iommu/vt-d: Fix possible recursive locking in intel_iommu_init()
Subsystems (2)
Lu Baolu
maintainerPATCH 1/1 iommu/vt-d: Fix possible recursive locking in intel_iommu_init()
The global rwsem dmar_global_lock was introduced by commit 3a5670e8ac932 ("iommu/vt-d: Introduce a rwsem to protect global data structures"). It is used to protect DMAR related global data from DMAR and PCI devices hotplug operations.
The dmar_global_lock used in the intel_iommu_init() might cause recursive locking issue, where intel_iommu_get_resv_regions() is taking the dmar_global_lock from within a section where intel_iommu_init() already holds it via probe_acpi_namespace_devices().
Using dmar_global_lock in intel_iommu_init() is unnecessary since it is unlikely that any IO board must be hot added during the early boot stage. This fixes the possible recursive locking issue by moving down DMAR and PCI devices hotplug support after the IOMMU is initialized and removing the dmar_global_lock uses in intel_iommu_init().
--- include/linux/dmar.h | 4 +++- drivers/iommu/intel/dmar.c | 7 +++++++ drivers/iommu/intel/iommu.c | 27 ++------------------------- 3 files changed, 12 insertions(+), 26 deletions(-)
include/linux/dmar.h +3 -1
@@ -65,6 +65,7 @@ struct dmar_pci_notify_info {extern struct rw_semaphore dmar_global_lock;extern struct list_head dmar_drhd_units; +extern int intel_iommu_enabled;#define for_each_drhd_unit (drhd) \ list_for_each_entry_rcu(drhd, &dmar_drhd_units, list, \
@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units;static inline bool dmar_rcu_check (void ) {return rwsem_is_locked (&dmar_global_lock) || - system_state ==SYSTEM_BOOTING ; + system_state ==SYSTEM_BOOTING || + (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled); }#define dmar_rcu_dereference (p) rcu_dereference_check((p), dmar_rcu_check())
drivers/iommu/intel/dmar.c +7 -0
@@ -2349,6 +2349,13 @@ static int dmar_device_hotplug(acpi_handle handle, bool insert)if (!dmar_in_use ())return 0 ; +/* +* It's unlikely that any I/O board is hot added before the IOMMU +* subsystem is initialized. +*/ +if (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled) +return -EOPNOTSUPP ; +if (dmar_detect_dsm (handle,DMAR_DSM_FUNC_DRHD )) { tmp = handle; }else {
drivers/iommu/intel/iommu.c +2 -25
@@ -3013,13 +3013,7 @@ static int __init init_dmars(void)#ifdef CONFIG_INTEL_IOMMU_SVM if (pasid_supported (iommu) &&ecap_prs (iommu->ecap)) { -/* -* Call dmar_alloc_hwirq() with dmar_global_lock held, -* could cause possible lock race condition. -*/ -up_write (&dmar_global_lock); ret =intel_svm_enable_prq (iommu); -down_write (&dmar_global_lock);if (ret) goto free_iommu; }
@@ -3932,7 +3926,6 @@ int __init intel_iommu_init(void) force_on = (!intel_iommu_tboot_noforce &&tboot_force_iommu ()) ||platform_optin_force_iommu (); -down_write (&dmar_global_lock);if (dmar_table_init ()) {if (force_on)panic ("tboot: Failed to initialize DMAR table\n" );
@@ -3945,16 +3938,6 @@ int __init intel_iommu_init(void) goto out_free_dmar; } -up_write (&dmar_global_lock); - -/* -* The bus notifier takes the dmar_global_lock, so lockdep will -* complain later when we register it under the lock. -*/ -dmar_register_bus_notifier (); - -down_write (&dmar_global_lock); -if (!no_iommu)intel_iommu_debugfs_init ();
@@ -3999,11 +3982,9 @@ int __init intel_iommu_init(void)pr_err ("Initialization failed\n" ); goto out_free_dmar; } -up_write (&dmar_global_lock);init_iommu_pm_ops (); -down_read (&dmar_global_lock);for_each_active_iommu (iommu, drhd) {/* * The flush queue implementation does not perform
