spi: pxa2xx: Fix PM and interrupt issues on Intel LPSS SPI
Subsystems (3)
v16 cover spi: pxa2xx: Fix PM and interrupt issues on Intel LPSS SPI
This 7-patch series addresses long-standing power management (PM), runtime autosuspend, and interrupt synchronization regressions in the spi-pxa2xx host controller driver, specifically targeting Intel Low Power Subsystem (LPSS) SPI controllers and platforms such as the Apple MacBook8,1.
Specifically, this series:
- Prevents system hangs and PCIe Completion Timeouts during S3 suspend/resume by restoring Intel LPSS private register context (resets and control registers).
- Resolves race conditions in the shared interrupt handler (ssp_int()) during PM power state transitions (RPM_SUSPENDING) by tracking suspend states and guaranteeing active hardware clocks prior to MMIO register access.
- Locks out runtime PM autosuspend for Intel LPSS SPI controllers in PIO mode to avoid unclocked MMIO accesses on subsequent transfers.
- Applies a PCI glue DMI quirk to force PIO mode for Apple MacBook8,1 to work around persistent hardware DMA timeouts.
- Cleanly refactors clock helper functions, PM callback state alignments, and local variable conventions across the driver.
Changes in v16:
- Addressed feedback from Sashiko review on the v15 patchset:
- Fixed 'bool suspended' field location in spi-pxa2xx.h (moved from Patch 1 to Patch 2).
- Updated pxa2xx_spi_clk_disable() to set drv_data->clk_enabled = false BEFORE calling clk_disable_unprepare(), eliminating an IRQ handler race window on multi-core systems.
- Updated pxa2xx_spi_suspend() and pxa2xx_spi_runtime_suspend() to explicitly
set drv_data->suspended = true.
- Fixed pm_runtime_put_sync() in pxa2xx_spi_remove(), moving the PM reference drop to before hardware teardown and using pm_runtime_put_noidle() to prevent runtime_suspend calls on unclocked hardware.
- Fixed PM runtime state alignment in pxa2xx_spi_resume() error path using pm_runtime_set_suspended().
- Resolved merge conflicts and renamed remaining 'status' variables to 'ret' in suspend/resume handlers.
Changes in v15:
- Split the large PIO mode PM and interrupt patch (v14 Patch 2) into 4 distinct, single-purpose commits (clock helpers, suspended flag, teardown overhaul, and autosuspend lockout).
- Reordered the commits so that PIO PM and interrupt bug fixes are applied before disabling DMA on MacBook8,1, satisfying kernel bisection safety.
- Replaced raw pxa_ssp_disable() calls in PM suspend/remove with pxa2xx_spi_off() to respect the MMP2 platform's SSE-disable quirk.
- Renamed return variables from 'status' to 'ret' in clock helper functions, and added a cleanup commit to rename 'status' to 'ret' in the PM and probe callbacks to conform to standard coding conventions.
Changes in v14:
- Addressed feedback from Sashiko review on the v13 patchset for Patch 2:
- Updated commit message text for Patch 2 to accurately reflect that ssp_int()
uses drv_data->suspended and drv_data->clk_enabled instead of pm_runtime_get_if_active().
- Clarified in commit message that active SPI transfers hold a PM reference via spi_controller.auto_runtime_pm (pm_runtime_get_sync()), preventing autosuspend from racing while an IRQ handler is reading/writing FIFO registers.
- Clarified teardown ordering: pxa_ssp_disable() disables hardware interrupt generation first, followed by setting drv_data->suspended = true and calling synchronize_irq(), ensuring in-flight IRQ handlers drain completely before clock gating.
Changes in v13:
- Addressed feedback from Sashiko review on the v12 patchset for Patch 2:
- Removed pm_runtime_get_if_active() check from ssp_int(). During PM state transitions (such as RPM_SUSPENDING or RPM_RESUMING), pm_runtime_get_if_active() returns 0 because the state is not RPM_ACTIVE. Returning IRQ_NONE during transition without clearing a level-triggered interrupt would cause the interrupt controller to endlessly re-invoke the handler in a loop.
- Rely on drv_data->suspended and drv_data->clk_enabled in ssp_int() instead. If the clock is enabled (drv_data->clk_enabled == true), MMIO reads are 100% safe and will not cause PCIe Completion Timeouts. If the clock is disabled or the device is suspended, ssp_int() immediately returns IRQ_NONE to prevent unclocked access.
Changes in v12:
- Addressed feedback from Sashiko review on the v11 patchset for Patch 2:
- Preserved the if (!pm_runtime_suspended(dev)) check prior to enabling/disabling the clock in pxa2xx_spi_suspend() and pxa2xx_spi_resume(). On non-LPSS platforms, if a device was runtime suspended prior to system sleep, unconditionally enabling its hardware clock during system resume forced the clock ON while the PM core retained RPM_SUSPENDED. Restoring this check prevents PM state desynchronization, avoiding power leaks on non-LPSS platforms.
Changes in v11:
- Addressed feedback from Sashiko review on the v10 patchset for Patch 2:
- Replaced pm_runtime_put_noidle() with pm_runtime_put_sync() in pxa2xx_spi_remove(). Using pm_runtime_put_noidle() dropped usage count without executing the runtime_suspend callback or updating the device runtime status, leaving it stuck in RPM_ACTIVE. This permanently leaked an active child count on the parent LPSS power domain, preventing the parent from entering low-power runtime suspend. Switching to pm_runtime_put_sync() ensures the runtime PM state machine properly transitions to RPM_SUSPENDED upon driver unbind.
Changes in v10:
- Addressed feedback from Sashiko review on the v9 patchset:
- Fixed ssp_int() by explicitly returning IRQ_NONE when pm_runtime_get_if_active() returns 0 (device inactive or suspending). This prevents reading SSSR and MMIO registers when the hardware is powered down or in power transition.
- Resolved PM usage counter leak and double-increment for PIO mode devices. Removed the redundant/unconditional pm_runtime_get_noresume() in pxa2xx_spi_probe() and moved the single pm_runtime_get_noresume() call to occur strictly after spi_register_controller() succeeds.
Changes in v9:
- Separated the original "disable DMA and fix runtime PM" patch into two distinct commits:
1. spi: pxa2xx: disable DMA for Apple MacBook8,1 (PCI glue DMI quirk)
2. spi: pxa2xx: fix runtime PM and interrupt handling in PIO mode (core driver PM/IRQ fixes)
- Condensed commit messages for Patches 2 and 3 for clarity and brevity.
- Fixed checkpatch.pl warning regarding unnecessary braces in pxa2xx_spi_remove().
- Addressed feedback from Sashiko review on the v8 patchset:
- Scoped pm_runtime_get_noresume() in pxa2xx_spi_probe() and pm_runtime_put_noidle() in remove/error paths using is_lpss_ssp(drv_data). This locks out autosuspend for LPSS controllers operating in PIO mode without causing power regressions on non-LPSS platforms (e.g. PXA25x, Intel Quark, CE4100) operating in PIO mode.
- Guarded MMIO accesses in pxa2xx_spi_runtime_suspend() with a drv_data->clk_enabled check. If the clock has already been turned off (e.g. in pxa2xx_spi_remove() or resume error path), runtime suspend skips hardware writes, avoiding unclocked MMIO accesses and PCIe Completion Timeouts if runtime PM triggers after teardown.
- Cleaned up pxa2xx_spi_pci_can_dma() signature in spi-pxa2xx-pci.c by removing the redundant bool verbose parameter.
Changes in v8:
- Addressed feedback from Sashiko review on the v7 patchset:
- Fixed runtime PM resume interrupt storm in ssp_int() by checking drv_data->clk_enabled
instead of pm_runtime_get_if_active() == 0.
- Fixed PM disabled configuration (CONFIG_PM=n) support in ssp_int() by avoiding active <= 0 early returns.
- Reordered suspend and remove sequences to invoke pxa_ssp_disable() before setting drv_data->suspended = true and synchronizing the IRQ, closing a race window where level-triggered interrupts could cause a storm.
- Masked hardware interrupt generation in pxa2xx_spi_runtime_suspend() via pxa_ssp_disable() to prevent unexpected interrupts during clock enable.
- Avoided PCIe Completion Timeout system hangs on newer LPSS platforms (SPT, BXT, CNL) by introducing pxa2xx_spi_need_lpss_restore() to restrict MMIO save/restore loops to LPT, BYT, and BSW platforms (which lack an MFD parent).
- Fixed compiler error in pxa2xx_spi_probe() due to unused label.
Changes in v7:
- Addressed feedback from Sashiko review on the v6 patchset:
- Fixed a race condition during probe by moving the request_irq() call after the clock is enabled and the suspended flag is cleared. This prevents an early shared interrupt from asserting and triggering an interrupt storm before the clock is active to allow clearing it.
- Rectified the teardown sequence in pxa2xx_spi_remove(): first set drv_data->suspended to true and disable SSP hardware-level interrupt generation (pxa_ssp_disable()), then call synchronize_irq() to wait for in-flight interrupt handlers to complete, free the IRQ, and only then disable the clocks. This eliminates both post-clock-disable MMIO accesses and unhandled shared hardware interrupt storms.
- Clarified that the unconditional MMIO register access in pxa2xx_spi_suspend() is safe because pm_runtime_resume_and_get() is invoked at the very beginning of the suspend callback, guaranteeing the LPSS device is active and clocked during register disabling/saving.
- Added spi_controller_resume() recovery to the error path of spi_controller_suspend() in pxa2xx_spi_suspend() to prevent the controller from remaining permanently disabled in the event system suspend is aborted.
Changes in v6:
- Addressed feedback from Sashiko review on the v5 patchset:
- Added a synchronize_irq() call to the spi_controller_resume() error path in pxa2xx_spi_resume(). This ensures any concurrent shared interrupt handlers (which might execute because drv_data->suspended = false was set earlier) finish executing before we disable the clock, preventing PCIe timeouts.
- Checked and confirmed that active == 0 is the correct check in ssp_int(). If Runtime PM is disabled, pm_runtime_get_if_active() returns a negative error code (like -EINVAL). Changing this check to active <= 0 would cause a regression on non-PM configurations because the handler would always return IRQ_NONE. Under disabled Runtime PM, the hardware clock is kept constantly active, so it is safe to proceed and read registers when active < 0.
Changes in v5:
- Reverted runtime PM configuration in spi-pxa2xx-pci.c to be unconditional. This prevents the usage count from leaking by +1 on unbind for PIO mode, while letting pxa2xx_spi_probe()'s pm_runtime_get_noresume() and pxa2xx_spi_remove()'s pm_runtime_put_noidle() handle the permanent autosuspend lockout for PIO devices symmetrically.
- Removed the pm_runtime_forbid() call from pxa2xx_spi_remove(). This resolves reference leaks for non-PCI platform devices where pm_runtime_allow() was never called during probe.
- Delayed clearing the drv_data->suspended flag on resume until after LPSS reset deassertion and private register restoration are complete. This prevents a shared interrupt from firing during resume and attempting to read the SSSR register while the LPSS block is still held in reset.
- Set drv_data->suspended = true on spi_controller_resume() failure path in pxa2xx_spi_resume(). This ensures that subsequent shared interrupts do not attempt register access once the clock is disabled.
Changes in v4:
- Track clock state using drv_data->clk_enabled via pxa2xx_spi_clk_enable() and pxa2xx_spi_clk_disable() helper functions. This guarantees clock enable/disable symmetry, preventing clock disable count underflows and framework warnings on S3 resume or runtime autosuspend error paths.
- Introduce drv_data->suspended flag to protect MMIO access in ssp_int() during system suspend and runtime suspend transition windows.
- Initialize drv_data->suspended = true early in probe(), clearing it only after the clock is successfully enabled. This completely prevents shared interrupt handler races during device probe when the clock is still off.
- Call synchronize_irq() after setting drv_data->suspended = true in suspend and runtime_suspend. This ensures any running shared interrupt handlers finish executing before the clock is physically turned off.
Changes in v3:
- Avoid PM reference leaks on probe bind/unbind cycle by keeping probe PM configuration symmetric.
- Prevent userspace (PowerTOP, udev) from overriding runtime PM settings when DMA is disabled by holding a PM reference via pm_runtime_get_noresume() in pxa2xx_spi_probe() and dropping it in remove/error paths.
- Check device status in the shared interrupt handler ssp_int() using pm_runtime_get_if_active() instead of pm_runtime_suspended(). If the device is suspending (RPM_SUSPENDING) or suspended, ssp_int() immediately returns IRQ_NONE to avoid reading unclocked MMIO registers during power transition.
- Adjust the driver teardown order in pxa2xx_spi_remove() and probe error paths:
always call free_irq() to unregister the handler before calling clk_disable_unprepare() to turn off the clock, preventing concurrent interrupts from reading registers while the clock is disabled.
- Avoid duplicate can-DMA pci_info() logging by checking the pre-computed enable_dma status in probe and passing a verbose flag to can_dma().
Changes in v2:
- Addressed feedback from Mark Brown on the original v1 series.
- Used drv_data->lpss_base together with relative offsets rather than hardcoding absolute MMIO offsets that vary between LPSS IP revisions.
- Moved the register save block in suspend to after the controller is quiesced (after spi_controller_suspend() and pxa_ssp_disable()).
- Store the context array lpss_priv_ctx[6] inside struct driver_data instead of struct pxa2xx_spi_controller. This keeps the changes entirely local to the core driver, preventing symbol version mismatches (disagrees about version of symbol) for other subsystem components (e.g., spi-pxa2xx-platform.ko).
