CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

Russell, Catalin,

In arch/arm/include/asm/io.h, we have the following piece of code:

#ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE
#define __iormb()               rmb()
#define __iowmb()               wmb()
#else
#define __iormb()               do { } while (0)
#define __iowmb()               do { } while (0)
#endif

#define readb(c)                ({ u8  __v = readb_relaxed(c); 
__iormb(); __v; })
#define readw(c)                ({ u16 __v = readw_relaxed(c); 
__iormb(); __v; })
#define readl(c)                ({ u32 __v = readl_relaxed(c); 
__iormb(); __v; })

#define writeb(v,c)             ({ __iowmb(); writeb_relaxed(v,c); })
#define writew(v,c)             ({ __iowmb(); writew_relaxed(v,c); })
#define writel(v,c)             ({ __iowmb(); writel_relaxed(v,c); })

If I'm not missing some magic, this would mean that 
"CONFIG_ARM_DMA_MEM_BUFFERABLE" determines if readl(s)/writel(s) get to 
have a built in mb() or not.

The rest of the emails is under the assumption that my statement above 
is true.

This seems like a real bad thing to do.

There are so many other drivers that don't use or care about DMA and 
might still want to ensure some ordering constraints between their 
readl(s)/writel(s). They can't depend on readl/writel taking care of it 
for them since their code could be used in a kernel configuration that 
doesn't enable this config.

Firstly, I don't know how many people noticed this and realize they 
can't depend on readl/writel to take care of the mb()s for them. Seems 
like an unnecessary encouragement to make mistakes when it didn't need 
to be so.

Secondly, even if they realize they have to take care of it, they will 
have to continue using mb()s in to force ordering between their 
reads/writes. So, are we depending on the compiler to optimize these 
extra mb() out in the case where the config is enabled? I'm not sure it 
will be able to optimize out the extra mb()s in all cases.

Please let me know if I'm missing something and there is a good reason 
for writing the code as it is today. If there is not much of a reason 
for this, do you have any objections if I send a patch to split out the 
"readl/writel with built in mb()" as a separate config and 'select'ing 
it in ARM_DMA_MEM_BUFFERABLE?

Thanks,
Saravana
-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Arnd Bergmann]

On Wednesday 02 March 2011 02:23:15 Saravana Kannan wrote:
> There are so many other drivers that don't use or care about DMA and 
> might still want to ensure some ordering constraints between their 
> readl(s)/writel(s). They can't depend on readl/writel taking care of it 
> for them since their code could be used in a kernel configuration that 
> doesn't enable this config.

What exactly are the ordering constraints, do you need the full
dmb() for readl() and dsb() for writel(), or just a compiler barrier?

I think we need the barrier() even for the relaxed variant, but 
that is fairly lightweight. What would be a reason to have more?

> Firstly, I don't know how many people noticed this and realize they 
> can't depend on readl/writel to take care of the mb()s for them. Seems 
> like an unnecessary encouragement to make mistakes when it didn't need 
> to be so.
> 
> Secondly, even if they realize they have to take care of it, they will 
> have to continue using mb()s in to force ordering between their 
> reads/writes. So, are we depending on the compiler to optimize these 
> extra mb() out in the case where the config is enabled? I'm not sure it 
> will be able to optimize out the extra mb()s in all cases.

The compiler certainly won't merge multiple inline assembly statements,
but multiple barrier() statements don't make it worse than just one.
barrier() just forces accessing variables from memory that could otherwise
be kept in registers.

The other problems we still need to fix are the complete absence of
barriers in inb()/outb() style accessors, which are meant to be stricter
than readl()/writel(), and the fact that we rely on undefined behavior
in gcc regarding the atomicity of multi-byte accesses, as we recently
found out when a new gcc turned a readl into byte accesses.

	Arnd

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

On 03/02/2011 12:23 AM, Arnd Bergmann wrote:
> On Wednesday 02 March 2011 02:23:15 Saravana Kannan wrote:
>> There are so many other drivers that don't use or care about DMA and
>> might still want to ensure some ordering constraints between their
>> readl(s)/writel(s). They can't depend on readl/writel taking care of it
>> for them since their code could be used in a kernel configuration that
>> doesn't enable this config.
>
> What exactly are the ordering constraints, do you need the full
> dmb() for readl() and dsb() for writel(), or just a compiler barrier?

I wasn't referring to a compiler barrier. I'm referring to one of the 
real barrier instructions.

