From: Johannes Berg <johannes.berg@xxxxxxxxx>
We found that that the following race is possible if userspace
enables runtime PM/auto-suspend immediately when a device shows
up in sysfs, if there's any call to pci_rescan_bus() during
normal system state (i.e. userspace is already active):
- we rescan the PCI bus (*)
- this creates the new PCI device including its sysfs
representation
- udev sees the new device, and the (OS-specific) scripting
enables runtime PM by writing to power/control; this can
happen _before_ the next step - this will runtime-suspend
the device which saves the config space, including the BARs
that weren't assigned yet
- the bus rescan assigns resources to the devices and writes
them to the config space of the device
(but not the runtime-pm saved copy, course)
- the driver binds and this disallows runtime PM, so the device
is resumed, restoring the (incomplete!) config space
- the device cannot work due to BARs not being configured
Fix this by allowing runtime PM only once the device has been
fully added. Also, with a warning, reject runtime PM on a not-
added device; this shouldn't happen anymore now.
Note that the comment that was there (that I'm replacing) was
indicating that pci_device_add() wouldn't be called at this
place yet, but in fact it's called much earlier during the whole
scan/probe process, which in part causes this problem, but it
doesn't seem possible to defer it until here either.
(*) In the case we encountered, this happened due to some reset
of the iwlwifi device that the driver then needs to recover
from by rescanning the bus since the device was reset and
the system doesn't know about it yet.
Signed-off-by: Johannes Berg <johannes.berg@xxxxxxxxx>
---
v2: use pm_runtime_get_noresume()/pm_runtime_put_noidle()
instead as advised by Rafael
---
drivers/pci/bus.c | 8 ++++++--
drivers/pci/pci-driver.c | 3 +++
drivers/pci/pci.c | 1 +
3 files changed, 10 insertions(+), 2 deletions(-)
diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
index 5bc81cc0a2de..e06ea5449be9 100644
--- a/drivers/pci/bus.c
+++ b/drivers/pci/bus.c
@@ -13,6 +13,7 @@
#include <linux/ioport.h>
#include <linux/proc_fs.h>
#include <linux/slab.h>
+#include <linux/pm_runtime.h>
#include "pci.h"
@@ -335,9 +336,12 @@ void pci_bus_add_device(struct pci_dev *dev)
int retval;
/*
- * Can not put in pci_device_add yet because resources
- * are not assigned yet for some devices.
+ * Allow runtime PM only here, since otherwise we may
+ * try to suspend a device that isn't fully configured
+ * yet, which causes problems.
*/
+ pm_runtime_put_noidle(&dev->dev);
+
pcibios_bus_add_device(dev);
pci_fixup_device(pci_fixup_final, dev);
pci_create_sysfs_dev_files(dev);
diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c
index ae9baf801681..8d82b4abb169 100644
--- a/drivers/pci/pci-driver.c
+++ b/drivers/pci/pci-driver.c
@@ -1278,6 +1278,9 @@ static int pci_pm_runtime_suspend(struct device *dev)
pci_power_t prev = pci_dev->current_state;
int error;
+ if (WARN_ON(!pci_dev_is_added(pci_dev)))
+ return -EBUSY;
+
pci_suspend_ptm(pci_dev);
/*
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c
index 5ede93222bc1..808906ad14b9 100644
--- a/drivers/pci/pci.c
+++ b/drivers/pci/pci.c
@@ -3139,6 +3139,7 @@ void pci_pm_init(struct pci_dev *dev)
u16 pmc;
pm_runtime_forbid(&dev->dev);
+ pm_runtime_get_noresume(&dev->dev);
pm_runtime_set_active(&dev->dev);
pm_runtime_enable(&dev->dev);
device_enable_async_suspend(&dev->dev);
--
2.40.1
