[PATCH char-misc-next v2 02/13] misc: mic: SCIF ring buffer infrastructure

From: Sudeep Dutt
Date: Mon Jan 12 2015 - 18:12:32 EST


SCIF ring buffer is a single producer, single consumer byte stream
ring buffer optimized for avoiding reads across the PCIe bus while
adding the required barriers and hardware workarounds for the MIC
Coprocessor. The ring buffer is used to implement a receive queue for
SCIF driver messaging between two nodes and for byte stream messaging
between SCIF endpoints. The existing in-kernel ring buffer was not
reused since it has not been designed for our use across the PCIe bus
where each node runs an independent OS. Each SCIF node has a receive
queue for every other SCIF node, and each connected endpoint has a
receive queue for messages from its peer. This pair of receive
queues is referred to as a SCIF queue pair.

Reviewed-by: Nikhil Rao <nikhil.rao@xxxxxxxxx>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@xxxxxxxxx>
Signed-off-by: Sudeep Dutt <sudeep.dutt@xxxxxxxxx>
---
drivers/misc/mic/scif/scif_rb.h | 100 ++++++++++++++++
drivers/misc/mic/scif/scif_rb.c | 248 ++++++++++++++++++++++++++++++++++++++++
2 files changed, 348 insertions(+)
create mode 100644 drivers/misc/mic/scif/scif_rb.h
create mode 100644 drivers/misc/mic/scif/scif_rb.c

