Re: [PATCH v7] crypto: qce - Add runtime PM and interconnect bandwidth scaling support
From: Udit Tiwari
Date: Wed Mar 04 2026 - 00:31:00 EST
Thanks for the review.
I’ll be posting a v8 soon with oss.qualcomm.com.
On 2/21/2026 5:09 AM, Bjorn Andersson wrote:
On Fri, Feb 20, 2026 at 12:58:18PM +0530, quic_utiwari@xxxxxxxxxxx wrote:
From: Udit Tiwari <quic_utiwari@xxxxxxxxxxx>
The Qualcomm Crypto Engine (QCE) driver currently lacks support for
runtime power management (PM) and interconnect bandwidth control.
As a result, the hardware remains fully powered and clocks stay
enabled even when the device is idle. Additionally, static
interconnect bandwidth votes are held indefinitely, preventing the
system from reclaiming unused bandwidth.
Address this by enabling runtime PM and dynamic interconnect
bandwidth scaling to allow the system to suspend the device when idle
and scale interconnect usage based on actual demand. Improve overall
system efficiency by reducing power usage and optimizing interconnect
resource allocation.
Make the following changes as part of this integration:
"this integration" is internal lingo. In fact I think you can omit this
whole paragraph, because the bullets here are expected.
- Add support for pm_runtime APIs to manage device power state
transitions.
- Implement runtime_suspend() and runtime_resume() callbacks to gate
clocks and vote for interconnect bandwidth only when needed.
- Replace devm_clk_get_optional_enabled() with devm_pm_clk_create() +
pm_clk_add() and let the PM core manage device clocks during runtime
PM and system sleep.
- Register dev_pm_ops with the platform driver to hook into the PM
framework.
Tested:
This isn't very useful to carry in the git history, please move this
down below '---'.
- Verify that ICC votes drop to zero after probe and upon request
completion.
- Confirm that runtime PM usage count increments during active
requests and decrements afterward.
- Observe that the device correctly enters the suspended state when
idle.
Signed-off-by: Udit Tiwari <quic_utiwari@xxxxxxxxxxx>
Please switch to oss.qualcomm.com
---
Changes in v7:
- Use ACQUIRE guard in probe to simplify runtime PM management and error paths.
- Drop redundant icc_enable() call in runtime resume path.
- Explicitly call pm_clk_suspend(dev) and pm_clk_resume(dev) within the
custom runtime PM callbacks. Since custom callbacks are provided to handle
interconnect scaling, the standard PM clock helpers must be invoked manually
to ensure clocks are gated/ungated.
Changes in v6:
- Adopt ACQUIRE(pm_runtime_active_try, ...) for scoped runtime PM management
in qce_handle_queue(). This removes the need for manual put calls and
goto labels in the error paths, as suggested by Konrad.
- Link to v6: https://lore.kernel.org/lkml/20260210061437.2293654-1-quic_utiwari@xxxxxxxxxxx/
Changes in v5:
- Drop Reported-by and Closes tags for kernel test robot W=1 warnings, as
the issue was fixed within the same patch series.
- Fix a minor comment indentation/style issue.
- Link to v5: https://lore.kernel.org/lkml/20251120062443.2016084-1-quic_utiwari@xxxxxxxxxxx/
Changes in v4:
- Annotate runtime PM callbacks with __maybe_unused to silence W=1 warnings.
- Add Reported-by and Closes tags for kernel test robot warning.
- Link to v4: https://lore.kernel.org/lkml/20251117062737.3946074-1-quic_utiwari@xxxxxxxxxxx/
Changes in v3:
- Switch from manual clock management to PM clock helpers
(devm_pm_clk_create() + pm_clk_add()); no direct clk_* enable/disable
in runtime callbacks.
- Replace pm_runtime_get_sync() with pm_runtime_resume_and_get(); remove
pm_runtime_put_noidle() on error.
- Define PM ops using helper macros and reuse runtime callbacks for system
sleep via pm_runtime_force_suspend()/pm_runtime_force_resume().
- Link to v2: https://lore.kernel.org/lkml/20250826110917.3383061-1-quic_utiwari@xxxxxxxxxxx/
Changes in v2:
- Extend suspend/resume support to include runtime PM and ICC scaling.
