arm64: barriers: allow dsb macro to take option parameter
Subsystems (5)
PATCH 1/6 arm64: barriers: allow dsb macro to take option parameter
The dsb instruction takes an option specifying both the target access types and shareability domain.
This patch allows such an option to be passed to the dsb macro, resulting in potentially more efficient code. Currently the option is ignored until all callers are updated (unlike ARM, the option is mandated by the assembler).
---
Catalin -- I'd like to get this simple change in for 3.14, since it's a
blocker for actually implementing the barrier options (we need to move all of the callers over before making the switch).
arch/arm64/include/asm/barrier.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm64/include/asm/barrier.h +1 -1
@@ -25,7 +25,7 @@#define wfi () asm volatile("wfi" : : : "memory")#define isb () asm volatile("isb" : : : "memory") -#define dsb () asm volatile("dsb sy" : : : "memory") +#define dsb (opt) asm volatile("dsb sy" : : : "memory")#define mb () dsb()#define rmb () asm volatile("dsb ld" : : : "memory")
Signature · 1 line
1.8.2.2
PATCH 2/6 arm64: barriers: make use of barrier options with explicit barriers
When calling our low-level barrier macros directly, we can often suffice with more relaxed behaviour than the default "all accesses, full system" option.
This patch updates the users of dsb() to specify the option whith they actually require. The sev() macro is also updated so that a sevl can be emitted if necessary.
--- arch/arm64/include/asm/barrier.h | 4 ++-- arch/arm64/include/asm/cacheflush.h | 2 +- arch/arm64/include/asm/pgtable.h | 4 ++-- arch/arm64/include/asm/tlbflush.h | 14 +++++++------- arch/arm64/kernel/process.c | 2 +- arch/arm64/kvm/sys_regs.c | 4 ++-- 6 files changed, 15 insertions(+), 15 deletions(-)
arch/arm64/include/asm/barrier.h +2 -2
@@ -20,14 +20,14 @@#ifndef __ASSEMBLY__ -#define sev () asm volatile("sev" : : : "memory") +#define sev (l) asm volatile("sev" #l : : : "memory")#define wfe () asm volatile("wfe" : : : "memory")#define wfi () asm volatile("wfi" : : : "memory")#define isb () asm volatile("isb" : : : "memory")#define dsb (opt) asm volatile("dsb sy" : : : "memory") -#define mb () dsb() +#define mb () dsb(sy)#define rmb () asm volatile("dsb ld" : : : "memory")#define wmb () asm volatile("dsb st" : : : "memory")
arch/arm64/include/asm/cacheflush.h +1 -1
@@ -142,7 +142,7 @@ static inline void flush_cache_vmap(unsigned long start, unsigned long end)* set_pte_at() called from vmap_pte_range() does not * have a DSB after cleaning the cache line. */ -dsb (); +dsb (ishst); }static inline void flush_cache_vunmap (unsigned long start,unsigned long end)
arch/arm64/include/asm/pgtable.h +2 -2
@@ -310,7 +310,7 @@ extern pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,static inline void set_pmd (pmd_t *pmdp,pmd_t pmd) { *pmdp = pmd; -dsb (); +dsb (ishst); }static inline void pmd_clear (pmd_t *pmdp)
@@ -340,7 +340,7 @@ static inline pte_t *pmd_page_vaddr(pmd_t pmd)static inline void set_pud (pud_t *pudp,pud_t pud) { *pudp = pud; -dsb (); +dsb (ishst); }static inline void pud_clear (pud_t *pudp)
arch/arm64/include/asm/tlbflush.h +7 -7
@@ -72,9 +72,9 @@ extern struct cpu_tlb_fns cpu_tlb;*/ static inline void flush_tlb_all (void ) { -dsb (); +dsb (ishst); asm("tlbi vmalle1is" ); -dsb (); +dsb (ish);isb (); }
