IB/{hfi1, qib, rdmavt}: Move copy SGE logic into rdmavt

This patch moves hfi1_copy_sge() into rdmavt for sharing with qib.
This patch also moves all the wss_*() functions into rdmavt as
several wss_*() functions are called from hfi1_copy_sge()

When SGE copy mode is adaptive, cacheless copy may be done in some cases
for performance reasons. In those cases, X86 cacheless copy function
is called since the drivers that use rdmavt and may set SGE copy mode
to adaptive are X86 only. For this reason, this patch adds
"depends on X86_64" to rdmavt/Kconfig.

Reviewed-by: Ashutosh Dixit <[e-mail]>
Reviewed-by: Michael J. Ruhl <[e-mail]>
Reviewed-by: Mike Marciniszyn <[e-mail]>
Reviewed-by: Dennis Dalessandro <[e-mail]>
Signed-off-by: Brian Welty <[e-mail]>
Signed-off-by: Harish Chegondi <[e-mail]>
Signed-off-by: Dennis Dalessandro <[e-mail]>
Signed-off-by: Jason Gunthorpe <[e-mail]>

drivers/infiniband/hw/hfi1/init.c

···
 	idr_init(&hfi1_unit_table);
 
 	hfi1_dbg_init();
-	ret = hfi1_wss_init();
-	if (ret < 0)
-		goto bail_wss;
 	ret = pci_register_driver(&hfi1_pci_driver);
 	if (ret < 0) {
 		pr_err("Unable to register driver: error %d\n", -ret);
···
 	goto bail; /* all OK */
 
 bail_dev:
-	hfi1_wss_exit();
-bail_wss:
 	hfi1_dbg_exit();
 	idr_destroy(&hfi1_unit_table);
 	dev_cleanup();
···
 {
 	pci_unregister_driver(&hfi1_pci_driver);
 	node_affinity_destroy_all();
-	hfi1_wss_exit();
 	hfi1_dbg_exit();
 
 	idr_destroy(&hfi1_unit_table);

drivers/infiniband/hw/hfi1/rc.c

···
 		qp->s_rdma_read_len -= pmtu;
 		update_last_psn(qp, psn);
 		spin_unlock_irqrestore(&qp->s_lock, flags);
-		hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, false, false);
+		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
+			     data, pmtu, false, false);
 		goto bail;
 
 	case OP(RDMA_READ_RESPONSE_ONLY):
···
 		if (unlikely(tlen != qp->s_rdma_read_len))
 			goto ack_len_err;
 		aeth = be32_to_cpu(ohdr->u.aeth);
-		hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, false, false);
+		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
+			     data, tlen, false, false);
 		WARN_ON(qp->s_rdma_read_sge.num_sge);
 		(void)do_rc_ack(qp, aeth, psn,
 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto nack_inv;
-		hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
 		break;
 
 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
···
 		wc.byte_len = tlen + qp->r_rcv_len;
 		if (unlikely(wc.byte_len > qp->r_len))
 			goto nack_inv;
-		hfi1_copy_sge(&qp->r_sge, data, tlen, true, copy_last);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, copy_last);
 		rvt_put_ss(&qp->r_sge);
 		qp->r_msn++;
 		if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))

drivers/infiniband/hw/hfi1/ruc.c

···
 		if (len > sge->sge_length)
 			len = sge->sge_length;
 		WARN_ON_ONCE(len == 0);
-		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
+		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr,
+			     len, release, copy_last);
 		sge->vaddr += len;
 		sge->length -= len;
 		sge->sge_length -= len;

drivers/infiniband/hw/hfi1/uc.c

···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto rewind;
-		hfi1_copy_sge(&qp->r_sge, data, pmtu, false, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, false, false);
 		break;
 
 	case OP(SEND_LAST_WITH_IMMEDIATE):
···
 		if (unlikely(wc.byte_len > qp->r_len))
 			goto rewind;
 		wc.opcode = IB_WC_RECV;
-		hfi1_copy_sge(&qp->r_sge, data, tlen, false, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, false, false);
 		rvt_put_ss(&qp->s_rdma_read_sge);
 last_imm:
 		wc.wr_id = qp->r_wr_id;
···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto drop;
-		hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
 		break;
 
 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
···
 		}
 		wc.byte_len = qp->r_len;
 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
-		hfi1_copy_sge(&qp->r_sge, data, tlen, true, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
 		rvt_put_ss(&qp->r_sge);
 		goto last_imm;
 
