[linux-block.git] / net / sunrpc / xprtrdma / fmr_ops.c

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (c) 2015 Oracle.  All rights reserved.
 * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
 */

/* Lightweight memory registration using Fast Memory Regions (FMR).
 * Referred to sometimes as MTHCAFMR mode.
 *
 * FMR uses synchronous memory registration and deregistration.
 * FMR registration is known to be fast, but FMR deregistration
 * can take tens of usecs to complete.
 */

/* Normal operation
 *
 * A Memory Region is prepared for RDMA READ or WRITE using the
 * ib_map_phys_fmr verb (fmr_op_map). When the RDMA operation is
 * finished, the Memory Region is unmapped using the ib_unmap_fmr
 * verb (fmr_op_unmap).
 */

#include "xprt_rdma.h"

#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
# define RPCDBG_FACILITY	RPCDBG_TRANS
#endif

/* Maximum scatter/gather per FMR */
#define RPCRDMA_MAX_FMR_SGES	(64)

/* Access mode of externally registered pages */
enum {
	RPCRDMA_FMR_ACCESS_FLAGS	= IB_ACCESS_REMOTE_WRITE |
					  IB_ACCESS_REMOTE_READ,
};

bool
fmr_is_supported(struct rpcrdma_ia *ia)
{
	if (!ia->ri_device->alloc_fmr) {
		pr_info("rpcrdma: 'fmr' mode is not supported by device %s\n",
			ia->ri_device->name);
		return false;
	}
	return true;
}

static int
fmr_op_init_mr(struct rpcrdma_ia *ia, struct rpcrdma_mw *mw)
{
	static struct ib_fmr_attr fmr_attr = {
		.max_pages	= RPCRDMA_MAX_FMR_SGES,
		.max_maps	= 1,
		.page_shift	= PAGE_SHIFT
	};

	mw->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
				       sizeof(u64), GFP_KERNEL);
	if (!mw->fmr.fm_physaddrs)
		goto out_free;

	mw->mw_sg = kcalloc(RPCRDMA_MAX_FMR_SGES,
			    sizeof(*mw->mw_sg), GFP_KERNEL);
	if (!mw->mw_sg)
		goto out_free;

	sg_init_table(mw->mw_sg, RPCRDMA_MAX_FMR_SGES);

	mw->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
				     &fmr_attr);
	if (IS_ERR(mw->fmr.fm_mr))
		goto out_fmr_err;

	return 0;

out_fmr_err:
	dprintk("RPC:       %s: ib_alloc_fmr returned %ld\n", __func__,
		PTR_ERR(mw->fmr.fm_mr));

out_free:
	kfree(mw->mw_sg);
	kfree(mw->fmr.fm_physaddrs);
	return -ENOMEM;
}

static int
__fmr_unmap(struct rpcrdma_mw *mw)
{
	LIST_HEAD(l);
	int rc;

	list_add(&mw->fmr.fm_mr->list, &l);
	rc = ib_unmap_fmr(&l);
	list_del(&mw->fmr.fm_mr->list);
	return rc;
}

static void
fmr_op_release_mr(struct rpcrdma_mw *r)
{
	LIST_HEAD(unmap_list);
	int rc;

	/* Ensure MW is not on any rl_registered list */
	if (!list_empty(&r->mw_list))
		list_del(&r->mw_list);

	kfree(r->fmr.fm_physaddrs);
	kfree(r->mw_sg);

	/* In case this one was left mapped, try to unmap it
	 * to prevent dealloc_fmr from failing with EBUSY
	 */
	rc = __fmr_unmap(r);
	if (rc)
		pr_err("rpcrdma: final ib_unmap_fmr for %p failed %i\n",
		       r, rc);

	rc = ib_dealloc_fmr(r->fmr.fm_mr);
	if (rc)
		pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
		       r, rc);

	kfree(r);
}

/* Reset of a single FMR.
 */
static void
fmr_op_recover_mr(struct rpcrdma_mw *mw)
{
	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
	int rc;

	/* ORDER: invalidate first */
	rc = __fmr_unmap(mw);