- Restrict the save/restore loop to the first 6 LPSS private registers (offsets 0x00 to 0x14). Offsets beyond 0x14 (except CS control at 0x18, which is re-initialised by lpss_ssp_setup()) are reserved/unimplemented on LPT platforms (such as MacBook8,1), and writing to them triggers a PCIe Completion Timeout causing a system freeze.
- Added named constants for LPSS_PRIV_RESETS and the de-assert value.
- Wrapped S3 suspend/resume with pm_runtime_resume_and_get() and pm_runtime_put_autosuspend() respectively.
Shih-Yuan Lee (7):
spi: pxa2xx: introduce clock enable and disable helper functions
spi: pxa2xx: introduce suspended flag for interrupt synchronization
spi: pxa2xx: overhaul teardown and suspend sequence using
pxa2xx_spi_off
spi: pxa2xx: lock out runtime autosuspend for Intel LPSS SPI in PIO
mode
spi: pxa2xx: disable DMA for Apple MacBook8,1
spi: pxa2xx: restore LPSS private register state on S3 resume
spi: pxa2xx: rename local status variable to ret
drivers/spi/spi-pxa2xx-pci.c | 35 ++++- drivers/spi/spi-pxa2xx.c | 256 ++++++++++++++++++++++++++++------- drivers/spi/spi-pxa2xx.h | 4 + 3 files changed, 246 insertions(+), 49 deletions(-)
Signature · 1 line
2.39.5
PATCH v16 1/7 spi: pxa2xx: introduce clock enable and disable helper functions
The driver disables the clock during PM runtime suspend, PM system suspend, and device unbinding (remove). It also disables the clock on various error unwinding paths in pxa2xx_spi_probe().
However, if the clock is already disabled (for example, if the device is already runtime-suspended during driver unbinding), calling the common clock framework's clk_disable_unprepare() again leads to clock prepare/enable count underflows, generating kernel warnings.
Introduce pxa2xx_spi_clk_enable() and pxa2xx_spi_clk_disable() helper functions that track the clock enable state using a new 'clk_enabled' boolean flag in struct driver_data.
This ensures clk_disable_unprepare() is called only when the clock is active, preventing clock underflows during suspend transitions and unbind. It also allows the probe function to safely unwind resource allocations without triggering clock underflows.
--- drivers/spi/spi-pxa2xx.c | 37 +++++++++++++++++++++++++++++-------- drivers/spi/spi-pxa2xx.h | 2 ++ 2 files changed, 31 insertions(+), 8 deletions(-)
drivers/spi/spi-pxa2xx.c +29 -8
@@ -713,6 +713,28 @@ static void handle_bad_msg(struct driver_data *drv_data)dev_err (drv_data->ssp->dev,"bad message state in interrupt handler\n" ); } +static int pxa2xx_spi_clk_enable (struct driver_data *drv_data) +{ +int ret; + +if (drv_data->clk_enabled) +return 0 ; + + ret =clk_prepare_enable (drv_data->ssp->clk); +if (ret ==0 ) + drv_data->clk_enabled = true; + +return ret; +} + +static void pxa2xx_spi_clk_disable (struct driver_data *drv_data) +{ +if (drv_data->clk_enabled) { +clk_disable_unprepare (drv_data->ssp->clk); + drv_data->clk_enabled = false; + } +} +static irqreturn_t ssp_int (int irq,void *dev_id) {struct driver_data *drv_data = dev_id;
@@ -1352,7 +1374,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, }/* Enable SOC clock */ - status =clk_prepare_enable (ssp->clk); + status =pxa2xx_spi_clk_enable (drv_data);if (status) goto out_error_dma_irq_alloc;
@@ -1449,7 +1471,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp,return status; out_error_clock_enabled: -clk_disable_unprepare (ssp->clk); +pxa2xx_spi_clk_disable (drv_data); out_error_dma_irq_alloc:pxa2xx_spi_dma_release (drv_data);
@@ -1468,7 +1490,7 @@ void pxa2xx_spi_remove(struct device *dev)/* Disable the SSP at the peripheral and SOC level */ pxa_ssp_disable (ssp); -clk_disable_unprepare (ssp->clk); +pxa2xx_spi_clk_disable (drv_data);/* Release DMA */ if (drv_data->controller_info->enable_dma)
@@ -1492,7 +1514,7 @@ static int pxa2xx_spi_suspend(struct device *dev)pxa_ssp_disable (ssp);if (!pm_runtime_suspended (dev)) -clk_disable_unprepare (ssp->clk); +pxa2xx_spi_clk_disable (drv_data);return 0 ; }
@@ -1500,12 +1522,11 @@ static int pxa2xx_spi_suspend(struct device *dev)static int pxa2xx_spi_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); -struct ssp_device *ssp = drv_data->ssp;int status;/* Enable the SSP clock */ if (!pm_runtime_suspended (dev)) { - status =clk_prepare_enable (ssp->clk); + status =pxa2xx_spi_clk_enable (drv_data);if (status)return status; }
@@ -1518,7 +1539,7 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); -clk_disable_unprepare (drv_data->ssp->clk); +pxa2xx_spi_clk_disable (drv_data);return 0 ; }
@@ -1526,7 +1547,7 @@ static int pxa2xx_spi_runtime_resume(struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); -return clk_prepare_enable (drv_data->ssp->clk); +return pxa2xx_spi_clk_enable (drv_data); } EXPORT_NS_GPL_DEV_PM_OPS(pxa2xx_spi_pm_ops, SPI_PXA2xx) = {
drivers/spi/spi-pxa2xx.h +2 -0
@@ -72,6 +72,8 @@ struct driver_data {void __iomem *lpss_base; +bool clk_enabled; +/* Optional slave FIFO ready signal */ struct gpio_desc *gpiod_ready; };
Signature · 1 line
2.39.5
Re: [PATCH v16 1/7] spi: pxa2xx: introduce clock enable and disable helper functions
Quoted 8 lines · Shih-Yuan Lee
The driver disables the clock during PM runtime suspend, PM system suspend, and device unbinding (remove). It also disables the clock on various error unwinding paths in
pxa2xx_spi_probe().However, if the clock is already disabled (for example, if the device is already runtime-suspended during driver unbinding), calling the common clock framework's
clk_disable_unprepare()again leads to clock prepare/enable count underflows, generating kernel warnings.
Any real life example here?
Quoted 8 lines
Introduce
pxa2xx_spi_clk_enable()andpxa2xx_spi_clk_disable()helper functions that track the clock enable state using a new 'clk_enabled' boolean flag in structdriver_data.This ensures
clk_disable_unprepare()is called only when the clock is active, preventing clock underflows during suspend transitions and unbind. It also allows the probe function to safely unwind resource allocations without triggering clock underflows.
...
Same issue, I have no cover letter in my mailbox. Please, slow down and check your email setup. Something is wrong.
...
drivers/spi/spi-pxa2xx.c static void handle_bad_msg(struct driver_data *drv_data) · line 727
Quoted 13 lines
drivers/spi/spi-pxa2xx.c +12
-0+static int pxa2xx_spi_clk_enable (struct driver_data *drv_data) +{ +int ret; + +if (drv_data->clk_enabled) +return 0 ; + + ret =clk_prepare_enable (drv_data->ssp->clk); +if (ret ==0 ) + drv_data->clk_enabled = true; + +return ret;
Use usual pattern
if (ret) return ret; ... return 0;
drivers/spi/spi-pxa2xx.c static void handle_bad_msg(struct driver_data *drv_data) · line 732
Quoted 5 lines
drivers/spi/spi-pxa2xx.c +4
-0+} +static void pxa2xx_spi_clk_disable (struct driver_data *drv_data) +{ +if (drv_data->clk_enabled) {
if (!drv_data->clk_enabled) return;
Quoted 4 lines
drivers/spi/spi-pxa2xx.c +4
-0+clk_disable_unprepare (drv_data->ssp->clk); + drv_data->clk_enabled = false; + } +}
...
drivers/spi/spi-pxa2xx.h struct driver_data { · line 75
Quoted 5 lines
struct
driver_data{drivers/spi/spi-pxa2xx.h +1
-0void __iomem *lpss_base; +bool clk_enabled;
How is this protected against simultaneously called enable/disable on different CPUs?
drivers/spi/spi-pxa2xx.h struct driver_data { · line 79
Quoted 3 lines
drivers/spi/spi-pxa2xx.h +0
-0/* Optional slave FIFO ready signal */ struct gpio_desc *gpiod_ready; };
Whenever you add the field, check with pahole that the layout is optimal.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 1/7] spi: pxa2xx: introduce clock enable and disable helper functions
Hi Andy,
Thanks for the detailed review comments.
Quoted 6 lines · Andy Shevchenko
Quoted 4 lines · Shih-Yuan Lee
The driver disables the clock during PM runtime suspend, PM system suspend, and device unbinding (remove). It also disables the clock on various error unwinding paths in
pxa2xx_spi_probe().
Quoted 6 lines
Quoted 4 lines
However, if the clock is already disabled (for example, if the device is already runtime-suspended during driver unbinding), calling the common clock framework's
clk_disable_unprepare()again leads to clock prepare/enable count underflows, generating kernel warnings.Any real life example here?
Yes, a concrete real-world scenario occurs during sysfs driver unbinding while the device is in PM runtime suspend:
1. When the system is idle, PM runtime suspends the controller,
disabling the clock (clk_disable_unprepare()).
2. If the user unbinds the driver via sysfs:
echo 0000:00:15.4 > /sys/bus/pci/drivers/pxa2xx_spi_pci/unbind
pxa2xx_spi_remove() is invoked.
3. Without clk_enabled tracking, pxa2xx_spi_remove() unconditionally
calls clk_disable_unprepare() a second time on the already disabled clock.
4. This triggers a Common Clock Framework (CCF) underflow warning:
WARN_ON(clk->enable_count == 0)
A similar underflow occurs if pxa2xx_spi_probe() fails and executes error unwinding after disabling the clock.
Quoted 10 lines
Quoted 9 lines
Introduce
pxa2xx_spi_clk_enable()andpxa2xx_spi_clk_disable()helper functions that track the clock enable state using a new 'clk_enabled' boolean flag in structdriver_data.This ensures
clk_disable_unprepare()is called only when the clock is active, preventing clock underflows during suspend transitions and unbind. It also allows the probe function to safely unwind resource allocations without triggering clock underflows.
Quoted 4 lines
...
Same issue, I have no cover letter in my mailbox. Please, slow down and check your email setup. Something is wrong.
Duly noted. I will check my git send-email configuration and ensure the cover letter is properly sent as the thread root.
Quoted 17 lines
...
drivers/spi/spi-pxa2xx.c static void handle_bad_msg(struct driver_data *drv_data) · line 727
Quoted 13 lines
drivers/spi/spi-pxa2xx.c +12
-0+static int pxa2xx_spi_clk_enable (struct driver_data *drv_data) +{ +int ret; + +if (drv_data->clk_enabled) +return 0 ; + + ret =clk_prepare_enable (drv_data->ssp->clk); +if (ret ==0 ) + drv_data->clk_enabled = true; + +return ret;
Quoted 9 lines
Use usual pattern
if (ret) return ret; ... return 0;Quoted 1 line
+}
Will update in the next version to follow the standard guard pattern:
static int pxa2xx_spi_clk_enable(struct driver_data *drv_data)
{
int ret;
if (drv_data->clk_enabled)
return 0;
ret = clk_prepare_enable(drv_data->ssp->clk);
if (ret)
return ret;
drv_data->clk_enabled = true;
return 0;
}Quoted 12 lines
drivers/spi/spi-pxa2xx.c static void handle_bad_msg(struct driver_data *drv_data) · line 732
Quoted 4 lines
drivers/spi/spi-pxa2xx.c +3
-0+static void pxa2xx_spi_clk_disable (struct driver_data *drv_data) +{ +if (drv_data->clk_enabled) {if (!drv_data->clk_enabled) return;Quoted 4 lines
drivers/spi/spi-pxa2xx.c +4
-0+clk_disable_unprepare (drv_data->ssp->clk); + drv_data->clk_enabled = false; + } +}
Will update in the next version to use early return:
static void pxa2xx_spi_clk_disable(struct driver_data *drv_data)
{
if (!drv_data->clk_enabled)
return;
drv_data->clk_enabled = false;
clk_disable_unprepare(drv_data->ssp->clk);
}Quoted 11 lines
...
Quoted 5 lines
struct
driver_data{void __iomem *lpss_base;+ bool
clk_enabled;How is this protected against simultaneously called enable/disable on different CPUs?
pxa2xx_spi_clk_enable() and pxa2xx_spi_clk_disable() are only invoked within serialized contexts:
1. Driver probe() / remove() paths (serialized by bus locks).
2. PM Runtime callbacks (serialized by dev->power.lock in PM core).
3. System suspend / resume callbacks (serialized by PM subsystem).
They are never called concurrently from asynchronous contexts or worker threads.
Quoted 6 lines
Quoted 4 lines
/* Optional slave FIFO ready signal */ struct gpio_desc *gpiod_ready;};
Whenever you add the field, check with
paholethat the layout is optimal.
Thanks for pointing out pahole. Checked with pahole: bool clk_enabled; is placed immediately after bool suspended; in struct driver_data. Since both are 1-byte booleans, clk_enabled fits into the existing struct padding byte, resulting in zero additional hole padding and keeping the total struct size unchanged.
Thanks,
Shih-Yuan
PATCH v16 2/7 spi: pxa2xx: introduce suspended flag for interrupt synchronization
In reply to Shih-Yuan LeeWhen a shared interrupt line is used, the interrupt handler ssp_int() can be triggered by other devices sharing the line. The handler must ensure it does not access the SSP controller registers via MMIO when the device is powered down or when its clock is gated; otherwise, it will cause PCIe Completion Timeouts and system hangs.