···
 		tlen -= (hdrsize + extra_bytes);
 		if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
 			goto drop;
-		hfi1_copy_sge(&qp->r_sge, data, tlen, true, false);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
 		rvt_put_ss(&qp->r_sge);
 		break;
 

drivers/infiniband/hw/hfi1/ud.c

···
 		}
 
 		hfi1_make_grh(ibp, &grh, &grd, 0, 0);
-		hfi1_copy_sge(&qp->r_sge, &grh,
-			      sizeof(grh), true, false);
+		rvt_copy_sge(qp, &qp->r_sge, &grh,
+			     sizeof(grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else {
 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
···
 		if (len > sge->sge_length)
 			len = sge->sge_length;
 		WARN_ON_ONCE(len == 0);
-		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, true, false);
+		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, true, false);
 		sge->vaddr += len;
 		sge->length -= len;
 		sge->sge_length -= len;
···
 		goto drop;
 	}
 	if (packet->grh) {
-		hfi1_copy_sge(&qp->r_sge, packet->grh,
-			      sizeof(struct ib_grh), true, false);
+		rvt_copy_sge(qp, &qp->r_sge, packet->grh,
+			     sizeof(struct ib_grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else if (packet->etype == RHF_RCV_TYPE_BYPASS) {
 		struct ib_grh grh;
···
 		 * out when creating 16B, add back the GRH here.
 		 */
 		hfi1_make_ext_grh(packet, &grh, slid, dlid);
-		hfi1_copy_sge(&qp->r_sge, &grh,
-			      sizeof(struct ib_grh), true, false);
+		rvt_copy_sge(qp, &qp->r_sge, &grh,
+			     sizeof(struct ib_grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else {
 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
 	}
-	hfi1_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
-		      true, false);
+	rvt_copy_sge(qp, &qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
+		     true, false);
 	rvt_put_ss(&qp->r_sge);
 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
 		return;

drivers/infiniband/hw/hfi1/verbs.c

···
 module_param(piothreshold, ushort, S_IRUGO);
 MODULE_PARM_DESC(piothreshold, "size used to determine sdma vs. pio");
 
-#define COPY_CACHELESS 1
-#define COPY_ADAPTIVE  2
 static unsigned int sge_copy_mode;
 module_param(sge_copy_mode, uint, S_IRUGO);
 MODULE_PARM_DESC(sge_copy_mode,
···
 /* 16B trailing buffer */
 static const u8 trail_buf[MAX_16B_PADDING];
 