> I think we need the barrier() even for the relaxed variant, but
> that is fairly lightweight. What would be a reason to have more?

Please see my response to Russell in other other email. I think it 
should answer your question by clarifying my point.

>> Firstly, I don't know how many people noticed this and realize they
>> can't depend on readl/writel to take care of the mb()s for them. Seems
>> like an unnecessary encouragement to make mistakes when it didn't need
>> to be so.
>>
>> Secondly, even if they realize they have to take care of it, they will
>> have to continue using mb()s in to force ordering between their
>> reads/writes. So, are we depending on the compiler to optimize these
>> extra mb() out in the case where the config is enabled? I'm not sure it
>> will be able to optimize out the extra mb()s in all cases.
>
> The compiler certainly won't merge multiple inline assembly statements,
> but multiple barrier() statements don't make it worse than just one.
> barrier() just forces accessing variables from memory that could otherwise
> be kept in registers.

Yes, I was aware of how compiler barriers work and how multiple 
consecutive compiler barriers are the same as one compiler barrier. The 
email was about the real barrier instructions.

> The other problems we still need to fix are the complete absence of
> barriers in inb()/outb() style accessors, which are meant to be stricter
> than readl()/writel(), and the fact that we rely on undefined behavior
> in gcc regarding the atomicity of multi-byte accesses, as we recently
> found out when a new gcc turned a readl into byte accesses.

I haven't gotten that far :-) I really appreciate response!

Thanks,
Saravana
-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Russell King - ARM Linux]

On Tue, Mar 01, 2011 at 05:23:15PM -0800, Saravana Kannan wrote:
> If I'm not missing some magic, this would mean that  
> "CONFIG_ARM_DMA_MEM_BUFFERABLE" determines if readl(s)/writel(s) get to  
> have a built in mb() or not.

You're missing that CONFIG_ARM_DMA_MEM_BUFFERABLE not only changes
readl/writel but also the type for DMA coherent memory from strongly
ordered to memory, non-cacheable.

The barriers are required to ensure that reads and writes to DMA
coherent memory are visible to the DMA device before the write
completes, and any value read from DMA coherent memory will not
bypass a read from a DMA device.

The barriers in the IO macros have nothing to do with whether reads/writes
to normal cacheable memory are visible to DMA devices.  That is what the
streaming DMA API is for.

In any case, the IO macros are always ordered with respect to other
device writes irrespective of CONFIG_ARM_DMA_MEM_BUFFERABLE.

> There are so many other drivers that don't use or care about DMA and  
> might still want to ensure some ordering constraints between their  
> readl(s)/writel(s). They can't depend on readl/writel taking care of it  
> for them since their code could be used in a kernel configuration that  
> doesn't enable this config.

The majority of device drivers don't need ordering on their IO macros.
However, Linus refuses to introduce relaxed IO ordering to the kernel,
saying that architectures must reflect the x86 behaviour as much as
possible.

So, we're stuck with device drivers which use readl/writel both where
they don't need the ordering constraints _and_ where they do need the
ordering constraints.  That means we must provide the ordering in these
macros to ensure proper system functioning.

> Firstly, I don't know how many people noticed this and realize they  
> can't depend on readl/writel to take care of the mb()s for them. Seems  
> like an unnecessary encouragement to make mistakes when it didn't need  
> to be so.

I think you misunderstand what's going on.  IO accesses are always ordered
with respect to themselves.  The barriers are there to ensure ordering
between DMA coherent memory (normal non-cached memory) and IO accesses
(device).

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

Sorry that it took a while for me to get back. Had to contact our 
resident ARM expert to reconfirm my points and gather all the data in an 
externally available format.

On 03/02/2011 12:39 AM, Russell King - ARM Linux wrote:
> On Tue, Mar 01, 2011 at 05:23:15PM -0800, Saravana Kannan wrote:
>> If I'm not missing some magic, this would mean that
>> "CONFIG_ARM_DMA_MEM_BUFFERABLE" determines if readl(s)/writel(s) get to
>> have a built in mb() or not.
>
> You're missing that CONFIG_ARM_DMA_MEM_BUFFERABLE not only changes
> readl/writel but also the type for DMA coherent memory from strongly
> ordered to memory, non-cacheable.

Yeah, I noticed that after I sent the email, but my questions in my 
previous email still remains valid. See below.

