lkml.org 
[lkml]   [2019]   [Mar]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] arm64/io: Don't use WZR in writel
From
Date
On 18/03/2019 17:00, Russell King - ARM Linux admin wrote:
> On Mon, Mar 18, 2019 at 04:04:03PM +0000, Robin Murphy wrote:
>> On 12/03/2019 12:36, Marc Gonzalez wrote:
>>> On 24/02/2019 04:53, Bjorn Andersson wrote:
>>>
>>>> On Sat 23 Feb 10:37 PST 2019, Marc Zyngier wrote:
>>>>
>>>>> On Sat, 23 Feb 2019 18:12:54 +0000, Bjorn Andersson wrote:
>>>>>>
>>>>>> On Mon 11 Feb 06:59 PST 2019, Marc Zyngier wrote:
>>>>>>
>>>>>>> On 11/02/2019 14:29, AngeloGioacchino Del Regno wrote:
>>>>>>>
>>>>>>>> Also, just one more thing: yes this thing is going ARM64-wide and
>>>>>>>> - from my findings - it's targeting certain Qualcomm SoCs, but...
>>>>>>>> I'm not sure that only QC is affected by that, others may as well
>>>>>>>> have the same stupid bug.
>>>>>>>
>>>>>>> At the moment, only QC SoCs seem to be affected, probably because
>>>>>>> everyone else has debugged their hypervisor (or most likely doesn't
>>>>>>> bother with shipping one).
>>>>>>>
>>>>>>> In all honesty, we need some information from QC here: which SoCs are
>>>>>>> affected, what is the exact nature of the bug, can it be triggered from
>>>>>>> EL0. Randomly papering over symptoms is not something I really like
>>>>>>> doing, and is likely to generate problems on unaffected systems.
>>>>>>
>>>>>> The bug at hand is that the XZR is not deemed a valid source in the
>>>>>> virtualization of the SMMU registers. It was identified and fixed for
>>>>>> all platforms that are shipping kernels based on v4.9 or later.
>>>>>
>>>>> When you say "fixed": Do you mean fixed in the firmware? Or by adding
>>>>> a workaround in the shipped kernel?
>>>>
>>>> I mean that it's fixed in the firmware.
>>>>
>>>>> If the former, is this part of an official QC statement, with an
>>>>> associated erratum number?
>>>>
>>>> I don't know, will get back to you on this one.
>>>>
>>>>> Is this really limited to the SMMU accesses?
>>>>
>>>> Yes.
>>>>
>>>>>> As such Angelo's list of affected platforms covers the high-profile
>>>>>> ones. In particular MSM8996 and MSM8998 is getting pretty good support
>>>>>> upstream, if we can figure out a way around this issue.
>>>>>
>>>>> We'd need an exhaustive list of the affected SoCs, and work out if we
>>>>> can limit the hack to the SMMU driver (cc'ing Robin, who's the one
>>>>> who'd know about it).
>>>>
>>>> I will try to compose a list.
>>>
>>> FWIW, I have just been bitten by this issue. I needed to enable an SMMU to
>>> filter PCIe EP accesses to system RAM (or something). I'm using an APQ8098
>>> MEDIABOX dev board. My system hangs in arm_smmu_device_reset() doing:
>>>
>>> /* Invalidate the TLB, just in case */
>>> writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLH);
>>> writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLNSNH);
>>>
>>>
>>> With the 'Z' constraint, gcc generates:
>>>
>>> str wzr, [x0]
>>>
>>> without the 'Z' constraint, gcc generates:
>>>
>>> mov w1, 0
>>> str w1, [x0]
>>>
>>>
>>> I can work around the problem using the following patch:
>>>
>>> diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
>>> index 045d93884164..93117519aed8 100644
>>> --- a/drivers/iommu/arm-smmu.c
>>> +++ b/drivers/iommu/arm-smmu.c
>>> @@ -59,6 +59,11 @@
>>> #include "arm-smmu-regs.h"
>>> +static inline void qcom_writel(u32 val, volatile void __iomem *addr)
>>> +{
>>> + asm volatile("str %w0, [%1]" : : "r" (val), "r" (addr));
>>> +}
>>> +
>>> #define ARM_MMU500_ACTLR_CPRE (1 << 1)
>>> #define ARM_MMU500_ACR_CACHE_LOCK (1 << 26)