-static uint wss_threshold;
+static uint wss_threshold = 80;
 module_param(wss_threshold, uint, S_IRUGO);
 MODULE_PARM_DESC(wss_threshold, "Percentage (1-100) of LLC to use as a threshold for a cacheless copy");
 static uint wss_clean_period = 256;
 module_param(wss_clean_period, uint, S_IRUGO);
 MODULE_PARM_DESC(wss_clean_period, "Count of verbs copies before an entry in the page copy table is cleaned");
-
-/* memory working set size */
-struct hfi1_wss {
-	unsigned long *entries;
-	atomic_t total_count;
-	atomic_t clean_counter;
-	atomic_t clean_entry;
-
-	int threshold;
-	int num_entries;
-	long pages_mask;
-};
-
-static struct hfi1_wss wss;
-
-int hfi1_wss_init(void)
-{
-	long llc_size;
-	long llc_bits;
-	long table_size;
-	long table_bits;
-
-	/* check for a valid percent range - default to 80 if none or invalid */
-	if (wss_threshold < 1 || wss_threshold > 100)
-		wss_threshold = 80;
-	/* reject a wildly large period */
-	if (wss_clean_period > 1000000)
-		wss_clean_period = 256;
-	/* reject a zero period */
-	if (wss_clean_period == 0)
-		wss_clean_period = 1;
-
-	/*
-	 * Calculate the table size - the next power of 2 larger than the
-	 * LLC size.  LLC size is in KiB.
-	 */
-	llc_size = wss_llc_size() * 1024;
-	table_size = roundup_pow_of_two(llc_size);
-
-	/* one bit per page in rounded up table */
-	llc_bits = llc_size / PAGE_SIZE;
-	table_bits = table_size / PAGE_SIZE;
-	wss.pages_mask = table_bits - 1;
-	wss.num_entries = table_bits / BITS_PER_LONG;
-
-	wss.threshold = (llc_bits * wss_threshold) / 100;
-	if (wss.threshold == 0)
-		wss.threshold = 1;
-
-	atomic_set(&wss.clean_counter, wss_clean_period);
-
-	wss.entries = kcalloc(wss.num_entries, sizeof(*wss.entries),
-			      GFP_KERNEL);
-	if (!wss.entries) {
-		hfi1_wss_exit();
-		return -ENOMEM;
-	}
-
-	return 0;
-}
-
-void hfi1_wss_exit(void)
-{
-	/* coded to handle partially initialized and repeat callers */
-	kfree(wss.entries);
-	wss.entries = NULL;
-}
-
-/*
- * Advance the clean counter.  When the clean period has expired,
- * clean an entry.
- *
- * This is implemented in atomics to avoid locking.  Because multiple
- * variables are involved, it can be racy which can lead to slightly
- * inaccurate information.  Since this is only a heuristic, this is
- * OK.  Any innaccuracies will clean themselves out as the counter
- * advances.  That said, it is unlikely the entry clean operation will
- * race - the next possible racer will not start until the next clean
- * period.
- *
- * The clean counter is implemented as a decrement to zero.  When zero
- * is reached an entry is cleaned.
- */
-static void wss_advance_clean_counter(void)
-{
-	int entry;
-	int weight;
-	unsigned long bits;
-
-	/* become the cleaner if we decrement the counter to zero */
-	if (atomic_dec_and_test(&wss.clean_counter)) {
-		/*
-		 * Set, not add, the clean period.  This avoids an issue
-		 * where the counter could decrement below the clean period.
-		 * Doing a set can result in lost decrements, slowing the
-		 * clean advance.  Since this a heuristic, this possible
-		 * slowdown is OK.
-		 *
-		 * An alternative is to loop, advancing the counter by a
-		 * clean period until the result is > 0. However, this could
-		 * lead to several threads keeping another in the clean loop.
-		 * This could be mitigated by limiting the number of times
-		 * we stay in the loop.
-		 */
-		atomic_set(&wss.clean_counter, wss_clean_period);
-
-		/*
-		 * Uniquely grab the entry to clean and move to next.
-		 * The current entry is always the lower bits of
-		 * wss.clean_entry.  The table size, wss.num_entries,
-		 * is always a power-of-2.
-		 */
-		entry = (atomic_inc_return(&wss.clean_entry) - 1)
-			& (wss.num_entries - 1);
-
-		/* clear the entry and count the bits */
-		bits = xchg(&wss.entries[entry], 0);
-		weight = hweight64((u64)bits);
-		/* only adjust the contended total count if needed */
-		if (weight)
-			atomic_sub(weight, &wss.total_count);
-	}
-}
-
-/*
- * Insert the given address into the working set array.
- */
-static void wss_insert(void *address)
-{
-	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss.pages_mask;
-	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
-	u32 nr = page & (BITS_PER_LONG - 1);
-
-	if (!test_and_set_bit(nr, &wss.entries[entry]))
-		atomic_inc(&wss.total_count);
-
-	wss_advance_clean_counter();
-}
-
-/*
- * Is the working set larger than the threshold?
- */
-static inline bool wss_exceeds_threshold(void)
-{
-	return atomic_read(&wss.total_count) >= wss.threshold;
-}
 
 /*
  * Translate ib_wr_opcode into ib_wc_opcode.
···
  * System image GUID.
  */
 __be64 ib_hfi1_sys_image_guid;
-
-/**
- * hfi1_copy_sge - copy data to SGE memory
- * @ss: the SGE state
- * @data: the data to copy
- * @length: the length of the data
- * @release: boolean to release MR
- * @copy_last: do a separate copy of the last 8 bytes
- */
-void hfi1_copy_sge(
-	struct rvt_sge_state *ss,
-	void *data, u32 length,
-	bool release,
-	bool copy_last)
-{
-	struct rvt_sge *sge = &ss->sge;
-	int i;
-	bool in_last = false;
-	bool cacheless_copy = false;
-
-	if (sge_copy_mode == COPY_CACHELESS) {
-		cacheless_copy = length >= PAGE_SIZE;
-	} else if (sge_copy_mode == COPY_ADAPTIVE) {
-		if (length >= PAGE_SIZE) {
-			/*
-			 * NOTE: this *assumes*:
-			 * o The first vaddr is the dest.
-			 * o If multiple pages, then vaddr is sequential.
-			 */
-			wss_insert(sge->vaddr);
-			if (length >= (2 * PAGE_SIZE))
-				wss_insert(sge->vaddr + PAGE_SIZE);
-
-			cacheless_copy = wss_exceeds_threshold();
-		} else {
-			wss_advance_clean_counter();
-		}
-	}
-	if (copy_last) {
-		if (length > 8) {
-			length -= 8;
-		} else {
-			copy_last = false;
-			in_last = true;
-		}
-	}
-
-again:
-	while (length) {
-		u32 len = rvt_get_sge_length(sge, length);
-
-		WARN_ON_ONCE(len == 0);
-		if (unlikely(in_last)) {
-			/* enforce byte transfer ordering */
-			for (i = 0; i < len; i++)
-				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
-		} else if (cacheless_copy) {
-			cacheless_memcpy(sge->vaddr, data, len);
-		} else {
-			memcpy(sge->vaddr, data, len);
-		}
-		rvt_update_sge(ss, len, release);
-		data += len;
-		length -= len;
-	}
-
-	if (copy_last) {
-		copy_last = false;
-		in_last = true;
-		length = 8;
-		goto again;
-	}
-}
 