@@ -82,9 +82,9 @@ static inline void flush_tlb_mm(struct mm_struct *mm) {unsigned long asid = (unsigned long )ASID (mm) <<48 ; -dsb (); +dsb (ishst); asm("tlbi aside1is, %0" : :"r" (asid)); -dsb (); +dsb (ish); } static inline void flush_tlb_page(struct vm_area_struct *vma,
@@ -93,9 +93,9 @@ static inline void flush_tlb_page(struct vm_area_struct *vma,unsigned long addr = uaddr >>12 | ((unsigned long )ASID (vma->vm_mm) <<48 ); -dsb (); +dsb (ishst); asm("tlbi vae1is, %0" : :"r" (addr)); -dsb (); +dsb (ish); }/*
@@ -114,7 +114,7 @@ static inline void update_mmu_cache(struct vm_area_struct *vma,* set_pte() does not have a DSB, so make sure that the page table * write is visible. */ -dsb (); +dsb (ishst); }#define update_mmu_cache_pmd (vma, address, pmd) do { } while (0)
arch/arm64/kernel/process.c +1 -1
@@ -294,7 +294,7 @@ struct task_struct *__switch_to(struct task_struct *prev,* Complete any pending TLB or cache maintenance on this CPU in case * the thread migrates to a different CPU. */ -dsb (); +dsb (ish);/* the actual thread switch */ last =cpu_switch_to (prev, next);
arch/arm64/kvm/sys_regs.c +2 -2
@@ -70,13 +70,13 @@ static u32 get_ccsidr(u32 csselr)static void do_dc_cisw (u32 val) { asmvolatile ("dc cisw, %x0" : :"r" (val)); -dsb (); +dsb (ishst); }static void do_dc_csw (u32 val) { asmvolatile ("dc csw, %x0" : :"r" (val)); -dsb (); +dsb (ishst); }/* See note@ARM ARM B1.14.4 */
Signature · 1 line
1.8.2.2
Catalin Marinas
maintainerPATCH 2/6 arm64: barriers: make use of barrier options with explicit barriers
arch/arm64/include/asm/barrier.h line 23
Quoted 8 lines · Will Deacon
arch/arm64/include/asm/barrier.h +1
-1@@ -20,14 +20,14 @@#ifndef __ASSEMBLY__ -#define sev () asm volatile("sev" : : : "memory") +#define sev (l) asm volatile("sev" #l : : : "memory")
Would we actually ever use sev(l) form C? And it's a new instruction rather than an argument to the existing sev.
Signature · 1 line
Catalin
PATCH 2/6 arm64: barriers: make use of barrier options with explicit barriers
Quoted 5 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
arch/arm64/include/asm/barrier.h +0
-0@@ -20,14 +20,14 @@
Quoted 7 lines
arch/arm64/include/asm/barrier.h line 23
Quoted 4 lines
arch/arm64/include/asm/barrier.h +1
-1#ifndef __ASSEMBLY__ -#define sev () asm volatile("sev" : : : "memory") +#define sev (l) asm volatile("sev" #l : : : "memory")Would we actually ever use sev(l) form C? And it's a new instruction rather than an argument to the existing sev.
I don't know, but if it's not there then I'm pretty sure people will always use sev, even if sevl could be used. In fact, sev and sevl are both aliases of hint, so they're the same instruction with different immediate operands.
Will
Catalin Marinas
maintainerPATCH 2/6 arm64: barriers: make use of barrier options with explicit barriers
Quoted 14 lines · Will Deacon
Quoted 5 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
arch/arm64/include/asm/barrier.h +0
-0@@ -20,14 +20,14 @@Quoted 7 lines
arch/arm64/include/asm/barrier.h line 23
Quoted 4 lines
arch/arm64/include/asm/barrier.h +1
-1#ifndef __ASSEMBLY__ -#define sev () asm volatile("sev" : : : "memory") +#define sev (l) asm volatile("sev" #l : : : "memory")Would we actually ever use sev(l) form C? And it's a new instruction rather than an argument to the existing sev.