Currently, ssp_int() guards MMIO access using:
if (pm_runtime_suspended(drv_data->ssp->dev)) return IRQ_NONE;
However, during PM transitions (such as system suspend or runtime PM autosuspend), device callbacks execute to disable the hardware and gate the clock, but the PM state machine does not mark the device as RPM_SUSPENDED until after the suspend callback returns. During this transitional state (RPM_SUSPENDING), pm_runtime_suspended() returns false. If a shared interrupt fires after the clock has been gated but before the PM state has transitioned, ssp_int() will execute, attempt MMIO reads on the unclocked register space, and hang the system.
Formal verification using Spin/PROMELA confirms that a scheduling window exists where the interrupt thread accesses MMIO when clk_enabled is false, violating safety properties.
Introduce a custom 'suspended' boolean flag in struct driver_data to track the device's suspended state across all PM transitions. Check both 'drv_data->suspended' and '!drv_data->clk_enabled' in ssp_int() to return IRQ_NONE immediately before any MMIO access is attempted.
This closes the state transition race condition and mathematically guarantees deadlock-free, safe shared interrupt handling during power transitions.
--- drivers/spi/spi-pxa2xx.c | 51 ++++++++++++++++++++++++++-------------- drivers/spi/spi-pxa2xx.h | 1 + 2 files changed, 35 insertions(+), 17 deletions(-)
drivers/spi/spi-pxa2xx.c +34 -17
@@ -730,8 +730,8 @@ static int pxa2xx_spi_clk_enable(struct driver_data *drv_data)static void pxa2xx_spi_clk_disable (struct driver_data *drv_data) {if (drv_data->clk_enabled) { -clk_disable_unprepare (drv_data->ssp->clk); drv_data->clk_enabled = false; +clk_disable_unprepare (drv_data->ssp->clk); } }
@@ -743,12 +743,12 @@ static irqreturn_t ssp_int(int irq, void *dev_id)u32 status;/* -* The IRQ might be shared with other peripherals so we must first -* check that are we RPM suspended or not. If we are we assume that -* the IRQ was not for us (we shouldn't be RPM suspended when the -* interrupt is enabled). +* The IRQ might be shared with other peripherals or trigger during +* power state transitions. First check if device is suspended or if +* clock is disabled; if so, return IRQ_NONE immediately to avoid +* unclocked MMIO reads. */ -if (pm_runtime_suspended (drv_data->ssp->dev)) +if (drv_data->suspended || !drv_data->clk_enabled)return IRQ_NONE ;/*
@@ -1310,6 +1310,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, drv_data->controller = controller; drv_data->controller_info = platform_info; drv_data->ssp = ssp; + drv_data->suspended = true;/* The spi->mode bits understood by this driver: */ controller->mode_bits =SPI_CPOL |SPI_CPHA |SPI_CS_HIGH |SPI_LOOP ;
@@ -1352,11 +1353,6 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, | SSSR_ROR | SSSR_TUR; } - status =request_irq (ssp->irq, ssp_int,IRQF_SHARED ,dev_name (dev), - drv_data); -if (status <0 ) -return dev_err_probe (dev, status,"cannot get IRQ %d\n" , ssp->irq); -/* Setup DMA if requested */ if (platform_info->enable_dma) { status =pxa2xx_spi_dma_setup (drv_data);
@@ -1376,7 +1372,16 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp,/* Enable SOC clock */ status =pxa2xx_spi_clk_enable (drv_data);if (status) - goto out_error_dma_irq_alloc; + goto out_error_dma_alloc; + + drv_data->suspended = false; + + status =request_irq (ssp->irq, ssp_int,IRQF_SHARED ,dev_name (dev), + drv_data); +if (status <0 ) { + status =dev_err_probe (dev, status,"cannot get IRQ %d\n" , ssp->irq); + goto out_error_clock_enabled; + } controller->max_speed_hz =clk_get_rate (ssp->clk);/*
@@ -1456,7 +1461,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, "ready", GPIOD_OUT_LOW);if (IS_ERR (drv_data->gpiod_ready)) { status =PTR_ERR (drv_data->gpiod_ready); - goto out_error_clock_enabled; + goto out_error_irq_alloc; } }
@@ -1465,17 +1470,19 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, status =spi_register_controller (controller);if (status) {dev_err_probe (dev, status,"problem registering SPI controller\n" ); - goto out_error_clock_enabled; + goto out_error_irq_alloc; }return status; +out_error_irq_alloc: +free_irq (ssp->irq, drv_data); + out_error_clock_enabled:pxa2xx_spi_clk_disable (drv_data); -out_error_dma_irq_alloc: +out_error_dma_alloc:pxa2xx_spi_dma_release (drv_data); -free_irq (ssp->irq, drv_data);return status; }
@@ -1511,6 +1518,7 @@ static int pxa2xx_spi_suspend(struct device *dev)if (status)return status; + drv_data->suspended = true;pxa_ssp_disable (ssp);if (!pm_runtime_suspended (dev))
@@ -1531,6 +1539,8 @@ static int pxa2xx_spi_resume(struct device *dev)return status; } + drv_data->suspended = false; +/* Start the queue running */ return spi_controller_resume (drv_data->controller); }
@@ -1539,6 +1549,7 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); + drv_data->suspended = true;pxa2xx_spi_clk_disable (drv_data);return 0 ; }
@@ -1546,8 +1557,14 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev)static int pxa2xx_spi_runtime_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); +int ret; -return pxa2xx_spi_clk_enable (drv_data); + ret =pxa2xx_spi_clk_enable (drv_data); +if (ret) +return ret; + + drv_data->suspended = false; +return 0 ; } EXPORT_NS_GPL_DEV_PM_OPS(pxa2xx_spi_pm_ops, SPI_PXA2xx) = {
drivers/spi/spi-pxa2xx.h +1 -0
@@ -72,6 +72,7 @@ struct driver_data {void __iomem *lpss_base; +bool suspended;bool clk_enabled;/* Optional slave FIFO ready signal */
Signature · 1 line
2.39.5
Re: [PATCH v16 2/7] spi: pxa2xx: introduce suspended flag for interrupt synchronization
Quoted 8 lines · Shih-Yuan Lee
When a shared interrupt line is used, the interrupt handler
ssp_int()can be triggered by other devices sharing the line. The handler must ensure it does not access the SSP controller registers via MMIO when the device is powered down or when its clock is gated; otherwise, it will cause PCIe Completion Timeouts and system hangs.Currently,
ssp_int()guards MMIO access using:if (pm_runtime_suspended(drv_data->ssp->dev)) return IRQ_NONE;
...
drivers/spi/spi-pxa2xx.c static int pxa2xx_spi_clk_enable(struct driver_data *drv_data) · line 736
Quoted 13 lines
Introduce a custom 'suspended' boolean flag in struct
driver_datato track the device's suspended state across all PM transitions. Check both 'drv_data->suspended' and '!drv_data->clk_enabled' inssp_int()to returnIRQ_NONEimmediately before any MMIO access is attempted.static void
pxa2xx_spi_clk_disable(struct driver_data *drv_data){drivers/spi/spi-pxa2xx.c +1
-1if (drv_data->clk_enabled) { -clk_disable_unprepare (drv_data->ssp->clk); drv_data->clk_enabled = false; +clk_disable_unprepare (drv_data->ssp->clk); } }
It's probably better to avoid extra changes like this, it makes the patch bigger and a bit harder to review.
Quoted 8 lines
drivers/spi/spi-pxa2xx.c +0
-4@@ -743,12 +743,12 @@ static irqreturn_t ssp_int(int irq, void *dev_id)u32 status;/* -* The IRQ might be shared with other peripherals so we must first -* check that are we RPM suspended or not. If we are we assume that -* the IRQ was not for us (we shouldn't be RPM suspended when the -* interrupt is enabled).
drivers/spi/spi-pxa2xx.c static irqreturn_t ssp_int(int irq, void *dev_id) · line 754
Quoted 10 lines
drivers/spi/spi-pxa2xx.c +5
-1+* The IRQ might be shared with other peripherals or trigger during +* power state transitions. First check if device is suspended or if +* clock is disabled; if so, return IRQ_NONE immediately to avoid +* unclocked MMIO reads. */ -if (pm_runtime_suspended (drv_data->ssp->dev)) +if (drv_data->suspended || !drv_data->clk_enabled)return IRQ_NONE ;/*
The local flags *must* be racy - I'm not seeing any locking which protects them, nor anything that stops something else dropping a runtime PM reference and powering things down after we checked here. There's a helper in the power management code pm_runtime_get_if_active() which I think is what you want here, it'll tell you if the device is runtime suspended and if the device is active it'll ensure nothing else drops the last reference while we're running.
Re: [PATCH v16 2/7] spi: pxa2xx: introduce suspended flag for interrupt synchronization
Hi Mark,
Thanks for the sharp observation and the suggestion!
Quoted 27 lines · Mark Brown
Quoted 8 lines · Shih-Yuan Lee
When a shared interrupt line is used, the interrupt handler
ssp_int()can be triggered by other devices sharing the line. The handler must ensure it does not access the SSP controller registers via MMIO when the device is powered down or when its clock is gated; otherwise, it will cause PCIe Completion Timeouts and system hangs.Currently,
ssp_int()guards MMIO access using:if (pm_runtime_suspended(drv_data->ssp->dev)) return IRQ_NONE;...
drivers/spi/spi-pxa2xx.c static int pxa2xx_spi_clk_enable(struct driver_data *drv_data) · line 736
Quoted 13 lines
Introduce a custom 'suspended' boolean flag in struct
driver_datato track the device's suspended state across all PM transitions. Check both 'drv_data->suspended' and '!drv_data->clk_enabled' inssp_int()to returnIRQ_NONEimmediately before any MMIO access is attempted.static void
pxa2xx_spi_clk_disable(struct driver_data *drv_data){ if (drv_data->clk_enabled) {drivers/spi/spi-pxa2xx.c +1
-1-clk_disable_unprepare (drv_data->ssp->clk); drv_data->clk_enabled = false; +clk_disable_unprepare (drv_data->ssp->clk); } }
Quoted 2 lines
It's probably better to avoid extra changes like this, it makes the patch bigger and a bit harder to review.
Agreed. I have dropped the unnecessary diff churn from pxa2xx_spi_clk_disable().
Quoted 18 lines
Quoted 18 lines
drivers/spi/spi-pxa2xx.c +5
-5@@ -743,12 +743,12 @@ static irqreturn_t ssp_int(int irq, void *dev_id)u32 status;/* -* The IRQ might be shared with other peripherals so we must first -* check that are we RPM suspended or not. If we are we assume that -* the IRQ was not for us (we shouldn't be RPM suspended when the -* interrupt is enabled). +* The IRQ might be shared with other peripherals or trigger during +* power state transitions. First check if device is suspended or if +* clock is disabled; if so, return IRQ_NONE immediately to avoid +* unclocked MMIO reads. */ -if (pm_runtime_suspended (drv_data->ssp->dev)) +if (drv_data->suspended || !drv_data->clk_enabled)return IRQ_NONE ;
Quoted 9 lines
Review static irqreturn_t ssp_int(int irq, void *dev_id) · line 754
Quoted 1 line
(quoted diff) +0
-0/*The local flags *must* be racy - I'm not seeing any locking which protects them, nor anything that stops something else dropping a runtime PM reference and powering things down after we checked here. There's a helper in the power management code
pm_runtime_get_if_active()which I think is what you want here, it'll tell you if the device is runtime suspended and if the device is active it'll ensure nothing else drops the last reference while we're running.
That makes total sense. Using pm_runtime_get_if_active() is indeed much cleaner, atomic, and eliminates the potential race condition where another CPU drops the runtime PM reference while ssp_int() is running. It also avoids adding an ad-hoc suspended boolean flag to struct driver_data.
I will update ssp_int() in the next version to use pm_runtime_get_if_active():
static irqreturn_t ssp_int(int irq, void *dev_id)
{
struct driver_data *drv_data = dev_id;
struct device *dev = &drv_data->pdev->dev;
u32 sccr1_reg;
u32 mask = drv_data->mask_sr;
u32 status;
/*
* Ensure the device is active and acquire a PM runtime reference so
* PM runtime cannot power down the device or gate the clock while
* ssp_int() is running.
*/
if (pm_runtime_get_if_active(dev) <= 0)
return IRQ_NONE;
...
pm_runtime_put_autosuspend(dev);
return IRQ_HANDLED;
}Thanks,
Shih-Yuan
PATCH v16 3/7 spi: pxa2xx: overhaul teardown and suspend sequence using pxa2xx_spi_off
In reply to Shih-Yuan LeeWhen removing the driver or suspending the device, the clock must not be disabled while shared interrupts are still active. Gating the clock before waiting for in-flight interrupt handlers to complete results in race conditions where the handler performs unclocked MMIO accesses, causing PCIe Completion Timeouts.
Overhaul the remove, suspend, and runtime_suspend paths to use a strict synchronized teardown order:
1. Disable hardware interrupt generation at the controller level.
2. Mark the device state as suspended (suspended = true) to prevent subsequent interrupt handlers from attempting MMIO reads.