	/* ORDER: then DMA unmap */
	ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
			mw->mw_sg, mw->mw_nents, mw->mw_dir);
	if (rc)
		goto out_release;

	rpcrdma_put_mw(r_xprt, mw);
	r_xprt->rx_stats.mrs_recovered++;
	return;

out_release:
	pr_err("rpcrdma: FMR reset failed (%d), %p released\n", rc, mw);
	r_xprt->rx_stats.mrs_orphaned++;

	spin_lock(&r_xprt->rx_buf.rb_mwlock);
	list_del(&mw->mw_all);
	spin_unlock(&r_xprt->rx_buf.rb_mwlock);

	fmr_op_release_mr(mw);
}

static int
fmr_op_open(struct rpcrdma_ia *ia, struct rpcrdma_ep *ep,
	    struct rpcrdma_create_data_internal *cdata)
{
	ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
				RPCRDMA_MAX_FMR_SGES);
	return 0;
}

/* FMR mode conveys up to 64 pages of payload per chunk segment.
 */
static size_t
fmr_op_maxpages(struct rpcrdma_xprt *r_xprt)
{
	return min_t(unsigned int, RPCRDMA_MAX_DATA_SEGS,
		     RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
}

/* Use the ib_map_phys_fmr() verb to register a memory region
 * for remote access via RDMA READ or RDMA WRITE.
 */
static struct rpcrdma_mr_seg *
fmr_op_map(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mr_seg *seg,
	   int nsegs, bool writing, struct rpcrdma_mw **out)
{
	struct rpcrdma_mr_seg *seg1 = seg;
	int len, pageoff, i, rc;
	struct rpcrdma_mw *mw;
	u64 *dma_pages;

	mw = rpcrdma_get_mw(r_xprt);
	if (!mw)
		return ERR_PTR(-ENOBUFS);

	pageoff = offset_in_page(seg1->mr_offset);
	seg1->mr_offset -= pageoff;	/* start of page */
	seg1->mr_len += pageoff;
	len = -pageoff;
	if (nsegs > RPCRDMA_MAX_FMR_SGES)
		nsegs = RPCRDMA_MAX_FMR_SGES;
	for (i = 0; i < nsegs;) {
		if (seg->mr_page)
			sg_set_page(&mw->mw_sg[i],
				    seg->mr_page,
				    seg->mr_len,
				    offset_in_page(seg->mr_offset));
		else
			sg_set_buf(&mw->mw_sg[i], seg->mr_offset,
				   seg->mr_len);
		len += seg->mr_len;
		++seg;
		++i;
		/* Check for holes */
		if ((i < nsegs && offset_in_page(seg->mr_offset)) ||
		    offset_in_page((seg-1)->mr_offset + (seg-1)->mr_len))
			break;
	}
	mw->mw_dir = rpcrdma_data_dir(writing);

	mw->mw_nents = ib_dma_map_sg(r_xprt->rx_ia.ri_device,
				     mw->mw_sg, i, mw->mw_dir);
	if (!mw->mw_nents)
		goto out_dmamap_err;

	for (i = 0, dma_pages = mw->fmr.fm_physaddrs; i < mw->mw_nents; i++)
		dma_pages[i] = sg_dma_address(&mw->mw_sg[i]);
	rc = ib_map_phys_fmr(mw->fmr.fm_mr, dma_pages, mw->mw_nents,
			     dma_pages[0]);
	if (rc)
		goto out_maperr;

	mw->mw_handle = mw->fmr.fm_mr->rkey;
	mw->mw_length = len;
	mw->mw_offset = dma_pages[0] + pageoff;

	*out = mw;
	return seg;

out_dmamap_err:
	pr_err("rpcrdma: failed to DMA map sg %p sg_nents %d\n",
	       mw->mw_sg, i);
	rpcrdma_put_mw(r_xprt, mw);
	return ERR_PTR(-EIO);

out_maperr:
	pr_err("rpcrdma: ib_map_phys_fmr %u@0x%llx+%i (%d) status %i\n",
	       len, (unsigned long long)dma_pages[0],
	       pageoff, mw->mw_nents, rc);
	ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
			mw->mw_sg, mw->mw_nents, mw->mw_dir);
	rpcrdma_put_mw(r_xprt, mw);
	return ERR_PTR(-EIO);
}

/* Invalidate all memory regions that were registered for "req".
 *
 * Sleeps until it is safe for the host CPU to access the
 * previously mapped memory regions.
 *
 * Caller ensures that @mws is not empty before the call. This
 * function empties the list.
 */
static void
fmr_op_unmap_sync(struct rpcrdma_xprt *r_xprt, struct list_head *mws)
{
	struct rpcrdma_mw *mw;
	LIST_HEAD(unmap_list);
	int rc;