 /*
  * Make sure the QP is ready and able to accept the given opcode.
···
 	dd->verbs_dev.rdi.dparms.lkey_table_size = hfi1_lkey_table_size;
 	dd->verbs_dev.rdi.dparms.nports = dd->num_pports;
 	dd->verbs_dev.rdi.dparms.npkeys = hfi1_get_npkeys(dd);
+	dd->verbs_dev.rdi.dparms.sge_copy_mode = sge_copy_mode;
+	dd->verbs_dev.rdi.dparms.wss_threshold = wss_threshold;
+	dd->verbs_dev.rdi.dparms.wss_clean_period = wss_clean_period;
 
 	/* post send table */
 	dd->verbs_dev.rdi.post_parms = hfi1_post_parms;

drivers/infiniband/hw/hfi1/verbs.h

···
 
 int hfi1_verbs_send(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
 
-void hfi1_copy_sge(struct rvt_sge_state *ss, void *data, u32 length,
-		   bool release, bool copy_last);
-
 void hfi1_cnp_rcv(struct hfi1_packet *packet);
 
 void hfi1_uc_rcv(struct hfi1_packet *packet);
···
 
 int hfi1_verbs_send_pio(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
 			u64 pbc);
-
-int hfi1_wss_init(void);
-void hfi1_wss_exit(void);
-
-/* platform specific: return the lowest level cache (llc) size, in KiB */
-static inline int wss_llc_size(void)
-{
-	/* assume that the boot CPU value is universal for all CPUs */
-	return boot_cpu_data.x86_cache_size;
-}
-
-/* platform specific: cacheless copy */
-static inline void cacheless_memcpy(void *dst, void *src, size_t n)
-{
-	/*
-	 * Use the only available X64 cacheless copy.  Add a __user cast
-	 * to quiet sparse.  The src agument is already in the kernel so
-	 * there are no security issues.  The extra fault recovery machinery
-	 * is not invoked.
-	 */
-	__copy_user_nocache(dst, (void __user *)src, n, 0);
-}
 
 static inline bool opa_bth_is_migration(struct ib_other_headers *ohdr)
 {

drivers/infiniband/hw/qib/qib_rc.c

···
 		qp->s_rdma_read_len -= pmtu;
 		update_last_psn(qp, psn);
 		spin_unlock_irqrestore(&qp->s_lock, flags);
-		qib_copy_sge(&qp->s_rdma_read_sge, data, pmtu, 0);
+		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
+			     data, pmtu, false, false);
 		goto bail;
 
 	case OP(RDMA_READ_RESPONSE_ONLY):
···
 		if (unlikely(tlen != qp->s_rdma_read_len))
 			goto ack_len_err;
 		aeth = be32_to_cpu(ohdr->u.aeth);
-		qib_copy_sge(&qp->s_rdma_read_sge, data, tlen, 0);
+		rvt_copy_sge(qp, &qp->s_rdma_read_sge,
+			     data, tlen, false, false);
 		WARN_ON(qp->s_rdma_read_sge.num_sge);
 		(void) do_rc_ack(qp, aeth, psn,
 				 OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto nack_inv;
-		qib_copy_sge(&qp->r_sge, data, pmtu, 1);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
 		break;
 