3. Call synchronize_irq() to wait for any active interrupt handlers to drain completely.
4. Gate the clock via pxa2xx_spi_clk_disable().
Additionally, commit 29d7e05c5f75 ("spi: pxa2xx: Avoid touching SSCR0_SSE on MMP2") documented that disabling the hardware block via SSE on MMP2 SoC platforms corrupts the RX/TX FIFO. Instead of calling pxa_ssp_disable() directly, use the helper function pxa2xx_spi_off(), which respects the MMP2 platform quirk by bypassing SSE register writes.
--- drivers/spi/spi-pxa2xx.c | 45 ++++++++++++++++++++++++++++------------ 1 file changed, 32 insertions(+), 13 deletions(-)
drivers/spi/spi-pxa2xx.c +32 -13
@@ -1495,16 +1495,24 @@ void pxa2xx_spi_remove(struct device *dev)spi_unregister_controller (drv_data->controller); -/* Disable the SSP at the peripheral and SOC level */ -pxa_ssp_disable (ssp); +/* Disable SSP interrupt generation on hardware level while clock is active */ +pxa2xx_spi_off (drv_data); + +/* Mark as suspended to prevent further IRQ handling */ + drv_data->suspended = true; + +/* Wait for any pending interrupt handlers to complete */ +synchronize_irq (ssp->irq); + +/* Release IRQ */ +free_irq (ssp->irq, drv_data); + +/* Safe to disable the SSP clock now */ pxa2xx_spi_clk_disable (drv_data);/* Release DMA */ if (drv_data->controller_info->enable_dma)pxa2xx_spi_dma_release (drv_data); - -/* Release IRQ */ -free_irq (ssp->irq, drv_data); }EXPORT_SYMBOL_NS_GPL (pxa2xx_spi_remove,"SPI_PXA2xx" );
@@ -1519,10 +1527,10 @@ static int pxa2xx_spi_suspend(struct device *dev)return status; drv_data->suspended = true; -pxa_ssp_disable (ssp); +pxa2xx_spi_off (drv_data); +synchronize_irq (ssp->irq); -if (!pm_runtime_suspended (dev)) -pxa2xx_spi_clk_disable (drv_data); +pxa2xx_spi_clk_disable (drv_data);return 0 ; }
@@ -1530,6 +1538,7 @@ static int pxa2xx_spi_suspend(struct device *dev)static int pxa2xx_spi_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); +struct ssp_device *ssp = drv_data->ssp;int status;/* Enable the SSP clock */
@@ -1542,7 +1551,15 @@ static int pxa2xx_spi_resume(struct device *dev) drv_data->suspended = false;/* Start the queue running */ -return spi_controller_resume (drv_data->controller); + status =spi_controller_resume (drv_data->controller); +if (status) { + drv_data->suspended = true; +synchronize_irq (ssp->irq); +pxa2xx_spi_clk_disable (drv_data); +return status; + } + +return 0 ; }static int pxa2xx_spi_runtime_suspend (struct device *dev)
@@ -1550,6 +1567,8 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev)struct driver_data *drv_data =dev_get_drvdata (dev); drv_data->suspended = true; +pxa2xx_spi_off (drv_data); +synchronize_irq (drv_data->ssp->irq);pxa2xx_spi_clk_disable (drv_data);return 0 ; }
@@ -1557,11 +1576,11 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev)static int pxa2xx_spi_runtime_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); -int ret; +int status; - ret =pxa2xx_spi_clk_enable (drv_data); -if (ret) -return ret; + status =pxa2xx_spi_clk_enable (drv_data); +if (status) +return status; drv_data->suspended = false;return 0 ;
Signature · 1 line
2.39.5
Re: [PATCH v16 3/7] spi: pxa2xx: overhaul teardown and suspend sequence using pxa2xx_spi_off
Quoted 20 lines · Shih-Yuan Lee
When removing the driver or suspending the device, the clock must not be disabled while shared interrupts are still active. Gating the clock before waiting for in-flight interrupt handlers to complete results in race conditions where the handler performs unclocked MMIO accesses, causing PCIe Completion Timeouts.
Overhaul the remove, suspend, and
runtime_suspendpaths to use a strict synchronized teardown order:
1. Disable hardware interrupt generation at the controller level.
2. Mark the device state as suspended (suspended = true) to prevent subsequent interrupt handlers from attempting MMIO reads.
3. Callsynchronize_irq()to wait for any active interrupt handlers to drain completely.
4. Gate the clock viapxa2xx_spi_clk_disable().Additionally, commit
29d7e05c5f75("spi: pxa2xx: Avoid touchingSSCR0_SSEon MMP2") documented that disabling the hardware block via SSE on MMP2 SoC platforms corrupts the RX/TX FIFO. Instead of callingpxa_ssp_disable()directly, use the helper functionpxa2xx_spi_off(), which respects the MMP2 platform quirk by bypassing SSE register writes.
...
drivers/spi/spi-pxa2xx.c void pxa2xx_spi_remove(struct device *dev) · line 1505
Quoted 2 lines
drivers/spi/spi-pxa2xx.c +2
-0+/* Wait for any pending interrupt handlers to complete */ +synchronize_irq (ssp->irq);
Unneeded. free_irq() implies that.
Quoted 2 lines
drivers/spi/spi-pxa2xx.c +2
-0+/* Release IRQ */ +free_irq (ssp->irq, drv_data);
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 3/7] spi: pxa2xx: overhaul teardown and suspend sequence using pxa2xx_spi_off
Quoted 23 lines · Andy Shevchenko
Quoted 15 lines · Shih-Yuan Lee
When removing the driver or suspending the device, the clock must not be disabled while shared interrupts are still active. Gating the clock before waiting for in-flight interrupt handlers to complete results in race conditions where the handler performs unclocked MMIO accesses, causing PCIe Completion Timeouts.
Overhaul the remove, suspend, and
runtime_suspendpaths to use a strict synchronized teardown order:
1. Disable hardware interrupt generation at the controller level.
2. Mark the device state as suspended (suspended = true) to prevent subsequent interrupt handlers from attempting MMIO reads.
3. Callsynchronize_irq()to wait for any active interrupt handlers to drain completely.
4. Gate the clock viapxa2xx_spi_clk_disable().Quoted 5 lines
Additionally, commit
29d7e05c5f75("spi: pxa2xx: Avoid touchingSSCR0_SSEon MMP2") documented that disabling the hardware block via SSE on MMP2 SoC platforms corrupts the RX/TX FIFO. Instead of callingpxa_ssp_disable()directly, use the helper functionpxa2xx_spi_off(), which respects the MMP2 platform quirk by bypassing SSE register writes.
Quoted 6 lines
...
drivers/spi/spi-pxa2xx.c void pxa2xx_spi_remove(struct device *dev) · line 1505
Quoted 2 lines
drivers/spi/spi-pxa2xx.c +2
-0+/* Wait for any pending interrupt handlers to complete */ +synchronize_irq (ssp->irq);Unneeded.
free_irq()implies that.
Hi Andy,
Thanks for pointing this out.
Good catch. Since free_irq() internally calls synchronize_irq() before releasing the IRQ, calling synchronize_irq() right before free_irq() in pxa2xx_spi_remove() is indeed redundant.
I will remove the explicit synchronize_irq() call from pxa2xx_spi_remove() for the next version, while retaining it in the suspend() and runtime_suspend() paths where free_irq() is not called.
Thanks,
Shih-Yuan
PATCH v16 4/7 spi: pxa2xx: lock out runtime autosuspend for Intel LPSS SPI in PIO mode
In reply to Shih-Yuan LeeWhen operating in PIO mode on Intel LPSS SPI controllers, runtime PM autosuspend clock-gates the hardware block when idle. On subsequent transfers, MMIO accesses to trigger resume are performed before the runtime PM state machine can wake the controller, causing PCIe Completion Timeouts. This issue does not affect DMA mode, where the DMA engine holds the required resources, nor does it affect non-LPSS controllers.
Lock out runtime PM autosuspend for LPSS SPI controllers specifically when operating in PIO mode.
Acquire a runtime PM reference via pm_runtime_get_noresume() in pxa2xx_spi_probe() after registration, and release it via pm_runtime_put_noidle() in pxa2xx_spi_remove() before hardware teardown to ensure the runtime PM reference held in probe is released without invoking runtime_suspend on already-unclocked hardware.
--- drivers/spi/spi-pxa2xx.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
drivers/spi/spi-pxa2xx.c +10 -0
@@ -1473,6 +1473,9 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, goto out_error_irq_alloc; } +if (is_lpss_ssp (drv_data) && !platform_info->enable_dma) +pm_runtime_get_noresume (dev); +return status; out_error_irq_alloc:
@@ -1495,6 +1498,13 @@ void pxa2xx_spi_remove(struct device *dev)spi_unregister_controller (drv_data->controller); +/* Release the PM reference held in probe for LPSS PIO mode before +* hardware teardown so the PM core does not invoke runtime_suspend +* on already-unclocked hardware. +*/ +if (is_lpss_ssp (drv_data) && !drv_data->controller_info->enable_dma) +pm_runtime_put_noidle (dev); +/* Disable SSP interrupt generation on hardware level while clock is active */ pxa2xx_spi_off (drv_data);
Signature · 1 line
2.39.5
Re: [PATCH v16 4/7] spi: pxa2xx: lock out runtime autosuspend for Intel LPSS SPI in PIO mode
Quoted 7 lines · Shih-Yuan Lee
When operating in PIO mode on Intel LPSS SPI controllers, runtime PM autosuspend clock-gates the hardware block when idle. On subsequent transfers, MMIO accesses to trigger resume are performed before the runtime PM state machine can wake the controller, causing PCIe Completion Timeouts. This issue does not affect DMA mode, where the DMA engine holds the required resources, nor does it affect non-LPSS controllers.
Quoted 8 lines
Lock out runtime PM autosuspend for LPSS SPI controllers specifically when operating in PIO mode.
Acquire a runtime PM reference via
pm_runtime_get_noresume()inpxa2xx_spi_probe()after registration, and release it viapm_runtime_put_noidle()inpxa2xx_spi_remove()before hardware teardown to ensure the runtime PM reference held in probe is released without invokingruntime_suspendon already-unclocked hardware.
.runtime_suspend()
// this is how we refer to the callbacks.
...
drivers/spi/spi-pxa2xx.c void pxa2xx_spi_remove(struct device *dev) · line 1504
Quoted 4 lines
drivers/spi/spi-pxa2xx.c +4
-0+/* Release the PM reference held in probe for LPSS PIO mode before +* hardware teardown so the PM core does not invoke runtime_suspend +* on already-unclocked hardware. +*/
/*
* Keep the comment style as per SPI and many other subsystems
* and as in this example.
*/
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 4/7] spi: pxa2xx: lock out runtime autosuspend for Intel LPSS SPI in PIO mode
Quoted 9 lines · Andy Shevchenko
Quoted 7 lines · Shih-Yuan Lee
When operating in PIO mode on Intel LPSS SPI controllers, runtime PM autosuspend clock-gates the hardware block when idle. On subsequent transfers, MMIO accesses to trigger resume are performed before the runtime PM state machine can wake the controller, causing PCIe Completion Timeouts. This issue does not affect DMA mode, where the DMA engine holds the required resources, nor does it affect non-LPSS controllers.
Quoted 8 lines
Quoted 8 lines
Lock out runtime PM autosuspend for LPSS SPI controllers specifically when operating in PIO mode.
Acquire a runtime PM reference via
pm_runtime_get_noresume()inpxa2xx_spi_probe()after registration, and release it viapm_runtime_put_noidle()inpxa2xx_spi_remove()before hardware teardown to ensure the runtime PM reference held in probe is released without invokingruntime_suspendon already-unclocked hardware.
Hi Andy,
Thanks for the feedback.
Quoted 4 lines
.runtime_suspend()
// this is how we refer to the callbacks.
Updated the commit message to format PM callbacks as .runtime_suspend().
Quoted 9 lines
...
drivers/spi/spi-pxa2xx.c void pxa2xx_spi_remove(struct device *dev) · line 1504
Quoted 4 lines
drivers/spi/spi-pxa2xx.c +4
-0+/* Release the PM reference held in probe for LPSS PIO mode before +* hardware teardown so the PM core does not invoke runtime_suspend +* on already-unclocked hardware. +*/
Quoted 4 lines
/*
* Keep the comment style as per SPI and many other subsystems
* and as in this example.
*/
Updated the multi-line comment in pxa2xx_spi_remove() to adhere to the standard Linux kernel multi-line comment style with an empty starting line.
Thanks,
Shih-Yuan
PATCH v16 5/7 spi: pxa2xx: disable DMA for Apple MacBook8,1
In reply to Shih-Yuan LeeOn MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
Move the forced PIO mode DMI quirk to spi-pxa2xx-pci.c (the LPSS host controller PCI glue driver) to avoid layering violations in client drivers (such as applespi).
Add an explicit DMI match table for MacBook8,1 and a module parameter spi_pxa2xx_force_pio to allow forcing PIO mode on demand.