> The barriers are required to ensure that reads and writes to DMA
> coherent memory are visible to the DMA device before the write
> completes, and any value read from DMA coherent memory will not
> bypass a read from a DMA device.
>
> The barriers in the IO macros have nothing to do with whether reads/writes
> to normal cacheable memory are visible to DMA devices.  That is what the
> streaming DMA API is for.
>
> In any case, the IO macros are always ordered with respect to other
> device writes irrespective of CONFIG_ARM_DMA_MEM_BUFFERABLE.

<snip>

> I think you misunderstand what's going on.  IO accesses are always ordered
> with respect to themselves.  The barriers are there to ensure ordering
> between DMA coherent memory (normal non-cached memory) and IO accesses
> (device).

Unfortunately this is not correct. The ARM spec doesn't guarantee that 
all IO accesses should be ordered with respect to themselves. It only 
requires that the ordering should be guaranteed at least within a 1KB 
region.

You can find this info in ARMv7 ARM spec[1] named 
"DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on 
page A3-45. There is a para that goes:

"Accesses must arrive at any particular memory-mapped peripheral or 
block of memory in program order, that is, A1 must arrive before A2. 
There are no ordering restrictions about when accesses arrive at 
different peripherals or blocks of memory, provided that the accesses 
follow the general ordering rules given in this section."

And the most critical point is hidden in a comment that goes:
"The size of a memory mapped peripheral, or a block of memory, is 
IMPLEMENTATION DEFINED, but is not smaller than 1KByte."

I guess most of the confusion is due to the ARM spec not being very 
obvious about the 1KB limitation.

So, going back to my point, I think it's wrong for 
CONFIG_ARM_DMA_MEM_BUFFERABLE to control how stuff unrelated to DMA behaves.

I have also encountered a few people who kept went "but readl/writel was 
recently changed to add mem barriers, so we can all remove the mb()s in 
our driver (unrelated to DMA) code". That would have made their code 
incorrect for two reasons:
1. readl/writel doesn't always have a mem barrier because of config that 
can be turned off.
2. In cases where readl/writel didn't have mb(), there is not enough 
ordering guarantee without an explicit mb().

I think as a community, we should stop saying that readl/writel ensures 
ordering with respect to all IO accesses. It doesn't even guarantee 
ordering within the same device (when their register regions are > 1KB).

After reading the above, please let me know if a patch to decouple the 
"readl/writel with builtin mb()" from CONFIG_ARM_DMA_MEM_BUFFERABLE 
would be accepted. If so, I can go ahead and send it out soon.

[1] 
http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0406b/index.html

Thanks,
Saravana

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Catalin Marinas]

On Thu, 2011-03-03 at 07:49 +0000, Saravana Kannan wrote:
> On 03/02/2011 12:39 AM, Russell King - ARM Linux wrote:
> > On Tue, Mar 01, 2011 at 05:23:15PM -0800, Saravana Kannan wrote:
> >> If I'm not missing some magic, this would mean that
> >> "CONFIG_ARM_DMA_MEM_BUFFERABLE" determines if readl(s)/writel(s) get to
> >> have a built in mb() or not.
[...]
> > I think you misunderstand what's going on.  IO accesses are always ordered
> > with respect to themselves.  The barriers are there to ensure ordering
> > between DMA coherent memory (normal non-cached memory) and IO accesses
> > (device).
> 
> Unfortunately this is not correct. The ARM spec doesn't guarantee that
> all IO accesses should be ordered with respect to themselves. It only
> requires that the ordering should be guaranteed at least within a 1KB
> region.

That's because the CPU does not have control of the delays on various
buses. But a device connected to the same bus receives the accesses in
order.

> And the most critical point is hidden in a comment that goes:
> "The size of a memory mapped peripheral, or a block of memory, is
> IMPLEMENTATION DEFINED, but is not smaller than 1KByte."
> 
> I guess most of the confusion is due to the ARM spec not being very
> obvious about the 1KB limitation.

What that means is that the hardware shouldn't have two different buses
(possibly with different delays) within a 1KB range.

Even if accesses to all peripherals are ordered, you still cannot
guarantee that a writel() would change the state of a device (and that's
specific to all architectures). Sometimes if you want to make sure the
device state changed you need a readl() back.

> So, going back to my point, I think it's wrong for
> CONFIG_ARM_DMA_MEM_BUFFERABLE to control how stuff unrelated to DMA behaves.

