* Derived from ivtv-queue.c
*
* Copyright (C) 2007 Hans Verkuil <hverkuil@xs4all.nl>
+ * Copyright (C) 2008 Andy Walls <awalls@md.metrocast.net>
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
*/
#include "cx18-driver.h"
-#include "cx18-streams.h"
#include "cx18-queue.h"
+#include "cx18-streams.h"
#include "cx18-scb.h"
-
-int cx18_buf_copy_from_user(struct cx18_stream *s, struct cx18_buffer *buf,
- const char __user *src, int copybytes)
-{
- if (s->buf_size - buf->bytesused < copybytes)
- copybytes = s->buf_size - buf->bytesused;
- if (copy_from_user(buf->buf + buf->bytesused, src, copybytes))
- return -EFAULT;
- buf->bytesused += copybytes;
- return copybytes;
-}
+#include "cx18-io.h"
void cx18_buf_swap(struct cx18_buffer *buf)
{
swab32s((u32 *)(buf->buf + i));
}
+void _cx18_mdl_swap(struct cx18_mdl *mdl)
+{
+ struct cx18_buffer *buf;
+
+ list_for_each_entry(buf, &mdl->buf_list, list) {
+ if (buf->bytesused == 0)
+ break;
+ cx18_buf_swap(buf);
+ }
+}
+
void cx18_queue_init(struct cx18_queue *q)
{
INIT_LIST_HEAD(&q->list);
- q->buffers = 0;
- q->length = 0;
+ atomic_set(&q->depth, 0);
q->bytesused = 0;
}
-void cx18_enqueue(struct cx18_stream *s, struct cx18_buffer *buf,
- struct cx18_queue *q)
+struct cx18_queue *_cx18_enqueue(struct cx18_stream *s, struct cx18_mdl *mdl,
+ struct cx18_queue *q, int to_front)
+{
+ /* clear the mdl if it is not to be enqueued to the full queue */
+ if (q != &s->q_full) {
+ mdl->bytesused = 0;
+ mdl->readpos = 0;
+ mdl->m_flags = 0;
+ mdl->skipped = 0;
+ mdl->curr_buf = NULL;
+ }
+
+ /* q_busy is restricted to a max buffer count imposed by firmware */
+ if (q == &s->q_busy &&
+ atomic_read(&q->depth) >= CX18_MAX_FW_MDLS_PER_STREAM)
+ q = &s->q_free;
+
+ spin_lock(&q->lock);
+
+ if (to_front)
+ list_add(&mdl->list, &q->list); /* LIFO */
+ else
+ list_add_tail(&mdl->list, &q->list); /* FIFO */
+ q->bytesused += mdl->bytesused - mdl->readpos;
+ atomic_inc(&q->depth);
+
+ spin_unlock(&q->lock);
+ return q;
+}
+
+struct cx18_mdl *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q)
{
- unsigned long flags = 0;
+ struct cx18_mdl *mdl = NULL;
+
+ spin_lock(&q->lock);
+ if (!list_empty(&q->list)) {
+ mdl = list_first_entry(&q->list, struct cx18_mdl, list);
+ list_del_init(&mdl->list);
+ q->bytesused -= mdl->bytesused - mdl->readpos;
+ mdl->skipped = 0;
+ atomic_dec(&q->depth);
+ }
+ spin_unlock(&q->lock);
+ return mdl;
+}
+
+static void _cx18_mdl_update_bufs_for_cpu(struct cx18_stream *s,
+ struct cx18_mdl *mdl)
+{
+ struct cx18_buffer *buf;
+ u32 buf_size = s->buf_size;
+ u32 bytesused = mdl->bytesused;
- /* clear the buffer if it is going to be enqueued to the free queue */
- if (q == &s->q_free) {
- buf->bytesused = 0;
+ list_for_each_entry(buf, &mdl->buf_list, list) {
buf->readpos = 0;
- buf->b_flags = 0;
+ if (bytesused >= buf_size) {
+ buf->bytesused = buf_size;
+ bytesused -= buf_size;
+ } else {
+ buf->bytesused = bytesused;
+ bytesused = 0;
+ }
+ cx18_buf_sync_for_cpu(s, buf);
}
- spin_lock_irqsave(&s->qlock, flags);
- list_add_tail(&buf->list, &q->list);
- q->buffers++;
- q->length += s->buf_size;
- q->bytesused += buf->bytesused - buf->readpos;
