The pasid_lock is used to synchronize different threads from modifying a
same pasid directory entry at the same time. It causes below lockdep splat.
[ 83.296538] ========================================================
[ 83.296538] WARNING: possible irq lock inversion dependency detected
[ 83.296539] 5.12.0-rc3+ #25 Tainted: G W
[ 83.296539] --------------------------------------------------------
[ 83.296540] bash/780 just changed the state of lock:
[ 83.296540] ffffffff82b29c98 (device_domain_lock){..-.}-{2:2}, at:
iommu_flush_dev_iotlb.part.0+0x32/0x110
[ 83.296547] but this lock took another, SOFTIRQ-unsafe lock in the past:
[ 83.296547] (pasid_lock){+.+.}-{2:2}
[ 83.296548]
and interrupts could create inverse lock ordering between them.
[ 83.296549] other info that might help us debug this:
[ 83.296549] Chain exists of:
device_domain_lock --> &iommu->lock --> pasid_lock
[ 83.296551] Possible interrupt unsafe locking scenario:
[ 83.296551] CPU0 CPU1
[ 83.296552] ---- ----
[ 83.296552] lock(pasid_lock);
[ 83.296553] local_irq_disable();
[ 83.296553] lock(device_domain_lock);
[ 83.296554] lock(&iommu->lock);
[ 83.296554] <Interrupt>
[ 83.296554] lock(device_domain_lock);
[ 83.296555]
*** DEADLOCK ***
Fix it by replacing the pasid_lock with an atomic exchange operation.
Reported-and-tested-by: Dave Jiang <dave.ji...@intel.com>
Signed-off-by: Lu Baolu <baolu...@linux.intel.com>
---
drivers/iommu/intel/pasid.c | 21 +++++++++++++--------
1 file changed, 13 insertions(+), 8 deletions(-)
Log:
v1->v2:
- v1:
https://lore.kernel.org/linux-iommu/20210317005834.173503-1-baolu...@linux.intel.com/
- Use retry to make code neat;
- Add a comment about no clear case, hence no race.
diff --git a/drivers/iommu/intel/pasid.c b/drivers/iommu/intel/pasid.c
index 7a73385edcc0..f2c747e62c6a 100644
--- a/drivers/iommu/intel/pasid.c
+++ b/drivers/iommu/intel/pasid.c
@@ -24,7 +24,6 @@
/*
* Intel IOMMU system wide PASID name space:
*/
-static DEFINE_SPINLOCK(pasid_lock);
u32 intel_pasid_max_id = PASID_MAX;
int vcmd_alloc_pasid(struct intel_iommu *iommu, u32 *pasid)
@@ -259,19 +258,25 @@ struct pasid_entry *intel_pasid_get_entry(struct device
*dev, u32 pasid)
dir_index = pasid >> PASID_PDE_SHIFT;
index = pasid & PASID_PTE_MASK;
- spin_lock(&pasid_lock);
+retry:
entries = get_pasid_table_from_pde(&dir[dir_index]);
if (!entries) {
entries = alloc_pgtable_page(info->iommu->node);
- if (!entries) {
- spin_unlock(&pasid_lock);
+ if (!entries)
return NULL;
- }
- WRITE_ONCE(dir[dir_index].val,
- (u64)virt_to_phys(entries) | PASID_PTE_PRESENT);
+ /*
+ * The pasid directory table entry won't be freed after
+ * allocation. No worry about the race with free and
+ * clear. However, this entry might be populated by others
+ * while we are preparing it. Use theirs with a retry.
+ */
+ if (cmpxchg64(&dir[dir_index].val, 0ULL,
+ (u64)virt_to_phys(entries) | PASID_PTE_PRESENT)) {
+ free_pgtable_page(entries);
+ goto retry;
+ }
}
- spin_unlock(&pasid_lock);
return &entries[index];
}
--
2.25.1
