lkml.org 
[lkml]   [2021]   [Jul]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH v4 15/16] [RFC][REWORK] firmware: arm_scmi: make virtio-scmi use delegated xfers
Hi Peter,

I agree with your remarks down below about raw_payload helpers, and
especially I did not like the need for the memmove and the awkard
rx_raw_len that I introduced, but in order to fix that I should have
changed deeply the way the core uses rx.buf since I cannot just use
the notion of the msg type here in the transport to discriminate between
command and notification. (because we still do not want to peek inside
the message structure from within this virtio transport)

But then I realized that I could have get rid of all of this raw_payload
machinery and simplify further, because now I have also management of
concurrent and out of order responses/delayed_responses and the xfer
obtained here is assured to be valid and sane.

So basically in V5 I get rid of these changes of mine about raw_payloads
and instead go like this in scmi_process_vqueue_input():

ret = scmi_transfer_acquire(vioch->cinfo, &msg_hdr, &xfer);
if (ret)
scmi_finalize_message(vioch, msg)

xfer->priv = msg;

scmi_transfer_release(vioch->cinfo, xfer);

if (vioch->is_rx || !xfer->hdr.poll_completion)
scmi_rx_callback(vioch->cinfo, msg_hdr);

scmi_finalize_message(vioch, msg);

and then let your original fetch_response/notification proceed as usual
picking the msg from xfer->priv instead.

Because, anyway, the xfer I've got is ensured to be valid and the new V5
serialization/ot-of-order logic in the core takes care of the case of
concurrent time-outs or out of order DRESPs reception (maybe this will be
more apparent looking at the code in V5...)

As a side consideration, I also dropped support for polling mode as a whole
in V5, since it seemed to not make much sense for virtio transport, where
there's always some sort of virtio core interrupt that triggers the complete
callbacks and this logic. (while polling support is meant in the core to
address the case in which a transport has no completion IRQ so it
cannnot ever call into scmi_rx_callback(), and such useless support here
would have stopped me from this further simplification.)

On Thu, Jul 01, 2021 at 10:43:12AM +0200, Peter Hilber wrote:
> On 11.06.21 18:59, Cristian Marussi wrote:
> > Draft changes to virtio-scmi to use new support for core delegated xfers
> > in an attempt to simplify the interactions between virtio-scmi transport
> > and the SCMI core transport layer.
> >
>
> These changes seem to make xfers delegation mandatory for message-passing
> transports, so that might be documented.

Not sure about this, since even it is true for virtio but it does not seem
to be necessarily true for any future message-passing based transport, imagining
as an example a message based transport with a queue of just one in flight message,
so with no state to keep (in V5 with raw_payload helpers dropped...), or a transport
that do use message-passing but it keeps its own states in a different way.
Anyway, I'll double check if these assumptions of mine hold true.

>
> > TODO:
> > - Polling is still not supported.
> > - Probe/remove sequence still to be reviewed.
> > - Concurrent or inverted reception of related responses and delayed
> > responses is still not addressed.
> > (it will be addressed in the SCMI core anyway most probably)
> >
> > Signed-off-by: Cristian Marussi <cristian.marussi@arm.com>
> > ---
> > drivers/firmware/arm_scmi/common.h | 5 +
> > drivers/firmware/arm_scmi/msg.c | 35 +++++
> > drivers/firmware/arm_scmi/virtio.c | 212 +++++++++++++++--------------
> > 3 files changed, 152 insertions(+), 100 deletions(-)
> >
> > diff --git a/drivers/firmware/arm_scmi/common.h b/drivers/firmware/arm_scmi/common.h
> > index c143a449d278..22e5532fc698 100644
> > --- a/drivers/firmware/arm_scmi/common.h
> > +++ b/drivers/firmware/arm_scmi/common.h
> > @@ -428,6 +428,11 @@ void msg_fetch_response(struct scmi_msg_payld *msg, size_t len,
> > void msg_fetch_notification(struct scmi_msg_payld *msg, size_t len,
> > size_t max_len, struct scmi_xfer *xfer);
> > +void msg_fetch_raw_payload(struct scmi_msg_payld *msg, size_t msg_len,
> > + size_t max_len, struct scmi_xfer *xfer);
> > +void msg_fetch_raw_response(struct scmi_xfer *xfer);
> > +void msg_fetch_raw_notification(struct scmi_xfer *xfer);
> > +
> > void scmi_notification_instance_data_set(const struct scmi_handle *handle,
> > void *priv);
> > void *scmi_notification_instance_data_get(const struct scmi_handle *handle);
> > diff --git a/drivers/firmware/arm_scmi/msg.c b/drivers/firmware/arm_scmi/msg.c
> > index 8a2d3303d281..3ed3ad0961ae 100644
> > --- a/drivers/firmware/arm_scmi/msg.c
> > +++ b/drivers/firmware/arm_scmi/msg.c
> > @@ -74,6 +74,17 @@ u32 msg_read_header(struct scmi_msg_payld *msg)
> > return le32_to_cpu(msg->msg_header);
> > }
> > +void msg_fetch_raw_payload(struct scmi_msg_payld *msg, size_t msg_len,
> > + size_t max_len, struct scmi_xfer *xfer)
> > +{
> > + xfer->rx_raw_len = min_t(size_t, max_len,
> > + msg_len >= sizeof(*msg) ?
> > + msg_len - sizeof(*msg) : 0);
> > +
> > + /* Take a copy to the rx buffer.. */
> > + memcpy(xfer->rx.buf, msg->msg_payload, xfer->rx_raw_len);
>
> In the usage throughout arm-scmi so far, scmi_desc.max_msg_size seems to
> consistently refer to the payload excluding the return status.
> scmi_desc.max_msg_size is the actual max_len parameter to this function.
>
> So the logic here would reduce the maximum payload length for responses to
> (scmi_desc.max_msg_size - sizeof(return status)).
>

