| Message ID | 20221115225502.2240227-1-oliver.upton@linux.dev (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v2] KVM: arm64: Don't acquire RCU read lock for exclusive table walks | expand |
On Tue, 15 Nov 2022 22:55:02 +0000, Oliver Upton <oliver.upton@linux.dev> wrote: > > Marek reported a BUG resulting from the recent parallel faults changes, > as the hyp stage-1 map walker attempted to allocate table memory while > holding the RCU read lock: > > BUG: sleeping function called from invalid context at > include/linux/sched/mm.h:274 > in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 1, name: swapper/0 > preempt_count: 0, expected: 0 > RCU nest depth: 1, expected: 0 > 2 locks held by swapper/0/1: > #0: ffff80000a8a44d0 (kvm_hyp_pgd_mutex){+.+.}-{3:3}, at: > __create_hyp_mappings+0x80/0xc4 > #1: ffff80000a927720 (rcu_read_lock){....}-{1:2}, at: > kvm_pgtable_walk+0x0/0x1f4 > CPU: 2 PID: 1 Comm: swapper/0 Not tainted 6.1.0-rc3+ #5918 > Hardware name: Raspberry Pi 3 Model B (DT) > Call trace: > dump_backtrace.part.0+0xe4/0xf0 > show_stack+0x18/0x40 > dump_stack_lvl+0x8c/0xb8 > dump_stack+0x18/0x34 > __might_resched+0x178/0x220 > __might_sleep+0x48/0xa0 > prepare_alloc_pages+0x178/0x1a0 > __alloc_pages+0x9c/0x109c > alloc_page_interleave+0x1c/0xc4 > alloc_pages+0xec/0x160 > get_zeroed_page+0x1c/0x44 > kvm_hyp_zalloc_page+0x14/0x20 > hyp_map_walker+0xd4/0x134 > kvm_pgtable_visitor_cb.isra.0+0x38/0x5c > __kvm_pgtable_walk+0x1a4/0x220 > kvm_pgtable_walk+0x104/0x1f4 > kvm_pgtable_hyp_map+0x80/0xc4 > __create_hyp_mappings+0x9c/0xc4 > kvm_mmu_init+0x144/0x1cc > kvm_arch_init+0xe4/0xef4 > kvm_init+0x3c/0x3d0 > arm_init+0x20/0x30 > do_one_initcall+0x74/0x400 > kernel_init_freeable+0x2e0/0x350 > kernel_init+0x24/0x130 > ret_from_fork+0x10/0x20 > > Since the hyp stage-1 table walkers are serialized by kvm_hyp_pgd_mutex, > RCU protection really doesn't add anything. Don't acquire the RCU read > lock for an exclusive walk. While at it, add a warning which codifies > the lack of support for shared walks in the hypervisor code. > > Reported-by: Marek Szyprowski <m.szyprowski@samsung.com> > Signed-off-by: Oliver Upton <oliver.upton@linux.dev> > --- > > Applies on top of the parallel faults series that was picked up last > week. Tested with kvm-arm.mode={nvhe,protected} on an Ampere Altra > system. > > v1 -> v2: > - Took Will's suggestion of conditioning RCU on a flag, small tweak to > use existing bit instead (Thanks!) > > arch/arm64/include/asm/kvm_pgtable.h | 22 ++++++++++++++++------ > arch/arm64/kvm/hyp/pgtable.c | 5 +++-- > 2 files changed, 19 insertions(+), 8 deletions(-) > > diff --git a/arch/arm64/include/asm/kvm_pgtable.h b/arch/arm64/include/asm/kvm_pgtable.h > index a874ce0ce7b5..d4c7321fa652 100644 > --- a/arch/arm64/include/asm/kvm_pgtable.h > +++ b/arch/arm64/include/asm/kvm_pgtable.h > @@ -51,8 +51,16 @@ static inline kvm_pte_t *kvm_dereference_pteref(kvm_pteref_t pteref, bool shared > return pteref; > } > > -static inline void kvm_pgtable_walk_begin(void) {} > -static inline void kvm_pgtable_walk_end(void) {} > +static inline void kvm_pgtable_walk_begin(bool shared) > +{ > + /* > + * Due to the lack of RCU (or a similar protection scheme), only > + * non-shared table walkers are allowed in the hypervisor. > + */ > + WARN_ON(shared); > +} > + > +static inline void kvm_pgtable_walk_end(bool shared) {} > > static inline bool kvm_pgtable_walk_lock_held(void) > { > @@ -68,14 +76,16 @@ static inline kvm_pte_t *kvm_dereference_pteref(kvm_pteref_t pteref, bool shared > return rcu_dereference_check(pteref, !shared); > } > > -static inline void kvm_pgtable_walk_begin(void) > +static inline void kvm_pgtable_walk_begin(bool shared) I'm not crazy about this sort of parameters. I think it would make a lot more sense to pass a pointer to the walker structure and do the flag check inside the helper. That way, we avoid extra churn if/when we need extra state or bookkeeping around the walk. Thanks, M.
