On my Linux system with over 16 GB of RAM, I would like to allocate about 100 MB of space, which:

- is physically contiguous. Ideally, all the 100MB would be physically contiguous, but I'm also OK with using HUGEPAGEs so that each 2MB is physically contiguous.
- can be used as a DMA target, with the 32 bit PCI space, not the 64 bit.This will allow us to use the shorter header for the PCIe transactions which saves a bit of bandwidth.
- is accessible from userspace programs (not just from kernel drivers).

I think the use of HUGEPAGEs and mmap can achieve the 1st (partial contiguity up to 2 MB at a time) and 3rd constraints (user space access through mmap), but I'm not sure how to ensure that the 2nd constraint (32 bit DMA) is guaranteed.

Thanks in advance for your help.

If you are running a 64 bit kernel then just use 64 bit dma. Otherwise plan on some late nights and hair loss.

Thanks. I don't have much hair left, so it's good advice to keep in mind.

All the same, I would have thought that pci_set_dma_mask(gDev, DMA_BIT_MASK(32)) would allow me to use only 32-bit DMA.
What am I missing?

I don't argue with people on the internet, so let me just ask a question. If the device doing the DMA handles 64 bit addresses, what are you planning on putting in the top 32?

The upper 32 bits will contain 0. The device (an FPGA) will pick the appropriate addressing cycle, based on the value of the upper 32 bits.

When I try using pci_set_dma_mask(32), and then pci_map_sg() on a page, it returns a failure, even for small sizes. So, it doesn't sound like this scheme will work, although I can't explain why.