Skip to contentSkip to conversation

PCI/sysfs: Avoid spurious runtime PM wakeup on config space accesses

PATCH Applied Krzysztof Wilczyński 2 participants 3 messages Jul 20 – Aug 1 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PCI/sysfs: Avoid spurious runtime PM wakeup on config space accesses

Currently, the boundary checks in pci_read_config() and pci_write_config() reject only offsets beyond the effective configuration space size.

Thus, an access at an offset exactly equal to that size passes the check, has its length clamped to zero, and then invokes pci_config_pm_runtime_get() and pci_config_pm_runtime_put() around transfer blocks that do nothing.

This is a problem because pci_config_pm_runtime_get() synchronously resumes the upstream bridge through pm_runtime_get_sync(), and resumes the device itself through pm_runtime_resume() when it is in D3cold, only for the handler to return zero immediately afterwards. Such a spurious wakeup wastes power and adds needless resume latency.

The sysfs core already clamps accesses against the attribute size set through the bin_size() callback, which reports either 256 or 4096 bytes. As such, the affected accesses are reads at offset 64 (or 128 for CardBus devices) through files opened without CAP_SYS_ADMIN, and reads and writes at the exact configuration space size on devices where a quirk sets a non-standard size.

Therefore, reject accesses at the boundary offset as well, so that they return early before any runtime PM involvement, matching the procfs implementations in proc_bus_pci_read() and proc_bus_pci_write().

The value returned to userspace at these offsets remains zero, so the change is not visible to userspace.

---
 drivers/pci/pci-sysfs.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/pci/pci-sysfs.c +2 -2
index 5ec0b245a69b..a4f6559c619c 100644
@@ -718,7 +718,7 @@ static ssize_t pci_read_config(struct file *filp, struct kobject *kobj,
 	else if (dev->hdr_type == PCI_HEADER_TYPE_CARDBUS)
 		size = 128;
 
-	if (off > size)
+	if (off >= size)
 		return 0;
 	if (off + count > size) {
 		size -= off;
@@ -799,7 +799,7 @@ static ssize_t pci_write_config(struct file *filp, struct kobject *kobj,
 		add_taint(TAINT_USER, LOCKDEP_STILL_OK);
 	}
 
-	if (off > dev->cfg_size)
+	if (off >= dev->cfg_size)
 		return 0;
 	if (off + count > dev->cfg_size) {
 		size = dev->cfg_size - off;
Signature · 1 line
2.55.0
#1 lore raw
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 22 lines

Currently, the boundary checks in pci_read_config() and pci_write_config() reject only offsets beyond the effective configuration space size.

Thus, an access at an offset exactly equal to that size passes the check, has its length clamped to zero, and then invokes pci_config_pm_runtime_get() and pci_config_pm_runtime_put() around transfer blocks that do nothing.

This is a problem because pci_config_pm_runtime_get() synchronously resumes the upstream bridge through pm_runtime_get_sync(), and resumes the device itself through pm_runtime_resume() when it is in D3cold, only for the handler to return zero immediately afterwards. Such a spurious wakeup wastes power and adds needless resume latency.

The sysfs core already clamps accesses against the attribute size set through the bin_size() callback, which reports either 256 or 4096 bytes. As such, the affected accesses are reads at offset 64 (or 128 for CardBus devices) through files opened without CAP_SYS_ADMIN, and reads and writes at the exact configuration space size on devices where a quirk sets a non-standard size.

Quoted 8 lines

Therefore, reject accesses at the boundary offset as well, so that they return early before any runtime PM involvement, matching the procfs implementations in proc_bus_pci_read() and proc_bus_pci_write().

The value returned to userspace at these offsets remains zero, so the change is not visible to userspace.

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author Applied
#2 lore raw
In reply to Krzysztof Wilczyński

Hello,

Quoted 22 lines

Currently, the boundary checks in pci_read_config() and pci_write_config() reject only offsets beyond the effective configuration space size.

Thus, an access at an offset exactly equal to that size passes the check, has its length clamped to zero, and then invokes pci_config_pm_runtime_get() and pci_config_pm_runtime_put() around transfer blocks that do nothing.

This is a problem because pci_config_pm_runtime_get() synchronously resumes the upstream bridge through pm_runtime_get_sync(), and resumes the device itself through pm_runtime_resume() when it is in D3cold, only for the handler to return zero immediately afterwards. Such a spurious wakeup wastes power and adds needless resume latency.

The sysfs core already clamps accesses against the attribute size set through the bin_size() callback, which reports either 256 or 4096 bytes. As such, the affected accesses are reads at offset 64 (or 128 for CardBus devices) through files opened without CAP_SYS_ADMIN, and reads and writes at the exact configuration space size on devices where a quirk sets a non-standard size.

Quoted 6 lines

Therefore, reject accesses at the boundary offset as well, so that they return early before any runtime PM involvement, matching the procfs implementations in proc_bus_pci_read() and proc_bus_pci_write().

The value returned to userspace at these offsets remains zero, so the change is not visible to userspace.

Applied to the sysfs branch.

Thank you!

Krzysztof