@@ -119,3 +119,6 @@ config PCI_IOAPIC
config PCI_LABEL
def_bool y if (DMI || ACPI)
select NLS
+
+config PCI_BUS_LOCK
+ def_bool n
@@ -15,9 +15,12 @@
#include <linux/proc_fs.h>
#include <linux/init.h>
#include <linux/slab.h>
+#include <linux/sched.h>
#include "pci.h"
+static DECLARE_WAIT_QUEUE_HEAD(pci_bus_state_wait_queue);
+
void pci_add_resource_offset(struct list_head *resources, struct resource *res,
resource_size_t offset)
{
@@ -297,6 +300,254 @@ void pci_bus_put(struct pci_bus *bus)
}
EXPORT_SYMBOL(pci_bus_put);
+/**
+ * pci_bus_change_state - Change state of @bus from @old_state to @new_state
+ * @bus: the PCI bus to change state
+ * @old_state: old state of @bus
+ * @new_state: new state for @bus
+ * @unlock: unlock @bus if true
+ *
+ * Change state of @bus from @old_state to @new_state, and unlock @bus
+ * if @unlock is true.
+ * Note:
+ * - @new_state must be bigger than @old_state
+ * - must be called with @bus locked.
+ */
+void pci_bus_change_state(struct pci_bus *bus, int old_state, int new_state,
+ bool unlock)
+{
+ int t;
+
+ BUG_ON(!pci_bus_is_locked(bus));
+ BUG_ON(new_state < old_state || pci_bus_get_state(bus) != old_state ||
+ (new_state & ~PCI_BUS_STATE_MASK));
+
+ old_state |= PCI_BUS_STATE_LOCK;
+ if (!unlock)
+ new_state |= PCI_BUS_STATE_LOCK;
+
+ do {
+ t = atomic_read(&bus->state);
+ t &= ~(PCI_BUS_STATE_MASK | PCI_BUS_STATE_LOCK);
+ } while (atomic_cmpxchg(&bus->state, t | old_state, t | new_state)
+ != (t | old_state));
+
+ wake_up_all(&pci_bus_state_wait_queue);
+}
+
+/* Wait for bus to reach one of the specified states. */
+static bool pci_bus_wait_for_states(struct pci_bus *bus, int states)
+{
+ int t = atomic_read(&bus->state);
+
+ /* Bus state is bigger than any of the specified states. */
+ if ((t & PCI_BUS_STATE_MASK) > states)
+ return true;
+
+ /* Bus is in one of the specified states and unlocked. */
+ if ((t & states) && !(t & PCI_BUS_STATE_LOCK))
+ return true;
+
+ return false;
+}
+
+static int pci_bus_lock_one(struct pci_bus *bus, int states, long *to)
+{
+ int t;
+
+ BUG_ON(states & ~PCI_BUS_STATE_MASK);
+ do {
+ do {
+ if (*to <= 0)
+ return -ETIMEDOUT;
+ *to = wait_event_timeout(pci_bus_state_wait_queue,
+ pci_bus_wait_for_states(bus, states), *to);
+ t = atomic_read(&bus->state);
+ if (((t & PCI_BUS_STATE_MASK) > states))
+ return -EBUSY;
+ } while (!(t & states));
+
+ t &= ~PCI_BUS_STATE_LOCK;
+ } while (atomic_cmpxchg(&bus->state, t , t | PCI_BUS_STATE_LOCK) != t);
+
+ return 0;
+}
+
+static int ___pci_bus_lock(struct pci_bus *bus, int states, bool recursive,
+ long *to, struct pci_bus **marker)
+{
+ int ret;
+ struct pci_bus *child;
+
+ ret = pci_bus_lock_one(bus, states, to);
+ if (ret < 0) {
+ *marker = bus;
+ } else if (recursive) {
+ list_for_each_entry(child, &bus->children, node) {
+ ret = ___pci_bus_lock(child, states, recursive,
+ to, marker);
+ if (ret < 0)
+ break;
+ }
+ }
+
+ return ret;
+}
+
+static int __pci_bus_unlock(struct pci_bus *bus, bool recursive,
+ struct pci_bus *mark)
+{
+ int t, ret = 0;
+ struct pci_bus *child;
+
+ if (unlikely(bus == mark))
+ return 1;
+
+ if (recursive)
+ list_for_each_entry(child, &bus->children, node) {
+ ret = __pci_bus_unlock(child, recursive, mark);
+ if (ret)
+ break;
+ }
+
+ BUG_ON(!pci_bus_is_locked(bus));
+ do {
+ t = atomic_read(&bus->state);
+ } while (atomic_cmpxchg(&bus->state, t, t & ~PCI_BUS_STATE_LOCK) != t);
+
+ return ret;
+}
+
+static int __pci_bus_lock(struct pci_bus *bus, int states, bool recursive,
+ long *to)
+{
+ int ret;
+ struct pci_bus *mark = NULL;
+
+ ret = ___pci_bus_lock(bus, states, recursive, to, &mark);
+ if (ret < 0) {
+ __pci_bus_unlock(bus, recursive, mark);
+ wake_up_all(&pci_bus_state_wait_queue);
+ }
+
+ return ret;
+}
+
+/**
+ * pci_bus_lock - wait for @bus to reach one of the specified states
+ * and then lock it.
