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 */
 32enum {
 33	RPCRDMA_FMR_ACCESS_FLAGS	= IB_ACCESS_REMOTE_WRITE |
 34					  IB_ACCESS_REMOTE_READ,
 35};
 36
 37bool
 38fmr_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
 48static int
 49fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *mw)
 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(ia->ri_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
 76out_fmr_err:
 77	dprintk("RPC:       %s: ib_alloc_fmr returned %ld\n", __func__,
 78		PTR_ERR(mw->fmr.fm_mr));
 79
 80out_free:
 81	kfree(mw->mw_sg);
 82	kfree(mw->fmr.fm_physaddrs);
 83	return -ENOMEM;
 84}
 85
 86static 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
 98static void
 99fmr_op_release_mr(struct rpcrdma_mw *r)
100{
101	LIST_HEAD(unmap_list);
102	int rc;
103
104	/* Ensure MW is not on any rl_registered list */
105	if (!list_empty(&r->mw_list))
106		list_del(&r->mw_list);
107
108	kfree(r->fmr.fm_physaddrs);
109	kfree(r->mw_sg);
110
111	/* In case this one was left mapped, try to unmap it
112	 * to prevent dealloc_fmr from failing with EBUSY
113	 */
114	rc = __fmr_unmap(r);
115	if (rc)
116		pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
117		       r, rc);
118
119	rc = ib_dealloc_fmr(r->fmr.fm_mr);
120	if (rc)
121		pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
122		       r, rc);
123
124	kfree(r);
125}
126
127/* Reset of a single FMR.
128 */
129static void
130fmr_op_recover_mr(struct rpcrdma_mw *mw)
131{
132	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
133	int rc;
134
135	/* ORDER: invalidate first */
136	rc = __fmr_unmap(mw);
137
138	/* ORDER: then DMA unmap */
139	ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
140			mw->mw_sg, mw->mw_nents, mw->mw_dir);
141	if (rc)
142		goto out_release;
143
144	rpcrdma_put_mw(r_xprt, mw);
145	r_xprt->rx_stats.mrs_recovered++;
146	return;
147
148out_release:
149	pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mw);
150	r_xprt->rx_stats.mrs_orphaned++;
151
152	spin_lock(&r_xprt->rx_buf.rb_mwlock);
153	list_del(&mw->mw_all);
154	spin_unlock(&r_xprt->rx_buf.rb_mwlock);
155
156	fmr_op_release_mr(mw);
157}
158
159static int
160fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
161	    struct rpcrdma_create_data_internal *cdata)
162{
163	ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
164				RPCRDMA_MAX_FMR_SGES);
165	return 0;
166}
167
168/* FMR mode conveys up to 64 pages of payload per chunk segment.
169 */
170static size_t
171fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
172{
173	return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
174		     RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
175}
176
177/* Use the ib_map_phys_fmr() verb to register a memory region
178 * for remote access via RDMA READ or RDMA WRITE.