On Wed, Nov 16, 2022 at 03:08:49AM +0000, Marc Zyngier wrote: > I'm not crazy about this sort of parameters. I think it would make a > lot more sense to pass a pointer to the walker structure and do the > flag check inside the helper. > > That way, we avoid extra churn if/when we need extra state or > bookkeeping around the walk. Sure, let's go that way instead. v3 on the way lol :) -- Thanks, Oliver
On Wed, Nov 16, 2022 at 07:27:27AM +0000, Oliver Upton wrote: > On Wed, Nov 16, 2022 at 03:08:49AM +0000, Marc Zyngier wrote: > > I'm not crazy about this sort of parameters. I think it would make a > > lot more sense to pass a pointer to the walker structure and do the > > flag check inside the helper. > > > > That way, we avoid extra churn if/when we need extra state or > > bookkeeping around the walk. > > Sure, let's go that way instead. v3 on the way lol :) Well, going this route is going to require hoisting around a few things. I'd very much prefer to keep the RCU indirection ifdeffery all in one place, but I'll need to haul it all after the definitions of kvm_pgtable_walker and kvm_pgtable_walk_flags but before the definition of kvm_pgtable (as it needs kvm_pteref_t). I'm not too bothered by it, but not quite as small of a bandaid this time around. -- Thanks, Oliver
diff --git a/arch/arm64/include/asm/kvm_pgtable.h b/arch/arm64/include/asm/kvm_pgtable.h index a874ce0ce7b5..d4c7321fa652 100644 --- a/arch/arm64/include/asm/kvm_pgtable.h +++ b/arch/arm64/include/asm/kvm_pgtable.h @@ -51,8 +51,16 @@ static inline kvm_pte_t *kvm_dereference_pteref(kvm_pteref_t pteref, bool shared return pteref; } -static inline void kvm_pgtable_walk_begin(void) {} -static inline void kvm_pgtable_walk_end(void) {} +static inline void kvm_pgtable_walk_begin(bool shared) +{ + /* + * Due to the lack of RCU (or a similar protection scheme), only + * non-shared table walkers are allowed in the hypervisor. + */ + WARN_ON(shared); +} + +static inline void kvm_pgtable_walk_end(bool shared) {} static inline bool kvm_pgtable_walk_lock_held(void) { @@ -68,14 +76,16 @@ static inline kvm_pte_t *kvm_dereference_pteref(kvm_pteref_t pteref, bool shared return rcu_dereference_check(pteref, !shared); } -static inline void kvm_pgtable_walk_begin(void) +static inline void kvm_pgtable_walk_begin(bool shared) { - rcu_read_lock(); + if (shared) + rcu_read_lock(); } -static inline void kvm_pgtable_walk_end(void) +static inline void kvm_pgtable_walk_end(bool shared) { - rcu_read_unlock(); + if (shared) + rcu_read_unlock(); } static inline bool kvm_pgtable_walk_lock_held(void) diff --git a/arch/arm64/kvm/hyp/pgtable.c b/arch/arm64/kvm/hyp/pgtable.c index 5bca9610d040..16aed10e7e59 