Ditto.

> > +}
> > +
> > /**
> > * msg_fetch_response() - Fetch response SCMI payload from transport SDU.
> > *
> > @@ -94,6 +105,25 @@ void msg_fetch_response(struct scmi_msg_payld *msg, size_t len,
> > memcpy(xfer->rx.buf, &msg->msg_payload[1], xfer->rx.len);
> > }
> > +void msg_fetch_raw_response(struct scmi_xfer *xfer)
> > +{
> > + __le32 *msg_payload = xfer->rx.buf;
> > +
> > + if (xfer->rx_raw_len < sizeof(xfer->hdr.status)) {
> > + xfer->rx.len = 0;
> > + return;
> > + }
> > +
> > + xfer->hdr.status = le32_to_cpu(msg_payload[0]);
> > + /*
> > + * rx.len has been already pre-calculated by fetch_raw
> > + * for the whole payload including status, so shrink it
> > + */
> > + xfer->rx.len = xfer->rx_raw_len - sizeof(xfer->hdr.status);
> > + /* Carveout status 4-byte field */
> > + memmove(xfer->rx.buf, &msg_payload[1], xfer->rx.len);
>
> Wouldn't it be feasible to align properly in msg_fetch_raw_payload()
> already? That could also get rid of .rx_raw_len.
>

Ditto.