In an ideal world, all driver authors know what memory ordering is and
add the necessary barriers. But since that's not the case, the only way
to get ordering between Normal Non-cacheable access (DMA buffer) and
Device access (via writel) is to add the mb() in the I/O accessors.

This has been discussed at length on several occasions on linux-arch and
LKML. We don't have other solution since adding barriers to drivers
wasn't found feasible by others. Of course, if you can optimised your
driver to use the relaxed accessors and add explicit barriers.

> I have also encountered a few people who kept went "but readl/writel was
> recently changed to add mem barriers, so we can all remove the mb()s in
> our driver (unrelated to DMA) code". 

But why did they have those barriers around I/O accessors in the first
place? As you say, nothing related to DMA. If you access two different
devices and want to ensure an ordering of the state changes, I doubt a
barrier would help. Most likely you need a read back from the device.

> That would have made their code incorrect for two reasons:
> 1. readl/writel doesn't always have a mem barrier because of config that
> can be turned off.
> 2. In cases where readl/writel didn't have mb(), there is not enough
> ordering guarantee without an explicit mb().

See my comment above, what do they try to achieve by using mb() around
already ordered I/O accessors?

> I think as a community, we should stop saying that readl/writel ensures
> ordering with respect to all IO accesses. It doesn't even guarantee
> ordering within the same device (when their register regions are > 1KB).

It definitely guarantees ordering to the same device. The 1KB is a
minimum limit but there is no upper bound (and it should definitely
cover a single device).

> After reading the above, please let me know if a patch to decouple the
> "readl/writel with builtin mb()" from CONFIG_ARM_DMA_MEM_BUFFERABLE
> would be accepted. If so, I can go ahead and send it out soon.

No, I don't think this would be acceptable. I can point you to past
discussion where Linus stated that Normal Non-cacheable memory and I/O
accesses should be ordered.

What you can do actually is make sure that all architectures support the
relaxed accessors and change the drivers to use these instead.

-- 
Catalin

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

Discussed this further with our resident ARM expert. The rest of the 
email is me understanding his comments and rephrasing them to suit this 
thread.

On 03/03/2011 02:11 AM, Catalin Marinas wrote:
> On Thu, 2011-03-03 at 07:49 +0000, Saravana Kannan wrote:
>> On 03/02/2011 12:39 AM, Russell King - ARM Linux wrote:
>>> On Tue, Mar 01, 2011 at 05:23:15PM -0800, Saravana Kannan wrote:
>>>> If I'm not missing some magic, this would mean that
>>>> "CONFIG_ARM_DMA_MEM_BUFFERABLE" determines if readl(s)/writel(s) get to
>>>> have a built in mb() or not.
> [...]
>>> I think you misunderstand what's going on.  IO accesses are always ordered
>>> with respect to themselves.  The barriers are there to ensure ordering
>>> between DMA coherent memory (normal non-cached memory) and IO accesses
>>> (device).
>>
>> Unfortunately this is not correct. The ARM spec doesn't guarantee that
>> all IO accesses should be ordered with respect to themselves. It only
>> requires that the ordering should be guaranteed at least within a 1KB
>> region.
>
> That's because the CPU does not have control of the delays on various
> buses. But a device connected to the same bus receives the accesses in
> order.

 From the ARM spec, at the least, we have established that IO accesses 
are not guaranteed to be ordered with respect to all IO accesses. So, I 
kindly request that we no longer perpetuate the incorrect statement that 
IO accesses are ordered wrt to all IO accesses.

The only continuing disagreement is on whether the minimum limit is 1KB 
or "the entire address space of a given device". More on that below.

>> And the most critical point is hidden in a comment that goes:
>> "The size of a memory mapped peripheral, or a block of memory, is
>> IMPLEMENTATION DEFINED, but is not smaller than 1KByte."
>>
>> I guess most of the confusion is due to the ARM spec not being very
>> obvious about the 1KB limitation.
>
> What that means is that the hardware shouldn't have two different buses
> (possibly with different delays) within a 1KB range.

Rephrasing comments from our resident expert:
This 1KB limit is not really tied to there being different "speeds" on 
the buses.  Rather it is a fundamental topological issue.  It has to do 
with how the "core" interleaves the interconnect. If the interleave is 
at the 1KB level, there are two different bus interfaces at the 1KB 
interleave.  So even if the downstream "device" has a single *input* 
bus, the processor and the interconnect have two different paths to get 
to that device, if the address space spans 1KB. So there is no 
guaranteed order across the 1KB boundary, even if in concept the two 
sides of the boundary are talking to the same "device".