diff --git a/drivers/misc/mic/scif/scif_rb.h b/drivers/misc/mic/scif/scif_rb.h
new file mode 100644
index 0000000..ba4bfce
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_rb.h
@@ -0,0 +1,100 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * This file is provided under a dual BSD/GPLv2 license. When using or
+ * redistributing this file, you may do so under either license.
+ *
+ * GPL LICENSE SUMMARY
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of version 2 of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * BSD LICENSE
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * * Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in
+ * the documentation and/or other materials provided with the
+ * distribution.
+ * * Neither the name of Intel Corporation nor the names of its
+ * contributors may be used to endorse or promote products derived
+ * from this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ * Intel SCIF driver.
+ */
+#ifndef SCIF_RB_H
+#define SCIF_RB_H
+/*
+ * This file describes a general purpose, byte based ring buffer. Writers to the
+ * ring buffer need to synchronize using a lock. The same is true for readers,
+ * although in practice, the ring buffer has a single reader. It is lockless
+ * between producer and consumer so it can handle being used across the PCIe
+ * bus. The ring buffer ensures that there are no reads across the PCIe bus for
+ * performance reasons. Two of these are used to form a single bidirectional
+ * queue-pair across PCIe.
+ */
+/*
+ * struct scif_rb - SCIF Ring Buffer
+ *
+ * @rb_base: The base of the memory used for storing RB messages
+ * @read_ptr: Pointer to the read offset
+ * @write_ptr: Pointer to the write offset
+ * @size: Size of the memory in rb_base
+ * @current_read_offset: Cached read offset for performance
+ * @current_write_offset: Cached write offset for performance
+ */
+struct scif_rb {
+ void *rb_base;
+ uint32_t *read_ptr;
+ uint32_t *write_ptr;
+ uint32_t size;
+ uint32_t current_read_offset;
+ uint32_t current_write_offset;
+};
+
+/* methods used by both */
+void scif_rb_init(struct scif_rb *rb, uint32_t *read_ptr, uint32_t *write_ptr,
+ void *rb_base, uint8_t size);
+/* writer only methods */
+/* write a new command, then scif_rb_commit() */
+int scif_rb_write(struct scif_rb *rb, void *msg, uint32_t size);
+/* after write(), then scif_rb_commit() */
+void scif_rb_commit(struct scif_rb *rb);
+/* query space available for writing to a RB. */
+uint32_t scif_rb_space(struct scif_rb *rb);
+
+/* reader only methods */
+/* read a new message from the ring buffer of size bytes */
+uint32_t scif_rb_get_next(struct scif_rb *rb, void *msg, uint32_t size);
+/* update the read pointer so that the space can be reused */
+void scif_rb_update_read_ptr(struct scif_rb *rb);
+/* count the number of bytes that can be read */
+uint32_t scif_rb_count(struct scif_rb *rb, uint32_t size);
+#endif
diff --git a/drivers/misc/mic/scif/scif_rb.c b/drivers/misc/mic/scif/scif_rb.c
new file mode 100644
index 0000000..28e99ea
--- /dev/null
+++ b/drivers/misc/mic/scif/scif_rb.c
@@ -0,0 +1,248 @@
+/*
+ * Intel MIC Platform Software Stack (MPSS)
+ *
+ * Copyright(c) 2014 Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License, version 2, as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License for more details.
+ *
+ * Intel SCIF driver.
+ *
+ */
+#include <linux/circ_buf.h>
+#include <linux/types.h>
+#include <linux/io.h>
+
+#include "scif_rb.h"
+
+#define scif_rb_ring_cnt(head, tail, size) CIRC_CNT(head, tail, size)
+#define scif_rb_ring_space(head, tail, size) CIRC_SPACE(head, tail, size)
+
+/**
+ * scif_rb_init - Initializes the ring buffer
+ * @rb: ring buffer
+ * @read_ptr: A pointer to the read offset
+ * @write_ptr: A pointer to the write offset
+ * @rb_base: A pointer to the base of the ring buffer
+ * @size: The size of the ring buffer in powers of two
+ */
+void scif_rb_init(struct scif_rb *rb, u32 *read_ptr, u32 *write_ptr,
+ void *rb_base, u8 size)
+{
+ rb->rb_base = rb_base;
+ rb->size = (1 << size);
+ rb->read_ptr = read_ptr;
+ rb->write_ptr = write_ptr;
+ rb->current_read_offset = *read_ptr;
+ rb->current_write_offset = *write_ptr;
+}
+
+/* Copies a message to the ring buffer -- handles the wrap around case */
+static void memcpy_torb(struct scif_rb *rb, void *header,
+ void *msg, u32 size)
+{
+ u32 size1, size2;
+
+ if (header + size >= rb->rb_base + rb->size) {
+ /* Need to call two copies if it wraps around */
+ size1 = (u32)(rb->rb_base + rb->size - header);
+ size2 = size - size1;
+ memcpy_toio((void __iomem __force *)header, msg, size1);
+ memcpy_toio((void __iomem __force *)rb->rb_base,
+ msg + size1, size2);
+ } else {
+ memcpy_toio((void __iomem __force *)header, msg, size);
+ }
+}
+
+/* Copies a message from the ring buffer -- handles the wrap around case */
+static void memcpy_fromrb(struct scif_rb *rb, void *header,
+ void *msg, u32 size)
+{
+ u32 size1, size2;
+
+ if (header + size >= rb->rb_base + rb->size) {
+ /* Need to call two copies if it wraps around */
+ size1 = (u32)(rb->rb_base + rb->size - header);
+ size2 = size - size1;