- spin_unlock_irqrestore(&s->qlock, flags);
}
-struct cx18_buffer *cx18_dequeue(struct cx18_stream *s, struct cx18_queue *q)
+static inline void cx18_mdl_update_bufs_for_cpu(struct cx18_stream *s,
+ struct cx18_mdl *mdl)
{
- struct cx18_buffer *buf = NULL;
- unsigned long flags = 0;
+ struct cx18_buffer *buf;
- spin_lock_irqsave(&s->qlock, flags);
- if (!list_empty(&q->list)) {
- buf = list_entry(q->list.next, struct cx18_buffer, list);
- list_del_init(q->list.next);
- q->buffers--;
- q->length -= s->buf_size;
- q->bytesused -= buf->bytesused - buf->readpos;
+ if (list_is_singular(&mdl->buf_list)) {
+ buf = list_first_entry(&mdl->buf_list, struct cx18_buffer,
+ list);
+ buf->bytesused = mdl->bytesused;
+ buf->readpos = 0;
+ cx18_buf_sync_for_cpu(s, buf);
+ } else {
+ _cx18_mdl_update_bufs_for_cpu(s, mdl);
}
- spin_unlock_irqrestore(&s->qlock, flags);
- return buf;
}
-struct cx18_buffer *cx18_queue_find_buf(struct cx18_stream *s, u32 id,
+struct cx18_mdl *cx18_queue_get_mdl(struct cx18_stream *s, u32 id,
u32 bytesused)
{
struct cx18 *cx = s->cx;
- struct list_head *p;
-
- list_for_each(p, &s->q_free.list) {
- struct cx18_buffer *buf =
- list_entry(p, struct cx18_buffer, list);
-
- if (buf->id != id)
+ struct cx18_mdl *mdl;
+ struct cx18_mdl *tmp;
+ struct cx18_mdl *ret = NULL;
+ LIST_HEAD(sweep_up);
+
+ /*
+ * We don't have to acquire multiple q locks here, because we are
+ * serialized by the single threaded work handler.
+ * MDLs from the firmware will thus remain in order as
+ * they are moved from q_busy to q_full or to the dvb ring buffer.
+ */
+ spin_lock(&s->q_busy.lock);
+ list_for_each_entry_safe(mdl, tmp, &s->q_busy.list, list) {
+ /*
+ * We should find what the firmware told us is done,
+ * right at the front of the queue. If we don't, we likely have
+ * missed an mdl done message from the firmware.
+ * Once we skip an mdl repeatedly, relative to the size of
+ * q_busy, we have high confidence we've missed it.
+ */
+ if (mdl->id != id) {
+ mdl->skipped++;
+ if (mdl->skipped >= atomic_read(&s->q_busy.depth)-1) {
+ /* mdl must have fallen out of rotation */
+ CX18_WARN("Skipped %s, MDL %d, %d "
+ "times - it must have dropped out of "
+ "rotation\n", s->name, mdl->id,
+ mdl->skipped);
+ /* Sweep it up to put it back into rotation */
+ list_move_tail(&mdl->list, &sweep_up);
+ atomic_dec(&s->q_busy.depth);
+ }
continue;
- buf->bytesused = bytesused;
- /* the transport buffers are handled differently,
- so there is no need to move them to the full queue */
- if (s->type == CX18_ENC_STREAM_TYPE_TS)
- return buf;
- s->q_free.buffers--;
- s->q_free.length -= s->buf_size;
- s->q_full.buffers++;
- s->q_full.length += s->buf_size;
- s->q_full.bytesused += buf->bytesused;
- list_move_tail(&buf->list, &s->q_full.list);
- return buf;
+ }
+ /*
+ * We pull the desired mdl off of the queue here. Something
+ * will have to put it back on a queue later.
+ */
+ list_del_init(&mdl->list);
+ atomic_dec(&s->q_busy.depth);
+ ret = mdl;
+ break;
+ }
+ spin_unlock(&s->q_busy.lock);
+
+ /*
+ * We found the mdl for which we were looking. Get it ready for
+ * the caller to put on q_full or in the dvb ring buffer.