> Even if accesses to all peripherals are ordered, you still cannot
> guarantee that a writel() would change the state of a device (and that's
> specific to all architectures). Sometimes if you want to make sure the
> device state changed you need a readl() back.

Yes, a device might take much longer to change the state or process the 
command after the write completes. I realize that has nothing to do with 
memory barriers.

>> So, going back to my point, I think it's wrong for
>> CONFIG_ARM_DMA_MEM_BUFFERABLE to control how stuff unrelated to DMA behaves.
>
> In an ideal world, all driver authors know what memory ordering is and
> add the necessary barriers. But since that's not the case, the only way
> to get ordering between Normal Non-cacheable access (DMA buffer) and
> Device access (via writel) is to add the mb() in the I/O accessors.
>
> This has been discussed at length on several occasions on linux-arch and
> LKML. We don't have other solution since adding barriers to drivers
> wasn't found feasible by others. Of course, if you can optimised your
> driver to use the relaxed accessors and add explicit barriers.

I think my comment for clean up was misunderstood to be more that what I 
intend. I will address this at the end of this email.

>> I have also encountered a few people who kept went "but readl/writel was
>> recently changed to add mem barriers, so we can all remove the mb()s in
>> our driver (unrelated to DMA) code".
>
> But why did they have those barriers around I/O accessors in the first
> place? As you say, nothing related to DMA. If you access two different
> devices and want to ensure an ordering of the state changes, I doubt a
> barrier would help. Most likely you need a read back from the device.

These barriers were needed because the device spanned more than 1KB of 
address space. See below.

>> That would have made their code incorrect for two reasons:
>> 1. readl/writel doesn't always have a mem barrier because of config that
>> can be turned off.
>> 2. In cases where readl/writel didn't have mb(), there is not enough
>> ordering guarantee without an explicit mb().
>
> See my comment above, what do they try to achieve by using mb() around
> already ordered I/O accessors?
>
>> I think as a community, we should stop saying that readl/writel ensures
>> ordering with respect to all IO accesses. It doesn't even guarantee
>> ordering within the same device (when their register regions are>  1KB).
 >
> It definitely guarantees ordering to the same device. The 1KB is a
> minimum limit but there is no upper bound (and it should definitely
> cover a single device).

Rephrasing comments from our resident expert:
That is not true.  It only guarantees ordering to the 1KB boundary. 
That is the implementation-defined limit of what constitutes a "device". 
  Any "device" which crosses that boundary is viewed as two separate 
devices from the core's perspective, and there is no guaranteed order, 
absent an explicit barrier operation.

The 1KB is the minimum size that the architecture allows the hardware to 
establish as its limit for guaranteed ordering.

It is not true that the hardware is obligated to go *above* that limit 
just because a given "device" happens to have address space that crosses 
the boundary.  As soon as you have two separate interconnect ports and 
thus two separate paths to get to the device, because it crossed the 1KB 
boundary, then you have the potential of being out-of-order.  Software 
is obligated to use explicit barriers on accesses that cross that 
threshold, regardless of whether *physically* the accesses end up at the 
same downstream "device".

>> After reading the above, please let me know if a patch to decouple the
>> "readl/writel with builtin mb()" from CONFIG_ARM_DMA_MEM_BUFFERABLE
>> would be accepted. If so, I can go ahead and send it out soon.
>
> No, I don't think this would be acceptable. I can point you to past
> discussion where Linus stated that Normal Non-cacheable memory and I/O
> accesses should be ordered.
>
> What you can do actually is make sure that all architectures support the
> relaxed accessors and change the drivers to use these instead.

There were a couple of reason for starting this thread:
* Check if my understanding of CONFIG_ARM_DMA_MEM_BUFFERABLE's effect on 
readl/writl were correct and if I was missing anything.
Result: Yes, I was missing some stuff about the config affecting DMA 
memory mapping, but I did understand the point about having mb()s in 
readl/writel correctly.

* Bring to attention that the ARM spec doesn't guarantee all IO accesses 
are ordered wrt each other and have the community update it's 
understanding of IO access ordering.
Result: It think we agree on this update. At least Russell seems to. The 
ARM spec is very clear on this. The only part we disagree on is the 1KB 
limit.

