diff options
| author | Srinath Mannam <srinath.mannam@broadcom.com> | 2017-08-18 21:50:48 -0500 |
|---|---|---|
| committer | Bjorn Helgaas <bhelgaas@google.com> | 2017-08-18 21:50:48 -0500 |
| commit | 40f11adc7cd9281227f0a6a627d966dd0a5f0cd9 (patch) | |
| tree | b79efdee9dbd7746d82644ee4070b16646539679 /drivers/pci | |
| parent | 62ce94a7a5a54aac80975f5e6731707225d4077e (diff) | |
| download | linux-40f11adc7cd9281227f0a6a627d966dd0a5f0cd9.tar.gz linux-40f11adc7cd9281227f0a6a627d966dd0a5f0cd9.tar.bz2 linux-40f11adc7cd9281227f0a6a627d966dd0a5f0cd9.zip | |
PCI: Avoid race while enabling upstream bridges
When we enable a device, we first enable any upstream bridges. If a bridge
has multiple downstream devices and we enable them simultaneously, the race
to enable the upstream bridge may cause problems. Consider this hierarchy:
bridge A --+-- device B
+-- device C
If drivers for B and C call pci_enable_device() simultaneously, both will
attempt to enable A, which involves setting PCI_COMMAND_MASTER via
pci_set_master() and PCI_COMMAND_MEMORY via pci_enable_resources().
In the following sequence, B's update to set A's PCI_COMMAND_MEMORY is
lost, and neither B nor C will work correctly:
B C
pci_set_master(A)
cmd = read(A, PCI_COMMAND)
cmd |= PCI_COMMAND_MASTER
pci_set_master(A)
cmd = read(A, PCI_COMMAND)
cmd |= PCI_COMMAND_MASTER
write(A, PCI_COMMAND, cmd)
pci_enable_device(A)
pci_enable_resources(A)
cmd = read(A, PCI_COMMAND)
cmd |= PCI_COMMAND_MEMORY
write(A, PCI_COMMAND, cmd)
write(A, PCI_COMMAND, cmd)
Avoid this race by holding a new pci_bridge_mutex while enabling a bridge.
This ensures that both PCI_COMMAND_MASTER and PCI_COMMAND_MEMORY will be
updated before another thread can start enabling the bridge.
Note that although pci_enable_bridge() is recursive, it enables any
upstream bridges *before* acquiring the mutex. When it acquires the mutex
and calls pci_set_master() and pci_enable_device(), any upstream bridges
have already been enabled so pci_enable_device() will not deadlock by
calling pci_enable_bridge() again.
Signed-off-by: Srinath Mannam <srinath.mannam@broadcom.com>
[bhelgaas: changelog, comment]
Signed-off-by: Bjorn Helgaas <bhelgaas@google.com>
Diffstat (limited to 'drivers/pci')
| -rw-r--r-- | drivers/pci/pci.c | 13 |
1 files changed, 11 insertions, 2 deletions
diff --git a/drivers/pci/pci.c b/drivers/pci/pci.c index af0cc3456dc1..7cb29a223b73 100644 --- a/drivers/pci/pci.c +++ b/drivers/pci/pci.c @@ -52,6 +52,7 @@ static void pci_pme_list_scan(struct work_struct *work); static LIST_HEAD(pci_pme_list); static DEFINE_MUTEX(pci_pme_list_mutex); static DECLARE_DELAYED_WORK(pci_pme_work, pci_pme_list_scan); +static DEFINE_MUTEX(pci_bridge_mutex); struct pci_pme_device { struct list_head list; @@ -1348,10 +1349,16 @@ static void pci_enable_bridge(struct pci_dev *dev) if (bridge) pci_enable_bridge(bridge); + /* + * Hold pci_bridge_mutex to prevent a race when enabling two + * devices below the bridge simultaneously. The race may cause a + * PCI_COMMAND_MEMORY update to be lost (see changelog). + */ + mutex_lock(&pci_bridge_mutex); if (pci_is_enabled(dev)) { if (!dev->is_busmaster) pci_set_master(dev); - return; + goto end; } retval = pci_enable_device(dev); @@ -1359,6 +1366,8 @@ static void pci_enable_bridge(struct pci_dev *dev) dev_err(&dev->dev, "Error enabling bridge (%d), continuing\n", retval); pci_set_master(dev); +end: + mutex_unlock(&pci_bridge_mutex); } static int pci_enable_device_flags(struct pci_dev *dev, unsigned long flags) @@ -1383,7 +1392,7 @@ static int pci_enable_device_flags(struct pci_dev *dev, unsigned long flags) return 0; /* already enabled */ bridge = pci_upstream_bridge(dev); - if (bridge) + if (bridge && !pci_is_enabled(bridge)) pci_enable_bridge(bridge); /* only skip sriov related */ |