100644 --- a/arch/arm64/kvm/hyp/pgtable.c +++ b/arch/arm64/kvm/hyp/pgtable.c @@ -287,11 +287,12 @@ int kvm_pgtable_walk(struct kvm_pgtable *pgt, u64 addr, u64 size, .end = PAGE_ALIGN(walk_data.addr + size), .walker = walker, }; + bool shared = walker->flags & KVM_PGTABLE_WALK_SHARED; int r; - kvm_pgtable_walk_begin(); + kvm_pgtable_walk_begin(shared); r = _kvm_pgtable_walk(pgt, &walk_data); - kvm_pgtable_walk_end(); + kvm_pgtable_walk_end(shared); return r; }
Marek reported a BUG resulting from the recent parallel faults changes, as the hyp stage-1 map walker attempted to allocate table memory while holding the RCU read lock: BUG: sleeping function called from invalid context at include/linux/sched/mm.h:274 in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 1, name: swapper/0 preempt_count: 0, expected: 0 RCU nest depth: 1, expected: 0 2 locks held by swapper/0/1: #0: ffff80000a8a44d0 (kvm_hyp_pgd_mutex){+.+.}-{3:3}, at: __create_hyp_mappings+0x80/0xc4 #1: ffff80000a927720 (rcu_read_lock){....}-{1:2}, at: kvm_pgtable_walk+0x0/0x1f4 CPU: 2 PID: 1 Comm: swapper/0 Not tainted 6.1.0-rc3+ #5918 Hardware name: Raspberry Pi 3 Model B (DT) Call trace: dump_backtrace.part.0+0xe4/0xf0 show_stack+0x18/0x40 dump_stack_lvl+0x8c/0xb8 dump_stack+0x18/0x34 __might_resched+0x178/0x220 __might_sleep+0x48/0xa0 prepare_alloc_pages+0x178/0x1a0 __alloc_pages+0x9c/0x109c alloc_page_interleave+0x1c/0xc4 alloc_pages+0xec/0x160 get_zeroed_page+0x1c/0x44 kvm_hyp_zalloc_page+0x14/0x20 hyp_map_walker+0xd4/0x134 kvm_pgtable_visitor_cb.isra.0+0x38/0x5c __kvm_pgtable_walk+0x1a4/0x220 kvm_pgtable_walk+0x104/0x1f4 kvm_pgtable_hyp_map+0x80/0xc4 __create_hyp_mappings+0x9c/0xc4 kvm_mmu_init+0x144/0x1cc kvm_arch_init+0xe4/0xef4 kvm_init+0x3c/0x3d0 arm_init+0x20/0x30 do_one_initcall+0x74/0x400 kernel_init_freeable+0x2e0/0x350 kernel_init+0x24/0x130 ret_from_fork+0x10/0x20 Since the hyp stage-1 table walkers are serialized by kvm_hyp_pgd_mutex, RCU protection really doesn't add anything. Don't acquire the RCU read lock for an exclusive walk. While at it, add a warning which codifies the lack of support for shared walks in the hypervisor code. Reported-by: Marek Szyprowski <m.szyprowski@samsung.com> Signed-off-by: Oliver Upton <oliver.upton@linux.dev> --- Applies on top of the parallel faults series that was picked up last week. Tested with kvm-arm.mode={nvhe,protected} on an Ampere Altra system. v1 -> v2: - Took Will's suggestion of conditioning RCU on a flag, small tweak to use existing bit instead (Thanks!) arch/arm64/include/asm/kvm_pgtable.h | 22 ++++++++++++++++------ arch/arm64/kvm/hyp/pgtable.c | 5 +++-- 2 files changed, 19 insertions(+), 8 deletions(-)