 224         .ro_destroy                     = fmr_op_destroy,
 225         .ro_displayname                 = "fmr",
 226 };