+ memcpy_fromio(msg, (void __iomem __force *)header, size1);
+ memcpy_fromio(msg+size1,
+ (void __iomem __force *)rb->rb_base, size2);
+ } else {
+ memcpy_fromio(msg, (void __iomem __force *)header, size);
+ }
+}
+
+/**
+ * scif_rb_space - Query space available for writing to the RB
+ * @rb: ring buffer
+ *
+ * Return: size available for writing to RB in bytes.
+ */
+u32 scif_rb_space(struct scif_rb *rb)
+{
+ rb->current_read_offset = *rb->read_ptr;
+ /*
+ * Update from the HW read pointer only once the peer has exposed the
+ * new empty slot. This barrier is paired with the memory barrier
+ * scif_rb_update_read_ptr()
+ */
+ mb();
+ return scif_rb_ring_space(rb->current_write_offset,
+ rb->current_read_offset, rb->size);
+}
+
+/**
+ * scif_rb_write - Write a message to the RB
+ * @rb: ring buffer
+ * @msg: buffer to send the message. Must be at least size bytes long
+ * @size: the size (in bytes) to be copied to the RB
+ *
+ * This API does not block if there isn't enough space in the RB.
+ * Returns: 0 on success or -ENOMEM on failure
+ */
+int scif_rb_write(struct scif_rb *rb, void *msg, u32 size)
+{
+ void *header;
+
+ if (scif_rb_space(rb) < size)
+ return -ENOMEM;
+ header = rb->rb_base + rb->current_write_offset;
+ memcpy_torb(rb, header, msg, size);
+ /*
+ * Wait until scif_rb_commit(). Update the local ring
+ * buffer data, not the shared data until commit.
+ */
+ rb->current_write_offset =
+ (rb->current_write_offset + size) & (rb->size - 1);
+ return 0;
+}
+
+/**
+ * scif_rb_commit - To submit the message to let the peer fetch it
+ * @rb: ring buffer
+ */
+void scif_rb_commit(struct scif_rb *rb)
+{
+ /*
+ * We must ensure ordering between the all the data committed
+ * previously before we expose the new message to the peer by
+ * updating the write_ptr. This write barrier is paired with
+ * the read barrier in scif_rb_count(..)
+ */
+ wmb();
+ ACCESS_ONCE(*rb->write_ptr) = rb->current_write_offset;
+#ifdef CONFIG_INTEL_MIC_CARD
+ /*
+ * X100 Si bug: For the case where a Core is performing an EXT_WR
+ * followed by a Doorbell Write, the Core must perform two EXT_WR to the
+ * same address with the same data before it does the Doorbell Write.
+ * This way, if ordering is violated for the Interrupt Message, it will
+ * fall just behind the first Posted associated with the first EXT_WR.
+ */
+ ACCESS_ONCE(*rb->write_ptr) = rb->current_write_offset;
+#endif
+}
+
+/**
+ * scif_rb_get - To get next message from the ring buffer
+ * @rb: ring buffer
+ * @size: Number of bytes to be read
+ *
+ * Return: NULL if no bytes to be read from the ring buffer, otherwise the
+ * pointer to the next byte
+ */
+static void *scif_rb_get(struct scif_rb *rb, u32 size)
+{
+ void *header = NULL;
+
+ if (scif_rb_count(rb, size) >= size)
+ header = rb->rb_base + rb->current_read_offset;
+ return header;
+}
+
+/*
+ * scif_rb_get_next - Read from ring buffer.
+ * @rb: ring buffer
+ * @msg: buffer to hold the message. Must be at least size bytes long
+ * @size: Number of bytes to be read
+ *
+ * Return: number of bytes read if available bytes are >= size, otherwise
+ * returns zero.
+ */
+u32 scif_rb_get_next(struct scif_rb *rb, void *msg, u32 size)
+{
+ void *header = NULL;
+ int read_size = 0;
+
+ header = scif_rb_get(rb, size);
+ if (header) {
+ u32 next_cmd_offset =
+ (rb->current_read_offset + size) & (rb->size - 1);
+
+ read_size = size;
+ rb->current_read_offset = next_cmd_offset;
+ memcpy_fromrb(rb, header, msg, size);
+ }
+ return read_size;
+}
+
+/**
+ * scif_rb_update_read_ptr
+ * @rb: ring buffer
+ */
+void scif_rb_update_read_ptr(struct scif_rb *rb)
+{
+ u32 new_offset;
+
+ new_offset = rb->current_read_offset;
+ /*
+ * We must ensure ordering between the all the data committed or read
+ * previously before we expose the empty slot to the peer by updating
+ * the read_ptr. This barrier is paired with the memory barrier in
+ * scif_rb_space(..)
+ */
+ mb();
+ ACCESS_ONCE(*rb->read_ptr) = new_offset;
+#ifdef CONFIG_INTEL_MIC_CARD
+ /*
+ * X100 Si Bug: For the case where a Core is performing an EXT_WR
+ * followed by a Doorbell Write, the Core must perform two EXT_WR to the
+ * same address with the same data before it does the Doorbell Write.
+ * This way, if ordering is violated for the Interrupt Message, it will
+ * fall just behind the first Posted associated with the first EXT_WR.
+ */
+ ACCESS_ONCE(*rb->read_ptr) = new_offset;
+#endif
+}
+
+/**
+ * scif_rb_count
+ * @rb: ring buffer
+ * @size: Number of bytes expected to be read
+ *
+ * Return: number of bytes that can be read from the RB
+ */
+u32 scif_rb_count(struct scif_rb *rb, u32 size)
+{
+ if (scif_rb_ring_cnt(rb->current_write_offset,
+ rb->current_read_offset,
+ rb->size) < size) {
+ rb->current_write_offset = *rb->write_ptr;
+ /*
+ * Update from the HW write pointer if empty only once the peer
+ * has exposed the new message. This read barrier is paired
+ * with the write barrier in scif_rb_commit(..)
+ */
+ smp_rmb();
+ }
+ return scif_rb_ring_cnt(rb->current_write_offset,
+ rb->current_read_offset,
+ rb->size);
+}
--
1.8.2.1

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/