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>

4 files changed, 272 insertions, 2 deletions

diff --git a/drivers/infiniband/sw/rdmavt/Kconfig b/drivers/infiniband/sw/rdmavt/Kconfig
index 98e798007f75..7df896a18d38 100644
--- a/drivers/infiniband/sw/rdmavt/Kconfig
+++ b/drivers/infiniband/sw/rdmavt/Kconfig
@@ -1,6 +1,6 @@
 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---
diff --git a/drivers/infiniband/sw/rdmavt/qp.c b/drivers/infiniband/sw/rdmavt/qp.c
index a036a5368103..d969b0803e6f 100644
--- a/drivers/infiniband/sw/rdmavt/qp.c
+++ b/drivers/infiniband/sw/rdmavt/qp.c
@@ -118,6 +118,187 @@ const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
 };
 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)
 {
@@ -2476,3 +2657,80 @@ void rvt_qp_iter(struct rvt_dev_info *rdi,
 	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);
diff --git a/drivers/infiniband/sw/rdmavt/qp.h b/drivers/infiniband/sw/rdmavt/qp.h
index 264811fdc530..6d883972e0b8 100644
--- a/drivers/infiniband/sw/rdmavt/qp.h
+++ b/drivers/infiniband/sw/rdmavt/qp.h
@@ -66,4 +66,6 @@ int rvt_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 		  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 */
diff --git a/drivers/infiniband/sw/rdmavt/vt.c b/drivers/infiniband/sw/rdmavt/vt.c
index e3249d46bcef..723d3daf2eba 100644
--- a/drivers/infiniband/sw/rdmavt/vt.c
+++ b/drivers/infiniband/sw/rdmavt/vt.c
@@ -774,6 +774,13 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 		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);
 
@@ -832,7 +839,7 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 				 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);
@@ -840,6 +847,8 @@ int rvt_register_device(struct rvt_dev_info *rdi, u32 driver_id)
 	rvt_pr_info(rdi, "Registration with rdmavt done.\n");
 	return ret;
 
+bail_wss:
+	rvt_wss_exit(rdi);
 bail_mr:
 	rvt_mr_exit(rdi);
 
@@ -863,6 +872,7 @@ void rvt_unregister_device(struct rvt_dev_info *rdi)
 	rvt_free_mad_agents(rdi);
 
 	ib_unregister_device(&rdi->ibdev);
+	rvt_wss_exit(rdi);
 	rvt_mr_exit(rdi);
 	rvt_qp_exit(rdi);
 }