[+ kvmarm]
Hi Heyi,
On Sun, 05 May 2019 03:26:18 +0100,
Heyi Guo <guoheyi@xxxxxxxxxx> wrote:
Hi folks,
We observed deadlocks after enabling GICv4 and PCI passthrough on
ARM64 virtual machines, when not pinning VCPU to physical CPU.
We observed below warnings after enabling lockdep debug in kernel:
[ 362.847021] =====================================================
[ 362.855643] WARNING: HARDIRQ-safe -> HARDIRQ-unsafe lock order detected
[ 362.864840] 4.19.34+ #7 Tainted: G W
[ 362.872314] -----------------------------------------------------
[ 362.881034] CPU 0/KVM/51468 [HC0[0]:SC0[0]:HE0:SE1] is trying to acquire:
[ 362.890504] 00000000659c1dc9 (fs_reclaim){+.+.}, at: fs_reclaim_acquire.part.22+0x0/0x48
[ 362.901413]
[ 362.901413] and this task is already holding:
[ 362.912976] 000000007318873f (&dev->event_map.vlpi_lock){....}, at: its_irq_set_vcpu_affinity+0x134/0x638
[ 362.928626] which would create a new lock dependency:
[ 362.936837] (&dev->event_map.vlpi_lock){....} -> (fs_reclaim){+.+.}
[ 362.946449]
[ 362.946449] but this new dependency connects a HARDIRQ-irq-safe lock:
[ 362.960877] (&irq_desc_lock_class){-.-.}
[ 362.960880]
[ 362.960880] ... which became HARDIRQ-irq-safe at:
[ 362.981234] lock_acquire+0xf0/0x258
[ 362.988337] _raw_spin_lock+0x54/0x90
[ 362.995543] handle_fasteoi_irq+0x2c/0x198
[ 363.003205] generic_handle_irq+0x34/0x50
[ 363.010787] __handle_domain_irq+0x68/0xc0
[ 363.018500] gic_handle_irq+0xf4/0x1e0
[ 363.025913] el1_irq+0xc8/0x180
[ 363.032683] _raw_spin_unlock_irq+0x40/0x60
[ 363.040512] finish_task_switch+0x98/0x258
[ 363.048254] __schedule+0x350/0xca8
[ 363.055359] schedule+0x40/0xa8
[ 363.062098] worker_thread+0xd8/0x410
[ 363.069340] kthread+0x134/0x138
[ 363.076070] ret_from_fork+0x10/0x18
[ 363.083111]
[ 363.083111] to a HARDIRQ-irq-unsafe lock:
[ 363.095213] (fs_reclaim){+.+.}
[ 363.095216]
[ 363.095216] ... which became HARDIRQ-irq-unsafe at:
[ 363.114527] ...
[ 363.114530] lock_acquire+0xf0/0x258
[ 363.126269] fs_reclaim_acquire.part.22+0x3c/0x48
[ 363.134206] fs_reclaim_acquire+0x2c/0x38
[ 363.141363] kmem_cache_alloc_trace+0x44/0x368
[ 363.148892] acpi_os_map_iomem+0x9c/0x208
[ 363.155934] acpi_os_map_memory+0x28/0x38
[ 363.162831] acpi_tb_acquire_table+0x58/0x8c
[ 363.170021] acpi_tb_validate_table+0x34/0x58
[ 363.177162] acpi_tb_get_table+0x4c/0x90
[ 363.183741] acpi_get_table+0x94/0xc4
[ 363.190020] find_acpi_cpu_topology_tag+0x54/0x240
[ 363.197404] find_acpi_cpu_topology_package+0x28/0x38
[ 363.204985] init_cpu_topology+0xdc/0x1e4
[ 363.211498] smp_prepare_cpus+0x2c/0x108
[ 363.217882] kernel_init_freeable+0x130/0x508
[ 363.224699] kernel_init+0x18/0x118
[ 363.230624] ret_from_fork+0x10/0x18
[ 363.236611]
[ 363.236611] other info that might help us debug this:
[ 363.236611]
[ 363.251604] Chain exists of:
[ 363.251604] &irq_desc_lock_class --> &dev->event_map.vlpi_lock --> fs_reclaim
[ 363.251604]
[ 363.270508] Possible interrupt unsafe locking scenario:
[ 363.270508]
[ 363.282238] CPU0 CPU1
[ 363.289228] ---- ----
[ 363.296189] lock(fs_reclaim);
[ 363.301726] local_irq_disable();
[ 363.310122] lock(&irq_desc_lock_class);
[ 363.319143] lock(&dev->event_map.vlpi_lock);
[ 363.328617] <Interrupt>
[ 363.333713] lock(&irq_desc_lock_class);
[ 363.340414]
[ 363.340414] *** DEADLOCK ***
[ 363.340414]
[ 363.353682] 5 locks held by CPU 0/KVM/51468:
[ 363.360412] #0: 00000000eeb852a5 (&vdev->igate){+.+.}, at: vfio_pci_ioctl+0x2f8/0xed0
[ 363.370915] #1: 000000002ab491f7 (lock#9){+.+.}, at: irq_bypass_register_producer+0x6c/0x1d0
[ 363.382139] #2: 000000000d9fd5c6 (&its->its_lock){+.+.}, at: kvm_vgic_v4_set_forwarding+0xd0/0x188
[ 363.396625] #3: 00000000232bdc47 (&irq_desc_lock_class){-.-.}, at: __irq_get_desc_lock+0x60/0xa0
[ 363.408486] #4: 000000007318873f (&dev->event_map.vlpi_lock){....}, at: its_irq_set_vcpu_affinity+0x134/0x638
Then we found that irq_set_vcpu_affinity() in kernel/irq/manage.c
aquires an antomic context by irq_get_desc_lock() at the beginning,
but in its_irq_set_vcpu_affinity()
(drivers/irqchip/irq-gic-v3-its.c) we are still using mutext_lock,
kcalloc, kfree, etc, which we think should be forbidden in atomic
context.
