vtd_find_add_as can be called by multiple threads which leads to a race condition. Taking the IOMMU lock ensures we avoid such a race. Moreover we also need to take the bql to avoid an assert to fail in memory_region_add_subregion_overlap when actually allocating a new address space.
Signed-off-by: Clement Mathieu--Drif <clement.mathieu--d...@eviden.com> --- hw/i386/intel_iommu.c | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/hw/i386/intel_iommu.c b/hw/i386/intel_iommu.c index dad1d9f300..144e25622a 100644 --- a/hw/i386/intel_iommu.c +++ b/hw/i386/intel_iommu.c @@ -4205,9 +4205,30 @@ VTDAddressSpace *vtd_find_add_as(IntelIOMMUState *s, PCIBus *bus, VTDAddressSpace *vtd_dev_as; char name[128]; + vtd_iommu_lock(s); vtd_dev_as = g_hash_table_lookup(s->vtd_address_spaces, &key); + vtd_iommu_unlock(s); + if (!vtd_dev_as) { - struct vtd_as_key *new_key = g_malloc(sizeof(*new_key)); + struct vtd_as_key *new_key; + /* Slow path */ + + /* + * memory_region_add_subregion_overlap requires the bql, + * make sure we own it. + */ + BQL_LOCK_GUARD(); + vtd_iommu_lock(s); + + /* Check again as we released the lock for a moment */ + vtd_dev_as = g_hash_table_lookup(s->vtd_address_spaces, &key); + if (vtd_dev_as) { + vtd_iommu_unlock(s); + return vtd_dev_as; + } + + /* Still nothing, allocate a new address space */ + new_key = g_malloc(sizeof(*new_key)); new_key->bus = bus; new_key->devfn = devfn; @@ -4298,6 +4319,8 @@ VTDAddressSpace *vtd_find_add_as(IntelIOMMUState *s, PCIBus *bus, vtd_switch_address_space(vtd_dev_as); g_hash_table_insert(s->vtd_address_spaces, new_key, vtd_dev_as); + + vtd_iommu_unlock(s); } return vtd_dev_as; } -- 2.49.0