+ * @bus: the PCI bus to lock
+ * @states: specified states to wait for
+ * @recursive: lock all the subtree if true
+ *
+ * Wait for @bus to reach one of the states specified by @states and then
+ * lock it. Resursively lock all the subtree if @recursive is true.
+ *
+ * Return values:
+ * - 0 on success
+ * - -EBUSY if failed to lock the bus
+ */
+int pci_bus_lock(struct pci_bus *bus, int states, bool recursive)
+{
+ int ret;
+ long to;
+
+ /*
+ * When recursively locking the sub-tree starting from @bus,
+ * __pci_bus_lock() may return -EBUSY if some child buses are under
+ * hot-removing, so keep retrying until success or @bus is out of
+ * requested @states.
+ */
+ do {
+ to = HZ / 10;
+ ret = __pci_bus_lock(bus, states, recursive, &to);
+ if (ret == 0)
+ return 0;
+
+ /*
+ * Prevent a deadlock scenario that thread A waits for
+ * all sysfs files to be released while holding PCI bus
+ * locks, and Thread B tries to acquire PCI bus locks
+ * in a sysfs handler. These checks break the deadlock
+ * condition.
+ */
+ if (pci_bus_get_state(bus) > states)
+ return -EBUSY;
+ if (to > 0)
+ schedule_timeout_interruptible(to);
+ } while (true);
+}
+EXPORT_SYMBOL(pci_bus_lock);
+
+/**
+ * pci_bus_lock_timeout - wait for @bus to reach one of the
+ * specified states and then lock it.
+ * @bus: the PCI bus to lock
+ * @states: specified states to wait for
+ * @recursive: lock all the subtree if true
+ * @timeout: maximum number of jiffies to wait
+ *
+ * Wait for @bus to reach one of the states specified by @states and then
+ * lock it. Resursively lock all the subtree if @recursive is true.
+ *
+ * Return values:
+ * - 0 on success
+ * - -ETIMEDOUT if timeouts
+ * - -EBUSY if failed to lock the bus
+ */
+int pci_bus_lock_timeout(struct pci_bus *bus, int states, bool recursive,
+ long timeout)
+{
+ int ret;
+ long to;
+
+ if (timeout <= 0)
+ return -ETIMEDOUT;
+
+ /*
+ * When recursively locking the sub-tree starting from @bus,
+ * __pci_bus_lock() may return -EBUSY if some child buses are under
+ * hot-removing, so keep retrying until success or @bus is out of
+ * requested @states.
+ */
+ do {
+ to = (timeout > HZ / 10) ? (HZ / 10) : timeout;
+ timeout -= to;
+ ret = __pci_bus_lock(bus, states, recursive, &to);
+ if (ret == 0)
+ return 0;
+
+ /*
+ * Prevent a deadlock scenario that thread A waits for
+ * all sysfs files to be released while holding PCI bus
+ * locks, and Thread B tries to acquire PCI bus locks
+ * in a sysfs handler. These checks break the deadlock
+ * condition.
+ */
+ if (pci_bus_get_state(bus) > states)
+ return -EBUSY;
+ if (to > 0)
+ schedule_timeout_interruptible(to);
+ } while (timeout > 0);
+
+ return -ETIMEDOUT;
+}
+EXPORT_SYMBOL(pci_bus_lock_timeout);
+
+/**
+ * pci_bus_unlock - unlock @bus and wake up waiters
+ * bus: PCI bus to unlock
+ * recursive: unlock subtree if recursive is true
+ *
+ * Unlock PCI bus @bus and wake up waiters. Must be called with @bus locked.