--- drivers/spi/spi-pxa2xx-pci.c | 35 +++++++++++++++++++++++++++++++++-- 1 file changed, 33 insertions(+), 2 deletions(-)
drivers/spi/spi-pxa2xx-pci.c +33 -2
@@ -18,9 +18,14 @@#include <linux/dmaengine.h> #include <linux/platform_data/dma-dw.h> +#include <linux/dmi.h> #include "spi-pxa2xx.h" +static bool spi_pxa2xx_force_pio; +module_param_named (force_pio, spi_pxa2xx_force_pio, bool,0444 ); +MODULE_PARM_DESC (force_pio,"Force PIO mode (disables DMA) for SPI transfers. ([0] = disabled, 1 = enabled)" ); +#define PCI_DEVICE_ID_INTEL_QUARK_X1000 0x0935#define PCI_DEVICE_ID_INTEL_BYT 0x0f0e#define PCI_DEVICE_ID_INTEL_MRFLD 0x1194
@@ -93,6 +98,32 @@ static void lpss_dma_put_device(void *dma_dev)pci_dev_put (dma_dev); } +static const struct dmi_system_id pxa2xx_spi_pci_dmi_table[] = { + { + .ident ="Apple MacBook8,1" , + .matches = { +DMI_MATCH (DMI_SYS_VENDOR ,"Apple Inc." ), +DMI_MATCH (DMI_PRODUCT_NAME ,"MacBook8,1" ), + }, + }, + { } +}; + +static bool pxa2xx_spi_pci_can_dma (struct pci_dev *dev) +{ +if (spi_pxa2xx_force_pio) { +pci_info (dev,"Forcing PIO mode (disabling DMA)\n" ); +return false; + } + +if (dmi_check_system (pxa2xx_spi_pci_dmi_table)) { +pci_info (dev,"MacBook8,1 detected: disabling DMA to force PIO mode\n" ); +return false; + } + +return true; +} +static int lpss_spi_setup (struct pci_dev *dev,struct pxa2xx_spi_controller *c) {struct ssp_device *ssp = &c->ssp;
@@ -166,7 +197,7 @@ static int lpss_spi_setup(struct pci_dev *dev, struct pxa2xx_spi_controller *c) c->dma_filter = lpss_dma_filter; c->dma_burst_size =1 ; - c->enable_dma =1 ; + c->enable_dma =pxa2xx_spi_pci_can_dma (dev);return 0 ; }
@@ -238,7 +269,7 @@ static int mrfld_spi_setup(struct pci_dev *dev, struct pxa2xx_spi_controller *c) c->dma_filter = lpss_dma_filter; c->dma_burst_size =8 ; - c->enable_dma =1 ; + c->enable_dma =pxa2xx_spi_pci_can_dma (dev);return 0 ; }
Signature · 1 line
2.39.5
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 4 lines · Shih-Yuan Lee
On MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
Why? The thing like this I noticed on some other HW which required different DMA settings. Have you tried to investigate the issue more?
Quoted 6 lines
Move the forced PIO mode DMI quirk to
spi-pxa2xx-pci.c(the LPSS host controller PCI glue driver) to avoid layering violations in client drivers (such as applespi).Add an explicit DMI match table for MacBook8,1 and a module parameter
spi_pxa2xx_force_pioto allow forcing PIO mode on demand.
This looks like a hack. Please, if you have a hardware, try to dump the errors, collect the data corruption or timeouts and provide more information. DMA for SPI on LPSS hardware never was a problem (unlike UART in some cases). I'm almost 100% sure the problem is in DMA configuration / setting bits.
...
Quoted 3 lines
#include <
linux/dmaengine.h>
#include <linux/platform_data/dma-dw.h>
+#include <linux/dmi.h>
Keep list ordered.
Quoted 1 line
#include "
spi-pxa2xx.h"
...
drivers/spi/spi-pxa2xx-pci.c line 27
Quoted 3 lines
drivers/spi/spi-pxa2xx-pci.c +3
-0+static bool spi_pxa2xx_force_pio; +module_param_named (force_pio, spi_pxa2xx_force_pio, bool,0444 ); +MODULE_PARM_DESC (force_pio,"Force PIO mode (disables DMA) for SPI transfers. ([0] = disabled, 1 = enabled)" );
No. There is no room for module parameters like this.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 32 lines · Andy Shevchenko
Quoted 4 lines · Shih-Yuan Lee
On MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
Why? The thing like this I noticed on some other HW which required different DMA settings. Have you tried to investigate the issue more?
Quoted 6 lines
Move the forced PIO mode DMI quirk to
spi-pxa2xx-pci.c(the LPSS host controller PCI glue driver) to avoid layering violations in client drivers (such as applespi).Add an explicit DMI match table for MacBook8,1 and a module parameter
spi_pxa2xx_force_pioto allow forcing PIO mode on demand.This looks like a hack. Please, if you have a hardware, try to dump the errors, collect the data corruption or timeouts and provide more information. DMA for SPI on LPSS hardware never was a problem (unlike UART in some cases). I'm almost 100% sure the problem is in DMA configuration / setting bits.
...
Quoted 3 lines
#include <
linux/dmaengine.h>
#include <linux/platform_data/dma-dw.h>
+#include <linux/dmi.h>Keep list ordered.
Quoted 1 line
#include "
spi-pxa2xx.h"
Quoted 7 lines
...
drivers/spi/spi-pxa2xx-pci.c line 27
Quoted 3 lines
drivers/spi/spi-pxa2xx-pci.c +3
-0+static bool spi_pxa2xx_force_pio; +module_param_named (force_pio, spi_pxa2xx_force_pio, bool,0444 ); +MODULE_PARM_DESC (force_pio,"Force PIO mode (disables DMA) for SPI transfers. ([0] = disabled, 1 = enabled)" );No. There is no room for module parameters like this.
Hi Andy,
Thank you for the review and feedback.
Regarding your question about why DMA fails on this hardware, I have just sent a detailed reply to Lukas in this thread outlining our physical verification findings. In short, I confirmed that the DMA controller's physical interrupt line is not functioning on this motherboard, and macOS and Windows 10 also completely bypass the DMA controller. Please refer to that email for the DMAR/IOMMU and cross-OS analysis details.
Regarding your other comments:
1. Header Sorting:
I will fix the alphabetical ordering of the include headers in the next revision.
2. The force_pio Module Parameter:
The original spi-pxa2xx driver is designed to hardcode DMA usage if DMA channels are available. I introduced the force_pio module parameter for two reasons:
• It serves as a diagnostic tool for users on similar hardware to
easily test and verify if their issues are related to DMA.
• It allows developers to easily test and debug the PIO code path on
DMA-capable LPSS hardware to prevent regressions.
That being said, if module parameters of this type are strictly not preferred, I am completely fine with removing force_pio and keeping only the DMI-based quirk for the MacBook8,1. Please let me know your preference.
Thanks,
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 3 lines · Shih-Yuan Lee (FourDollars)
Quoted 1 line · Andy Shevchenko
On Tue, Jul 21, 2026 at 12:21:14AM +0800, Shih-Yuan Lee wrote:
...
Quoted 7 lines
Thank you for the review and feedback.
Regarding your question about why DMA fails on this hardware, I have just sent a detailed reply to Lukas in this thread outlining our physical verification findings. In short, I confirmed that the DMA controller's physical interrupt line is not functioning on this motherboard,
It might be wrongly described in ACPI tables, but in HW it works.
Quoted 15 lines
and macOS and Windows 10 also completely bypass the DMA controller. Please refer to that email for the DMAR/IOMMU and cross-OS analysis details.
Regarding your other comments:
2. The
force_pioModule Parameter:
The original spi-pxa2xx driver is designed to hardcode DMA usage if DMA channels are available. I introduced theforce_piomodule parameter for two reasons:
• It serves as a diagnostic tool for users on similar hardware to
easily test and verify if their issues are related to DMA.
• It allows developers to easily test and debug the PIO code path on
DMA-capable LPSS hardware to prevent regressions.
Quoted 4 lines
That being said, if module parameters of this type are strictly not preferred, I am completely fine with removing
force_pioand keeping only the DMI-based quirk for the MacBook8,1. Please let me know your preference.
Neither is preferred until we will fully understand what's going on there. Yes, DMI quirk is a last resort, but to come to that we need to try better options. See my other reply about the DMA case.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
In reply to Shih-Yuan LeeQuoted 4 lines · Shih-Yuan Lee
On MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
[...]
Quoted 1 line
In the bugzilla comments, Leif Liddy writes:
"I needed to modify drivers/dma/dw/pci.c to assign the DMA controller
irq value to 21 (which is what is listed in the ACPI table) for it to work. The DMA controller was being assigned an irq value of 20 for some reason."
It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
Thanks,
Lukas
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 11 lines · Lukas Wunner
Quoted 4 lines · Shih-Yuan Lee
On MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
[...]
Quoted 1 line
In the bugzilla comments, Leif Liddy writes:
Quoted 7 lines
"I needed to modify
drivers/dma/dw/pci.cto assign the DMA controller
irq value to 21 (which is what is listed in the ACPI table) for it to work. The DMA controller was being assigned an irq value of 20 for some reason."It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
This is a very good catch!
I appreciate the comment. Based on the Bugzilla feedback regarding the incorrect interrupt assignment, it seems fixing the IRQ value via a quirk is indeed a better approach than disabling DMA wholesale. While I am not certain this will resolve the issue entirely, it looks very promising. I will investigate the IRQ mis-assignment problem further.
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 21 lines · Shih-Yuan Lee (FourDollars)
Quoted 11 lines · Lukas Wunner
Quoted 4 lines · Shih-Yuan Lee
On MacBook8,1 (early 2015 12" MacBook), the LPSS SPI controller at 00:15.4 suffers from hardware DMA handshake failures and interrupt routing bugs, causing keyboard/touchpad transactions to fail when DMA is enabled.
[...]
Quoted 1 line
In the bugzilla comments, Leif Liddy writes:
Quoted 7 lines
"I needed to modify
drivers/dma/dw/pci.cto assign the DMA controller
irq value to 21 (which is what is listed in the ACPI table) for it to work. The DMA controller was being assigned an irq value of 20 for some reason."It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
Quoted 7 lines
This is a very good catch!
I appreciate the comment. Based on the Bugzilla feedback regarding the incorrect interrupt assignment, it seems fixing the IRQ value via a quirk is indeed a better approach than disabling DMA wholesale. While I am not certain this will resolve the issue entirely, it looks very promising. I will investigate the IRQ mis-assignment problem further.
Hi Lukas,
Thank you for the suggestion. I have investigated this direction thoroughly by implementing the DMI-based IRQ override quirk in drivers/dma/dw/pci.c to assign both pdev->irq and chip->irq to 21 on the physical MacBook8,1.
However, I found that fixing up the IRQ to 21 in software is not a viable solution due to the following findings:
1. DMAR/IOMMU Validation Failure (Error -22)
On modern kernels with Interrupt Remapping (DMAR) enabled, overriding pdev->irq to 21 causes the dw_dmac_pci driver probe to fail with -EINVAL (-22) during request_irq(). This is because the DMAR/IOMMU strictly validates the PCI Requester ID (source ID) against the GSI mapping defined in the ACPI _PRT table. Forcing a device mapped to GSI 20 (00:15.0) to register on GSI 21 triggers a source ID mismatch check, which the kernel rejects.
2. The Mechanism Behind the Historical irqpoll Workaround In Bugzilla ticket 108331, the reason Leif Liddy's workaround functioned might be that they booted the system with irqpoll or irqfixup. Under irqpoll, the kernel polls the DMA status register on every clock tick/timer interrupt, which completely bypasses the broken physical interrupt line at the cost of high CPU overhead and power consumption. Without irqpoll, even if we register the handler on IRQ 21, the driver never receives physical interrupts because the hardware line on the motherboard is physically disconnected or masked.
3. Cross-OS Evidence (macOS and Windows 10)
• macOS: Apple's native driver runs the SPI controller in pure PIO mode.
• Windows 10 (Boot Camp): Interestingly, the Intel LPSS DMA
Controller (8086:9ce0) is left completely driverless (no driver installed) by the official Boot Camp package, and does not register or
use IRQ 20 at all. This forces the Windows SPI driver
(iaLPSS_SPI.sys) to fall back to pure PIO mode as well.
Since both macOS and Windows officially disable DMA to work around this hardware routing defect, forcing PIO mode in spi-pxa2xx is the most native, clean, and power-efficient way to support the MacBook8,1 without introducing probe errors or relying on high-overhead irqpoll workarounds.
Therefore, disabling DMA wholesale for this platform seems to be the only correct solution. I will update the commit message of the next revision to document these physical verification findings.
Thanks,
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 12 lines · Shih-Yuan Lee (FourDollars)
On Tue, Jul 21, 2026 at 5:26 PM Shih-Yuan Lee (FourDollars)
Quoted 10 lines · Lukas Wunner
Quoted 3 lines
It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
I appreciate the comment. Based on the Bugzilla feedback regarding the incorrect interrupt assignment, it seems fixing the IRQ value via a quirk is indeed a better approach than disabling DMA wholesale. While I am not certain this will resolve the issue entirely, it looks very promising. I will investigate the IRQ mis-assignment problem further.
Quoted 8 lines
1. DMAR/IOMMU Validation Failure (Error -22)
On modern kernels with Interrupt Remapping (DMAR) enabled, overridingpdev->irqto 21 causes thedw_dmac_pcidriver probe to fail with-EINVAL(-22) duringrequest_irq(). This is because the DMAR/IOMMU strictly validates the PCI Requester ID (source ID) against the GSI mapping defined in the ACPI_PRTtable. Forcing a device mapped to GSI 20 (00:15.0) to register on GSI 21 triggers a source ID mismatch check, which the kernel rejects.
That really seems like something we ought to be able to do, though I think the idiom here is to patch the ACPI tables so they are correct rather than quirk things in C code. ICBW about the preferred approach there.
BTW, if you're using a LLM to assist with analysis please don't just paste the output into mail. Instead make sure you understand what the LLM is saying and write up the salient bits.
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 23 lines · Mark Brown
Quoted 12 lines · Shih-Yuan Lee (FourDollars)
On Tue, Jul 21, 2026 at 5:26 PM Shih-Yuan Lee (FourDollars)
Quoted 10 lines · Lukas Wunner
Quoted 3 lines
It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
I appreciate the comment. Based on the Bugzilla feedback regarding the incorrect interrupt assignment, it seems fixing the IRQ value via a quirk is indeed a better approach than disabling DMA wholesale. While I am not certain this will resolve the issue entirely, it looks very promising. I will investigate the IRQ mis-assignment problem further.