Quoted 3 lines
I don't know, but if it's not there then I'm pretty sure people will always use sev, even if sevl could be used. In fact, sev and sevl are both aliases of hint, so they're the same instruction with different immediate operands.
SEVL was pretty much introduced to avoid a branch in the spinlock functions. I'm not aware of other uses (on arm64 we don't even use the wfe() macro directly). I would skip this for now.
Signature · 1 line
Catalin
PATCH 2/6 arm64: barriers: make use of barrier options with explicit barriers
Quoted 15 lines · Catalin Marinas
Quoted 14 lines · Will Deacon
Quoted 12 lines · Catalin Marinas
arch/arm64/include/asm/barrier.h line 23
Quoted 8 lines · Will Deacon
arch/arm64/include/asm/barrier.h +1
-1@@ -20,14 +20,14 @@#ifndef __ASSEMBLY__ -#define sev () asm volatile("sev" : : : "memory") +#define sev (l) asm volatile("sev" #l : : : "memory")Would we actually ever use sev(l) form C? And it's a new instruction rather than an argument to the existing sev.
Quoted 7 lines
Quoted 3 lines
I don't know, but if it's not there then I'm pretty sure people will always use sev, even if sevl could be used. In fact, sev and sevl are both aliases of hint, so they're the same instruction with different immediate operands.
SEVL was pretty much introduced to avoid a branch in the spinlock functions. I'm not aware of other uses (on arm64 we don't even use the
wfe()macro directly). I would skip this for now.
Okey doke, I'll drop this hunk for the time being.
Will
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
In reply to Will DeaconWhen sending an SGI to another CPU, we require a DSB to ensure that any pending stores to normal memory are made visible to the recipient before the interrupt arrives.
Rather than use a dsb() (which will soon cause an assembly error on arm64) followed by a writel_relaxed, we can use a writel instead, which will emit a dsb st prior to the str.
--- drivers/irqchip/irq-gic.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-)
drivers/irqchip/irq-gic.c +3 -4
@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/* * Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */ -dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );raw_spin_unlock_irqrestore (&irq_controller_lock, flags); }
Signature · 1 line
1.8.2.2
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 11 lines · Will Deacon
When sending an SGI to another CPU, we require a DSB to ensure that any pending stores to normal memory are made visible to the recipient before the interrupt arrives.
Rather than use a
dsb()(which will soon cause an assembly error on arm64) followed by awritel_relaxed, we can use a writel instead, which will emit a dsb st prior to the str.Cc Thomas Gleixner <[email protected]>Cc Marc Zyngier <[email protected]>Signed-off-by Will Deacon <[email protected]>
Looks sensible to me.
M.
Signature · 1 line
Jazz is not dead. It just smells funny...
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
In reply to Will DeaconQuoted 6 lines · Will Deacon
drivers/irqchip/irq-gic.c +0
-0@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/*
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 10 lines
drivers/irqchip/irq-gic.c +3
-4* Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */ -dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );
That's heavier than a dsb() since with outer caches on ARM we also get an outer_sync() call.
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 12 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
drivers/irqchip/irq-gic.c +2
-0@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/* * Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */
Quoted 8 lines
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 5 lines
drivers/irqchip/irq-gic.c +1
-4-dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.
Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
I can update the commit message if you like?
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 22 lines · Will Deacon
Quoted 12 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
drivers/irqchip/irq-gic.c +2
-0@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/* * Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */ Quoted 8 lines
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 5 lines
drivers/irqchip/irq-gic.c +1
-4-dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.
Quoted 2 lines
Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
If they are all in the inner shareable domain and with the caches enabled, we don't need to flush the outer cache (as in the PL310 case which is common to all CPUs; other saner outer caches propagate the barrier ;). The outer_sync is needed when the memory accesses are non-cacheable and we need to drain both the CPU write-buffer and the PL310 one.