+ */
+void pci_bus_unlock(struct pci_bus *bus, bool recursive)
+{
+ __pci_bus_unlock(bus, recursive, NULL);
+ wake_up_all(&pci_bus_state_wait_queue);
+}
+EXPORT_SYMBOL(pci_bus_unlock);
+
EXPORT_SYMBOL(pci_bus_alloc_resource);
EXPORT_SYMBOL_GPL(pci_bus_add_device);
EXPORT_SYMBOL(pci_bus_add_devices);
@@ -450,8 +450,55 @@ struct pci_bus {
struct bin_attribute *legacy_io; /* legacy I/O for this bus */
struct bin_attribute *legacy_mem; /* legacy mem */
unsigned int is_added:1;
+ atomic_t state;
};
+/*
+ * PCI bus state machine:
+ * _________________________________________________________________
+ * | v
+ * INITIALIZED->REGISTERED->STARTED->STOPPING->STOPPED->REMOVING->REMOVED->DESTOYED
+ * |_____________________________________^
+ */
+#define PCI_BUS_STATE_UNKNOWN 0x0 /* invalid state */
+#define PCI_BUS_STATE_INITIALIZED 0x1 /* bus initialized */
+#define PCI_BUS_STATE_REGISTERED 0x2 /* bus added into device tree */
+#define PCI_BUS_STATE_STARTED 0x4 /* working state */
+#define PCI_BUS_STATE_STOPPING 0x8 /* start to stop children */
+#define PCI_BUS_STATE_STOPPED 0x10 /* child devices stopped */
+#define PCI_BUS_STATE_REMOVING 0x20 /* start to remove PCI */
+#define PCI_BUS_STATE_REMOVED 0x40 /* bus removed */
+#define PCI_BUS_STATE_DESTROYED 0x80 /* used for debug only */
+#define PCI_BUS_STATE_MASK 0xFF
+
+#ifdef CONFIG_PCI_BUS_LOCK
+#define PCI_BUS_STATE_LOCK 0x10000 /* for pci core only */
+
+static inline bool pci_bus_is_locked(const struct pci_bus *bus)
+{
+ return !!(atomic_read(&bus->state) & PCI_BUS_STATE_LOCK);
+}
+#else /* CONFIG_PCI_BUS_LOCK */
+#define PCI_BUS_STATE_LOCK 0x00000 /* for pci core only */
+
+static inline bool pci_bus_is_locked(const struct pci_bus *bus)
+{
+ return true;
+}
+#endif /* CONFIG_PCI_BUS_LOCK */
+
+static inline int pci_bus_get_state(const struct pci_bus *bus)
+{
+ return atomic_read(&bus->state) & PCI_BUS_STATE_MASK;
+}
+
+void pci_bus_change_state(struct pci_bus *bus, int new_state,
+ int old_state, bool unlock);
+int pci_bus_lock(struct pci_bus *bus, int states, bool recursive);
+int pci_bus_lock_timeout(struct pci_bus *bus, int states,
+ bool recursive, long timeout);
+void pci_bus_unlock(struct pci_bus *bus, bool recursive);
+
#define pci_bus_b(n) list_entry(n, struct pci_bus, node)
#define to_pci_bus(n) container_of(n, struct pci_bus, dev)
#define for_each_pci_bus(b) for (b = NULL; (b = pci_get_next_bus(b)); )
There are multiple ways to trigger concurrent PCI hotplug operations for a specific PCI bus, but we have no way to serialize those PCI hotplug operations yet and thus it may break the PCI hotplug logic. This patch introduces a lock mechanism and a state machine for PCI bus to serialize PCI hotplug operations. The state machine is as below: _________________________________________________________________ | v INITIALIZED->REGISTERED->STARTED->STOPPING->STOPPED->REMOVING->REMOVED->DESTOYED |_____________________________________^ INITIALIZED: the bus structure has just been allocated REGISTERED: the device for the bus has been registered into the device tree STARTED: PCI devices under the bus has been started STOPPING: start to stop devices on the bus STOPPED: child PCI devices under the bus has been stopped REMOVING: start to remove the PCI bus REMOVED: the bus has been removed from the PCI tree DESTROYED: memory for the bus has been freed, just for debug The interfaces to manage PCI bus state are: static inline int pci_bus_get_state(const struct pci_bus *bus); extern void pci_bus_change_state(struct pci_bus *bus, int new_state, int old_state, bool unlock); And another three interfaces for PCI bus lock: extern int pci_bus_lock(struct pci_bus *bus, int states, bool recursive); extern int pci_bus_lock_timeout(struct pci_bus *bus, int states, bool recursive, long timeout); extern void pci_bus_unlock(struct pci_bus *bus, bool recursive); The locking rules are simple: 1) The bus must be locked when changing state of child devices. 2) The bus must be locked when changing state of itself. A typical usage looks like: if (pci_bus_lock(bus, PCI_BUS_STATE_STARTED, true) == 0) { do_pci_hotplug(); pci_bus_unlock(bus, true); } Note: The PCI_BUS_LOCK config option is a temporary solution to avoid breaking bisects, it will be removed when all architectures have been enhanced to support the new PCI bus state and lock mechanism. Signed-off-by: Jiang Liu <jiang.liu@huawei.com> Signed-off-by: Yijing Wang <wangyijing@huawei.com> Cc: Yinghai Lu <yinghai@kernel.org> Cc: linux-pci@vger.kernel.org Cc: linux-kernel@vger.kernel.org --- drivers/pci/Kconfig | 3 + drivers/pci/bus.c | 251 ++++++++++++++++++++++++++++++++++++++++++++++++++++ include/linux/pci.h | 47 ++++++++++ 3 files changed, 301 insertions(+)