On 2020-11-17 16:53, Vidya Sagar wrote:
Set DMA mask to 32-bit while allocating the MSI target address so that
the address is usable for both 32-bit and 64-bit MSI capable devices.
Throw a warning if it fails to set the mask to 32-bit to alert that
devices that are only 32-bit MSI capable may not work properly.
This is slightly wacky, but no more so than the rest of the not-DMA
shenanigans here... Ultimately it probably is the least-worst way to
avoid the issue, so in terms of functionality,
Reviewed-by: Robin Murphy <robin.murphy@xxxxxxx>
Signed-off-by: Vidya Sagar <vidyas@xxxxxxxxxx>
---
Given the other patch that I've pushed to the MSI sub-system
http://patchwork.ozlabs.org/project/linux-pci/patch/20201117145728.4516-1-vidyas@xxxxxxxxxx/
which is going to catch any mismatch between MSI capability (32-bit) of the
device and system's inability to allocate the required MSI target address,
I'm not sure how much sense is this patch going to be make. But, I can
certainly say that if the memory allocation mechanism gives the addresses
from 64-bit pool by default, this patch at least makes sure that MSI target
address is allocated from 32-bit pool.
Note that this doesn't change where anything is allocated as such, it
just means that on systems with most of their RAM above 4GB, those few
bytes of private data that you map "for free" will be copied into the
SWIOTLB buffer and hog 2KB of its typical 64MB capacity effectively for
ever.
drivers/pci/controller/dwc/pcie-designware-host.c | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/drivers/pci/controller/dwc/pcie-designware-host.c b/drivers/pci/controller/dwc/pcie-designware-host.c
index 44c2a6572199..e6a230eddf66 100644
--- a/drivers/pci/controller/dwc/pcie-designware-host.c
+++ b/drivers/pci/controller/dwc/pcie-designware-host.c
@@ -388,6 +388,14 @@ int dw_pcie_host_init(struct pcie_port *pp)
dw_chained_msi_isr,
pp);
+ ret = dma_set_mask(pci->dev, DMA_BIT_MASK(32));
+ if (!ret) {
+ dev_warn(pci->dev,
+ "Failed to set DMA mask to 32-bit. "
+ "Devices with only 32-bit MSI support"
+ " may not work properly\n");
+ }
Ironically, the only real reason for that dma_set_mask() to ever fail is
if the system had no 32-bit addressable memory, in which case you could
likely pick any 32-bit doorbell address with impunity, just not via this
mechanism (although whether it would be worthwhile is another matter).
Robin.
+
pp->msi_data = dma_map_single_attrs(pci->dev, &pp->msi_msg,
sizeof(pp->msi_msg),
DMA_FROM_DEVICE,