> > +}
> > +
> > /**
> > * msg_fetch_notification() - Fetch notification payload from transport SDU.
> > *
> > @@ -111,3 +141,8 @@ void msg_fetch_notification(struct scmi_msg_payld *msg, size_t len,
> > /* Take a copy to the rx buffer.. */
> > memcpy(xfer->rx.buf, msg->msg_payload, xfer->rx.len);
> > }
> > +
> > +void msg_fetch_raw_notification(struct scmi_xfer *xfer)
> > +{
> > + xfer->rx.len = xfer->rx_raw_len;
> > +}
> > diff --git a/drivers/firmware/arm_scmi/virtio.c b/drivers/firmware/arm_scmi/virtio.c
> > index 20972adf6dc7..4412bc590ca7 100644
> > --- a/drivers/firmware/arm_scmi/virtio.c
> > +++ b/drivers/firmware/arm_scmi/virtio.c
> > @@ -37,23 +37,24 @@
> > /**
> > * struct scmi_vio_channel - Transport channel information
> > *
> > - * @lock: Protects access to all members except ready.
> > - * @ready_lock: Protects access to ready. If required, it must be taken before
> > - * lock.
> > * @vqueue: Associated virtqueue
> > * @cinfo: SCMI Tx or Rx channel
> > * @free_list: List of unused scmi_vio_msg, maintained for Tx channels only
> > * @is_rx: Whether channel is an Rx channel
> > * @ready: Whether transport user is ready to hear about channel
> > + * @lock: Protects access to all members except ready.
> > + * @ready_lock: Protects access to ready. If required, it must be taken before
> > + * lock.
> > */
> > struct scmi_vio_channel {
> > - spinlock_t lock;
> > - spinlock_t ready_lock;
> > struct virtqueue *vqueue;
> > struct scmi_chan_info *cinfo;
> > struct list_head free_list;
> > - u8 is_rx;
> > - u8 ready;
> > + bool is_rx;
> > + bool ready;
> > + unsigned int max_msg;
> > + spinlock_t lock;
> > + spinlock_t ready_lock;
> > };
> > /**
> > @@ -100,6 +101,73 @@ static int scmi_vio_feed_vq_rx(struct scmi_vio_channel *vioch,
> > return rc;
> > }
> > +static void scmi_finalize_message(struct scmi_vio_channel *vioch,
> > + struct scmi_vio_msg *msg)
> > +{
> > + unsigned long flags;
> > +
> > + if (vioch->is_rx) {
> > + scmi_vio_feed_vq_rx(vioch, msg);
> > + } else {
> > + spin_lock_irqsave(&vioch->lock, flags);
> > + list_add(&msg->list, &vioch->free_list);
> > + spin_unlock_irqrestore(&vioch->lock, flags);
> > + }
> > +}
> > +
> > +static void scmi_process_vqueue_input(struct scmi_vio_channel *vioch,
> > + struct scmi_vio_msg *msg)
> > +{
> > + u32 msg_hdr;
> > + int ret;
> > + struct scmi_xfer *xfer = NULL;
> > +
> > + msg_hdr = msg_read_header(msg->input);
> > + /*
> > + * Acquire from the core transport layer a currently valid xfer
> > + * descriptor associated to the received msg_hdr: this could be a
> > + * previously allocated xfer for responses and delayed responses to
> > + * in-flight commands, or a freshly allocated new xfer for a just
> > + * received notification.
> > + *
> > + * In case of responses and delayed_responses the acquired xfer, at
> > + * the time scmi_transfer_acquire() succcessfully returns is guaranteed
> > + * to be still associated with a valid (not timed-out nor stale)
> > + * descriptor and proper refcounting is kept in the core along this xfer
> > + * so that should the core time out the xfer concurrently to this receive
> > + * path the xfer will be properly deallocated only once the last user is
> > + * done with it. (and this code path will terminate normally even though
> > + * all the processing related to the timed out xfer will be discarded).
> > + */
>
> This comment would better fit to scmi_transfer_acquire().
>

Yes indeed I strip down this comment in V5.