	/* ORDER: Invalidate all of the req's MRs first
	 *
	 * ib_unmap_fmr() is slow, so use a single call instead
	 * of one call per mapped FMR.
	 */
	list_for_each_entry(mw, mws, mw_list) {
		dprintk("RPC:       %s: unmapping fmr %p\n",
			__func__, &mw->fmr);
		list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
	}
	r_xprt->rx_stats.local_inv_needed++;
	rc = ib_unmap_fmr(&unmap_list);
	if (rc)
		goto out_reset;

	/* ORDER: Now DMA unmap all of the req's MRs, and return
	 * them to the free MW list.
	 */
	while (!list_empty(mws)) {
		mw = rpcrdma_pop_mw(mws);
		dprintk("RPC:       %s: DMA unmapping fmr %p\n",
			__func__, &mw->fmr);
		list_del(&mw->fmr.fm_mr->list);
		ib_dma_unmap_sg(r_xprt->rx_ia.ri_device,
				mw->mw_sg, mw->mw_nents, mw->mw_dir);
		rpcrdma_put_mw(r_xprt, mw);
	}

	return;

out_reset:
	pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);

	while (!list_empty(mws)) {
		mw = rpcrdma_pop_mw(mws);
		list_del(&mw->fmr.fm_mr->list);
		fmr_op_recover_mr(mw);
	}
}

/* Use a slow, safe mechanism to invalidate all memory regions
 * that were registered for "req".
 */
static void
fmr_op_unmap_safe(struct rpcrdma_xprt *r_xprt, struct rpcrdma_req *req,
		  bool sync)
{
	struct rpcrdma_mw *mw;

	while (!list_empty(&req->rl_registered)) {
		mw = rpcrdma_pop_mw(&req->rl_registered);
		if (sync)
			fmr_op_recover_mr(mw);
		else
			rpcrdma_defer_mr_recovery(mw);
	}
}