* If my understanding about how CONFIG_ARM_DMA_MEM_BUFFERABLE affects 
readl/writel was correct and we agree on the IO access ordering update, 
then it would show that mb()s matter for more than just DMA. In which 
case, I wanted to split that single config into two and clean up the 
names while still leaving CONFIG_ARM_DMA_MEM_BUFFERABLE functionally the 
same.
Result?: After the above discussion, I hope the community agrees that 
it's reasonable to reorganize and rename the config to make the config 
more understandable and less confusing to developers. Less confusion == 
less bugs in my opinion. I will send out a patch sometime this week to 
show what I mean.

But modifying the behaviour of DMA was not my intention. I will of 
course read the discussion you mention (will ping you if I can't find 
it) to get a better understanding of the DMA stuff.

Catalin,

To summarize,
Our internal ARM expert continues to state that the 1KB is the minimum 
limit even within a single device. Judging his expertise based on my 
previous interactions with him, I'm inclined to believe him. I think at 
the least we should give him the benefit of the doubt. Looks like a 
definitive way to settle this disagreement would be to check with the 
author of those updates/comment I referred to in the ARMv7 ARM.

If the 1KB limit is correct, it would certainly help improving the 
correctness of drivers for the ARM kernel. If the 1KB limit is wrong, we 
don't lose anything.

So, if you don't mind, could you please check with the author of those 
updates and comments in the ARMv7 ARM?

Thanks,
Saravana

-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Russell King - ARM Linux]

On Wed, Mar 02, 2011 at 11:49:47PM -0800, Saravana Kannan wrote:
>> I think you misunderstand what's going on.  IO accesses are always ordered
>> with respect to themselves.  The barriers are there to ensure ordering
>> between DMA coherent memory (normal non-cached memory) and IO accesses
>> (device).
>
> Unfortunately this is not correct. The ARM spec doesn't guarantee that  
> all IO accesses should be ordered with respect to themselves. It only  
> requires that the ordering should be guaranteed at least within a 1KB  
> region.
>
> You can find this info in ARMv7 ARM spec[1] named  
> "DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on  
> page A3-45. There is a para that goes:
>
> "Accesses must arrive at any particular memory-mapped peripheral or  
> block of memory in program order, that is, A1 must arrive before A2.  
> There are no ordering restrictions about when accesses arrive at  
> different peripherals or blocks of memory, provided that the accesses  
> follow the general ordering rules given in this section."

That is news to me.  My DDI0406B does not have this paragraph, so it's
something that ARM has sprung upon us without telling *anyone* about it.
It's not unreasonable or even unexpected.  That is exactly the same
condition which applies on buses like PCI due to write posting on bridges
downstream of the CPU, and issuing memory barriers will not help with
that.

Consider two PCI devices each behind their own P2P bridge.  Device A's
bridge is really lazy and takes time to empty its write post buffer.
Device B's bridge is really fast at getting writes.  If you write to
device A then device B, they'll arrive at device B before device A.
Again, let me stress that memory barriers will not allow you to solve
this problem.

The only way to solve this is to read back from the device, because reads
to device memory can not bypass writes to device memory, otherwise the
system is unpredictable.  With PCI, it's recommended to read back from
the exact same address which you've written to _provided_ there's no
side effects from doing so.  If there are, you have to find some other
solution to it.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

On 03/03/2011 02:24 AM, Russell King - ARM Linux wrote:
> On Wed, Mar 02, 2011 at 11:49:47PM -0800, Saravana Kannan wrote:
>>> I think you misunderstand what's going on.  IO accesses are always ordered
>>> with respect to themselves.  The barriers are there to ensure ordering
>>> between DMA coherent memory (normal non-cached memory) and IO accesses
>>> (device).
>>
>> Unfortunately this is not correct. The ARM spec doesn't guarantee that
>> all IO accesses should be ordered with respect to themselves. It only
>> requires that the ordering should be guaranteed at least within a 1KB
>> region.
>>
>> You can find this info in ARMv7 ARM spec[1] named
>> "DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on
>> page A3-45. There is a para that goes:
>>
>> "Accesses must arrive at any particular memory-mapped peripheral or
>> block of memory in program order, that is, A1 must arrive before A2.
>> There are no ordering restrictions about when accesses arrive at
>> different peripherals or blocks of memory, provided that the accesses
>> follow the general ordering rules given in this section."
>
> That is news to me.  My DDI0406B does not have this paragraph, so it's
> something that ARM has sprung upon us without telling *anyone* about it.
> It's not unreasonable or even unexpected.  That is exactly the same
> condition which applies on buses like PCI due to write posting on bridges
> downstream of the CPU, and issuing memory barriers will not help with
> that.

