Skip to content

Releases: NVIDIA/gdrcopy

v2.6

Choose a tag to compare

@pakmarkthub pakmarkthub released this 13 Jun 00:07
  • Introduce a DMA-BUF mmap backend for mapping GPU memory without the GDRCopy kernel module. GDRCopy still prefers gdrdrv when available, can fall back to DMA-BUF mmap with CUDA driver 13.3+, and can force the new backend with GDRCOPY_USE_DMABUF_MMAP.
  • Introduce GDR_ATTR_USING_DMA_BUF_MMAP so applications can query whether GDRCopy is using the DMA-BUF mmap backend.
  • