On 11/06/2024 12:37 pm, Stefan Wahren wrote:
Am 11.06.24 um 13:08 schrieb Arnd Bergmann:
On Tue, Jun 11, 2024, at 12:47, Stefan Wahren wrote:
Am 10.06.24 um 12:25 schrieb Robin Murphy:
On 2024-06-10 10:24 am, Phil Elwell wrote:
On Mon, 10 Jun 2024 at 10:20, Arnd Bergmann <arnd@xxxxxxxx> wrote:
Arnd
vchiq sends partial cache lines at the start and of reads (as seen
from the ARM host) out of band, so the only misaligned DMA transfers
should be from ARM to VPU. This should not require a bounce buffer.
Hmm, indeed the dma_kmalloc_safe() takes into account that unaligned
DMA_TO_DEVICE does not need bouncing, so it would suggest that
something's off in what vchiq is asking for.
I'm available to debug this further, but i need more guidance here.
At least i extend the output for the error case:
- WARN_ON(len == 0);
- WARN_ON(i && (i != (dma_buffers - 1)) && (len &
~PAGE_MASK));
- WARN_ON(i && (addr & ~PAGE_MASK));
+ if (len == 0)
+ pr_warn_once("%s: sg_dma_len() == 0\n",
__func__);
+ else if (i && (i != (dma_buffers - 1)) && (len &
~PAGE_MASK))
+ pr_warn_once("%s: following block not page
aligned\n", __func__);
+ else if (i && (addr & ~PAGE_MASK)) {
+ pr_warn_once("%s: block %u, DMA address %pad
doesn't align with PAGE_MASK 0x%lx\n", __func__, i, &addr, PAGE_MASK);
+ pr_warn_once("sg_dma_is_swiotlb: %d, dma_flags:
%x\n", sg_dma_is_swiotlb(sg), sg->dma_flags);
+ }
Example result:
[ 84.180527] create_pagelist: block 1, DMA address 0x00000000f5f74800
doesn't align with PAGE_MASK 0xfffffffffffff000
[ 84.180553] sg_dma_is_swiotlb: 0, dma_flags: 0
Is this helpful?
It's interesting that this does not have the SG_DMA_SWIOTLB
flag set, as the theory so far was that an unaligned
user address is what caused this to bounce.
dma-direct doesn't use that flag (at least for now; potentially it might
offer a tiny micro-optimisation, but not enough to really matter). At
this point, len and pagelistinfo->dma_dir (as passed to dma_map_sg())
are what would have defined the bounce condition...
I think the most helpful bit of information that is
currently missing is the 'ubuf' and 'count' arguments
that got passed down from userspace into create_pagelist(),
to see what alignment they have in the failure case.
Here is my attempt:
if (len == 0)
pr_warn_once("%s: sg_dma_len() == 0\n", __func__);
else if (i && (i != (dma_buffers - 1)) && (len & ~PAGE_MASK))
pr_warn_once("%s: following block not page aligned\n",
__func__);
else if (i && (addr & ~PAGE_MASK)) {
pr_warn_once("%s: block %u, DMA address %pad doesn't align
with PAGE_MASK 0x%lx\n", __func__, i, &addr, PAGE_MASK);
pr_warn_once("sg_dma_is_swiotlb: %d, dma_flags: %x\n",
sg_dma_is_swiotlb(sg), sg->dma_flags);
pr_warn_once("type = %s\n", (type == PAGELIST_WRITE) ?
"PAGELIST_WRITE" : "PAGELIST_READ");
if (buf)
pr_warn_once("buf = %p, count = %zu\n", buf, count);
else
pr_warn_once("ubuf = %p, count = %zu\n", ubuf, count);
}
Output:
[ 66.184030] create_pagelist: block 1, DMA address 0x00000000f5fc7800
doesn't align with PAGE_MASK 0xfffffffffffff000
[ 66.184056] sg_dma_is_swiotlb: 0, dma_flags: 0
[ 66.184063] type = PAGELIST_READ
...so from the dma_dir assignment, this would imply DMA_FROM_DEVICE,
(indicating vchiq writing to RAM, CPU reading from RAM) which with the
unaligned address below would indeed cause a bounce. However that
appears to contradict what Phil said is supposed to happen :/
Thanks,
Robin
[ 66.184066] ubuf = 00000000266a70a7, count = 0
Arnd
