lkml.org 
[lkml]   [2018]   [Mar]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v6 03/20] firmware: arm_scmi: add basic driver infrastructure for SCMI


> >> +/**
> >> + * scmi_one_xfer_get() - Allocate one message
> >> + *
> >> + * @handle: SCMI entity handle
> >> + *
> >> + * Helper function which is used by various command functions that are
> >> + * exposed to clients of this driver for allocating a message traffic event.
> >> + *
> >> + * This function can sleep depending on pending requests already in the system
> >> + * for the SCMI entity. Further, this also holds a spinlock to maintain
> >> + * integrity of internal data structures.
> >> + *
> >> + * Return: 0 if all went fine, else corresponding error.
> >> + */
> >> +static struct scmi_xfer *scmi_one_xfer_get(const struct scmi_handle *handle)
> > Maybe it's just me, but I think this would be more clearly named as
> > scmi_xfer_alloc.
> >
>
> Agreed to some extent. The reason I didn't have it as alloc as they are
> preallocated and this just returns a reference to free slot in that
> preallocated array.
>
> > I'd assume we were dealing with one anyway as it's not called scmi_xfers_get
> > and the get to my mind implies a reference counter rather than allocating
> > an xfer for use...
> >
>
> Ah OK, I get your concerne with _get/_put but _alloc/_free is equally
> bad then in the contect of preallocated buffers.
Sure, this is always a fun question. Lots of other options
_assign _deassign works but never feels nice.

>
...
>
> >> + .max_msg = 20, /* Limited by MBOX_TX_QUEUE_LEN */
> >> + .max_msg_size = 128,
> >> +};
> >> +
> >> +/* Each compatible listed below must have descriptor associated with it */
> >> +static const struct of_device_id scmi_of_match[] = {
> >> + { .compatible = "arm,scmi", .data = &scmi_generic_desc },
> >> + { /* Sentinel */ },
> >> +};
> >> +
> >> +MODULE_DEVICE_TABLE(of, scmi_of_match);
> >> +
> >> +static int scmi_xfer_info_init(struct scmi_info *sinfo)
> >> +{
> >> + int i;
> >> + struct scmi_xfer *xfer;
> >> + struct device *dev = sinfo->dev;
> >> + const struct scmi_desc *desc = sinfo->desc;
> >> + struct scmi_xfers_info *info = &sinfo->minfo;
> >> +
> >> + /* Pre-allocated messages, no more than what hdr.seq can support */
> >> + if (WARN_ON(desc->max_msg >= (MSG_TOKEN_ID_MASK + 1))) {
> >> + dev_err(dev, "Maximum message of %d exceeds supported %d\n",
> >> + desc->max_msg, MSG_TOKEN_ID_MASK + 1);
> >> + return -EINVAL;
> >> + }
> >> +
> >> + info->xfer_block = devm_kcalloc(dev, desc->max_msg,
> >> + sizeof(*info->xfer_block), GFP_KERNEL);
> >> + if (!info->xfer_block)
> >> + return -ENOMEM;
> >> +
> >> + info->xfer_alloc_table = devm_kcalloc(dev, BITS_TO_LONGS(desc->max_msg),
> >> + sizeof(long), GFP_KERNEL);
> >> + if (!info->xfer_alloc_table)
> >> + return -ENOMEM;
> > Hmm. I wonder if it is worth adding a devm_bitmap_alloc?
> >
>
> OK
>
> >> +
> >> + bitmap_zero(info->xfer_alloc_table, desc->max_msg);
> > kcalloc zeros the memory.
> >
> >> +
> >> + /* Pre-initialize the buffer pointer to pre-allocated buffers */
> >> + for (i = 0, xfer = info->xfer_block; i < desc->max_msg; i++, xfer++) {
> >> + xfer->rx.buf = devm_kcalloc(dev, sizeof(u8), desc->max_msg_size,
> >> + GFP_KERNEL);
> >> + if (!xfer->rx.buf)
> >> + return -ENOMEM;
> >> +
> >> + xfer->tx.buf = xfer->rx.buf;
> >> + init_completion(&xfer->done);
> >> + }
> >> +
> >> + spin_lock_init(&info->xfer_lock);
> >> +
> >> + return 0;
> >> +}
> >> +
> >> +static int scmi_mailbox_check(struct device_node *np)
> >> +{
> >> + struct of_phandle_args arg;
> >> +
> >> + return of_parse_phandle_with_args(np, "mboxes", "#mbox-cells", 0, &arg);
> >> +}
> >> +
> >> +static int scmi_mbox_free_channel(struct scmi_info *info)
> > Some of the naming in here could do with being adjusted to be obviously
> > 'balanced'. The moment I see a free I expect a matched alloc but in this
> > case the alloc is done in scmi_mbox_chan_setup which at very least
> > should be scmi_mbox_setup_channel and should really imply that it is
> > doing allocation as well.
> >
>
> That's inline with mailbox APIs.

oh goody. Fair enough if ugly
>
...
> OK

\
 
 \ /
  Last update: 2018-03-05 15:49    [W:0.210 / U:0.228 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site