 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
···
 		wc.byte_len = tlen + qp->r_rcv_len;
 		if (unlikely(wc.byte_len > qp->r_len))
 			goto nack_inv;
-		qib_copy_sge(&qp->r_sge, data, tlen, 1);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
 		rvt_put_ss(&qp->r_sge);
 		qp->r_msn++;
 		if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))

drivers/infiniband/hw/qib/qib_ruc.c

···
 		if (len > sge->sge_length)
 			len = sge->sge_length;
 		BUG_ON(len == 0);
-		qib_copy_sge(&qp->r_sge, sge->vaddr, len, release);
+		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, release, false);
 		sge->vaddr += len;
 		sge->length -= len;
 		sge->sge_length -= len;

drivers/infiniband/hw/qib/qib_uc.c

···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto rewind;
-		qib_copy_sge(&qp->r_sge, data, pmtu, 0);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, false, false);
 		break;
 
 	case OP(SEND_LAST_WITH_IMMEDIATE):
···
 		if (unlikely(wc.byte_len > qp->r_len))
 			goto rewind;
 		wc.opcode = IB_WC_RECV;
-		qib_copy_sge(&qp->r_sge, data, tlen, 0);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, false, false);
 		rvt_put_ss(&qp->s_rdma_read_sge);
 last_imm:
 		wc.wr_id = qp->r_wr_id;
···
 		qp->r_rcv_len += pmtu;
 		if (unlikely(qp->r_rcv_len > qp->r_len))
 			goto drop;
-		qib_copy_sge(&qp->r_sge, data, pmtu, 1);
+		rvt_copy_sge(qp, &qp->r_sge, data, pmtu, true, false);
 		break;
 
 	case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
···
 		}
 		wc.byte_len = qp->r_len;
 		wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
-		qib_copy_sge(&qp->r_sge, data, tlen, 1);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
 		rvt_put_ss(&qp->r_sge);
 		goto last_imm;
 
···
 		tlen -= (hdrsize + pad + 4);
 		if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
 			goto drop;
-		qib_copy_sge(&qp->r_sge, data, tlen, 1);
+		rvt_copy_sge(qp, &qp->r_sge, data, tlen, true, false);
 		rvt_put_ss(&qp->r_sge);
 		break;
 

drivers/infiniband/hw/qib/qib_ud.c

···
 		const struct ib_global_route *grd = rdma_ah_read_grh(ah_attr);
 
 		qib_make_grh(ibp, &grh, grd, 0, 0);
-		qib_copy_sge(&qp->r_sge, &grh,
-			     sizeof(grh), 1);
+		rvt_copy_sge(qp, &qp->r_sge, &grh,
+			     sizeof(grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else
 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
···
 		if (len > sge->sge_length)
 			len = sge->sge_length;
 		BUG_ON(len == 0);
-		qib_copy_sge(&qp->r_sge, sge->vaddr, len, 1);
+		rvt_copy_sge(qp, &qp->r_sge, sge->vaddr, len, true, false);
 		sge->vaddr += len;
 		sge->length -= len;
 		sge->sge_length -= len;
···
 		goto drop;
 	}
 	if (has_grh) {
-		qib_copy_sge(&qp->r_sge, &hdr->u.l.grh,
-			     sizeof(struct ib_grh), 1);
+		rvt_copy_sge(qp, &qp->r_sge, &hdr->u.l.grh,
+			     sizeof(struct ib_grh), true, false);
 		wc.wc_flags |= IB_WC_GRH;
 	} else
 		rvt_skip_sge(&qp->r_sge, sizeof(struct ib_grh), true);
-	qib_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh), 1);
+	rvt_copy_sge(qp, &qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
+		     true, false);
 	rvt_put_ss(&qp->r_sge);
 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
 		return;

drivers/infiniband/hw/qib/qib_verbs.c

···
  */
 __be64 ib_qib_sys_image_guid;
 
-/**
- * qib_copy_sge - copy data to SGE memory
- * @ss: the SGE state
- * @data: the data to copy
- * @length: the length of the data
- */
-void qib_copy_sge(struct rvt_sge_state *ss, void *data, u32 length, int release)
-{
-	struct rvt_sge *sge = &ss->sge;
-
-	while (length) {
-		u32 len = rvt_get_sge_length(sge, length);
-
-		WARN_ON_ONCE(len == 0);
-		memcpy(sge->vaddr, data, len);
-		rvt_update_sge(ss, len, release);
-		data += len;
-		length -= len;
-	}
-}
-
 /*
  * Count the number of DMA descriptors needed to send length bytes of data.
  * Don't modify the qib_sge_state to get the count.
···
 	dd->verbs_dev.rdi.dparms.node = dd->assigned_node_id;
 	dd->verbs_dev.rdi.dparms.core_cap_flags = RDMA_CORE_PORT_IBA_IB;
 	dd->verbs_dev.rdi.dparms.max_mad_size = IB_MGMT_MAD_SIZE;
+	dd->verbs_dev.rdi.dparms.sge_copy_mode = RVT_SGE_COPY_MEMCPY;
 
 	qib_fill_device_attr(dd);
 

drivers/infiniband/hw/qib/qib_verbs.h

···
 int qib_verbs_send(struct rvt_qp *qp, struct ib_header *hdr,
 		   u32 hdrwords, struct rvt_sge_state *ss, u32 len);
 