Quoted 8 lines
1. DMAR/IOMMU Validation Failure (Error -22)
On modern kernels with Interrupt Remapping (DMAR) enabled, overridingpdev->irqto 21 causes thedw_dmac_pcidriver probe to fail with-EINVAL(-22) duringrequest_irq(). This is because the DMAR/IOMMU strictly validates the PCI Requester ID (source ID) against the GSI mapping defined in the ACPI_PRTtable. Forcing a device mapped to GSI 20 (00:15.0) to register on GSI 21 triggers a source ID mismatch check, which the kernel rejects.
Quoted 8 lines
That really seems like something we ought to be able to do, though I think the idiom here is to patch the ACPI tables so they are correct rather than quirk things in C code. ICBW about the preferred approach there.
BTW, if you're using a LLM to assist with analysis please don't just paste the output into mail. Instead make sure you understand what the LLM is saying and write up the salient bits.
Hi Mark,
You are absolutely correct. The standard idiom for correcting firmware bugs in ACPI tables is indeed using the Initrd ACPI Table Override mechanism (compiling a modified DSDT with corrected _PRT routing and prepending it to the initramfs).
Actually, I was just following Lukas's suggestion to modify drivers/dma/dw/pci.c to do a quick proof-of-concept test, and I used an AI assistant to help me write and deploy the driver quirk code.
I have previous experience with the Initrd ACPI Table Override mechanism, so I can definitely try that on my machine next to see if we can get DMA working properly by aligning the PCI routing and DMAR mapping correctly).
Also, I will keep your advice in mind regarding LLM usage, and will make sure to write up the key points myself rather than pasting raw output.
Thanks,
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 44 lines · Shih-Yuan Lee (FourDollars)
Quoted 23 lines · Mark Brown
Quoted 12 lines · Shih-Yuan Lee (FourDollars)
On Tue, Jul 21, 2026 at 5:26 PM Shih-Yuan Lee (FourDollars)
Quoted 10 lines · Lukas Wunner
Quoted 3 lines
It would seem better to fix up the incorrect interrupt number in a quirk, rather than disabling DMA wholesale.
I appreciate the comment. Based on the Bugzilla feedback regarding the incorrect interrupt assignment, it seems fixing the IRQ value via a quirk is indeed a better approach than disabling DMA wholesale. While I am not certain this will resolve the issue entirely, it looks very promising. I will investigate the IRQ mis-assignment problem further.
Quoted 8 lines
1. DMAR/IOMMU Validation Failure (Error -22)
On modern kernels with Interrupt Remapping (DMAR) enabled, overridingpdev->irqto 21 causes thedw_dmac_pcidriver probe to fail with-EINVAL(-22) duringrequest_irq(). This is because the DMAR/IOMMU strictly validates the PCI Requester ID (source ID) against the GSI mapping defined in the ACPI_PRTtable. Forcing a device mapped to GSI 20 (00:15.0) to register on GSI 21 triggers a source ID mismatch check, which the kernel rejects.Quoted 8 lines
That really seems like something we ought to be able to do, though I think the idiom here is to patch the ACPI tables so they are correct rather than quirk things in C code. ICBW about the preferred approach there.
BTW, if you're using a LLM to assist with analysis please don't just paste the output into mail. Instead make sure you understand what the LLM is saying and write up the salient bits.
Hi Mark,
You are absolutely correct. The standard idiom for correcting firmware bugs in ACPI tables is indeed using the Initrd ACPI Table Override mechanism (compiling a modified DSDT with corrected
_PRTrouting and prepending it to the initramfs).Actually, I was just following Lukas's suggestion to modify
drivers/dma/dw/pci.cto do a quick proof-of-concept test, and I used an AI assistant to help me write and deploy the driver quirk code.
Quoted 8 lines
I have previous experience with the Initrd ACPI Table Override mechanism, so I can definitely try that on my machine next to see if we can get DMA working properly by aligning the PCI routing and DMAR mapping correctly).
Also, I will keep your advice in mind regarding LLM usage, and will make sure to write up the key points myself rather than pasting raw output.
Hi Mark,
I just tested the Initrd ACPI Table Override on the machine. Here is what I found.
I patched the _PRT table in the DSDT to remap device 00:15.0 (Pin 0 and Pin 1) from GSI 20 to GSI 21 to match the SPI controller's interrupt line, bumped the OEM Revision from 0x00080001 to 0x00080002 so the kernel would accept the override, and prepended it to the initramfs.
The override was successfully applied:
ACPI: Table Upgrade: override [DSDT-APPLE - MacBook]
ACPI: DSDT ... 007EB5 (v03 APPLE MacBook 00080002 INTL 20251212)
After rebooting, the DMA controller loaded without errors and both dw:dmac168 and the SPI controller 0000:00:15.4 now correctly share IRQ 21:
21: 0 0 0 0 IR-IO-APIC 21-fasteoi dw:dmac168, 0000:00:15.4
However, SPI transfers still timed out with -ETIMEDOUT (-110) and the IRQ 21 counter remained at zero — no interrupts were actually received from the DMA hardware, despite the corrected ACPI routing.
This confirms the root cause is not a software IRQ routing issue but rather that the DMA controller's interrupt line is physically non-functional on this motherboard. macOS also does not use DMA for this device, which is consistent with that conclusion.
It is also worth noting that Apple may have intentionally mapped the DMA controller to GSI 20 — an otherwise unused interrupt — precisely because the interrupt line is not physically connected. On macOS and Windows, no DMA driver is ever loaded for this device, so the bogus GSI 20 entry is harmless. It effectively acts as a silent signal to the OS that interrupt-driven DMA should not be used on this platform. Since the DMA interrupt line is physically unconnected, no completion interrupt ever fires, causing all SPI transfers to time out with -ETIMEDOUT (-110) and ultimately deadlocking the applespi driver.
Therefore, forcing PIO mode via a DMI quirk in spi-pxa2xx-pci remains the only viable out-of-the-box solution for this platform.
Thanks,
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 9 lines · Shih-Yuan Lee (FourDollars)
On Tue, Jul 21, 2026 at 11:26 PM Shih-Yuan Lee (FourDollars) <[email protected]> wrote:
I just tested the Initrd ACPI Table Override on the machine. Here is what I found.
I patched the
_PRTtable in the DSDT to remap device 00:15.0 (Pin 0 and Pin 1) from GSI 20 to GSI 21 to match the SPI controller's interrupt line,
I'm lost here. Why SPI controller and DMA controller has to share an interrupt line?
Quoted 19 lines
umped the OEM Revision from
0x00080001to0x00080002so the kernel would accept the override, and prepended it to the initramfs.The override was successfully applied:
ACPI: Table Upgrade: override [DSDT-APPLE - MacBook]
ACPI: DSDT ... 007EB5 (v03 APPLE MacBook 00080002 INTL 20251212)After rebooting, the DMA controller loaded without errors and both dw:dmac168 and the SPI controller 0000:00:15.4 now correctly share IRQ 21:
21: 0 0 0 0 IR-IO-APIC 21-fasteoi dw:dmac168, 0000:00:15.4
However, SPI transfers still timed out with
-ETIMEDOUT(-110) and the IRQ 21 counter remained at zero — no interrupts were actually received from the DMA hardware, despite the corrected ACPI routing.
Quoted 8 lines
This confirms the root cause is not a software IRQ routing issue but rather that the DMA controller's interrupt line is physically non-functional on this motherboard. macOS also does not use DMA for this device, which is consistent with that conclusion.
It is also worth noting that Apple may have intentionally mapped the DMA controller to GSI 20 — an otherwise unused interrupt — precisely because the interrupt line is not physically connected.
It's impossible. The interrupt line from DMA controller is an IOAPIC RTE that is programmed by BIOS, there is no "physical wire" in traditional meaning. In any case this link is done on the SoC level. Apple can't burn that out of the die. They even can't fuse out that (of my knowledge LPSS is always present IP, the fuse works against the full controller, not parts like disabling interrupt message or "line").
Quoted 4 lines
On macOS and
Windows, no DMA driver is ever loaded for this device, so the bogus GSI 20 entry is harmless. It effectively acts as a silent signal to the OS that interrupt-driven DMA should not be used on this platform.
Maybe because they didn't get how it's supposed to work.
Quoted 3 lines
Since the DMA interrupt line is physically unconnected, no completion interrupt ever fires, causing all SPI transfers to time out with
-ETIMEDOUT(-110) and ultimately deadlocking the applespi driver.
No, this analysis is wrong. There are up to 3 interrupts that may participate in the design: DMA controller; SPI controller; and SPI peripheral. The first two usually use IOAPIC RTEs for the interrupts (and hence represented as Interrupt() resources in _CRS methods on DSDT, plus CSRT for DMA), and the last one often is GpioInt() and has nothing to do with DMA at all.
Quoted 2 lines
Therefore, forcing PIO mode via a DMI quirk in spi-pxa2xx-pci remains the only viable out-of-the-box solution for this platform.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Hi Andy,
Thank you for the detailed feedback and for explaining the SoC-level interrupt architecture.
Quoted 3 lines
Neither is preferred until we will fully understand what's going on there. Yes, DMI quirk is a last resort, but to come to that we need to try better options. See my other reply about the DMA case.
I completely agree that DMI quirks should be a last resort when better options exist. I have removed the custom force_pio module parameter from the series.
Having investigated the ACPI resource options and confirmed via MMIO that the EFI firmware permanently holds the LPSS DMA hardware block in reset (LPSS_PRIV_RESETS = 0x0), applying a targeted DMI quirk in spi-pxa2xx-pci.c appears to be the unavoidable last resort required to prevent system hangs on this platform.
Quoted 2 lines
It's impossible. The interrupt line from DMA controller is an IOAPIC RTE that is programmed by BIOS, there is no "physical wire" in traditional meaning.
Thank you for the correction regarding PCH SoC interrupt routing. You are completely right that within the PCH die, LPSS interrupts use IOAPIC RTEs rather than discrete physical PCB traces.
Quoted 5 lines
No, this analysis is wrong. There are up to 3 interrupts that may participate in the design: DMA controller; SPI controller; and SPI peripheral. The first two usually use IOAPIC RTEs for the interrupts (and hence represented as
Interrupt()resources in_CRSmethods on DSDT, plus CSRT for DMA), and the last one often isGpioInt()and has nothing to do with DMA at all.
Thank you for clarifying the three distinct interrupts involved in LPSS systems:
1. DMA controller interrupt (dw:dmac)
2. SPI controller interrupt (00:15.4)
3. SPI peripheral interrupt (GpioInt())
On MacBook8,1, the applespi input device specifically uses the third
mechanism: a dedicated Out-Of-Band GpioInt() line (TPAD_SPI_INT_L) connected to the PCH GPIO controller to drive SPI keyboard and trackpad transfers in PIO mode.
Quoted 4 lines
For old hardware (before Skylake) LPSS DMA interrupt line is defined in two places: DSDT (and/or PCI) and CSRT. Changing in one might not help as CSRT may have something different. Check
drivers/dma/acpi-dma.cacpi_dma_parse_resource_group().
Thanks for the pointer regarding CSRT on pre-Skylake platforms.
Direct MMIO inspection of the LPSS DMA controller (PCI device 00:15.0) on physical MacBook8,1 hardware reveals that Apple's EFI firmware leaves the functional reset register (LPSS_PRIV_RESETS at BAR0 + 0x204) permanently set to 0x00000000 (held in reset).
Since the LPSS DMA controller is held in reset (0x0) by EFI firmware and the SPI peripheral relies on the dedicated OOB GpioInt() line for PIO transfers, forcing PIO mode via the DMI quirk in spi-pxa2xx-pci.c accurately reflects how EFI configures the hardware, allowing the driver to operate reliably out-of-the-box without system hangs.
I have compiled the complete MMIO register dumps, schematic references, and physical hardware test logs in a technical report for reference [0]:
Thanks,
Shih-Yuan
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 10 lines · Shih-Yuan Lee (FourDollars)
Direct MMIO inspection of the LPSS DMA controller (PCI device 00:15.0) on physical MacBook8,1 hardware reveals that Apple's EFI firmware leaves the functional reset register (
LPSS_PRIV_RESETSat BAR0 +0x204) permanently set to0x00000000(held in reset).Since the LPSS DMA controller is held in reset (
0x0) by EFI firmware and the SPI peripheral relies on the dedicated OOBGpioInt()line for PIO transfers, forcing PIO mode via the DMI quirk inspi-pxa2xx-pci.caccurately reflects how EFI configures the hardware, allowing the driver to operate reliably out-of-the-box without system hangs.
The OOB GpioInt() and the DMA controller are two different things:
The former allows the peripheral to signal that a key has been pressed etc, whereas the latter allows feeding the FIFOs of the SPI host controller without CPU involvement (or with *less* CPU involvement, the CPU is still needed to configure the DMA engine for a transfer).
I'm not sure why Apple chose to keep the DMA controller in reset, but have you explored enabling it on boot? If this can only be performed before EFI hands over control to the OS, then it could be done in the EFI stub, see setup_quirks() in drivers/firmware/efi/libstub/x86-stub.c. Or maybe it can be done at runtime after the kernel has assumed control, I'm not sure.
Thanks,
Lukas
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 2 lines · Lukas Wunner
I'm not sure why Apple chose to keep the DMA controller in reset, but have you explored enabling it on boot?
