Re: DMA Mapping Error in ppc64

On Fri, 2018-03-23 at 07:41 -0500, Jared Bents wrote:

Thank you for the advice.  Looks like I get to try to rewrite the ath9k and ath10k drivers to use dma_alloc_coherent() instead of kmemdup() and dev_alloc_skb()

Euh no... dev_alloc_skb() is the right thing to do for receive
packets for a device driver.

The arch should be able to map that for DMA, even if include
bounce buffers via swiotlb.

Cheers,
Ben.

On Thu, Mar 22, 2018 at 8:19 PM, Oliver [off-list ref] wrote:

quoted

On Fri, Mar 23, 2018 at 1:37 AM, Jared Bents
[off-list ref] wrote:

quoted

Thank you for the response but unfortunately, it looks like I already
have that and it is being used.  To verify, I commented that out and
got the failure "dma_direct_alloc_coherent: No suitable zone for pfn
0xe0000".  Below is the code flow for function
ath10k_pci_hif_exchange_bmi_msg which is showing the first dma mapping
error.

ath10k_pci_hif_exchange_bmi_msg -> dma_map_single ->
dma_map_single_attrs -> swiotlb_map_page -> dma_capable (returns
false)


dma_capable is what reports the failure in that flow.

static inline bool dma_capable(struct device *dev, dma_addr_t addr, size_t size)
{
#ifdef CONFIG_SWIOTLB
    struct dev_archdata *sd = &dev->archdata;

   if (sd->max_direct_dma_addr && addr + size > sd->max_direct_dma_addr)
        return false;
#endif

    if (!dev->dma_mask)
        return false;

    return addr + size - 1 <= *dev->dma_mask;
}
Getting the below values:
addr = 1ee376218
size = 4
sd->max_direct_dma_addr = e0000000 which is I believe DMA window size (e0000000)

when executed sd->max_direct_dma_addr(e0000000) && addr(1ee376218) +
size(4) becomes e0000004 which is > sd->max_direct_dma_addr (e0000000)


So even though limit_zone_pfn(ZONE_DMA32, 1UL << (31 - PAGE_SHIFT)) is
being used in arch/powerpc/platforms/85xx/corenet_generic.c,

quoted

kmemdup(req, req_len, GFP_KERNEL) is returning an address that when
sent to dma_map_single(), results in a bad map.

You need to use (GFP_KERNEL | GFP_DMA32) to constrain the allocations
to ZONE_DMA32. Without that the kmemdup() will allocate from any zone
so you'll probably get an unmappable address.

That said, the driver probably shouldn't be using kmemdup() here.
DMA-API.txt pretty explicitly says that drivers should not assume that
dma_map_single() will work with arbitrary memory. It should be using
dma_alloc_coherent() or a dma pool here.

quoted

- Jared

On Wed, Mar 21, 2018 at 11:54 PM, Oliver [off-list ref] wrote:

quoted

On Thu, Mar 22, 2018 at 8:00 AM, Jared Bents
[off-list ref] wrote:

quoted

Hi all,

Apologies for the amount of information but we've been debugging this
for a while and I wanted to get what we are seeing captured as much as
possible.  We are a T1042 processor and have a total 8GB DDR and our
kernel version is fsl-sdk-v2.0-1703 (linux v4.1.35) as that is the
latest version supplied by NXP.

A while ago we ported from 32 bit to 64 bit.  Everything continued to
work except the ath10k module we have.  So as a first step, we checked
to see if an ath9k module also failed to work and it was also no
longer working.  The ath10k is working fine on a 32 bit system but
it's not working on 64 bit system as we are getting dma mapping errors
when trying to initialize the wifi modules.

pci_bus 0002:01: bus scan returning with max=01
pci_bus 0002:01: busn_res: [bus 01] end is updated to 01
pci_bus 0002:00: bus scan returning with max=01
ath10k_pci 0000:01:00.0: unable to get target info from device
ath10k_pci 0000:01:00.0: could not get target info (-5)
ath10k_pci 0000:01:00.0: could not probe fw (-5)
ath10k_pci 0001:01:00.0: Direct firmware load for
ath10k/cal-pci-0001:01:00.0.bin failed with error -2


First, we have tried the mainline kernel (v4.15)  to see if that would
fix the issue, it did not.  So I made a patch for the ath10k driver to
restrict to just GFP_DMA areas when allocating memory or creating
sk_buffs and have attached it.  The ath10k wifi modules now initialize
correctly but when I try to connect them and send traffic, they get a
DMA mapping error from the sk_buff that it receives from elsewhere in
the kernel.  So while the driver appears to be fixable with the patch,
the modules are still unusable due to data being sent to the driver
when ath10k_tx is called and it tries to dma map with the provided
skb.  Also, according to the ath10k mailing list, GFP_DMA is not
supposed to be used in general.  The error below is the same sort of
dma mapping error that is seen when initializing the modules without
the patch to OR with GFP_DMA.

ath10k_pci 0001:01:00.0: failed to transmit packet, dropping: -5


We asked on the ath10k mailing list if anyone else is having this
problem and no one else seems to have the issue but they are using
different architectures (ARM or X86). As a result, it does not seem to
be a driver issue to us but something within the PowerPC arch.  So we
dug a little deeper to try to find what addresses being mapped are
working and what address being mapped are not working.

We found that when the virtual address of data pointer (a member of
sk_buff) is above ~3.7 GB RAM address range then return address from
dma_map_single API is failed to validate in dma_mapping_error
function.

We also noticed that in a 64bit machine sometimes ping is working and
because of the virtual address is under ~3.7GAM RAM address range.  So
if we set mem=2048M in the bootargs, the ath10k module works
perfectly, however this isn't a real solution since it cuts our
available RAM from 8GB to 2GB.

I think there's a known issue with the freescale PCIe root complex
where it can't DMA beyond the 4GB mark. There's a workaround in
the form of limit_zone_pfn() which you can use to put the lower 4GB into
ZONE_DMA32 and allocate from there rather than ZONE_NORMAL.
For details of how to use it have a look at corenet_gen_setup_arch() in
arch/powerpc/platforms/85xx/corenet_generic.c

Hope that helps,
Oliver