While the PCI stuff is true, as you say, it's not related to mb()s. The 
mb()s matter to the point of getting the writes to the  intended 
"devices addresses" in the program order. What happens after that is a 
separate issue.

So, going back to the discussion of mb()s and readl/writel (and 
variations), I think we should no longer state the all IO accesses are 
ordered wrt each other. Are we in agreement on this?

Thanks,
Saravana
-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Russell King - ARM Linux]

On Tue, Mar 08, 2011 at 08:58:20PM -0800, Saravana Kannan wrote:
> On 03/03/2011 02:24 AM, Russell King - ARM Linux wrote:
>> On Wed, Mar 02, 2011 at 11:49:47PM -0800, Saravana Kannan wrote:
>>>> I think you misunderstand what's going on.  IO accesses are always ordered
>>>> with respect to themselves.  The barriers are there to ensure ordering
>>>> between DMA coherent memory (normal non-cached memory) and IO accesses
>>>> (device).
>>>
>>> Unfortunately this is not correct. The ARM spec doesn't guarantee that
>>> all IO accesses should be ordered with respect to themselves. It only
>>> requires that the ordering should be guaranteed at least within a 1KB
>>> region.
>>>
>>> You can find this info in ARMv7 ARM spec[1] named
>>> "DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on
>>> page A3-45. There is a para that goes:
>>>
>>> "Accesses must arrive at any particular memory-mapped peripheral or
>>> block of memory in program order, that is, A1 must arrive before A2.
>>> There are no ordering restrictions about when accesses arrive at
>>> different peripherals or blocks of memory, provided that the accesses
>>> follow the general ordering rules given in this section."
>>
>> That is news to me.  My DDI0406B does not have this paragraph, so it's
>> something that ARM has sprung upon us without telling *anyone* about it.
>> It's not unreasonable or even unexpected.  That is exactly the same
>> condition which applies on buses like PCI due to write posting on bridges
>> downstream of the CPU, and issuing memory barriers will not help with
>> that.
>
> While the PCI stuff is true, as you say, it's not related to mb()s. The  
> mb()s matter to the point of getting the writes to the  intended  
> "devices addresses" in the program order. What happens after that is a  
> separate issue.
>
> So, going back to the discussion of mb()s and readl/writel (and  
> variations), I think we should no longer state the all IO accesses are  
> ordered wrt each other. Are we in agreement on this?

No, because you clearly haven't understood the point I made.

I still believe you are wrong on this point.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Saravana Kannan]

On Wed, March 9, 2011 12:05 am, Russell King - ARM Linux wrote:
> On Tue, Mar 08, 2011 at 08:58:20PM -0800, Saravana Kannan wrote:
>> On 03/03/2011 02:24 AM, Russell King - ARM Linux wrote:
>>> On Wed, Mar 02, 2011 at 11:49:47PM -0800, Saravana Kannan wrote:
>>>>> I think you misunderstand what's going on.  IO accesses are always
>>>>> ordered
>>>>> with respect to themselves.  The barriers are there to ensure
>>>>> ordering
>>>>> between DMA coherent memory (normal non-cached memory) and IO
>>>>> accesses
>>>>> (device).
>>>>
>>>> Unfortunately this is not correct. The ARM spec doesn't guarantee that
>>>> all IO accesses should be ordered with respect to themselves. It only
>>>> requires that the ordering should be guaranteed at least within a 1KB
>>>> region.
>>>>
>>>> You can find this info in ARMv7 ARM spec[1] named
>>>> "DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on
>>>> page A3-45. There is a para that goes:
>>>>
>>>> "Accesses must arrive at any particular memory-mapped peripheral or
>>>> block of memory in program order, that is, A1 must arrive before A2.
>>>> There are no ordering restrictions about when accesses arrive at
>>>> different peripherals or blocks of memory, provided that the accesses
>>>> follow the general ordering rules given in this section."
>>>
>>> That is news to me.  My DDI0406B does not have this paragraph, so it's
>>> something that ARM has sprung upon us without telling *anyone* about
>>> it.
>>> It's not unreasonable or even unexpected.  That is exactly the same
>>> condition which applies on buses like PCI due to write posting on
>>> bridges
>>> downstream of the CPU, and issuing memory barriers will not help with
>>> that.
>>
>> While the PCI stuff is true, as you say, it's not related to mb()s. The
>> mb()s matter to the point of getting the writes to the  intended
>> "devices addresses" in the program order. What happens after that is a
>> separate issue.
>>
>> So, going back to the discussion of mb()s and readl/writel (and
>> variations), I think we should no longer state the all IO accesses are
>> ordered wrt each other. Are we in agreement on this?
>
> No, because you clearly haven't understood the point I made.
>
> I still believe you are wrong on this point.