+ */
+ if (ret != NULL) {
+ ret->bytesused = bytesused;
+ ret->skipped = 0;
+ /* 0'ed readpos, m_flags & curr_buf when mdl went on q_busy */
+ cx18_mdl_update_bufs_for_cpu(s, ret);
+ if (s->type != CX18_ENC_STREAM_TYPE_TS)
+ set_bit(CX18_F_M_NEED_SWAP, &ret->m_flags);
}
- CX18_ERR("Cannot find buffer %d for stream %s\n", id, s->name);
- return NULL;
+
+ /* Put any mdls the firmware is ignoring back into normal rotation */
+ list_for_each_entry_safe(mdl, tmp, &sweep_up, list) {
+ list_del_init(&mdl->list);
+ cx18_enqueue(s, mdl, &s->q_free);
+ }
+ return ret;
}
-static void cx18_queue_move_buf(struct cx18_stream *s, struct cx18_queue *from,
- struct cx18_queue *to, int clear, int full)
+/* Move all mdls of a queue, while flushing the mdl */
+static void cx18_queue_flush(struct cx18_stream *s,
+ struct cx18_queue *q_src, struct cx18_queue *q_dst)
{
- struct cx18_buffer *buf =
- list_entry(from->list.next, struct cx18_buffer, list);
-
- list_move_tail(from->list.next, &to->list);
- from->buffers--;
- from->length -= s->buf_size;
- from->bytesused -= buf->bytesused - buf->readpos;
- /* special handling for q_free */
- if (clear)
- buf->bytesused = buf->readpos = buf->b_flags = 0;
- else if (full) {
- /* special handling for stolen buffers, assume
- all bytes are used. */
- buf->bytesused = s->buf_size;
- buf->readpos = buf->b_flags = 0;
+ struct cx18_mdl *mdl;
+
+ /* It only makes sense to flush to q_free or q_idle */
+ if (q_src == q_dst || q_dst == &s->q_full || q_dst == &s->q_busy)
+ return;
+
+ spin_lock(&q_src->lock);
+ spin_lock(&q_dst->lock);
+ while (!list_empty(&q_src->list)) {
+ mdl = list_first_entry(&q_src->list, struct cx18_mdl, list);
+ list_move_tail(&mdl->list, &q_dst->list);
+ mdl->bytesused = 0;
+ mdl->readpos = 0;
+ mdl->m_flags = 0;
+ mdl->skipped = 0;
+ mdl->curr_buf = NULL;
+ atomic_inc(&q_dst->depth);
}
- to->buffers++;
- to->length += s->buf_size;
- to->bytesused += buf->bytesused - buf->readpos;
+ cx18_queue_init(q_src);
+ spin_unlock(&q_src->lock);
+ spin_unlock(&q_dst->lock);
}
-/* Move 'needed_bytes' worth of buffers from queue 'from' into queue 'to'.
- If 'needed_bytes' == 0, then move all buffers from 'from' into 'to'.
- If 'steal' != NULL, then buffers may also taken from that queue if
- needed.
-
- The buffer is automatically cleared if it goes to the free queue. It is
- also cleared if buffers need to be taken from the 'steal' queue and
- the 'from' queue is the free queue.
-
- When 'from' is q_free, then needed_bytes is compared to the total
- available buffer length, otherwise needed_bytes is compared to the
- bytesused value. For the 'steal' queue the total available buffer
- length is always used.
-
- -ENOMEM is returned if the buffers could not be obtained, 0 if all
- buffers where obtained from the 'from' list and if non-zero then
- the number of stolen buffers is returned. */
-int cx18_queue_move(struct cx18_stream *s, struct cx18_queue *from,
- struct cx18_queue *steal, struct cx18_queue *to, int needed_bytes)
+void cx18_flush_queues(struct cx18_stream *s)
{
- unsigned long flags;
- int rc = 0;
- int from_free = from == &s->q_free;
- int to_free = to == &s->q_free;
- int bytes_available;
-
- spin_lock_irqsave(&s->qlock, flags);
- if (needed_bytes == 0) {
- from_free = 1;
- needed_bytes = from->length;
- }
+ cx18_queue_flush(s, &s->q_busy, &s->q_free);
+ cx18_queue_flush(s, &s->q_full, &s->q_free);
+}
- bytes_available = from_free ? from->length : from->bytesused;
- bytes_available += steal ? steal->length : 0;
+/*