- Register dev_pm_ops and implement runtime_suspend/resume callbacks.
- Link to v1: https://lore.kernel.org/lkml/20250606105808.2119280-1-quic_utiwari@xxxxxxxxxxx/
---
drivers/crypto/qce/core.c | 87 +++++++++++++++++++++++++++++++++------
1 file changed, 75 insertions(+), 12 deletions(-)
diff --git a/drivers/crypto/qce/core.c b/drivers/crypto/qce/core.c
index b966f3365b7d..776a08340b08 100644
--- a/drivers/crypto/qce/core.c
+++ b/drivers/crypto/qce/core.c
@@ -12,6 +12,9 @@
#include <linux/module.h>
#include <linux/mod_devicetable.h>
#include <linux/platform_device.h>
+#include <linux/pm.h>
+#include <linux/pm_runtime.h>
+#include <linux/pm_clock.h>
#include <linux/types.h>
#include <crypto/algapi.h>
#include <crypto/internal/hash.h>
@@ -90,6 +93,11 @@ static int qce_handle_queue(struct qce_device *qce,
struct crypto_async_request *async_req, *backlog;
int ret = 0, err;
First use of ret is now an assignment, so please drop the unnecessary
zero-initialization.
+ ACQUIRE(pm_runtime_active_try, pm)(qce->dev);
+ ret = ACQUIRE_ERR(pm_runtime_active_auto_try, &pm);
Luckily, this - to me - incomprehensible construct got some useful
wrappers in ef8057b07c72 ("PM: runtime: Wrapper macros for
ACQUIRE()/ACQUIRE_ERR()"), merged back in November. So, you should
instead use the form:
PM_RUNTIME_ACQUIRE(qce->dev, pm);
if ((ret = PM_RUNTIME_ACQUIRE_ERR(&pm)))
return ret;
or (I don't think we really care about the specific error returned):
PM_RUNTIME_ACQUIRE(qce->dev, pm);
if (PM_RUNTIME_ACQUIRE_ERR(&pm))
return -Esome_specific_error;
Although I presume that's PM_RUNTIME_ACQUIRE_AUTOSUSPEND() in your case.
+ if (ret)
+ return ret;
+
scoped_guard(mutex, &qce->lock) {
if (req)
ret = crypto_enqueue_request(&qce->queue, req);
@@ -207,23 +215,34 @@ static int qce_crypto_probe(struct platform_device *pdev)
if (ret < 0)
return ret;
- qce->core = devm_clk_get_optional_enabled(qce->dev, "core");
- if (IS_ERR(qce->core))
- return PTR_ERR(qce->core);
+ /* PM clock helpers: register device clocks */
+ ret = devm_pm_clk_create(dev);
+ if (ret)
+ return ret;
- qce->iface = devm_clk_get_optional_enabled(qce->dev, "iface");
- if (IS_ERR(qce->iface))
- return PTR_ERR(qce->iface);
+ ret = pm_clk_add(dev, "core");
+ if (ret)
+ return ret;
- qce->bus = devm_clk_get_optional_enabled(qce->dev, "bus");
- if (IS_ERR(qce->bus))
- return PTR_ERR(qce->bus);
+ ret = pm_clk_add(dev, "iface");
+ if (ret)
+ return ret;
- qce->mem_path = devm_of_icc_get(qce->dev, "memory");
+ ret = pm_clk_add(dev, "bus");
+ if (ret)
+ return ret;
+
+ qce->mem_path = devm_of_icc_get(dev, "memory");
if (IS_ERR(qce->mem_path))
return PTR_ERR(qce->mem_path);
- ret = icc_set_bw(qce->mem_path, QCE_DEFAULT_MEM_BANDWIDTH, QCE_DEFAULT_MEM_BANDWIDTH);
+ /* Enable runtime PM after clocks and ICC are acquired */
It wouldn't hurt to continue this sentence to include that you're doing
it like that so that clocks and interconnect votes are applied by the
resume callback (I assume that's your reason for the ordering at least).