> > + ret = scmi_transfer_acquire(vioch->cinfo, &msg_hdr, &xfer);
> > + if (ret) {
> > + dev_err(vioch->cinfo->dev,
> > + "Cannot find matching xfer for hdr:0x%X\n", msg_hdr);
> > + scmi_finalize_message(vioch, msg);
> > + return;
> > + }
> > +
> > + dev_dbg(vioch->cinfo->dev,
> > + "VQUEUE[%d] - INPUT MSG_RX_LEN:%d - HDR:0x%X TYPE:%d XFER_ID:%d XFER:%px\n",
> > + vioch->vqueue->index, msg->rx_len, msg_hdr, xfer->hdr.type,
> > + xfer->hdr.seq, xfer);
> > +
> > + msg_fetch_raw_payload(msg->input, msg->rx_len,
> > + scmi_virtio_desc.max_msg_size, xfer); > +
> > + /* Drop processed virtio message anyway */
> > + scmi_finalize_message(vioch, msg);
> > +
> > + if (vioch->is_rx || !xfer->hdr.poll_completion)
> > + scmi_rx_callback(vioch->cinfo, msg_hdr);
> > + else
> > + dev_warn(vioch->cinfo->dev,
> > + "Polling mode NOT supported. Dropped hdr:0X%X\n",
> > + msg_hdr);
> > +
> > + scmi_transfer_release(vioch->cinfo, xfer);
> > +}
> > +
> > static void scmi_vio_complete_cb(struct virtqueue *vqueue)
> > {
> > unsigned long ready_flags;
> > @@ -138,15 +206,9 @@ static void scmi_vio_complete_cb(struct virtqueue *vqueue)
> > if (msg) {
> > msg->rx_len = length;
> > -
> > - /*
> > - * Hold the ready_lock during the callback to avoid
> > - * races when the arm-scmi driver is unbinding while
> > - * the virtio device is not quiesced yet.
> > - */
> > - scmi_rx_callback(vioch->cinfo,
> > - msg_read_header(msg->input), msg);
> > + scmi_process_vqueue_input(vioch, msg);
> > }
> > +
> > spin_unlock_irqrestore(&vioch->ready_lock, ready_flags);
> > }
> > @@ -163,27 +225,11 @@ static vq_callback_t *scmi_vio_complete_callbacks[] = {
> > scmi_vio_complete_cb
> > };
> > -static unsigned int virtio_get_max_msg(bool tx,
> > - struct scmi_chan_info *base_cinfo)
> > +static unsigned int virtio_get_max_msg(struct scmi_chan_info *base_cinfo)
> > {
> > struct scmi_vio_channel *vioch = base_cinfo->transport_info;
> > - unsigned int ret;
> > - ret = virtqueue_get_vring_size(vioch->vqueue);
> > -
> > - /* Tx messages need multiple descriptors. */
> > - if (tx)
> > - ret /= DESCRIPTORS_PER_TX_MSG;
> > -
> > - if (ret > MSG_TOKEN_MAX) {
> > - dev_info_once(
> > - base_cinfo->dev,
> > - "Only %ld messages can be pending simultaneously, while the %s virtqueue could hold %d\n",
> > - MSG_TOKEN_MAX, tx ? "tx" : "rx", ret);
> > - ret = MSG_TOKEN_MAX;
> > - }
> > -
> > - return ret;
> > + return vioch->max_msg;
> > }
> > static int scmi_vio_match_any_dev(struct device *dev, const void *data)
> > @@ -195,13 +241,14 @@ static struct virtio_driver virtio_scmi_driver; /* Forward declaration */
> > static int virtio_link_supplier(struct device *dev)
> > {
> > - struct device *vdev = driver_find_device(
> > - &virtio_scmi_driver.driver, NULL, NULL, scmi_vio_match_any_dev);
> > + struct device *vdev;
> > +
> > + vdev = driver_find_device(&virtio_scmi_driver.driver,
> > + NULL, NULL, scmi_vio_match_any_dev);
> > if (!vdev) {
> > - dev_notice_once(
> > - dev,
> > - "Deferring probe after not finding a bound scmi-virtio device\n");
> > + dev_notice_once(dev,
> > + "Deferring probe after not finding a bound scmi-virtio device\n");
> > return -EPROBE_DEFER;
> > }
> > @@ -245,12 +292,8 @@ static int virtio_chan_setup(struct scmi_chan_info *cinfo, struct device *dev,
> > struct virtio_device *vdev;
> > struct scmi_vio_channel *vioch;
> > int index = tx ? VIRTIO_SCMI_VQ_TX : VIRTIO_SCMI_VQ_RX;
> > - int max_msg;
> > int i;
> > - if (!virtio_chan_available(dev, index))
> > - return -ENODEV;
> > -
> > vdev = scmi_get_transport_info(dev);
> > vioch = &((struct scmi_vio_channel *)vdev->priv)[index];
> > @@ -259,9 +302,7 @@ static int virtio_chan_setup(struct scmi_chan_info *cinfo, struct device *dev,
> > vioch->cinfo = cinfo;
> > spin_unlock_irqrestore(&vioch->lock, flags);
> > - max_msg = virtio_get_max_msg(tx, cinfo);
> > -
> > - for (i = 0; i < max_msg; i++) {
> > + for (i = 0; i < vioch->max_msg; i++) {
> > struct scmi_vio_msg *msg;
> > msg = devm_kzalloc(cinfo->dev, sizeof(*msg), GFP_KERNEL);
> > @@ -322,13 +363,6 @@ static int virtio_send_message(struct scmi_chan_info *cinfo,
> > int rc;
> > struct scmi_vio_msg *msg;
> > - /*
> > - * TODO: For now, we don't support polling. But it should not be
> > - * difficult to add support.
> > - */
> > - if (xfer->hdr.poll_completion)
> > - return -EINVAL;
> > -
> > spin_lock_irqsave(&vioch->lock, flags);
> > if (list_empty(&vioch->free_list)) {
> > @@ -351,6 +385,11 @@ static int virtio_send_message(struct scmi_chan_info *cinfo,
> > "%s() failed to add to virtqueue (%d)\n", __func__,
> > rc);
> > } else {
> > + dev_dbg(vioch->cinfo->dev,
> > + "VQUEUE[%d] - REQUEST - PROTO:0x%X ID:0x%X XFER_ID:%d XFER:%px RX_LEN:%zd\n",
> > + vioch->vqueue->index, xfer->hdr.protocol_id,
> > + xfer->hdr.id, xfer->hdr.seq, xfer, xfer->rx.len);
>
> Indentation appears to be inconsistent.
>

Right.


Thanks,
Cristian

\
 
 \ /
  Last update: 2021-07-01 13:27    [W:0.093 / U:1.244 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site