const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
	.ro_map				= fmr_op_map,
	.ro_unmap_sync			= fmr_op_unmap_sync,
	.ro_unmap_safe			= fmr_op_unmap_safe,
	.ro_recover_mr			= fmr_op_recover_mr,
	.ro_open			= fmr_op_open,
	.ro_maxpages			= fmr_op_maxpages,
	.ro_init_mr			= fmr_op_init_mr,
	.ro_release_mr			= fmr_op_release_mr,
	.ro_displayname			= "fmr",
	.ro_send_w_inv_ok		= 0,
};
Commit	Line	Data
b2441318	1	// SPDX-License-Identifier: GPL-2.0
a0ce85f5 CL	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
fc7fbb59 CL	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
a0ce85f5 CL	23	#include "xprt_rdma.h"
	24
	25	#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
	26	# define RPCDBG_FACILITY RPCDBG_TRANS
	27	#endif
	28
1c9351ee CL	29	/* Maximum scatter/gather per FMR */
	30	#define RPCRDMA_MAX_FMR_SGES (64)
	31
d48b1d29 CL	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
b54054ca CL	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
d48b1d29	49	static int
e2ac236c	50	fmr_op_init_mr(struct rpcrdma_ia ia, struct rpcrdma_mw mw)
d48b1d29 CL	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
88975ebe CL	58	mw->fmr.fm_physaddrs = kcalloc(RPCRDMA_MAX_FMR_SGES,
	59	sizeof(u64), GFP_KERNEL);
	60	if (!mw->fmr.fm_physaddrs)
d48b1d29 CL	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
e2ac236c	70	mw->fmr.fm_mr = ib_alloc_fmr(ia->ri_pd, RPCRDMA_FMR_ACCESS_FLAGS,
88975ebe CL	71	&fmr_attr);
88975ebe CL	72	if (IS_ERR(mw->fmr.fm_mr))
d48b1d29 CL	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__,
88975ebe	79	PTR_ERR(mw->fmr.fm_mr));
d48b1d29 CL	80
	81	out_free:
	82	kfree(mw->mw_sg);
88975ebe	83	kfree(mw->fmr.fm_physaddrs);
d48b1d29 CL	84	return -ENOMEM;
	85	}
	86
ead3f26e CL	87	static int
	88	__fmr_unmap(struct rpcrdma_mw *mw)
	89	{
	90	LIST_HEAD(l);
38f1932e	91	int rc;
ead3f26e	92
88975ebe	93	list_add(&mw->fmr.fm_mr->list, &l);
38f1932e	94	rc = ib_unmap_fmr(&l);
e2f6ef09	95	list_del(&mw->fmr.fm_mr->list);
38f1932e	96	return rc;
ead3f26e CL	97	}
ead3f26e CL	98
d48b1d29	99	static void
e2ac236c	100	fmr_op_release_mr(struct rpcrdma_mw *r)
d48b1d29	101	{
505bbe64	102	LIST_HEAD(unmap_list);
d48b1d29 CL	103	int rc;
d48b1d29 CL	104
9d6b0409 CL	105	/* Ensure MW is not on any rl_registered list */
	106	if (!list_empty(&r->mw_list))
	107	list_del(&r->mw_list);
	108
88975ebe	109	kfree(r->fmr.fm_physaddrs);
d48b1d29 CL	110	kfree(r->mw_sg);
d48b1d29 CL	111
505bbe64 CL	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
88975ebe	120	rc = ib_dealloc_fmr(r->fmr.fm_mr);
d48b1d29 CL	121	if (rc)
	122	pr_err("rpcrdma: final ib_dealloc_fmr for %p returned %i\n",
	123	r, rc);
e2ac236c CL	124
e2ac236c CL	125	kfree(r);
d48b1d29 CL	126	}
d48b1d29 CL	127
505bbe64	128	/* Reset of a single FMR.
ead3f26e CL	129	*/
ead3f26e CL	130	static void
505bbe64	131	fmr_op_recover_mr(struct rpcrdma_mw *mw)
ead3f26e	132	{
505bbe64 CL	133	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
505bbe64 CL	134	int rc;
ead3f26e	135
505bbe64 CL	136	/* ORDER: invalidate first */
505bbe64 CL	137	rc = __fmr_unmap(mw);
ead3f26e	138
505bbe64 CL	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);
2ffc871a CL	142	if (rc)
2ffc871a CL	143	goto out_release;
505bbe64 CL	144
	145	rpcrdma_put_mw(r_xprt, mw);
	146	r_xprt->rx_stats.mrs_recovered++;
2ffc871a CL	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);
ead3f26e CL	158	}
ead3f26e CL	159
3968cb58 CL	160	static int
	161	fmr_op_open(struct rpcrdma_ia ia, struct rpcrdma_ep ep,
	162	struct rpcrdma_create_data_internal *cdata)
	163	{
87cfb9a0 CL	164	ia->ri_max_segs = max_t(unsigned int, 1, RPCRDMA_MAX_DATA_SEGS /
87cfb9a0 CL	165	RPCRDMA_MAX_FMR_SGES);
3968cb58 CL	166	return 0;
	167	}
	168
1c9351ee CL	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,
94931746	175	RPCRDMA_MAX_HDR_SEGS * RPCRDMA_MAX_FMR_SGES);
1c9351ee CL	176	}
1c9351ee CL	177
9c1b4d77 CL	178	/* Use the ib_map_phys_fmr() verb to register a memory region
	179	* for remote access via RDMA READ or RDMA WRITE.
	180	*/
6748b0ca	181	static struct rpcrdma_mr_seg *
9c1b4d77	182	fmr_op_map(struct rpcrdma_xprt r_xprt, struct rpcrdma_mr_seg seg,