Though the issue is observed in 4.19.34, we don't find any related
fixes in the mainline yet.
Thanks for the report. Given that you're the only users of GICv4,
you're bound to find a number of these issues.
Can you try the patch below and let me know whether it helps? This is
the simplest thing I can think off to paper over the issue, but is
isn't pretty, and I'm looking at possible alternatives (ideally, we'd
be able to allocate the map outside of the irqdesc lock, but this
requires some API change between KVM, the GICv4 layer and the ITS
code).
Note that I'm travelling for the next two weeks without access to my
test rig, so I'm relying on you to test this stuff.
Thanks,
M.
diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
index 7577755bdcf4..18aa04b6a9f4 100644
--- a/drivers/irqchip/irq-gic-v3-its.c
+++ b/drivers/irqchip/irq-gic-v3-its.c
@@ -142,7 +142,7 @@ struct event_lpi_map {
u16 *col_map;
irq_hw_number_t lpi_base;
int nr_lpis;
- struct mutex vlpi_lock;
+ raw_spinlock_t vlpi_lock;
struct its_vm *vm;
struct its_vlpi_map *vlpi_maps;
int nr_vlpis;
@@ -1263,13 +1263,13 @@ static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
if (!info->map)
return -EINVAL;
- mutex_lock(&its_dev->event_map.vlpi_lock);
+ raw_spin_lock(&its_dev->event_map.vlpi_lock);
if (!its_dev->event_map.vm) {
struct its_vlpi_map *maps;
maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
- GFP_KERNEL);
+ GFP_ATOMIC);
if (!maps) {
ret = -ENOMEM;
goto out;
@@ -1312,7 +1312,7 @@ static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
}
out:
- mutex_unlock(&its_dev->event_map.vlpi_lock);
+ raw_spin_unlock(&its_dev->event_map.vlpi_lock);
return ret;
}
@@ -1322,7 +1322,7 @@ static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
u32 event = its_get_event_id(d);
int ret = 0;
- mutex_lock(&its_dev->event_map.vlpi_lock);
+ raw_spin_lock(&its_dev->event_map.vlpi_lock);
if (!its_dev->event_map.vm ||
!its_dev->event_map.vlpi_maps[event].vm) {
@@ -1334,7 +1334,7 @@ static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
*info->map = its_dev->event_map.vlpi_maps[event];
out:
- mutex_unlock(&its_dev->event_map.vlpi_lock);
+ raw_spin_unlock(&its_dev->event_map.vlpi_lock);
return ret;
}
@@ -1344,7 +1344,7 @@ static int its_vlpi_unmap(struct irq_data *d)
u32 event = its_get_event_id(d);
int ret = 0;
- mutex_lock(&its_dev->event_map.vlpi_lock);
+ raw_spin_lock(&its_dev->event_map.vlpi_lock);
if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
ret = -EINVAL;
@@ -1374,7 +1374,7 @@ static int its_vlpi_unmap(struct irq_data *d)
}
out:
- mutex_unlock(&its_dev->event_map.vlpi_lock);
+ raw_spin_unlock(&its_dev->event_map.vlpi_lock);
return ret;
}
@@ -2436,7 +2436,7 @@ static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
dev->event_map.col_map = col_map;
dev->event_map.lpi_base = lpi_base;
dev->event_map.nr_lpis = nr_lpis;
- mutex_init(&dev->event_map.vlpi_lock);
+ raw_spin_lock_init(&dev->event_map.vlpi_lock);
dev->device_id = dev_id;
INIT_LIST_HEAD(&dev->entry);