179 */
180static int
181fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
182	   int nsegs, bool writing, struct rpcrdma_mw **out)
183{
184	struct rpcrdma_mr_seg *seg1 = seg;
185	int len, pageoff, i, rc;
186	struct rpcrdma_mw *mw;
187	u64 *dma_pages;
188
189	mw = rpcrdma_get_mw(r_xprt);
190	if (!mw)
191		return -ENOBUFS;
192
193	pageoff = offset_in_page(seg1->mr_offset);
194	seg1->mr_offset -= pageoff;	/* start of page */
195	seg1->mr_len += pageoff;
196	len = -pageoff;
197	if (nsegs > RPCRDMA_MAX_FMR_SGES)
198		nsegs = RPCRDMA_MAX_FMR_SGES;
199	for (i = 0; i < nsegs;) {
200		if (seg->mr_page)
201			sg_set_page(&mw->mw_sg[i],
202				    seg->mr_page,
203				    seg->mr_len,
204				    offset_in_page(seg->mr_offset));
205		else
206			sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
207				   seg->mr_len);
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	mw->mw_nents = i;
217	mw->mw_dir = rpcrdma_data_dir(writing);
218	if (i == 0)
219		goto out_dmamap_err;
220
221	if (!ib_dma_map_sg(r_xprt->rx_ia.ri_device,
222			   mw->mw_sg, mw->mw_nents, mw->mw_dir))
223		goto out_dmamap_err;
224
225	for (i = 0, dma_pages = mw->fmr.fm_physaddrs; i < mw->mw_nents; i++)
226		dma_pages[i] = sg_dma_address(&mw->mw_sg[i]);
227	rc = ib_map_phys_fmr(mw->fmr.fm_mr, dma_pages, mw->mw_nents,
228			     dma_pages[0]);
229	if (rc)
230		goto out_maperr;
231
232	mw->mw_handle = mw->fmr.fm_mr->rkey;
233	mw->mw_length = len;
234	mw->mw_offset = dma_pages[0] + pageoff;
235
236	*out = mw;
237	return mw->mw_nents;
238
239out_dmamap_err:
240	pr_err("rpcrdma: failed to dma map sg %p sg_nents %u\n",
241	       mw->mw_sg, mw->mw_nents);
242	rpcrdma_defer_mr_recovery(mw);
243	return -EIO;
244
245out_maperr:
246	pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
247	       len, (unsigned long long)dma_pages[0],
248	       pageoff, mw->mw_nents, rc);
249	rpcrdma_defer_mr_recovery(mw);
250	return -EIO;
251}
252
253/* Invalidate all memory regions that were registered for "req".
254 *
255 * Sleeps until it is safe for the host CPU to access the
256 * previously mapped memory regions.
257 *
258 * Caller ensures that req->rl_registered is not empty.
259 */
260static void
261fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req)
262{
263	struct rpcrdma_mw *mw, *tmp;
264	LIST_HEAD(unmap_list);
265	int rc;
266
267	dprintk("RPC:       %s: req %p\n", __func__, req);
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, &req->rl_registered, mw_list)
275		list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
276	r_xprt->rx_stats.local_inv_needed++;
277	rc = ib_unmap_fmr(&unmap_list);
278	if (rc)
279		goto out_reset;
280
281	/* ORDER: Now DMA unmap all of the req's MRs, and return
282	 * them to the free MW list.
283	 */
284	list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
285		list_del_init(&mw->mw_list);
286		list_del_init(&mw->fmr.fm_mr->list);
287		ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
288				mw->mw_sg, mw->mw_nents, mw->mw_dir);
289		rpcrdma_put_mw(r_xprt, mw);
290	}
291
292	return;
293
294out_reset:
295	pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
296
297	list_for_each_entry_safe(mw, tmp, &req->rl_registered, mw_list) {
298		list_del_init(&mw->fmr.fm_mr->list);
299		fmr_op_recover_mr(mw);
300	}
301}
302
303/* Use a slow, safe mechanism to invalidate all memory regions
304 * that were registered for "req".
305 */
306static void
307fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
308		  bool sync)
309{
310	struct rpcrdma_mw *mw;
311
312	while (!list_empty(&req->rl_registered)) {
313		mw = list_first_entry(&req->rl_registered,
314				      struct rpcrdma_mw, mw_list);
315		list_del_init(&mw->mw_list);
316
317		if (sync)
318			fmr_op_recover_mr(mw);
319		else
320			rpcrdma_defer_mr_recovery(mw);
321	}
322}
323
324const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
325	.ro_map				= fmr_op_map,
326	.ro_unmap_sync			= fmr_op_unmap_sync,
327	.ro_unmap_safe			= fmr_op_unmap_safe,
328	.ro_recover_mr			= fmr_op_recover_mr,
329	.ro_open			= fmr_op_open,
330	.ro_maxpages			= fmr_op_maxpages,
331	.ro_init_mr			= fmr_op_init_mr,
332	.ro_release_mr			= fmr_op_release_mr,
333	.ro_displayname			= "fmr",
334	.ro_send_w_inv_ok		= 0,
335};