Draft: KMD control over sysmem #17
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This is used to configure address translation from NOC address (x=0, y=4, 0 <= address <= 0xFFFD_FFFF) to the system bus.
This is used to configure address translation from NOC address (x=0, y=3, 0x8_0000_0000 <= address <= 0x8_FFFD_FFFF) to the system bus.
joelsmithTT
force-pushed
the
joelsmith/tell-kmd-about-hugepages
branch
from
3321b69
to
3c04dd5
Compare
Tensix DMA refers to the ability of Tensix and Ethernet tiles on the SoC to access the system bus. This is distinct from DMA performed by the PCIe IP block's DMA controller. On Grayskull and Wormhole, there exists an (almost) 32-bit address range in the PCIe tile for Tensix DMA. - Grayskull: 0x0_0000_0000 to 0x0_FFFD_FFFF - Wormhole: 0x8_0000_0000 to 0x8_FFFD_FFFF Sysmem refers to the buffer(s) in host DRAM that are accessed via Tensix DMA. The convention as of Q1 2024 is to allocate one (Grayskull) or four (Wormhole) 1G hugepages for sysmem, with cooperation beween UMD, KMD, and ARC firmware to make this memory accessible to hardware. This scheme has drawbacks: - Use of the system IOMMU is unsupported. - UMD manipulates hardware address translation state at runtime, resulting in a situation where user applications (e.g. AI workloads) and platform software can not effectively share the Tensix DMA address space. Improvement goals: - Enable use of system IOMMU without requiring IOMMU. - Manage sysmem buffer(s) in the driver, in a NUMA-aware manner. Initial work allocated the buffer in the kernel using the DMA subsystem, without the need for huge pages. With IOMMU enabled, `dma_alloc_coherent` allocates IO virtually contiguous memory. With IOMMU disabled, CMA configured at boot allows `dma_alloc_coherent` to allocate physically contiguous memory. Even with large (i.e. 4GiB) buffers, both scenarios work as expected under Linux 6.5. Unfortunately, this in-kernel approach to allocation does not work under older kernels such as 5.4 (Ubuntu 20.04) and 5.15 (Ubuntu 22.04). Experiments with a carveout approach (pass memmap to the kernel at boot) were similarly unsuccessful on such older kernels with IOMMU enabled. The fallback technique implemented in this commit is to tell the driver about each 1G hugepage via an ioctl. This can be regarded as a one-time, per-device setup step. When the ioctl is invoked, the driver: - Checks that the page(s) are 1G huge - Pins the page(s) - Configures hardware address translation and IOMMU (if necessary) - Tracks the page(s) so that user software can mmap a sysmem buffer - Hangs onto the pages until the ioctl is called again or the driver is unloaded The driver therefore provides both hardware and software with the illusion that the sysmem for a device is contiguous. With the exception of this setup step, userspace is expected to access the sysmem buffer via the driver, rather than interacting with the hugetlbfs. Abandoning support for older kernels will allow the setup step to be omitted, as the driver will be able to handle allocation itself.
First, the test code performs hugepage setup: - Allocates 1G hugepages (one for GS; four for WH) - Tells the driver about the hugepages - Fills the hugepages with a pattern, then unmaps them Now the driver owns the page(s) and hardware has been configured to have access to them. Next, the test code: - Maps the sysmem buffer into its address space - Checks that the pattern is still valid Finally, the test code: - Uses the device to read random addresses in the buffer - Writes to random addresses in the buffer In both cases, the resulting state is checked for validity.
joelsmithTT
force-pushed
the
joelsmith/tell-kmd-about-hugepages
branch
from
3c04dd5
to
b038417
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This introduces:
The goals:
There is more detail in the commit messages.