+ * Note, s->buf_pool is not protected by a lock,
+ * the stream better not have *anything* going on when calling this
+ */
+void cx18_unload_queues(struct cx18_stream *s)
+{
+ struct cx18_queue *q_idle = &s->q_idle;
+ struct cx18_mdl *mdl;
+ struct cx18_buffer *buf;
- if (bytes_available < needed_bytes) {
- spin_unlock_irqrestore(&s->qlock, flags);
- return -ENOMEM;
+ /* Move all MDLS to q_idle */
+ cx18_queue_flush(s, &s->q_busy, q_idle);
+ cx18_queue_flush(s, &s->q_full, q_idle);
+ cx18_queue_flush(s, &s->q_free, q_idle);
+
+ /* Reset MDL id's and move all buffers back to the stream's buf_pool */
+ spin_lock(&q_idle->lock);
+ list_for_each_entry(mdl, &q_idle->list, list) {
+ while (!list_empty(&mdl->buf_list)) {
+ buf = list_first_entry(&mdl->buf_list,
+ struct cx18_buffer, list);
+ list_move_tail(&buf->list, &s->buf_pool);
+ buf->bytesused = 0;
+ buf->readpos = 0;
+ }
+ mdl->id = s->mdl_base_idx; /* reset id to a "safe" value */
+ /* all other mdl fields were cleared by cx18_queue_flush() */
}
- if (from_free) {
- u32 old_length = to->length;
-
- while (to->length - old_length < needed_bytes) {
- if (list_empty(&from->list))
- from = steal;
- if (from == steal)
- rc++; /* keep track of 'stolen' buffers */
- cx18_queue_move_buf(s, from, to, 1, 0);
+ spin_unlock(&q_idle->lock);
+}
+
+/*
+ * Note, s->buf_pool is not protected by a lock,
+ * the stream better not have *anything* going on when calling this
+ */
+void cx18_load_queues(struct cx18_stream *s)
+{
+ struct cx18 *cx = s->cx;
+ struct cx18_mdl *mdl;
+ struct cx18_buffer *buf;
+ int mdl_id;
+ int i;
+ u32 partial_buf_size;
+
+ /*
+ * Attach buffers to MDLs, give the MDLs ids, and add MDLs to q_free
+ * Excess MDLs are left on q_idle
+ * Excess buffers are left in buf_pool and/or on an MDL in q_idle
+ */
+ mdl_id = s->mdl_base_idx;
+ for (mdl = cx18_dequeue(s, &s->q_idle), i = s->bufs_per_mdl;
+ mdl != NULL && i == s->bufs_per_mdl;
+ mdl = cx18_dequeue(s, &s->q_idle)) {
+
+ mdl->id = mdl_id;
+
+ for (i = 0; i < s->bufs_per_mdl; i++) {
+ if (list_empty(&s->buf_pool))
+ break;
+
+ buf = list_first_entry(&s->buf_pool, struct cx18_buffer,
+ list);
+ list_move_tail(&buf->list, &mdl->buf_list);
+
+ /* update the firmware's MDL array with this buffer */
+ cx18_writel(cx, buf->dma_handle,
+ &cx->scb->cpu_mdl[mdl_id + i].paddr);
+ cx18_writel(cx, s->buf_size,
+ &cx->scb->cpu_mdl[mdl_id + i].length);
}
- } else {
- u32 old_bytesused = to->bytesused;
-
- while (to->bytesused - old_bytesused < needed_bytes) {
- if (list_empty(&from->list))
- from = steal;
- if (from == steal)
- rc++; /* keep track of 'stolen' buffers */
- cx18_queue_move_buf(s, from, to, to_free, rc);
+
+ if (i == s->bufs_per_mdl) {
+ /*
+ * The encoder doesn't honor s->mdl_size. So in the
+ * case of a non-integral number of buffers to meet
+ * mdl_size, we lie about the size of the last buffer
+ * in the MDL to get the encoder to really only send
+ * us mdl_size bytes per MDL transfer.
+ */
+ partial_buf_size = s->mdl_size % s->buf_size;
+ if (partial_buf_size) {
+ cx18_writel(cx, partial_buf_size,
+ &cx->scb->cpu_mdl[mdl_id + i - 1].length);
+ }
+ cx18_enqueue(s, mdl, &s->q_free);
+ } else {
+ /* Not enough buffers for this MDL; we won't use it */
+ cx18_push(s, mdl, &s->q_idle);
}
+ mdl_id += i;
}
- spin_unlock_irqrestore(&s->qlock, flags);
- return rc;
}
-void cx18_flush_queues(struct cx18_stream *s)
+void _cx18_mdl_sync_for_device(struct cx18_stream *s, struct cx18_mdl *mdl)
{
- cx18_queue_move(s, &s->q_io, NULL, &s->q_free, 0);
- cx18_queue_move(s, &s->q_full, NULL, &s->q_free, 0);
