Hi,
On 01/27/2015 11:37 PM, Johannes Berg wrote:
On Tue, 2015-01-27 at 21:39 -0800, Peter Oh wrote:
Ok, sure, but I/O ordering two different writes, and ensuring device
has seen a posted write, are related but different things, no?
yes, they are different and wmb guarantees both.
No, wmb() doesn't. I'd be very surprised if it had any side effect on
the PCI bus even on ARM. Read Documentation/memory-barriers.txt.
And to understand (PCI) posted writes, wikipedia helps:
http://en.wikipedia.org/wiki/Posted_write
I admit that I/O ordering and posted write are looked different in
theory and at glance since posted write could be related cache invalidate.
But wmb are still related to both.
As I addressed wmb uses dsb (in arm arch) and here is the description of
arm architecture.
* DSB drains write buffer.
* DSB is architecturally defined to include all cache, TLB and branch
prediction maintenance operations as well as explicit memory operations
These are the reasons why I mentioned wmb does both.
* captured from ARMv7 Architecture Manual
--- Notes ---
Historically, this operation was referred to as Drain Write Buffer or
Data Write Barrier (DWB). From ARMv6, these
names and the use of DWB were deprecated in favor of the new Data
Synchronization Barrier name and DSB
abbreviation. DSB better reflects the functionality provided from ARMv6,
because DSB is architecturally defined
to include all cache, TLB and branch prediction maintenance operations
as well as explicit memory operations
--- A DSB completes when: ---
? all explicit memory accesses that are observed by Pe before the DSB is
executed, are of the required access
types, and are from observers in the same required shareability domain
as Pe, are complete for the set of
observers in the required shareability domain.
? if the required accesses types of the DSB is reads and writes, all
cache and branch predictor maintenance
operations issued by Pe before the DSB are complete for the required
shareability domain.
? if the required accesses types of the DSB is reads and writes, all TLB
maintenance operations issued by Pe
before the DSB are complete for the required shareability domain.
--------------
Furthermore this is the comparison of the compiled assembly code between
ath10k_pci_read32 and wmb.
ath10k_pci_read32()
bac: e5932008 ldr r2, [r3, #8]
bb0: f57ff04f dsb sy
bb4: e2883d52 add r3, r8, #5248 ; 0x1480
bb8: e283303c add r3, r3, #60 ; 0x3c
bbc: e593300c ldr r3, [r3, #12]
bc0: e2833a09 add r3, r3, #36864 ; 0x9000
wmb();
b9c: f57ff04e dsb st
ath10k_pci_read32 does register operation except dsb and there is no
cache invalidate related commands.
So that if wmb is not enough for the purpose then ath10k_pci_read32 is
also not enough for that.
Also refer the section "ACQUIRES VS I/O ACCESSES" in memory-barriers.txt.
It gives an example with PCI bridge and introduces readl as an
alternative method to mmiowb which weaker form of wmb.
Please give your opinion.
johannes
_______________________________________________
ath10k mailing list
ath10k@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/ath10k
Regards,
Peter
--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