For our case here, we only need to ensure the visibility of writes on a CPU to another but assuming SMP and caches enabled, so DSB is enough.
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 33 lines · Catalin Marinas
Quoted 22 lines · Will Deacon
Quoted 12 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
drivers/irqchip/irq-gic.c +2
-0@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/* * Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */ Quoted 8 lines
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 5 lines
drivers/irqchip/irq-gic.c +1
-4-dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.Quoted 2 lines
Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
If they are all in the inner shareable domain and with the caches enabled, we don't need to flush the outer cache (as in the PL310 case which is common to all CPUs; other saner outer caches propagate the barrier ;). The
outer_syncis needed when the memory accesses are non-cacheable and we need to drain both the CPU write-buffer and the PL310 one.
Quoted 2 lines
For our case here, we only need to ensure the visibility of writes on a CPU to another but assuming SMP and caches enabled, so DSB is enough.
Hmm, but we *do* use this for boot and need to ensure that any mailboxes are visible. Maybe we have enough cacheflushing/barriers for that already, but I'm really uncomfortable making this a simple dsb(ishst).
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 34 lines · Will Deacon
Quoted 33 lines · Catalin Marinas
Quoted 22 lines · Will Deacon
Quoted 12 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
drivers/irqchip/irq-gic.c +2
-0@@ -662,11 +662,10 @@ void gic_raise_softirq(const struct cpumask *mask, unsigned int irq)/* * Ensure that stores to Normal memory are visible to the * other CPUs before issuing the IPI. +* +* This always happens on GIC0. */ Quoted 8 lines
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 5 lines
drivers/irqchip/irq-gic.c +1
-4-dsb (); - -/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.Quoted 2 lines
Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
If they are all in the inner shareable domain and with the caches enabled, we don't need to flush the outer cache (as in the PL310 case which is common to all CPUs; other saner outer caches propagate the barrier ;). The
outer_syncis needed when the memory accesses are non-cacheable and we need to drain both the CPU write-buffer and the PL310 one.
Quoted 6 lines
Quoted 2 lines
For our case here, we only need to ensure the visibility of writes on a CPU to another but assuming SMP and caches enabled, so DSB is enough.
Hmm, but we *do* use this for boot and need to ensure that any mailboxes are visible. Maybe we have enough cacheflushing/barriers for that already, but I'm really uncomfortable making this a simple dsb(ishst).
For boot we explicitly flush the caches for the shared data, so we don't need this. The dsb() here is for standard smp_call_* etc. We didn't have outer_sync() before, so you are slightly changing the functionality here (arguably, ishst is relaxing the requirements but I'm not worried about this, I consider that's the standard use-case for this function).
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 29 lines · Catalin Marinas
Quoted 21 lines · Will Deacon
Quoted 20 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
Quoted 7 lines · Catalin Marinas
drivers/irqchip/irq-gic.c void gic_raise_softirq(const struct cpumask *mask, unsigned int irq) · line 668
Quoted 3 lines · Will Deacon
drivers/irqchip/irq-gic.c +1
-2-/* this always happens on GIC0 */ -writel_relaxed (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT ); +writel (map <<16 | irq,gic_data_dist_base (&gic_data[0 ]) +GIC_DIST_SOFTINT );That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
If they are all in the inner shareable domain and with the caches enabled, we don't need to flush the outer cache (as in the PL310 case which is common to all CPUs; other saner outer caches propagate the barrier ;). The
outer_syncis needed when the memory accesses are non-cacheable and we need to drain both the CPU write-buffer and the PL310 one.Quoted 6 lines
Quoted 2 lines
For our case here, we only need to ensure the visibility of writes on a CPU to another but assuming SMP and caches enabled, so DSB is enough.
Hmm, but we *do* use this for boot and need to ensure that any mailboxes are visible. Maybe we have enough cacheflushing/barriers for that already, but I'm really uncomfortable making this a simple dsb(ishst).