+ int dma = s->dma;
+ u32 buf_size = s->buf_size;
+ struct pci_dev *pci_dev = s->cx->pci_dev;
+ struct cx18_buffer *buf;
+
+ list_for_each_entry(buf, &mdl->buf_list, list)
+ pci_dma_sync_single_for_device(pci_dev, buf->dma_handle,
+ buf_size, dma);
}
int cx18_stream_alloc(struct cx18_stream *s)
if (s->buffers == 0)
return 0;
- CX18_DEBUG_INFO("Allocate %s stream: %d x %d buffers (%dkB total)\n",
+ CX18_DEBUG_INFO("Allocate %s stream: %d x %d buffers "
+ "(%d.%02d kB total)\n",
s->name, s->buffers, s->buf_size,
- s->buffers * s->buf_size / 1024);
+ s->buffers * s->buf_size / 1024,
+ (s->buffers * s->buf_size * 100 / 1024) % 100);
- if (((char *)&cx->scb->cpu_mdl[cx->mdl_offset + s->buffers] -
- (char *)cx->scb) > SCB_RESERVED_SIZE) {
- unsigned bufsz = (((char *)cx->scb) + SCB_RESERVED_SIZE -
- ((char *)cx->scb->cpu_mdl));
+ if (((char __iomem *)&cx->scb->cpu_mdl[cx->free_mdl_idx + s->buffers] -
+ (char __iomem *)cx->scb) > SCB_RESERVED_SIZE) {
+ unsigned bufsz = (((char __iomem *)cx->scb) + SCB_RESERVED_SIZE -
+ ((char __iomem *)cx->scb->cpu_mdl));
CX18_ERR("Too many buffers, cannot fit in SCB area\n");
CX18_ERR("Max buffers = %zd\n",
- bufsz / sizeof(struct cx18_mdl));
+ bufsz / sizeof(struct cx18_mdl_ent));
return -ENOMEM;
}
- s->mdl_offset = cx->mdl_offset;
+ s->mdl_base_idx = cx->free_mdl_idx;
- /* allocate stream buffers. Initially all buffers are in q_free. */
+ /* allocate stream buffers and MDLs */
for (i = 0; i < s->buffers; i++) {
- struct cx18_buffer *buf =
- kzalloc(sizeof(struct cx18_buffer), GFP_KERNEL);
+ struct cx18_mdl *mdl;
+ struct cx18_buffer *buf;
+
+ /* 1 MDL per buffer to handle the worst & also default case */
+ mdl = kzalloc(sizeof(struct cx18_mdl), GFP_KERNEL|__GFP_NOWARN);
+ if (mdl == NULL)
+ break;
- if (buf == NULL)
+ buf = kzalloc(sizeof(struct cx18_buffer),
+ GFP_KERNEL|__GFP_NOWARN);
+ if (buf == NULL) {
+ kfree(mdl);
break;
- buf->buf = kmalloc(s->buf_size, GFP_KERNEL);
+ }
+
+ buf->buf = kmalloc(s->buf_size, GFP_KERNEL|__GFP_NOWARN);
if (buf->buf == NULL) {
+ kfree(mdl);
kfree(buf);
break;
}
- buf->id = cx->buffer_id++;
+
+ INIT_LIST_HEAD(&mdl->list);
+ INIT_LIST_HEAD(&mdl->buf_list);
+ mdl->id = s->mdl_base_idx; /* a somewhat safe value */
+ cx18_enqueue(s, mdl, &s->q_idle);
+
INIT_LIST_HEAD(&buf->list);
- buf->dma_handle = pci_map_single(s->cx->dev,
+ buf->dma_handle = pci_map_single(s->cx->pci_dev,
buf->buf, s->buf_size, s->dma);
cx18_buf_sync_for_cpu(s, buf);
- cx18_enqueue(s, buf, &s->q_free);
+ list_add_tail(&buf->list, &s->buf_pool);
}
if (i == s->buffers) {
- cx->mdl_offset += s->buffers;
+ cx->free_mdl_idx += s->buffers;
return 0;
}
CX18_ERR("Couldn't allocate buffers for %s stream\n", s->name);
void cx18_stream_free(struct cx18_stream *s)
{
+ struct cx18_mdl *mdl;
struct cx18_buffer *buf;
+ struct cx18 *cx = s->cx;
+
+ CX18_DEBUG_INFO("Deallocating buffers for %s stream\n", s->name);
+
+ /* move all buffers to buf_pool and all MDLs to q_idle */
+ cx18_unload_queues(s);
+
+ /* empty q_idle */
+ while ((mdl = cx18_dequeue(s, &s->q_idle)))
+ kfree(mdl);
- /* move all buffers to q_free */
- cx18_flush_queues(s);
+ /* empty buf_pool */
+ while (!list_empty(&s->buf_pool)) {
+ buf = list_first_entry(&s->buf_pool, struct cx18_buffer, list);
+ list_del_init(&buf->list);
- /* empty q_free */
- while ((buf = cx18_dequeue(s, &s->q_free))) {
- pci_unmap_single(s->cx->dev, buf->dma_handle,
+ pci_unmap_single(s->cx->pci_dev, buf->dma_handle,
s->buf_size, s->dma);
kfree(buf->buf);
kfree(buf);
git://ftp.safe.ca / safe / jmp / linux-2.6 / blobdiff