+ ret = devm_pm_runtime_enable(dev);
+ if (ret)
+ return ret;
+
+ ACQUIRE(pm_runtime_active_try, pm)(dev);
+ ret = ACQUIRE_ERR(pm_runtime_active_auto_try, &pm);
As above.
if (ret)
return ret;
@@ -245,9 +264,52 @@ static int qce_crypto_probe(struct platform_device *pdev)
qce->async_req_enqueue = qce_async_request_enqueue;
qce->async_req_done = qce_async_request_done;
- return devm_qce_register_algs(qce);
+ ret = devm_qce_register_algs(qce);
+ if (ret)
+ return ret;
+
+ /* Configure autosuspend after successful init */
+ pm_runtime_set_autosuspend_delay(dev, 100);
+ pm_runtime_use_autosuspend(dev);
+ pm_runtime_mark_last_busy(dev);
+
+ return 0;
}
+static int __maybe_unused qce_runtime_suspend(struct device *dev)
+{
+ struct qce_device *qce = dev_get_drvdata(dev);
+
+ icc_disable(qce->mem_path);
+
+ return pm_clk_suspend(dev);
+}
+
+static int __maybe_unused qce_runtime_resume(struct device *dev)
+{
+ struct qce_device *qce = dev_get_drvdata(dev);
+ int ret = 0;
+
+ ret = pm_clk_resume(dev);
What is the reason to use pm_clk_add() if you need to manually
enable/disable them anyways?
Regarding pm_clk_add(): the intent was to register the device clocks with the PM clock framework so the runtime PM callbacks could use pm_clk_suspend() / pm_clk_resume() instead of open-coding clock gating.
However, since this driver already needs explicit runtime PM callbacks anyway (for interconnect votes, etc.), and pm_clk_add() by itself doesn’t provide any “automatic” clock toggling, this doesn’t really buy us much. I’ll simplify the next revision by dropping the PM clock registration and handling the clocks directly in the runtime PM callbacks.
+ if (ret)
+ return ret;
+
+ ret = icc_set_bw(qce->mem_path, QCE_DEFAULT_MEM_BANDWIDTH, QCE_DEFAULT_MEM_BANDWIDTH);
icc_disable() will set the "enabled" flag on the internal interconnect
data structures, which causes the vote to be skipped in aggregation.
So, not only does it look unbalanced with icc_disable() vs icc_set_bw()
in suspend/resume, I don't think you have a bandwidth vote after the
first suspend, unless you call icc_enable() here.
Thanks for the note about icc_disable() affecting the internal enabled flag and aggregation.
To address this, I see two options:
(1) Keep icc_disable() in runtime suspend, but then explicitly call icc_enable() in runtime resume before restoring the bandwidth vote (similar to what I did here: https://lore.kernel.org/lkml/20260210061437.2293654-1-quic_utiwari@xxxxxxxxxxx/).
(2) Drop icc_disable() entirely and instead use icc_set_bw(path, 0, 0) in runtime suspend, restoring the BW in runtime resume.
Is option (2) the preferred approach for ICC consumers? My understanding is it avoids the “disabled path skipped during aggregation” issue and keeps the state machine balanced.
Udit
+ if (ret)
Just put pm_clk_suspend(dev) here and return ret; below. Skip the goto.
Regards,
Bjorn
+ goto err_icc;
+
+ return 0;
+
+err_icc:
+ pm_clk_suspend(dev);
+ return ret;
+}
+
+static const struct dev_pm_ops qce_crypto_pm_ops = {
+ SET_RUNTIME_PM_OPS(qce_runtime_suspend, qce_runtime_resume, NULL)
+ SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend, pm_runtime_force_resume)
+};
+
static const struct of_device_id qce_crypto_of_match[] = {
{ .compatible = "qcom,crypto-v5.1", },
{ .compatible = "qcom,crypto-v5.4", },
@@ -261,6 +323,7 @@ static struct platform_driver qce_crypto_driver = {
.driver = {
.name = KBUILD_MODNAME,
.of_match_table = qce_crypto_of_match,
+ .pm = &qce_crypto_pm_ops,
},
};
module_platform_driver(qce_crypto_driver);
--
2.34.1