>>> @@ -422,7 +427,7 @@ static void __arm_smmu_tlb_sync(struct arm_smmu_device *smmu,
>>> {
>>> unsigned int spin_cnt, delay;
>>> - writel_relaxed(0, sync);
>>> + qcom_writel(0, sync);
>>> for (delay = 1; delay < TLB_LOOP_TIMEOUT; delay *= 2) {
>>> for (spin_cnt = TLB_SPIN_COUNT; spin_cnt > 0; spin_cnt--) {
>>> if (!(readl_relaxed(status) & sTLBGSTATUS_GSACTIVE))
>>> @@ -1760,8 +1765,8 @@ static void arm_smmu_device_reset(struct arm_smmu_device *smmu)
>>> }
>>> /* Invalidate the TLB, just in case */
>>> - writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLH);
>>> - writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLNSNH);
>>> + qcom_writel(0, gr0_base + ARM_SMMU_GR0_TLBIALLH);
>>> + qcom_writel(0, gr0_base + ARM_SMMU_GR0_TLBIALLNSNH);
>>> reg = readl_relaxed(ARM_SMMU_GR0_NS(smmu) + ARM_SMMU_GR0_sCR0);
>>>
>>>
>>>
>>> Can a quirk be used to work around the issue?
>>> Or can we just "pessimize" the 3 writes for everybody?
>>> (Might be cheaper than a test anyway)
>>
>> If it really is just the SMMU driver which is affected, we can work around
>> it for free (not counting the 'cost' of slightly-weird-looking code, of
>> course). If the diff below works as expected, I'll write it up properly.
>>
>> Robin.
>> ----->8-----
>> diff --git a/drivers/iommu/arm-smmu.c b/drivers/iommu/arm-smmu.c
>> index 045d93884164..7ff29e33298f 100644
>> --- a/drivers/iommu/arm-smmu.c
>> +++ b/drivers/iommu/arm-smmu.c
>> @@ -422,7 +422,7 @@ static void __arm_smmu_tlb_sync(struct arm_smmu_device
>> *smmu,
>> {
>> unsigned int spin_cnt, delay;
>>
>> - writel_relaxed(0, sync);
>> + writel_relaxed((unsigned long)sync, sync);
>> for (delay = 1; delay < TLB_LOOP_TIMEOUT; delay *= 2) {
>> for (spin_cnt = TLB_SPIN_COUNT; spin_cnt > 0; spin_cnt--) {
>> if (!(readl_relaxed(status) & sTLBGSTATUS_GSACTIVE))
>> @@ -681,7 +681,12 @@ static void arm_smmu_write_context_bank(struct
>> arm_smmu_device *smmu, int idx)
>>
>> /* Unassigned context banks only need disabling */
>> if (!cfg) {
>> - writel_relaxed(0, cb_base + ARM_SMMU_CB_SCTLR);
>> + /*
>> + * For Qualcomm reasons, we want to guarantee that we write a
>> + * zero from a register which is not WZR. Fortunately, the cfg
>> + * logic here plays right into our hands...
>> + */
>> + writel_relaxed((unsigned long)cfg, cb_base + ARM_SMMU_CB_SCTLR);
>> return;
>> }
>>
>> @@ -1760,8 +1765,8 @@ static void arm_smmu_device_reset(struct
>> arm_smmu_device *smmu)
>> }
>>
>> /* Invalidate the TLB, just in case */
>> - writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLH);
>> - writel_relaxed(0, gr0_base + ARM_SMMU_GR0_TLBIALLNSNH);
>> + writel_relaxed(reg, gr0_base + ARM_SMMU_GR0_TLBIALLH);
>> + writel_relaxed(reg, gr0_base + ARM_SMMU_GR0_TLBIALLNSNH);
>>
>> reg = readl_relaxed(ARM_SMMU_GR0_NS(smmu) + ARM_SMMU_GR0_sCR0);
>>
>
> Given what we've seen from Clang for futex stuff in 32-bit ARM, are
> you really sure that the above will not result in Clang still spotting
> that the value is zero and using a wzr for all these cases?

The trick is that in the write-only TLBI cases the variable we're
passing in really is nonzero, so that can't possibly happen. For the
context bank reset, yes, I am assuming that no complier will ever be
perverse enough to detect that cfg is not written after the NULL check
and immediately reallocate it to XZR for no good reason. I'd like to
think that assumption is going to hold for the reasonable scope of this
particular workaround, though.

Robin.

\
 
 \ /
  Last update: 2019-03-18 18:20    [W:0.100 / U:1.396 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site