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 <ashutosh.dixit@intel.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Brian Welty <brian.welty@intel.com>
Signed-off-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

7 files changed, 26 insertions, 272 deletions

diff --git a/drivers/infiniband/hw/hfi1/init.c b/drivers/infiniband/hw/hfi1/init.c
index 1e770a133779..09044905284f 100644
--- a/drivers/infiniband/hw/hfi1/init.c
+++ b/drivers/infiniband/hw/hfi1/init.c
@@ -1504,9 +1504,6 @@ static int __init hfi1_mod_init(void)
 	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);
@@ -1515,8 +1512,6 @@ static int __init hfi1_mod_init(void)
 	goto bail; /* all OK */
 
 bail_dev:
-	hfi1_wss_exit();
-bail_wss:
 	hfi1_dbg_exit();
 	idr_destroy(&hfi1_unit_table);
 	dev_cleanup();
@@ -1533,7 +1528,6 @@ static void __exit hfi1_mod_cleanup(void)
 {
 	pci_unregister_driver(&hfi1_pci_driver);
 	node_affinity_destroy_all();
-	hfi1_wss_exit();
 	hfi1_dbg_exit();
 
 	idr_destroy(&hfi1_unit_table);
diff --git a/drivers/infiniband/hw/hfi1/rc.c b/drivers/infiniband/hw/hfi1/rc.c
index 9bd63abb2dfe..673b31ebf0ac 100644
--- a/drivers/infiniband/hw/hfi1/rc.c
+++ b/drivers/infiniband/hw/hfi1/rc.c
@@ -1644,7 +1644,8 @@ read_middle:
 		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):
@@ -1684,7 +1685,8 @@ read_last:
 		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);
@@ -2144,7 +2146,7 @@ send_middle:
 		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):
@@ -2200,7 +2202,7 @@ send_last:
 		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))
diff --git a/drivers/infiniband/hw/hfi1/ruc.c b/drivers/infiniband/hw/hfi1/ruc.c
index 17b49b4309a3..223eaf184934 100644
--- a/drivers/infiniband/hw/hfi1/ruc.c
+++ b/drivers/infiniband/hw/hfi1/ruc.c
@@ -361,7 +361,8 @@ do_write:
 		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;
diff --git a/drivers/infiniband/hw/hfi1/uc.c b/drivers/infiniband/hw/hfi1/uc.c
index e254dcec6f64..48a320c01552 100644
--- a/drivers/infiniband/hw/hfi1/uc.c
+++ b/drivers/infiniband/hw/hfi1/uc.c
@@ -426,7 +426,7 @@ send_first:
 		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):
@@ -449,7 +449,7 @@ send_last:
 		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;
@@ -523,7 +523,7 @@ rdma_first:
 		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):
@@ -550,7 +550,7 @@ rdma_last_imm:
 		}
 		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;
 
@@ -564,7 +564,7 @@ rdma_last:
 		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;
 
diff --git a/drivers/infiniband/hw/hfi1/ud.c b/drivers/infiniband/hw/hfi1/ud.c
index 70d39fc450a1..e55bc4280d58 100644
--- a/drivers/infiniband/hw/hfi1/ud.c
+++ b/drivers/infiniband/hw/hfi1/ud.c
@@ -210,8 +210,8 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 		}
 
 		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);
@@ -228,7 +228,7 @@ static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
 		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;
@@ -1019,8 +1019,8 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 		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;
@@ -1030,14 +1030,14 @@ void hfi1_ud_rcv(struct hfi1_packet *packet)
 		 * 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;
diff --git a/drivers/infiniband/hw/hfi1/verbs.c b/drivers/infiniband/hw/hfi1/verbs.c
index 16b88948383b..0a47b46f979e 100644
--- a/drivers/infiniband/hw/hfi1/verbs.c
+++ b/drivers/infiniband/hw/hfi1/verbs.c
@@ -129,8 +129,6 @@ unsigned short piothreshold = 256;
 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,
@@ -151,159 +149,13 @@ static int pio_wait(struct rvt_qp *qp,
 /* 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.
  */
@@ -438,79 +290,6 @@ static const u32 pio_opmask[BIT(3)] = {
  */
 __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.
  */
@@ -1949,6 +1728,9 @@ int hfi1_register_ib_device(struct hfi1_devdata *dd)
 	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;
diff --git a/drivers/infiniband/hw/hfi1/verbs.h b/drivers/infiniband/hw/hfi1/verbs.h
index d4164114396e..eb99e8df6251 100644
--- a/drivers/infiniband/hw/hfi1/verbs.h
+++ b/drivers/infiniband/hw/hfi1/verbs.h
@@ -315,9 +315,6 @@ void hfi1_put_txreq(struct verbs_txreq *tx);
 
 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);
@@ -393,28 +390,6 @@ int hfi1_verbs_send_dma(struct rvt_qp *qp, struct hfi1_pkt_state *ps,
 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)
 {
 	return ohdr->bth[1] & cpu_to_be32(OPA_BTH_MIG_REQ);