Re: [PATCH 2/3] ACPI: RISC-V: Enumerate CBQRI controllers from the RQSC table
From: Jonathan Cameron
Date: Wed Nov 18 2026 - 18:42:16 EST
On Tue, 17 Nov 2026 14:11:28 -0800
"Drew Fustini (Tenstorrent)" <fustini@xxxxxxxxxx> wrote:
> Walk the RQSC table at ACPI init and create a platform device for each
> CBQRI controller, so one platform driver can serve device tree and ACPI
> systems. Nothing selects ACPI_RQSC until the CBQRI driver can bind to
> these devices.
>
> Signed-off-by: Drew Fustini (Tenstorrent) <fustini@xxxxxxxxxx>
Hi Drew
Very much a drive by review as I haven't even opened the spec yet ;)
Anyhow, various comments inline. Most code readability suggestions.
Thanks,
Jonathan
> diff --git a/drivers/acpi/riscv/rqsc.c b/drivers/acpi/riscv/rqsc.c
> new file mode 100644
> index 000000000000..262fd5755d29
> --- /dev/null
> +++ b/drivers/acpi/riscv/rqsc.c
> @@ -0,0 +1,210 @@
"
> +
> +static struct platform_device * __init
> +rqsc_add_device(unsigned int idx, const struct acpi_rqsc_controller *ctrl,
> + const struct acpi_rqsc_resource *res)
> +{
> + bool is_cc = ctrl->type == ACPI_RQSC_CONTROLLER_TYPE_CAPACITY;
Not a fan of a bool when we might get more types at some point.
If you want to do something like this have a trivial helper
bool rqsc_ctrl_is_cc() and use it inline.
> + struct platform_device *pdev __free(platform_device_put) =
> + platform_device_alloc(is_cc ? "cbqri-capacity" : "cbqri-bandwidth",
> + idx);
Not particularly readable or extensible if more types show up.
I'd use a switch that sets the string. Declare the pdev and __free after that.
> + phys_addr_t addr = ctrl->address.address;
> + /* The GAS carries no length, so claim up to the end of the page. */
> + struct resource mem = DEFINE_RES_MEM(addr, PAGE_SIZE - offset_in_page(addr));
> + /* rcid, mcid, cache id and a sentinel */
> + struct property_entry props[4] = {};
> + unsigned int n = 0;
> + int ret;
> +
> + if (!pdev)
Drag the pdev alloc down to just above this. Having it burried
in allocations is not easy to read when you only check for
success all the way down here.
> + return ERR_PTR(-ENOMEM);
> +
> + ret = platform_device_add_resources(pdev, &mem, 1);
> + if (ret)
> + return ERR_PTR(ret);
> +
> + props[n++] = PROPERTY_ENTRY_U32("riscv,cbqri-rcid", ctrl->rcid_count);
> + props[n++] = PROPERTY_ENTRY_U32("riscv,cbqri-mcid", ctrl->mcid_count);
> + if (is_cc)
> + props[n++] = PROPERTY_ENTRY_U32("riscv,cbqri-cache-id",
> + lower_32_bits(res->id1));
> + else
> + set_dev_node(&pdev->dev, pxm_to_node(lower_32_bits(res->id1)));
> +
> + ret = device_create_managed_software_node(&pdev->dev, props, NULL);
> + if (ret)
> + return ERR_PTR(ret);
> +
> + ret = platform_device_add(pdev);
> + if (ret)
> + return ERR_PTR(ret);
> +
> + return_ptr(pdev);
> +}
> +
> +static void __init rqsc_parse(const struct acpi_table_rqsc *rqsc)
> +{
> + const void *end = (const void *)rqsc + rqsc->header.length;
> + const struct acpi_rqsc_controller *ctrl = (const void *)(rqsc + 1);
> + unsigned int idx = 0, added = 0;
> +
> + while ((const void *)ctrl < end) {
> + const struct acpi_rqsc_resource *res = (const void *)(ctrl + 1);
> + struct platform_device *pdev;
> + phys_addr_t addr;
> +
> + /*
> + * The controller structure must fit and stay inside the
> + * table. Anything else is corruption and the remaining
> + * entries cannot be trusted.
> + */
> + if ((const void *)ctrl + sizeof(*ctrl) > end ||
if ((const void *)(ctrl + 1) > end ?
Which looks familiar. Isn't that
if ((const void *)res > end)
Maybe that is a little confusing as res might be zero size.
For that matter I'd not set res until after this check because it
might take an invalid value. Doesn't matter as such but is confusing
to read.