I'm not going to pretend to be a PCI expert, but I think the ARMv7 ARM
text I quoted makes it pretty clear that all IO accesses are not ordered
wrt each other. So, why do you still disagree on that point?

-Saravana
-- 
Sent by an employee of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.

CONFIG_ARM_DMA_MEM_BUFFERABLE and readl/writel weirdness

[Russell King - ARM Linux]

On Wed, Mar 09, 2011 at 01:32:15AM -0800, Saravana Kannan wrote:
> On Wed, March 9, 2011 12:05 am, Russell King - ARM Linux wrote:
> > On Tue, Mar 08, 2011 at 08:58:20PM -0800, Saravana Kannan wrote:
> >> On 03/03/2011 02:24 AM, Russell King - ARM Linux wrote:
> >>> On Wed, Mar 02, 2011 at 11:49:47PM -0800, Saravana Kannan wrote:
> >>>>> I think you misunderstand what's going on.  IO accesses are always
> >>>>> ordered
> >>>>> with respect to themselves.  The barriers are there to ensure
> >>>>> ordering
> >>>>> between DMA coherent memory (normal non-cached memory) and IO
> >>>>> accesses
> >>>>> (device).
> >>>>
> >>>> Unfortunately this is not correct. The ARM spec doesn't guarantee that
> >>>> all IO accesses should be ordered with respect to themselves. It only
> >>>> requires that the ordering should be guaranteed at least within a 1KB
> >>>> region.
> >>>>
> >>>> You can find this info in ARMv7 ARM spec[1] named
> >>>> "DDI0406B_arm_architecture_reference_manual_errata_markup_8_0.pdf", on
> >>>> page A3-45. There is a para that goes:
> >>>>
> >>>> "Accesses must arrive at any particular memory-mapped peripheral or
> >>>> block of memory in program order, that is, A1 must arrive before A2.
> >>>> There are no ordering restrictions about when accesses arrive at
> >>>> different peripherals or blocks of memory, provided that the accesses
> >>>> follow the general ordering rules given in this section."
> >>>
> >>> That is news to me.  My DDI0406B does not have this paragraph, so it's
> >>> something that ARM has sprung upon us without telling *anyone* about
> >>> it.
> >>> It's not unreasonable or even unexpected.  That is exactly the same
> >>> condition which applies on buses like PCI due to write posting on
> >>> bridges
> >>> downstream of the CPU, and issuing memory barriers will not help with
> >>> that.
> >>
> >> While the PCI stuff is true, as you say, it's not related to mb()s. The
> >> mb()s matter to the point of getting the writes to the  intended
> >> "devices addresses" in the program order. What happens after that is a
> >> separate issue.
> >>
> >> So, going back to the discussion of mb()s and readl/writel (and
> >> variations), I think we should no longer state the all IO accesses are
> >> ordered wrt each other. Are we in agreement on this?
> >
> > No, because you clearly haven't understood the point I made.
> >
> > I still believe you are wrong on this point.
> 
> I'm not going to pretend to be a PCI expert, but I think the ARMv7 ARM
> text I quoted makes it pretty clear that all IO accesses are not ordered
> wrt each other. So, why do you still disagree on that point?

Because you're entirely missing my point, and you don't need to be a PCI
expert to understand it.  You're just not taking the time to think about
it because it says "PCI" and I guess you're not interested in PCI.

The point I made is that even on a strongly ordered CPU, accesses to PCI
devices on different busses can be _out_ _of_ _order_.  This is something
we expect.  This is something we deal with by reading back from the device
after a write _if_ the driver deems that the relative ordering matters.

That's a decision for the device driver to make - and the _only_ way to
do that is by reading back from the device.  Adding memory barriers do
*not* help, and if your 'ARM expert' thinks it does then he's wrong.

I will not change the CONFIG_ARM_DMA_MEM_BUFFERABLE text - this config
option is about the type of memory used for the DMA memory, and not
about IO access ordering.  To make it so confuses the two issues.