-void qib_copy_sge(struct rvt_sge_state *ss, void *data, u32 length,
-		  int release);
-
 void qib_uc_rcv(struct qib_ibport *ibp, struct ib_header *hdr,
 		int has_grh, void *data, u32 tlen, struct rvt_qp *qp);
 

drivers/infiniband/sw/rdmavt/Kconfig

···
 config INFINIBAND_RDMAVT
 	tristate "RDMA verbs transport library"
-	depends on 64BIT && ARCH_DMA_ADDR_T_64BIT
+	depends on X86_64 && ARCH_DMA_ADDR_T_64BIT
 	depends on PCI
 	select DMA_VIRT_OPS
 	---help---

drivers/infiniband/sw/rdmavt/qp.c

···
 };
 EXPORT_SYMBOL(ib_rvt_state_ops);
 
+/* platform specific: return the last level cache (llc) size, in KiB */
+static int rvt_wss_llc_size(void)
+{
+	/* assume that the boot CPU value is universal for all CPUs */
+	return boot_cpu_data.x86_cache_size;
+}
+
+/* platform specific: cacheless copy */
+static void cacheless_memcpy(void *dst, void *src, size_t n)
+{
+	/*
+	 * Use the only available X64 cacheless copy.  Add a __user cast
+	 * to quiet sparse.  The src agument is already in the kernel so
+	 * there are no security issues.  The extra fault recovery machinery
+	 * is not invoked.
+	 */
+	__copy_user_nocache(dst, (void __user *)src, n, 0);
+}
+
+void rvt_wss_exit(struct rvt_dev_info *rdi)
+{
+	struct rvt_wss *wss = rdi->wss;
+
+	if (!wss)
+		return;
+
+	/* coded to handle partially initialized and repeat callers */
+	kfree(wss->entries);
+	wss->entries = NULL;
+	kfree(rdi->wss);
+	rdi->wss = NULL;
+}
+
+/**
+ * rvt_wss_init - Init wss data structures
+ *
+ * Return: 0 on success
+ */
+int rvt_wss_init(struct rvt_dev_info *rdi)
+{
+	unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode;
+	unsigned int wss_threshold = rdi->dparms.wss_threshold;
+	unsigned int wss_clean_period = rdi->dparms.wss_clean_period;
+	long llc_size;
+	long llc_bits;
+	long table_size;
+	long table_bits;
+	struct rvt_wss *wss;
+	int node = rdi->dparms.node;
+
+	if (sge_copy_mode != RVT_SGE_COPY_ADAPTIVE) {
+		rdi->wss = NULL;
+		return 0;
+	}
+
+	rdi->wss = kzalloc_node(sizeof(*rdi->wss), GFP_KERNEL, node);
+	if (!rdi->wss)
+		return -ENOMEM;
+	wss = rdi->wss;
+
+	/* check for a valid percent range - default to 80 if none or invalid */
+	if (wss_threshold < 1 || wss_threshold > 100)
+		wss_threshold = 80;
+
+	/* reject a wildly large period */
+	if (wss_clean_period > 1000000)
+		wss_clean_period = 256;
+
+	/* reject a zero period */
+	if (wss_clean_period == 0)
+		wss_clean_period = 1;
+
+	/*
+	 * Calculate the table size - the next power of 2 larger than the
+	 * LLC size.  LLC size is in KiB.
+	 */
+	llc_size = rvt_wss_llc_size() * 1024;
+	table_size = roundup_pow_of_two(llc_size);
+
+	/* one bit per page in rounded up table */
+	llc_bits = llc_size / PAGE_SIZE;
+	table_bits = table_size / PAGE_SIZE;
+	wss->pages_mask = table_bits - 1;
+	wss->num_entries = table_bits / BITS_PER_LONG;
+
+	wss->threshold = (llc_bits * wss_threshold) / 100;
+	if (wss->threshold == 0)
+		wss->threshold = 1;
+
+	wss->clean_period = wss_clean_period;
+	atomic_set(&wss->clean_counter, wss_clean_period);
+
+	wss->entries = kcalloc_node(wss->num_entries, sizeof(*wss->entries),
+				    GFP_KERNEL, node);
+	if (!wss->entries) {
+		rvt_wss_exit(rdi);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+/*
+ * Advance the clean counter.  When the clean period has expired,
+ * clean an entry.
+ *
+ * This is implemented in atomics to avoid locking.  Because multiple
+ * variables are involved, it can be racy which can lead to slightly
+ * inaccurate information.  Since this is only a heuristic, this is
+ * OK.  Any innaccuracies will clean themselves out as the counter
+ * advances.  That said, it is unlikely the entry clean operation will
+ * race - the next possible racer will not start until the next clean
+ * period.
+ *
+ * The clean counter is implemented as a decrement to zero.  When zero
+ * is reached an entry is cleaned.
+ */
+static void wss_advance_clean_counter(struct rvt_wss *wss)
+{
+	int entry;
