iommu/vt-d: Fix lockdep splat in intel_pasid_get_entry()

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.jiang@intel.com>
Signed-off-by: Lu Baolu <baolu.lu@linux.intel.com>
Link: https://lore.kernel.org/r/20210320020916.640115-1-baolu.lu@linux.intel.com
Signed-off-by: Joerg Roedel <jroedel@suse.de>

diff --git a/drivers/iommu/intel/pasid.c b/drivers/iommu/intel/pasid.c
index 0bf7e0a..1a4105a 100644 (file)
--- 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];
 }