@@ -4021,13 +4002,11 @@ int __init intel_iommu_init(void) "%s", iommu->name);iommu_device_register (&iommu->iommu, &intel_iommu_ops,NULL ); } -up_read (&dmar_global_lock);bus_set_iommu (&pci_bus_type, &intel_iommu_ops);if (si_domain && !hw_pass_through)register_memory_notifier (&intel_iommu_memory_nb); -down_read (&dmar_global_lock);if (probe_acpi_namespace_devices ())pr_warn ("ACPI name space devices didn't probe correctly\n" );
@@ -4038,17 +4017,15 @@ int __init intel_iommu_init(void)iommu_disable_protect_mem_regions (iommu); } -up_read (&dmar_global_lock); - -pr_info ("Intel(R) Virtualization Technology for Directed I/O\n" ); intel_iommu_enabled =1 ; +dmar_register_bus_notifier (); +pr_info ("Intel(R) Virtualization Technology for Directed I/O\n" );return 0 ; out_free_dmar:intel_iommu_free_dmars (); -up_write (&dmar_global_lock);return ret; }
Signature · 1 line
2.25.1
include/linux/dmar.h extern struct list_head dmar_drhd_units; · line 95
Quoted 12 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +2
-1@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units;static inline bool dmar_rcu_check (void ) {return rwsem_is_locked (&dmar_global_lock) || - system_state ==SYSTEM_BOOTING ; + system_state ==SYSTEM_BOOTING || + (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled); }
intel_iommu_enabled is 0 if CONFIG_INTEL_IOMMU is not set.
same for other similar checks.
Baolu Lu
maintainerHi Kevin,
Quoted 15 lines · Kevin
include/linux/dmar.h extern struct list_head dmar_drhd_units; · line 95
Quoted 12 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +2
-1@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units;static inline bool dmar_rcu_check (void ) {return rwsem_is_locked (&dmar_global_lock) || - system_state ==SYSTEM_BOOTING ; + system_state ==SYSTEM_BOOTING || + (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled); }
intel_iommu_enabledis 0 ifCONFIG_INTEL_IOMMUis not set.
Quoted 1 line
same for other similar checks.
Sorry that I didn't get your point. If CONFIG_INTEL_IOMMU is not set, IS_ENABLED(CONFIG_INTEL_IOMMU) is 0. The adding check has no effect. Did I miss anything?
Best regards,
baolu
Quoted 21 lines
From: Baolu Lu <[email protected]>
Sent: Sunday, July 24, 2022 11:00 AMHi Kevin,
Quoted 15 lines · Kevin
include/linux/dmar.h extern struct list_head dmar_drhd_units; · line 95
Quoted 12 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +2
-1@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units;static inline bool dmar_rcu_check (void ) {return rwsem_is_locked (&dmar_global_lock) || - system_state ==SYSTEM_BOOTING ; + system_state ==SYSTEM_BOOTING || + (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled); }
intel_iommu_enabledis 0 ifCONFIG_INTEL_IOMMUis not set.
Quoted 7 lines
Quoted 1 line
same for other similar checks.
Sorry that I didn't get your point. If
CONFIG_INTEL_IOMMUis not set,IS_ENABLED(CONFIG_INTEL_IOMMU)is 0. The adding check has no effect. Did
I miss anything?
My point was that the check on CONFIG_INTEL_IOMMU is unnecessary.
Baolu Lu
maintainerQuoted 23 lines · Kevin
Quoted 23 lines
From: Baolu Lu <[email protected]>
Sent: Sunday, July 24, 2022 11:00 AMHi Kevin,
Quoted 16 lines · Kevin
include/linux/dmar.h extern struct list_head dmar_drhd_units; · line 95
Quoted 12 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +2
-1@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units;static inline bool dmar_rcu_check (void ) {return rwsem_is_locked (&dmar_global_lock) || - system_state ==SYSTEM_BOOTING ; + system_state ==SYSTEM_BOOTING || + (IS_ENABLED (CONFIG_INTEL_IOMMU ) && !intel_iommu_enabled); }
intel_iommu_enabledis 0 ifCONFIG_INTEL_IOMMUis not set.same for other similar checks.
Quoted 7 lines
Quoted 5 lines
Sorry that I didn't get your point. If
CONFIG_INTEL_IOMMUis not set,IS_ENABLED(CONFIG_INTEL_IOMMU)is 0. The adding check has no effect. Did
I miss anything?My point was that the check on
CONFIG_INTEL_IOMMUis unnecessary.