> + ctrl->length < sizeof(*ctrl) ||
> + (const void *)ctrl + ctrl->length > end) {
> + pr_err("entry %u is malformed, stopping\n", idx);
> + break;
> + }
> +
> + addr = ctrl->address.address;
> +
> + if (ctrl->type != ACPI_RQSC_CONTROLLER_TYPE_CAPACITY &&
> + ctrl->type != ACPI_RQSC_CONTROLLER_TYPE_BANDWIDTH) {
> + pr_warn("controller at %pa: unknown type %u, skipping\n",
> + &addr, ctrl->type);
> + goto next;
> + }
> +
> + if (ctrl->address.space_id != ACPI_ADR_SPACE_SYSTEM_MEMORY || !addr) {
> + pr_warn("entry %u: no usable register interface address, skipping\n",
> + idx);
> + goto next;
> + }
> +
> + if (!ctrl->rcid_count && !ctrl->mcid_count) {
> + pr_warn("controller at %pa: no RCIDs and no MCIDs, skipping\n",
> + &addr);
> + goto next;
> + }
> +
> + if (!ctrl->num_resources ||
Little fiddly but I'd split this. The reason being that the next
check if kind of checking for only one resource, but you only confirm
that is ok with pr_warn() path that follows. So maybe
if (!ctrl->num_resources) {
pr_warn("controller at %pa: no resource, skipping\n",
&addr);
I'd go a bit long on lines - not sure ACPI is that fussy as subsystem go.
goto next;
}
if (ctrl->num_resources > 1)
pr_warn("controller at %pa: %u resources, using the first\n",
&addr, ctrl->num_resources);
//now we have clarity that we only care about first resource.
if (ctrl->length < sizeof(*ctrl) + sizeof(*res) ||
res->length < sizeof(*res)) {
pr_warn("controller at %pa: no usable resource, skipping\n",
&addr);
goto next;
}
> + ctrl->length < sizeof(*ctrl) + sizeof(*res) ||
> + res->length < sizeof(*res)) {
> + pr_warn("controller at %pa: no usable resource, skipping\n",
> + &addr);
> + goto next;
> + }
> +
> + if (ctrl->num_resources > 1)
> + pr_warn("controller at %pa: %u resources, using the first\n",
> + &addr, ctrl->num_resources);
> +
> + if (!rqsc_resource_ok(ctrl, res, addr))
> + goto next;
> +
> + pdev = rqsc_add_device(idx, ctrl, res);
> + if (IS_ERR(pdev)) {
> + pr_warn("controller at %pa: failed to add device: %pe\n",
> + &addr, pdev);
This feels like the sort of thing you should just fail init on.
Papering over bits of the table that don't like you expect
might be dealing with changes that have been made. Failing to add
a device is something much worse.
> + goto next;
> + }
> +
> + added++;
> +next:
> + ctrl = (const void *)ctrl + ctrl->length;
> + idx++;
Could you turn this whole thing into a do { } while() and
update these conditions in the while bit? Then the various
gotos become continues.
> + }
> +
> + if (idx != rqsc->num_controllers)
> + pr_warn("table announces %u controllers, found %u\n",
> + rqsc->num_controllers, idx);
> +
> + if (added)
> + pr_info("%u CBQRI controllers\n", added);
> + else
> + pr_warn("no usable CBQRI controller\n");
> +}
> +
> +void __init riscv_acpi_rqsc_init(void)
> +{
> + struct acpi_table_header *table __free(acpi_put_table) = NULL;
See cleanup.h That needs to move to next to the constructor.
> + acpi_status status;
> +
> + BUILD_BUG_ON(sizeof(struct acpi_rqsc_resource) != 20);
> + BUILD_BUG_ON(sizeof(struct acpi_rqsc_controller) != 24);
> + BUILD_BUG_ON(sizeof(struct acpi_table_rqsc) != 40);
It's an ACPI table. You should be updating it only in a safe
forwards compatible way. Hence I'd not normally expect hard checks
that will fail if it changes in acpica.
> +
Declare the struct .. * table _free() down here.
> + status = acpi_get_table(ACPI_SIG_RQSC, 0, &table);
> + if (status == AE_NOT_FOUND)
> + return;
> +
> + if (ACPI_FAILURE(status)) {
> + pr_err("failed to get table: %s\n",
> + acpi_format_exception(status));
> + return;
> + }
> +
> + /*
> + * A later revision could extend the fixed header, which would
> + * shift the first controller structure. Stop rather than guess.
That would be an ACPI spec bug. Sure they have happened a few times
(usually when assumption was no one had implemented something that needed
a fix) but the rules for ACPI are very strict that any change even with
a version tick should not make incompatible changes. You add stuff
and sometime reserve things that previously had values, but you don't
move fields around unless there is an offset indicator or size value
in the structure to allow you to do that.
If riscv isn't keeping to that, I'd suggest working to make sure
people understand this. We can't have a regression because someone
put that sort of change in the spec.
> + */
> + if (table->revision != ACPI_RQSC_REVISION) {
> + pr_err("table revision %u, expected %u\n",
> + table->revision, ACPI_RQSC_REVISION);
Likewise, that rings alarm bells about the spec process if there
is any likelihood of that being needed. The only similar
instances that I've seen were where things accidentally broke
and we had to deal with that - and HMAT where this happened
we at least landed support for v1 and v2 at the same time.
> + return;
> + }
> +
> + if (table->length < sizeof(struct acpi_table_rqsc)) {
> + pr_err("table truncated: length %u\n", table->length);
> + return;
> + }
> +
> + rqsc_parse((const struct acpi_table_rqsc *)table);
> +}
>