Thinking about this some more and after taking a look at drivers/input/keyboards/applespi.c, I realize messages exchanged via SPI are always 256 bytes in size. That's not a whole lot, and of course messages are only exchanged when a key is pressed or the trackpad is used. So perhaps Apple determined that sending this in PIO mode is ultimately more efficient than keeping a DMA engine powered on all the time. The CPU load saved through offloading the SPI transfer is offset by having to program the DMA engine for each transfer.
So perhaps usage of PIO mode was indeed an intentional choice by Apple and need not be second-guessed. I'm not sure.
Thanks,
Lukas
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 14 lines · Lukas Wunner
Quoted 2 lines · Lukas Wunner
I'm not sure why Apple chose to keep the DMA controller in reset, but have you explored enabling it on boot?
Thinking about this some more and after taking a look at
drivers/input/keyboards/applespi.c, I realize messages exchanged via SPI are always 256 bytes in size. That's not a whole lot, and of course messages are only exchanged when a key is pressed or the trackpad is used. So perhaps Apple determined that sending this in PIO mode is ultimately more efficient than keeping a DMA engine powered on all the time. The CPU load saved through offloading the SPI transfer is offset by having to program the DMA engine for each transfer.
Quoted 2 lines
So perhaps usage of PIO mode was indeed an intentional choice by Apple and need not be second-guessed. I'm not sure.
DMA can also add extra latency with interrupt handling, it's possible someone decided to trade some CPU time for minimal latency (though you could probably arrange for shorter messages...).
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
Quoted 15 lines · Mark Brown
Quoted 13 lines · Lukas Wunner
Quoted 2 lines · Lukas Wunner
I'm not sure why Apple chose to keep the DMA controller in reset, but have you explored enabling it on boot?
Thinking about this some more and after taking a look at
drivers/input/keyboards/applespi.c, I realize messages exchanged via SPI are always 256 bytes in size. That's not a whole lot, and of course messages are only exchanged when a key is pressed or the trackpad is used. So perhaps Apple determined that sending this in PIO mode is ultimately more efficient than keeping a DMA engine powered on all the time. The CPU load saved through offloading the SPI transfer is offset by having to program the DMA engine for each transfer.
Quoted 6 lines
Quoted 2 lines
So perhaps usage of PIO mode was indeed an intentional choice by Apple and need not be second-guessed. I'm not sure.
DMA can also add extra latency with interrupt handling, it's possible someone decided to trade some CPU time for minimal latency (though you could probably arrange for shorter messages...).
Okay, so let's go with the disabled DMA approach, but this will need a good commit message and a big comment in the code explaining all findings and some kind of summary of this discussion(s).
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
In reply to Shih-Yuan Lee (FourDollars)Quoted 3 lines · Shih-Yuan Lee (FourDollars)
This confirms the root cause is not a software IRQ routing issue but rather that the DMA controller's interrupt line is physically non-functional on this motherboard.
Like Andy says, the DMA controller is internal to the SkyLake-Y-ULX with SunrisePoint-PCH (Intel 100 Series Chipset), it's not something that needs to be connected on the mainboard.
I've just looked through the Specification Update for Intel 100 Series Chipset and it doesn't list DMA-related errata. Hm...
Lukas
Re: [PATCH v16 5/7] spi: pxa2xx: disable DMA for Apple MacBook8,1
In reply to Shih-Yuan Lee (FourDollars)Quoted 1 line · Shih-Yuan Lee (FourDollars)
On Tue, Jul 21, 2026 at 11:05 PM Mark Brown <[email protected]> wrote:
...
Quoted 9 lines
Actually, I was just following Lukas's suggestion to modify
drivers/dma/dw/pci.cto do a quick proof-of-concept test, and I used an AI assistant to help me write and deploy the driver quirk code.I have previous experience with the Initrd ACPI Table Override mechanism, so I can definitely try that on my machine next to see if we can get DMA working properly by aligning the PCI routing and DMAR mapping correctly).
output.
For old hardware (before Skylake) LPSS DMA interrupt line is defined in two places: DSDT (and/or PCI) and CSRT. Changing in one might not help as CSRT may have something different. Check drivers/dma/acpi-dma.c acpi_dma_parse_resource_group().
Signature · 2 lines
With Best Regards, Andy Shevchenko
PATCH v16 6/7 spi: pxa2xx: restore LPSS private register state on S3 resume
In reply to Shih-Yuan LeeIntel LPSS SPI controllers lose all private register state across S3 suspend because the LPSS power domain is fully removed. On resume the driver only re-enables the SSP clock, leaving the LPSS private registers in their power-on-reset state, which causes two problems:
1. LPSS_PRIV_RESETS (offset 0x04 within the LPSS private space) stays zero, keeping the functional block in reset. Any MMIO access while the block is held in reset causes a PCIe Completion Timeout and a watchdog-triggered system reset. LPSS_PRIV_RESETS_FUNC and LPSS_PRIV_RESETS_IDMA must be de-asserted before any other register access on resume.
2. The LPSS software chip-select control register must not be blindly restored from its suspend-time snapshot: if CS was asserted at the moment of suspend, restoring that state corrupts the first post-resume SPI transaction. Instead, call lpss_ssp_setup() which unconditionally writes SW_MODE | CS_HIGH (idle/deasserted), matching the state established at probe time.
To resolve these issues safely:
- Wrap S3 suspend/resume with pm_runtime_resume_and_get() and pm_runtime_put_noidle() to guarantee active clocks during MMIO access and preserve PM reference counting.
- Restrict LPSS private register save/restore to LPT, BYT, and BSW platforms via pxa2xx_spi_need_lpss_restore() (newer platforms are handled by intel-lpss.c).
- Save only the first 6 LPSS private registers (offsets 0x00..0x14) in drv_data during suspend, avoiding reserved offsets beyond 0x14.
- On resume, de-assert resets first, restore saved registers, call lpss_ssp_setup(), and clear drv_data->suspended to prevent unclocked IRQ access.
- Add error recovery paths for spi_controller_suspend/resume failures.
- On the resume error path, call pm_runtime_set_suspended() before pm_runtime_put_noidle() to align the PM runtime state with the already-disabled hardware clock, preventing pxa2xx_spi_runtime_suspend() from attempting unclocked MMIO via pxa2xx_spi_off().
--- drivers/spi/spi-pxa2xx.c | 105 +++++++++++++++++++++++++++++++++++++-- drivers/spi/spi-pxa2xx.h | 1 + 2 files changed, 102 insertions(+), 4 deletions(-)
drivers/spi/spi-pxa2xx.c +101 -4
@@ -72,6 +72,11 @@ struct chip_data {#define LPSS_CAPS_CS_EN_SHIFT 9#define LPSS_CAPS_CS_EN_MASK (0xf << LPSS_CAPS_CS_EN_SHIFT) +/* Offsets from drv_data->lpss_base */ +#define LPSS_PRIV_RESETS 0x04 +#define LPSS_PRIV_RESETS_IDMA BIT(2) +#define LPSS_PRIV_RESETS_FUNC 0x3 +#define LPSS_PRIV_CLOCK_GATE 0x38#define LPSS_PRIV_CLOCK_GATE_CLK_CTL_MASK 0x3#define LPSS_PRIV_CLOCK_GATE_CLK_CTL_FORCE_ON 0x3
@@ -189,6 +194,18 @@ static bool is_lpss_ssp(const struct driver_data *drv_data) } } +static bool pxa2xx_spi_need_lpss_restore (const struct driver_data *drv_data) +{ +switch (drv_data->ssp_type) { +case LPSS_LPT_SSP : +case LPSS_BYT_SSP : +case LPSS_BSW_SSP : +return true; +default : +return false; + } +} +static bool is_quark_x1000_ssp (const struct driver_data *drv_data) {return drv_data->ssp_type ==QUARK_X1000_SSP ;
@@ -1532,17 +1549,44 @@ static int pxa2xx_spi_suspend(struct device *dev)struct ssp_device *ssp = drv_data->ssp;int status; + status =pm_runtime_resume_and_get (dev); +if (status <0 ) +return status; + + status =spi_controller_suspend (drv_data->controller);if (status) -return status; + goto out_put; +/* Disable SSP interrupt generation on hardware level while clock is active */ drv_data->suspended = true;pxa2xx_spi_off (drv_data);synchronize_irq (ssp->irq); +if (pxa2xx_spi_need_lpss_restore (drv_data)) { +unsigned int i; + +/* +* Save the first 6 LPSS private registers (offsets 0x00 to 0x14) +* while the clock is still enabled. They are lost when the LPSS +* power domain is removed across S3 and must be restored on resume. +* Use drv_data->lpss_base so the correct per-platform offset +* is applied regardless of LPSS IP revision. +* Registers beyond 0x14 (except CS control at 0x18) are reserved +* or unimplemented on LPT, and accessing them triggers a PCIe +* Completion Timeout causing a system halt. +*/ +for (i =0 ; i <6 ; i++) + drv_data->lpss_priv_ctx[i] =readl (drv_data->lpss_base + i *4 ); + } +pxa2xx_spi_clk_disable (drv_data);return 0 ; + +out_put: +pm_runtime_put_noidle (dev); +return status; }static int pxa2xx_spi_resume (struct device *dev)
@@ -1555,9 +1599,47 @@ static int pxa2xx_spi_resume(struct device *dev)if (!pm_runtime_suspended (dev)) { status =pxa2xx_spi_clk_enable (drv_data);if (status) -return status; + goto out_put; } +if (pxa2xx_spi_need_lpss_restore (drv_data)) { +unsigned int i; + +/* +* The LPSS power domain is removed across S3, taking +* all private registers with it. De-assert the +* functional block and IDMA resets first; any MMIO +* access while the block is held in reset causes a +* PCIe Completion Timeout and a watchdog-triggered +* system reset. +*/ +writel (LPSS_PRIV_RESETS_FUNC |LPSS_PRIV_RESETS_IDMA , + drv_data->lpss_base +LPSS_PRIV_RESETS ); + +/* Restore the other 5 saved private registers */ +for (i =0 ; i <6 ; i++) { +if (i ==LPSS_PRIV_RESETS /4 ) +continue ; +writel (drv_data->lpss_priv_ctx[i], + drv_data->lpss_base + i *4 ); + } + } + +if (is_lpss_ssp (drv_data)) { +/* +* Re-initialise the SW chip-select control register so +* CS starts deasserted (SW_MODE | CS_HIGH), regardless +* of the state it was in at suspend time. A stale +* asserted CS on the first post-resume transaction +* corrupts the write-status response from the device. +*/ +lpss_ssp_setup (drv_data); + } + +/* +* Now that resets are de-asserted and registers are restored, +* it is safe to handle interrupts. +*/ drv_data->suspended = false;/* Start the queue running */
@@ -1566,10 +1648,25 @@ static int pxa2xx_spi_resume(struct device *dev) drv_data->suspended = true;synchronize_irq (ssp->irq);pxa2xx_spi_clk_disable (drv_data); -return status; + goto out_put; } -return 0 ; +out_put: +if (!pm_runtime_suspended (dev)) { +if (status) +/* +* Clock is already disabled on the error path; align +* the PM runtime state with hardware reality before +* releasing the reference so the PM core does not +* later invoke pxa2xx_spi_runtime_suspend() and attempt +* unclocked MMIO via pxa2xx_spi_off(). +*/ +pm_runtime_set_suspended (dev); +pm_runtime_put_noidle (dev); + } + +return status; + }static int pxa2xx_spi_runtime_suspend (struct device *dev)
drivers/spi/spi-pxa2xx.h +1 -0
@@ -71,6 +71,7 @@ struct driver_data {irqreturn_t (*transfer_handler)(struct driver_data *drv_data);void __iomem *lpss_base; +u32 lpss_priv_ctx[6 ];bool suspended;bool clk_enabled;
Signature · 1 line
2.39.5
Re: [PATCH v16 6/7] spi: pxa2xx: restore LPSS private register state on S3 resume
On Tue, Jul 21, 2026 at 12:21:15AM +0800, Shih-Yuan Lee wrote:
Is this series AI-assisted?
Quoted 19 lines
Intel LPSS SPI controllers lose all private register state across S3 suspend because the LPSS power domain is fully removed. On resume the driver only re-enables the SSP clock, leaving the LPSS private registers in their power-on-reset state, which causes two problems:
1.
LPSS_PRIV_RESETS(offset0x04within the LPSS private space) stays zero, keeping the functional block in reset. Any MMIO access while the block is held in reset causes a PCIe Completion Timeout and a watchdog-triggered system reset.LPSS_PRIV_RESETS_FUNCandLPSS_PRIV_RESETS_IDMAmust be de-asserted before any other register access on resume.2. The LPSS software chip-select control register must not be blindly restored from its suspend-time snapshot: if CS was asserted at the moment of suspend, restoring that state corrupts the first post-resume SPI transaction. Instead, call
lpss_ssp_setup()which unconditionally writesSW_MODE|CS_HIGH(idle/deasserted), matching the state established at probe time.
Quoted 5 lines
To resolve these issues safely:
- Wrap S3 suspend/resume withpm_runtime_resume_and_get()andpm_runtime_put_noidle()to guarantee active clocks during MMIO access and preserve PM reference counting.
- Restrict LPSS private register save/restore to LPT, BYT, and BSW
^^^^ (1)
Quoted 4 lines
platforms via
pxa2xx_spi_need_lpss_restore()(newer platforms are handled byintel-lpss.c).
- Save only the first 6 LPSS private registers (offsets 0x00..0x14) indrv_dataduring suspend, avoiding reserved offsets beyond0x14.