Quoted 5 lines
For boot we explicitly flush the caches for the shared data, so we don't need this. The
dsb()here is for standard smp_call_* etc. We didn't haveouter_sync()before, so you are slightly changing the functionality here (arguably, ishst is relaxing the requirements but I'm not worried about this, I consider that's the standard use-case for this function).
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 36 lines · Will Deacon
Quoted 29 lines · Catalin Marinas
Quoted 21 lines · Will Deacon
Quoted 20 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
Quoted 7 lines · Catalin Marinas
Review
Quoted 3 lines · Will Deacon
(quoted diff) +1
-2- /* this always happens on GIC0 */ - writel_relaxed(map << 16 | irq, gic_data_dist_base(&gic_data[0]) + GIC_DIST_SOFTINT); + writel(map << 16 | irq, gic_data_dist_base(&gic_data[0]) + GIC_DIST_SOFTINT);That's heavier than a
dsb()since with outer caches on ARM we also get anouter_sync()call.Yes, which I think we actually need in this case, since we're trying to make normal writes visible to a CPU before a device write hits the GIC.
If they are all in the inner shareable domain and with the caches enabled, we don't need to flush the outer cache (as in the PL310 case which is common to all CPUs; other saner outer caches propagate the barrier ;). The
outer_syncis needed when the memory accesses are non-cacheable and we need to drain both the CPU write-buffer and the PL310 one.Quoted 6 lines
Quoted 2 lines
For our case here, we only need to ensure the visibility of writes on a CPU to another but assuming SMP and caches enabled, so DSB is enough.
Hmm, but we *do* use this for boot and need to ensure that any mailboxes are visible. Maybe we have enough cacheflushing/barriers for that already, but I'm really uncomfortable making this a simple dsb(ishst).
Quoted 5 lines
For boot we explicitly flush the caches for the shared data, so we don't need this. The
dsb()here is for standard smp_call_* etc. We didn't haveouter_sync()before, so you are slightly changing the functionality here (arguably, ishst is relaxing the requirements but I'm not worried about this, I consider that's the standard use-case for this function).
Quoted 4 lines
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Because you want to guarantee the ordering between a store to Normal Cacheable memory vs store to Device for the IPI (see the mailbox example in the Barrier Litmus section ;)). The second is just a slave access, DMB guarantees observability from the master access perspective.
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 10 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Because you want to guarantee the ordering between a store to Normal Cacheable memory vs store to Device for the IPI (see the mailbox example in the Barrier Litmus section ;)). The second is just a slave access, DMB guarantees observability from the master access perspective.
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
so the only way this can break is if the GICD write reaches the distributor before being observed by CPU1 (otherwise, we know the mailbox write was observed by CPU1). I dread to think how you would build such a beast (dual-ported GICD with no serialisation to the same locations?)...
Furthermore, if we decide that device writes can reach their endpoints before being observed by other inner-shareable observers, then doesn't that pose a potential problem for spinlocks? If I take a lock and write to a device, the write can hit the device before the lock appears to be taken. That doesn't sound right to me.
Using a dsb(ishst) will ensure that we don't issue the GICD write until the mailbox is visible to CPU1, but may be overkill.
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 17 lines · Will Deacon
Quoted 10 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Because you want to guarantee the ordering between a store to Normal Cacheable memory vs store to Device for the IPI (see the mailbox example in the Barrier Litmus section ;)). The second is just a slave access, DMB guarantees observability from the master access perspective.
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
Quoted 6 lines
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
Quoted 4 lines
so the only way this can break is if the GICD write reaches the distributor before being observed by CPU1 (otherwise, we know the mailbox write was observed by CPU1). I dread to think how you would build such a beast (dual-ported GICD with no serialisation to the same locations?)...
The above is possible because the CPU1 would never "observe" the GICD write. It observes a side-effect of that write (interrupt) which is not covered by DMB. You could argue that CPU1 reads GICC in the interrupt handler but I'm not sure GICC vs GICD ordering is guaranteed (and I know of hardware where not even accesses to the same GICD are guaranteed ;)).