+	int weight;
+	unsigned long bits;
+
+	/* become the cleaner if we decrement the counter to zero */
+	if (atomic_dec_and_test(&wss->clean_counter)) {
+		/*
+		 * Set, not add, the clean period.  This avoids an issue
+		 * where the counter could decrement below the clean period.
+		 * Doing a set can result in lost decrements, slowing the
+		 * clean advance.  Since this a heuristic, this possible
+		 * slowdown is OK.
+		 *
+		 * An alternative is to loop, advancing the counter by a
+		 * clean period until the result is > 0. However, this could
+		 * lead to several threads keeping another in the clean loop.
+		 * This could be mitigated by limiting the number of times
+		 * we stay in the loop.
+		 */
+		atomic_set(&wss->clean_counter, wss->clean_period);
+
+		/*
+		 * Uniquely grab the entry to clean and move to next.
+		 * The current entry is always the lower bits of
+		 * wss.clean_entry.  The table size, wss.num_entries,
+		 * is always a power-of-2.
+		 */
+		entry = (atomic_inc_return(&wss->clean_entry) - 1)
+			& (wss->num_entries - 1);
+
+		/* clear the entry and count the bits */
+		bits = xchg(&wss->entries[entry], 0);
+		weight = hweight64((u64)bits);
+		/* only adjust the contended total count if needed */
+		if (weight)
+			atomic_sub(weight, &wss->total_count);
+	}
+}
+
+/*
+ * Insert the given address into the working set array.
+ */
+static void wss_insert(struct rvt_wss *wss, void *address)
+{
+	u32 page = ((unsigned long)address >> PAGE_SHIFT) & wss->pages_mask;
+	u32 entry = page / BITS_PER_LONG; /* assumes this ends up a shift */
+	u32 nr = page & (BITS_PER_LONG - 1);
+
+	if (!test_and_set_bit(nr, &wss->entries[entry]))
+		atomic_inc(&wss->total_count);
+
+	wss_advance_clean_counter(wss);
+}
+
+/*
+ * Is the working set larger than the threshold?
+ */
+static inline bool wss_exceeds_threshold(struct rvt_wss *wss)
+{
+	return atomic_read(&wss->total_count) >= wss->threshold;
+}
+
 static void get_map_page(struct rvt_qpn_table *qpt,
 			 struct rvt_qpn_map *map)
 {
···
 	rcu_read_unlock();
 }
 EXPORT_SYMBOL(rvt_qp_iter);
+
+/**
+ * rvt_copy_sge - copy data to SGE memory
+ * @qp: associated QP
+ * @ss: the SGE state
+ * @data: the data to copy
+ * @length: the length of the data
+ * @release: boolean to release MR
+ * @copy_last: do a separate copy of the last 8 bytes
+ */
+void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
+		  void *data, u32 length,
+		  bool release, bool copy_last)
+{
+	struct rvt_sge *sge = &ss->sge;
+	int i;
+	bool in_last = false;
+	bool cacheless_copy = false;
+	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
+	struct rvt_wss *wss = rdi->wss;
+	unsigned int sge_copy_mode = rdi->dparms.sge_copy_mode;
+
+	if (sge_copy_mode == RVT_SGE_COPY_CACHELESS) {
+		cacheless_copy = length >= PAGE_SIZE;
+	} else if (sge_copy_mode == RVT_SGE_COPY_ADAPTIVE) {
+		if (length >= PAGE_SIZE) {
+			/*
+			 * NOTE: this *assumes*:
+			 * o The first vaddr is the dest.
+			 * o If multiple pages, then vaddr is sequential.
+			 */
+			wss_insert(wss, sge->vaddr);
+			if (length >= (2 * PAGE_SIZE))
+				wss_insert(wss, (sge->vaddr + PAGE_SIZE));
+
+			cacheless_copy = wss_exceeds_threshold(wss);
+		} else {
+			wss_advance_clean_counter(wss);
+		}
+	}
+
+	if (copy_last) {
+		if (length > 8) {
+			length -= 8;
+		} else {
+			copy_last = false;
+			in_last = true;
+		}
+	}
+
+again:
+	while (length) {
+		u32 len = rvt_get_sge_length(sge, length);
+
+		WARN_ON_ONCE(len == 0);
+		if (unlikely(in_last)) {
+			/* enforce byte transfer ordering */
+			for (i = 0; i < len; i++)
+				((u8 *)sge->vaddr)[i] = ((u8 *)data)[i];
+		} else if (cacheless_copy) {
+			cacheless_memcpy(sge->vaddr, data, len);
+		} else {
+			memcpy(sge->vaddr, data, len);
+		}
+		rvt_update_sge(ss, len, release);
+		data += len;
+		length -= len;
+	}
+
+	if (copy_last) {
+		copy_last = false;
+		in_last = true;
+		length = 8;
+		goto again;
+	}
+}
+EXPORT_SYMBOL(rvt_copy_sge);