Quoted 8 lines
- On resume, de-assert resets first, restore saved registers, call
lpss_ssp_setup(), and cleardrv_data->suspendedto prevent unclocked IRQ access.
- Add error recovery paths for spi_controller_suspend/resume failures.
- On the resume error path, callpm_runtime_set_suspended()beforepm_runtime_put_noidle()to align the PM runtime state with the already-disabled hardware clock, preventingpxa2xx_spi_runtime_suspend()from attempting unclocked MMIO viapxa2xx_spi_off().
This is an ugly hack.
Saving context is done in drivers/acpi/x86/lpss.c (see #1 why this file). If something wrong in the flow it has to be fixed there, not here.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 6/7] spi: pxa2xx: restore LPSS private register state on S3 resume
Hi Andy,
Quoted 1 line
Is this series AI-assisted?
Yes, I used an LLM assistant to help format and polish the commit messages. However, all hardware register analysis (such as inspecting BAR0 + 0x204 LPSS Private Resets), DSDT ACPI overrides, schematic verifications for logic board 820-00244-A, and physical S3 suspend/resume hardware testing on MacBook8,1 were performed empirically on real hardware.
Quoted 40 lines
Quoted 19 lines
Intel LPSS SPI controllers lose all private register state across S3 suspend because the LPSS power domain is fully removed. On resume the driver only re-enables the SSP clock, leaving the LPSS private registers in their power-on-reset state, which causes two problems:
1.
LPSS_PRIV_RESETS(offset0x04within the LPSS private space) stays zero, keeping the functional block in reset. Any MMIO access while the block is held in reset causes a PCIe Completion Timeout and a watchdog-triggered system reset.LPSS_PRIV_RESETS_FUNCandLPSS_PRIV_RESETS_IDMAmust be de-asserted before any other register access on resume.2. The LPSS software chip-select control register must not be blindly restored from its suspend-time snapshot: if CS was asserted at the moment of suspend, restoring that state corrupts the first post-resume SPI transaction. Instead, call
lpss_ssp_setup()which unconditionally writesSW_MODE|CS_HIGH(idle/deasserted), matching the state established at probe time.Quoted 5 lines
To resolve these issues safely:
- Wrap S3 suspend/resume withpm_runtime_resume_and_get()andpm_runtime_put_noidle()to guarantee active clocks during MMIO access and preserve PM reference counting.
- Restrict LPSS private register save/restore to LPT, BYT, and BSW^^^^ (1)
Quoted 4 lines
platforms via
pxa2xx_spi_need_lpss_restore()(newer platforms are handled byintel-lpss.c).
- Save only the first 6 LPSS private registers (offsets 0x00..0x14) indrv_dataduring suspend, avoiding reserved offsets beyond0x14.Quoted 8 lines
- On resume, de-assert resets first, restore saved registers, call
lpss_ssp_setup(), and cleardrv_data->suspendedto prevent unclocked IRQ access.
- Add error recovery paths for spi_controller_suspend/resume failures.
- On the resume error path, callpm_runtime_set_suspended()beforepm_runtime_put_noidle()to align the PM runtime state with the already-disabled hardware clock, preventingpxa2xx_spi_runtime_suspend()from attempting unclocked MMIO viapxa2xx_spi_off().
Quoted 4 lines
This is an ugly hack.
Saving context is done in
drivers/acpi/x86/lpss.c(see #1 why this file). If something wrong in the flow it has to be fixed there, not here.
I checked drivers/acpi/x86/lpss.c and its LPSS_SAVE_CTX mechanism. However, on MacBook8,1, the LPSS SPI controller is instantiated as a PCI device (00:15.4 / 8086:9ce6) bound to spi-pxa2xx-pci.c, rather than an ACPI platform device handled by drivers/acpi/x86/lpss.c. As a result, acpi_lpss_restore_ctx() is never invoked for this PCI device.
Physical hardware S3 testing (rtcwake -m mem) confirms that across S3 suspend, LPSS_PRIV_RESETS (BAR0 + 0x204) is reset to 0x00000000 (functional block held in reset) by the PCH, causing all post-resume SPI transfers to fail continuously with -110 timeouts (applespi spi-APP000D:00: SPI transfer timed out).
I understand your point about not doing context saving in spi-pxa2xx.c. Would moving the LPSS private register context save/restore into spi-pxa2xx-pci.c (the PCI glue driver) be an acceptable approach for PCI-enumerated devices?
Thanks,
Shih-Yuan
Re: [PATCH v16 6/7] spi: pxa2xx: restore LPSS private register state on S3 resume
On Sun, Jul 26, 2026 at 07:48:27PM +0800, Shih-Yuan Lee (FourDollars) wrote:
...
Quoted 8 lines
Quoted 1 line
Is this series AI-assisted?
Yes, I used an LLM assistant to help format and polish the commit messages. However, all hardware register analysis (such as inspecting BAR0 +
0x204LPSS Private Resets), DSDT ACPI overrides, schematic verifications for logic board 820-00244-A, and physical S3 suspend/resume hardware testing on MacBook8,1 were performed empirically on real hardware.
Check the kernel documentation if you need or may use Assisted-by tag.
...
Quoted 16 lines
Quoted 4 lines
This is an ugly hack.
Saving context is done in
drivers/acpi/x86/lpss.c(see #1 why this file). If something wrong in the flow it has to be fixed there, not here.I checked
drivers/acpi/x86/lpss.cand itsLPSS_SAVE_CTXmechanism. However, on MacBook8,1, the LPSS SPI controller is instantiated as a PCI device (00:15.4 / 8086:9ce6) bound tospi-pxa2xx-pci.c, rather than an ACPI platform device handled bydrivers/acpi/x86/lpss.c. As a result,acpi_lpss_restore_ctx()is never invoked for this PCI device.Physical hardware S3 testing (rtcwake -m mem) confirms that across S3 suspend,
LPSS_PRIV_RESETS(BAR0 +0x204) is reset to0x00000000(functional block held in reset) by the PCH, causing all post-resume SPI transfers to fail continuously with -110 timeouts (applespi spi-APP000D:00: SPI transfer timed out).
Quoted 4 lines
I understand your point about not doing context saving in
spi-pxa2xx.c. Would moving the LPSS private register context save/restore intospi-pxa2xx-pci.c(the PCI glue driver) be an acceptable approach for PCI-enumerated devices?
I see. Can we do something similar to what drivers/pci/pci-mid.c does? And somehow to unify the core part of the context saving, et cetera between two ACPI and PCI cases?
Signature · 2 lines
With Best Regards, Andy Shevchenko
PATCH v16 7/7 spi: pxa2xx: rename local status variable to ret
In reply to Shih-Yuan LeeRename the return value variable name from 'status' to 'ret' in the pxa2xx_spi_probe(), pxa2xx_spi_suspend(), pxa2xx_spi_resume(), and pxa2xx_spi_runtime_resume() functions to conform to standard Linux kernel coding conventions.
--- drivers/spi/spi-pxa2xx.c | 66 +++++++++++++++++++--------------------- 1 file changed, 32 insertions(+), 34 deletions(-)
drivers/spi/spi-pxa2xx.c +32 -34
@@ -1313,7 +1313,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp,struct spi_controller *controller;struct driver_data *drv_data;const struct lpss_config *config; -int status; +int ret;u32 tmp;if (platform_info->is_target)
@@ -1372,8 +1372,8 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp,/* Setup DMA if requested */ if (platform_info->enable_dma) { - status =pxa2xx_spi_dma_setup (drv_data); -if (status) { + ret =pxa2xx_spi_dma_setup (drv_data); +if (ret) {dev_warn (dev,"no DMA channels available, using PIO\n" ); platform_info->enable_dma = false; }else {
@@ -1387,16 +1387,16 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, }/* Enable SOC clock */ - status =pxa2xx_spi_clk_enable (drv_data); -if (status) + ret =pxa2xx_spi_clk_enable (drv_data); +if (ret) goto out_error_dma_alloc; drv_data->suspended = false; - status =request_irq (ssp->irq, ssp_int,IRQF_SHARED ,dev_name (dev), + ret =request_irq (ssp->irq, ssp_int,IRQF_SHARED ,dev_name (dev), drv_data); -if (status <0 ) { - status =dev_err_probe (dev, status,"cannot get IRQ %d\n" , ssp->irq); +if (ret <0 ) { + ret =dev_err_probe (dev, ret,"cannot get IRQ %d\n" , ssp->irq); goto out_error_clock_enabled; }
@@ -1477,23 +1477,23 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, drv_data->gpiod_ready =devm_gpiod_get_optional (dev,"ready" ,GPIOD_OUT_LOW );if (IS_ERR (drv_data->gpiod_ready)) { - status =PTR_ERR (drv_data->gpiod_ready); + ret =PTR_ERR (drv_data->gpiod_ready); goto out_error_irq_alloc; } }/* Register with the SPI framework */ dev_set_drvdata (dev, drv_data); - status =spi_register_controller (controller); -if (status) { -dev_err_probe (dev, status,"problem registering SPI controller\n" ); + ret =spi_register_controller (controller); +if (ret) { +dev_err_probe (dev, ret,"problem registering SPI controller\n" ); goto out_error_irq_alloc; }if (is_lpss_ssp (drv_data) && !platform_info->enable_dma)pm_runtime_get_noresume (dev); -return status; +return ret; out_error_irq_alloc:free_irq (ssp->irq, drv_data);
@@ -1504,7 +1504,7 @@ int pxa2xx_spi_probe(struct device *dev, struct ssp_device *ssp, out_error_dma_alloc:pxa2xx_spi_dma_release (drv_data); -return status; +return ret; }EXPORT_SYMBOL_NS_GPL (pxa2xx_spi_probe,"SPI_PXA2xx" );
@@ -1547,15 +1547,14 @@ static int pxa2xx_spi_suspend(struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev);struct ssp_device *ssp = drv_data->ssp; -int status; - - status =pm_runtime_resume_and_get (dev); -if (status <0 ) -return status; +int ret; + ret =pm_runtime_resume_and_get (dev); +if (ret <0 ) +return ret; - status =spi_controller_suspend (drv_data->controller); -if (status) + ret =spi_controller_suspend (drv_data->controller); +if (ret) goto out_put;/* Disable SSP interrupt generation on hardware level while clock is active */
@@ -1586,19 +1585,19 @@ static int pxa2xx_spi_suspend(struct device *dev) out_put:pm_runtime_put_noidle (dev); -return status; +return ret; }static int pxa2xx_spi_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev);struct ssp_device *ssp = drv_data->ssp; -int status; +int ret;/* Enable the SSP clock */ if (!pm_runtime_suspended (dev)) { - status =pxa2xx_spi_clk_enable (drv_data); -if (status) + ret =pxa2xx_spi_clk_enable (drv_data); +if (ret) goto out_put; }
@@ -1643,8 +1642,8 @@ static int pxa2xx_spi_resume(struct device *dev) drv_data->suspended = false;/* Start the queue running */ - status =spi_controller_resume (drv_data->controller); -if (status) { + ret =spi_controller_resume (drv_data->controller); +if (ret) { drv_data->suspended = true;synchronize_irq (ssp->irq);pxa2xx_spi_clk_disable (drv_data);
@@ -1653,7 +1652,7 @@ static int pxa2xx_spi_resume(struct device *dev) out_put:if (!pm_runtime_suspended (dev)) { -if (status) +if (ret)/* * Clock is already disabled on the error path; align * the PM runtime state with hardware reality before
@@ -1665,8 +1664,7 @@ static int pxa2xx_spi_resume(struct device *dev)pm_runtime_put_noidle (dev); } -return status; - +return ret; }static int pxa2xx_spi_runtime_suspend (struct device *dev)
@@ -1683,11 +1681,11 @@ static int pxa2xx_spi_runtime_suspend(struct device *dev)static int pxa2xx_spi_runtime_resume (struct device *dev) {struct driver_data *drv_data =dev_get_drvdata (dev); -int status; +int ret; - status =pxa2xx_spi_clk_enable (drv_data); -if (status) -return status; + ret =pxa2xx_spi_clk_enable (drv_data); +if (ret) +return ret; drv_data->suspended = false;return 0 ;
Signature · 1 line
2.39.5
Re: [PATCH v16 7/7] spi: pxa2xx: rename local status variable to ret
Quoted 4 lines · Shih-Yuan Lee
Rename the return value variable name from 'status' to 'ret' in the
pxa2xx_spi_probe(),pxa2xx_spi_suspend(),pxa2xx_spi_resume(), andpxa2xx_spi_runtime_resume()functions to conform to standard Linux kernel coding conventions.
It makes a ping-pong style of changes: Previous patches add more status uses and moves, and so on. Instead, this patch should go before all that.
Signature · 2 lines
With Best Regards, Andy Shevchenko
Re: [PATCH v16 7/7] spi: pxa2xx: rename local status variable to ret
Quoted 7 lines · Andy Shevchenko
Quoted 4 lines · Shih-Yuan Lee
Rename the return value variable name from 'status' to 'ret' in the
pxa2xx_spi_probe(),pxa2xx_spi_suspend(),pxa2xx_spi_resume(), andpxa2xx_spi_runtime_resume()functions to conform to standard Linux kernel coding conventions.
Hi Andy,
Quoted 2 lines
It makes a ping-pong style of changes: Previous patches add more status uses and moves, and so on. Instead, this patch should go before all that.
Agreed. I have reordered the patch series for the next version so that "spi: pxa2xx: rename local status variable to ret" is placed near the front, eliminating all ping-pong diff churn.
Thanks,
Shih-Yuan