Quoted 1 line
Furthermore, if we decide that device writes can reach their endpoints
Device writes are not observed according to the ARM ARM meaning of "observability" (master accesses only; well, there is something about Strongly Ordered memory and devices but it's not the case here and IIRC refers to when a _device_ observes the write rather than a CPU observing the write to that device, I need to read it again).
Quoted 4 lines
before being observed by other inner-shareable observers, then doesn't that pose a potential problem for spinlocks? If I take a lock and write to a device, the write can hit the device before the lock appears to be taken. That doesn't sound right to me.
For simplicity, assuming that the lock acquiring was a simple STR, the write to device can indeed hit the device before the STR is observed by another CPU or device (that's why writel() has a DSB). This is however not relevant as you don't care when it hit the device (unless you issue IPIs). If considering master accesses, the below is valid even though B could hit the device before A is observed by CPU1:
CPU0:
STR A, [Normal]
DMB
STR B, [Device]
CPU1:
LDR C, [Device]
DMB
LDR D, [Normal]
If on CPU1 C == B then D == A. The key here is that observability of STR B, [Device] is done via another master access on CPU1 (LDR C, [Device]) and not just the change to the device state.
When we talk about LDREX/STREX loop, I don't think the write to the device can be issued before the STREX has been guaranteed to succeed (not necessarily observed) and therefore the spinlock acquired (we don't issue writes speculatively and the spinlock loop has a conditional branch).
If we go to the definition of the STR observability (in short):
1. A load from a location returns the value written by the observed store.
2. A store to a location changes the value written by the observed store.
The "not observability" means _any_ of the above conditions is false. So going back to your example, the write to the device can hit the device _and_ the load of the spinlock value on another CPU1 still return the unlocked value. However, a subsequent STREX on CPU1 would fail and the LDREX restarted, eventually observing the STREX on CPU0.
If we talk about lock acquiring on another CPU and issuing of device accesses, the DMB guarantees ordering (master accesses).
Another interesting scenario is the write to device followed by spin_unlock and on another CPU spin_lock and device write. As per my example above, I don't see any issue (change Normal with Device).
Quoted 2 lines
Using a dsb(ishst) will ensure that we don't issue the GICD write until the mailbox is visible to CPU1, but may be overkill.
In the ARM ARM examples, the mailbox write generates the interrupt (it is not visible to the other CPU at all).
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 22 lines · Catalin Marinas
Quoted 21 lines · Will Deacon
Quoted 10 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Because you want to guarantee the ordering between a store to Normal Cacheable memory vs store to Device for the IPI (see the mailbox example in the Barrier Litmus section ;)). The second is just a slave access, DMB guarantees observability from the master access perspective.
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
Quoted 6 lines
Quoted 2 lines
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
No, that's not how observability is defined, unfortunately.
Rather than attempt to solve this via email (your examples below are already getting hard to follow :), how about we sit down with $drink_of_choice and post back here with our conclusions?
In the meantime, I'll use dsb(ishst) because I think we now agree that works. The question is whether it can be relaxed further to a dmb.
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 23 lines · Will Deacon
Quoted 22 lines · Catalin Marinas
Quoted 21 lines · Will Deacon
Quoted 10 lines · Catalin Marinas
Quoted 4 lines · Will Deacon
Ok, so if we assume that a dsb(ishst) is sufficient because the CPU we're talking to is either (a) coherent in the inner-shareable domain or (b) incoherent, and we flushed everything to PoC, then why wouldn't a dmb(ishst) work?
Because you want to guarantee the ordering between a store to Normal Cacheable memory vs store to Device for the IPI (see the mailbox example in the Barrier Litmus section ;)). The second is just a slave access, DMB guarantees observability from the master access perspective.
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
Quoted 8 lines
Quoted 6 lines
Quoted 2 lines
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
No, that's not how observability is defined, unfortunately.