drivers/infiniband/sw/rdmavt/qp.h

···
 		  const struct ib_send_wr **bad_wr);
 int rvt_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
 		      const struct ib_recv_wr **bad_wr);
+int rvt_wss_init(struct rvt_dev_info *rdi);
+void rvt_wss_exit(struct rvt_dev_info *rdi);
 #endif          /* DEF_RVTQP_H */

drivers/infiniband/sw/rdmavt/vt.c

···
 		goto bail_no_mr;
 	}
 
+	/* Memory Working Set Size */
+	ret = rvt_wss_init(rdi);
+	if (ret) {
+		rvt_pr_err(rdi, "Error in WSS init.\n");
+		goto bail_mr;
+	}
+
 	/* Completion queues */
 	spin_lock_init(&rdi->n_cqs_lock);
 
···
 				 rdi->driver_f.port_callback);
 	if (ret) {
 		rvt_pr_err(rdi, "Failed to register driver with ib core.\n");
-		goto bail_mr;
+		goto bail_wss;
 	}
 
 	rvt_create_mad_agents(rdi);
···
 	rvt_pr_info(rdi, "Registration with rdmavt done.\n");
 	return ret;
 
+bail_wss:
+	rvt_wss_exit(rdi);
 bail_mr:
 	rvt_mr_exit(rdi);
 
···
 	rvt_free_mad_agents(rdi);
 
 	ib_unregister_device(&rdi->ibdev);
+	rvt_wss_exit(rdi);
 	rvt_mr_exit(rdi);
 	rvt_qp_exit(rdi);
 }

include/rdma/rdma_vt.h

···
 
 #define RVT_CQN_MAX 16 /* maximum length of cq name */
 
+#define RVT_SGE_COPY_MEMCPY	0
+#define RVT_SGE_COPY_CACHELESS	1
+#define RVT_SGE_COPY_ADAPTIVE	2
+
 /*
  * Things that are driver specific, module parameters in hfi1 and qib
  */
···
 	 */
 	unsigned int lkey_table_size;
 	unsigned int qp_table_size;
+	unsigned int sge_copy_mode;
+	unsigned int wss_threshold;
+	unsigned int wss_clean_period;
 	int qpn_start;
 	int qpn_inc;
 	int qpn_res_start;
···
 	atomic_t refcount;
 	u8 vl;
 	u8 log_pmtu;
+};
+
+/* memory working set size */
+struct rvt_wss {
+	unsigned long *entries;
+	atomic_t total_count;
+	atomic_t clean_counter;
+	atomic_t clean_entry;
+
+	int threshold;
+	int num_entries;
+	long pages_mask;
+	unsigned int clean_period;
 };
 
 struct rvt_dev_info;
···
 	u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
 	spinlock_t n_mcast_grps_lock;
 
+	/* Memory Working Set Size */
+	struct rvt_wss *wss;
 };
 
 /**

include/rdma/rdmavt_qp.h

···
 void rvt_stop_rc_timers(struct rvt_qp *qp);
 void rvt_add_retry_timer(struct rvt_qp *qp);
 
+void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
+		  void *data, u32 length,
+		  bool release, bool copy_last);
+
 /**
  * struct rvt_qp_iter - the iterator for QPs
  * @qp - the current QP