9d6b0409	183	int nsegs, bool writing, struct rpcrdma_mw **out)
9c1b4d77	184	{
9c1b4d77	185	struct rpcrdma_mr_seg *seg1 = seg;
9c1b4d77	186	int len, pageoff, i, rc;
fc7fbb59	187	struct rpcrdma_mw *mw;
fcdfb968	188	u64 *dma_pages;
fc7fbb59	189
505bbe64 CL	190	mw = rpcrdma_get_mw(r_xprt);
505bbe64 CL	191	if (!mw)
6748b0ca	192	return ERR_PTR(-ENOBUFS);
9c1b4d77 CL	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;) {
fcdfb968 CL	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);
9c1b4d77 CL	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	}
fcdfb968 CL	217	mw->mw_dir = rpcrdma_data_dir(writing);
fcdfb968 CL	218
1f541895 CL	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)
fcdfb968	222	goto out_dmamap_err;
9c1b4d77	223
fcdfb968 CL	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]);
9c1b4d77 CL	228	if (rc)
	229	goto out_maperr;
	230
9d6b0409 CL	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;
6748b0ca	236	return seg;
fcdfb968 CL	237
fcdfb968 CL	238	out_dmamap_err:
1f541895 CL	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);
6748b0ca	242	return ERR_PTR(-EIO);
9c1b4d77 CL	243
9c1b4d77 CL	244	out_maperr:
fcdfb968 CL	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);
1f541895 CL	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);
6748b0ca	251	return ERR_PTR(-EIO);
9c1b4d77 CL	252	}
9c1b4d77 CL	253
7c7a5390 CL	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.
9d6b0409	258	*
451d26e1 CL	259	* Caller ensures that @mws is not empty before the call. This
451d26e1 CL	260	* function empties the list.
7c7a5390 CL	261	*/
7c7a5390 CL	262	static void
451d26e1	263	fmr_op_unmap_sync(struct rpcrdma_xprt r_xprt, struct list_head mws)
7c7a5390	264	{
e2f6ef09	265	struct rpcrdma_mw *mw;
7c7a5390 CL	266	LIST_HEAD(unmap_list);
	267	int rc;
	268
7c7a5390 CL	269	/* ORDER: Invalidate all of the req's MRs first
	270	*
	271	* ib_unmap_fmr() is slow, so use a single call instead
505bbe64	272	* of one call per mapped FMR.
7c7a5390	273	*/
451d26e1 CL	274	list_for_each_entry(mw, mws, mw_list) {
	275	dprintk("RPC: %s: unmapping fmr %p\n",
	276	__func__, &mw->fmr);
88975ebe	277	list_add_tail(&mw->fmr.fm_mr->list, &unmap_list);
451d26e1	278	}
c8b920bb	279	r_xprt->rx_stats.local_inv_needed++;
7c7a5390 CL	280	rc = ib_unmap_fmr(&unmap_list);
7c7a5390 CL	281	if (rc)
505bbe64	282	goto out_reset;
7c7a5390 CL	283
	284	/* ORDER: Now DMA unmap all of the req's MRs, and return
	285	* them to the free MW list.
	286	*/
e2f6ef09 CL	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);
505bbe64 CL	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);
7c7a5390 CL	295	}
7c7a5390 CL	296
505bbe64 CL	297	return;
	298
	299	out_reset:
	300	pr_err("rpcrdma: ib_unmap_fmr failed (%i)\n", rc);
	301
e2f6ef09 CL	302	while (!list_empty(mws)) {
	303	mw = rpcrdma_pop_mw(mws);
	304	list_del(&mw->fmr.fm_mr->list);
505bbe64	305	fmr_op_recover_mr(mw);
505bbe64	306	}
7c7a5390 CL	307	}
7c7a5390 CL	308
ead3f26e CL	309	/* Use a slow, safe mechanism to invalidate all memory regions
ead3f26e CL	310	* that were registered for "req".
ead3f26e CL	311	*/
	312	static void
	313	fmr_op_unmap_safe(struct rpcrdma_xprt r_xprt, struct rpcrdma_req req,
	314	bool sync)
	315	{
ead3f26e	316	struct rpcrdma_mw *mw;
ead3f26e	317
9d6b0409	318	while (!list_empty(&req->rl_registered)) {
9a5c63e9	319	mw = rpcrdma_pop_mw(&req->rl_registered);
fcdfb968	320	if (sync)
505bbe64	321	fmr_op_recover_mr(mw);
fcdfb968	322	else
505bbe64	323	rpcrdma_defer_mr_recovery(mw);
ead3f26e CL	324	}
	325	}
	326
a0ce85f5	327	const struct rpcrdma_memreg_ops rpcrdma_fmr_memreg_ops = {
9c1b4d77	328	.ro_map = fmr_op_map,
7c7a5390	329	.ro_unmap_sync = fmr_op_unmap_sync,
ead3f26e	330	.ro_unmap_safe = fmr_op_unmap_safe,
505bbe64	331	.ro_recover_mr = fmr_op_recover_mr,
3968cb58	332	.ro_open = fmr_op_open,
1c9351ee	333	.ro_maxpages = fmr_op_maxpages,
e2ac236c CL	334	.ro_init_mr = fmr_op_init_mr,
e2ac236c CL	335	.ro_release_mr = fmr_op_release_mr,
a0ce85f5	336	.ro_displayname = "fmr",
c8b920bb	337	.ro_send_w_inv_ok = 0,
a0ce85f5	338	};