According to the ARM ARM, "an observer is a master in the system that is capable of observing memory accesses". GICD is not memory, hence the assumptions about DMB in this context no longer work.
Quoted 3 lines
Rather than attempt to solve this via email (your examples below are already getting hard to follow :), how about we sit down with $drink_of_choice and post back here with our conclusions?
Sounds good (though not a wide choice of drinks).
Quoted 2 lines
In the meantime, I'll use dsb(ishst) because I think we now agree that works. The question is whether it can be relaxed further to a dmb.
I'm fine with dsb(ishst).
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
In reply to Will DeaconWell, the results are in (*drum roll*)...
Quoted 14 lines · Will Deacon
Quoted 13 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
Quoted 7 lines
Quoted 3 lines
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
Rather than attempt to solve this via email (your examples below are already getting hard to follow :), how about we sit down with $drink_of_choice and post back here with our conclusions?
... and it turns out that a dmb(ishst) is sufficient!
This counter-intuitive result was brought to you by ARM Ltd, purveyors of fine grained memory barrier instructions.
I'll spin the patch again.
Will
Catalin Marinas
maintainerPATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 21 lines · Will Deacon
Well, the results are in (*drum roll*)...
Quoted 18 lines · Will Deacon
Quoted 16 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
Quoted 5 lines
Quoted 3 lines
Rather than attempt to solve this via email (your examples below are already getting hard to follow :), how about we sit down with $drink_of_choice and post back here with our conclusions?
... and it turns out that a dmb(ishst) is sufficient!
Until we hear otherwise ;)
Signature · 1 line
Catalin
PATCH 3/6 irqchip: gic: use writel instead of dsb + writel_relaxed
Quoted 22 lines · Catalin Marinas
Quoted 21 lines · Will Deacon
Well, the results are in (*drum roll*)...
Quoted 18 lines · Will Deacon
Quoted 16 lines · Catalin Marinas
Quoted 11 lines · Will Deacon
Ok, my reasoning is as follows:
- CPU0 tries to message CPU1. It writes to a location in normal memory, then writes to the GICD to send the SGI
- We need to ensure that CPU1 observes the write to normal memory before the write to GICD reaches the distributor. This is *not* about end-point ordering (the usual non-coherent DMA example).
- A dmb ishst ensures that the two writes are observed in order by CPU1 (and, in fact, the inner-shareable domain containing CPU0).
The last bullet point is not correct. DMB would only guarantee that the two writes (memory and GICD) are observed by CPU1 if CPU1 actually read the GICD (observability is defined for master accesses).
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Rather than attempt to solve this via email (your examples below are already getting hard to follow :), how about we sit down with $drink_of_choice and post back here with our conclusions?
... and it turns out that a dmb(ishst) is sufficient!
Until we hear otherwise ;)
It is a shame that all this thought and effort will just end with barrier related chicken bits getting set anyway. Maybe the $drink_of_choice was worth it. ;)
Rob
PATCH 4/6 iommu/arm-smmu: provide option to dsb macro when publishing tables
In reply to Will DeaconOn coherent systems, publishing new page tables to the SMMU walker is achieved with a dsb instruction. In fact, this can be a dsb(ishst) which also provides the mandatory barrier option for arm64.
--- drivers/iommu/arm-smmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/iommu/arm-smmu.c +1 -1
@@ -1417,7 +1417,7 @@ out_unlock:/* Ensure new page tables are visible to the hardware walker */ if (smmu->features &ARM_SMMU_FEAT_COHERENT_WALK ) -dsb (); +dsb (ishst);return ret; }
Signature · 1 line
1.8.2.2
Catalin Marinas
maintainerPATCH 4/6 iommu/arm-smmu: provide option to dsb macro when publishing tables
Quoted 5 lines · Will Deacon
On coherent systems, publishing new page tables to the SMMU walker is achieved with a dsb instruction. In fact, this can be a dsb(ishst) which also provides the mandatory barrier option for arm64.