Oh, if INTEL_IOMMU is not configured, the interrupt remapping could also be supported, so we still need the rcu protection. We only relax the rcu check when INTEL_IOMMU is configured, but not enabled yet.
Best regards,
baolu
Baolu Lu
maintainerHi Kevin,
Quoted 30 lines · Baolu Lu
Quoted 29 lines · Kevin
Quoted 28 lines
From: Baolu Lu <[email protected]>
Sent: Sunday, July 24, 2022 11:00 AMHi Kevin,
Quoted 15 lines · Kevin
Quoted 6 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +0
-0@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units; static inline bool dmar_rcu_check(void) {Quoted 6 lines
return
rwsem_is_locked(&dmar_global_lock)|| -system_state==SYSTEM_BOOTING; +system_state==SYSTEM_BOOTING|| + (IS_ENABLED(CONFIG_INTEL_IOMMU)&& !intel_iommu_enabled);
}
intel_iommu_enabledis 0 ifCONFIG_INTEL_IOMMUis not set.Quoted 1 line
same for other similar checks.
Sorry that I didn't get your point. If
CONFIG_INTEL_IOMMUis not set,IS_ENABLED(CONFIG_INTEL_IOMMU)is 0. The adding check has no effect. Did
I miss anything?
Quoted 5 lines
Quoted 1 line
My point was that the check on
CONFIG_INTEL_IOMMUis unnecessary.Oh, if
INTEL_IOMMUis not configured, the interrupt remapping could also be supported, so we still need the rcu protection. We only relax the rcu check whenINTEL_IOMMUis configured, but not enabled yet.
The next stepping, we will tie INTEL_IOMMU and VT-d IRQ_REMAPPING together, that will make the VT-d software simpler.
Joerg also proposed this in another discussion thread:
https://lore.kernel.org/all/YrVPelnOi9nql%[email protected]/
After that, we have no need to add above check anymore. As this is a quick fix for lockdep splat, we don't need to wait until that done.
Does this work for you?
Best regards,
baolu
Tian, Kevin
Reviewed-byQuoted 46 lines
From: Baolu Lu <[email protected]>
Sent: Thursday, September 8, 2022 2:13 PMHi Kevin,
Quoted 22 lines · Baolu Lu
Quoted 21 lines · Kevin
Quoted 21 lines
From: Baolu Lu <[email protected]>
Sent: Sunday, July 24, 2022 11:00 AMHi Kevin,
Quoted 15 lines · Kevin
Quoted 6 lines
From: Lu Baolu <[email protected]>
Sent: Tuesday, July 19, 2022 7:53 AMinclude/linux/dmar.h +0
-0@@ -88,7 +89,8 @@ extern struct list_head dmar_drhd_units; static inline bool dmar_rcu_check(void) {Quoted 6 lines
return
rwsem_is_locked(&dmar_global_lock)|| -system_state==SYSTEM_BOOTING; +system_state==SYSTEM_BOOTING|| + (IS_ENABLED(CONFIG_INTEL_IOMMU)&& !intel_iommu_enabled);
}
intel_iommu_enabledis 0 ifCONFIG_INTEL_IOMMUis not set.Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line
same for other similar checks.
Sorry that I didn't get your point. If
CONFIG_INTEL_IOMMUis not set,IS_ENABLED(CONFIG_INTEL_IOMMU)is 0. The adding check has noeffect.
Quoted 9 lines
Quoted 5 lines
Quoted 3 lines
Did
I miss anything?My point was that the check on
CONFIG_INTEL_IOMMUis unnecessary.Oh, if
INTEL_IOMMUis not configured, the interrupt remapping could also be supported, so we still need the rcu protection. We only relax the rcu check whenINTEL_IOMMUis configured, but not enabled yet.The next stepping, we will tie
INTEL_IOMMUand VT-dIRQ_REMAPPINGtogether, that will make the VT-d software simpler.
Quoted 9 lines
Joerg also proposed this in another discussion thread:
https://lore.kernel.org/all/YrVPelnOi9nql%[email protected]/
After that, we have no need to add above check anymore. As this is a quick fix for lockdep splat, we don't need to wait until that done.
Does this work for you?
Yes. With above I'm fine with it given Robin's series has been merged.