Linux Kernel Expands GPU Mmap Offset Range to 8 TiB
The Linux DRM subsystem raises its fake mmap offset range on 64-bit kernels to prevent exhaustion on multi-GPU systems with large VRAM.
Updated on Nov. 20, 2026 in Computers

The Linux kernel's Direct Rendering Manager subsystem is increasing the fake mmap offset range on 64-bit kernels from 1 TiB to 8 TiB. The change addresses failures in buffer creation and dma-buf imports when the current range is exhausted.
Why it matters
Peer-to-peer dma-buf imports across devices and processes can exhaust the existing offset range on multi-GPU systems with large VRAM, causing allocations to fail. Expanding the range keeps such configurations working while preserving compatibility with existing userspace.
The change raises DRM_FILE_PAGE_OFFSET_SIZE on 64-bit kernels to SZ_8T, eightfold the current 1 TiB, and stays below panthor's 1ULL << 43 MMIO carveout and the 16 TiB 32-bit compat limit. Its effect on systems beyond those tested configurations is not addressed in the source.
The players
Linux kernel DRM subsystem
The Direct Rendering Manager component of the Linux kernel that manages graphics devices and buffer memory for GPU drivers.
Intel compute-runtime
Intel's open-source compute runtime that provides OpenCL and Level Zero support for its GPUs, and which queries mmap offsets on imported buffers.
The details
Every GEM and TTM buffer object receives a fake mmap offset from a per-device offset manager, and on multi-GPU systems peer-to-peer imports can consume the 1 TiB range. When exhausted, buffer creation or PRIME import fails with -ENOSPC. The change expands the range rather than skipping offset allocation for imported buffers, because skipping it was judged to risk regressions for existing userspace such as Intel compute-runtime, which queries offsets on imported host dma-bufs.
Timeline
• November 20, 2026: Article published describing the Linux 7.5 DRM change.
Legacy Systems
The change extends the GEM/TTM fake mmap offset design that DRM drivers have relied on, widening its 1 TiB address range on 64-bit kernels to 8 TiB. It remains constrained by the 16 TiB limit of the 32-bit compat mmap2 pgoff interface when 4K pages are used.
Users running multi-GPU workstations with large VRAM are less likely to hit buffer allocation failures once the expanded range ships. Most desktop users will not notice a difference, since the change affects only configurations that exhaust the existing range.
The takeaway
Developers building GPU software for multi-device setups should not assume the old 1 TiB offset range is sufficient on high-memory systems. Testing peer-to-peer buffer imports across large VRAM configurations can surface exhaustion before the updated kernel reaches production.
Further reading
Explore more Computers coverage for context on operating system kernel development.
Source note: This article includes information reported by Phoronix.