Signed-off-by Will Deacon <[email protected]>
PATCH 5/6 arm64: barriers: wire up new barrier options
In reply to Will DeaconNow that all callers of the barrier macros are updated to pass the mandatory options, update the macros so the option is actually used.
--- arch/arm64/include/asm/barrier.h | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)
arch/arm64/include/asm/barrier.h +7 -6
@@ -25,11 +25,12 @@#define wfi () asm volatile("wfi" : : : "memory")#define isb () asm volatile("isb" : : : "memory") -#define dsb (opt) asm volatile("dsb sy" : : : "memory") +#define dmb (opt) asm volatile("dmb " #opt : : : "memory") +#define dsb (opt) asm volatile("dsb " #opt : : : "memory")#define mb () dsb(sy) -#define rmb () asm volatile("dsb ld" : : : "memory") -#define wmb () asm volatile("dsb st" : : : "memory") +#define rmb () dsb(ld) +#define wmb () dsb(st)#ifndef CONFIG_SMP #define smp_mb () barrier()
@@ -53,9 +54,9 @@ do { \#else -#define smp_mb () asm volatile("dmb ish" : : : "memory") -#define smp_rmb () asm volatile("dmb ishld" : : : "memory") -#define smp_wmb () asm volatile("dmb ishst" : : : "memory") +#define smp_mb () dmb(ish) +#define smp_rmb () dmb(ishld) +#define smp_wmb () dmb(ishst)#define smp_store_release (p, v) \ do { \
Signature · 1 line
1.8.2.2
Catalin Marinas
maintainerPATCH 5/6 arm64: barriers: wire up new barrier options
Quoted 4 lines · Will Deacon
Now that all callers of the barrier macros are updated to pass the mandatory options, update the macros so the option is actually used.
Signed-off-by Will Deacon <[email protected]>
PATCH 6/6 arm64: barriers: use barrier() instead of smp_mb() when !SMP
In reply to Will DeaconThe recently introduced acquire/release accessors refer to smp_mb() in the !CONFIG_SMP case. This is confusing when reading the code, so use barrier() directly when we know we're UP.
--- arch/arm64/include/asm/barrier.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/arm64/include/asm/barrier.h +2 -2
@@ -40,7 +40,7 @@#define smp_store_release (p, v) \ do { \ compiletime_assert_atomic_type(*p); \ - smp_mb(); \ + barrier(); \ ACCESS_ONCE(*p) = (v); \ } while (0)
@@ -48,7 +48,7 @@ do { \ ({ \ typeof(*p) ___p1 = ACCESS_ONCE(*p); \compiletime_assert_atomic_type (*p); \ -smp_mb (); \ +barrier (); \ ___p1; \ })
Signature · 1 line
1.8.2.2
Catalin Marinas
maintainerPATCH 6/6 arm64: barriers: use barrier() instead of smp_mb() when !SMP
Quoted 5 lines · Will Deacon
The recently introduced acquire/release accessors refer to
smp_mb()in the !CONFIG_SMP case. This is confusing when reading the code, so usebarrier()directly when we know we're UP.Signed-off-by Will Deacon <[email protected]>
Catalin Marinas
maintainerPATCH 1/6 arm64: barriers: allow dsb macro to take option parameter
In reply to Will DeaconQuoted 8 lines · Will Deacon
The dsb instruction takes an option specifying both the target access types and shareability domain.
This patch allows such an option to be passed to the dsb macro, resulting in potentially more efficient code. Currently the option is ignored until all callers are updated (unlike ARM, the option is mandated by the assembler).
Quoted 6 lines
Signed-off-by Will Deacon <[email protected]>---
Catalin -- I'd like to get this simple change in for 3.14, since it's a
blocker for actually implementing the barrier options (we need to move all of the callers over before making the switch).
OK, no problem. Applied.
Signature · 1 line
Catalin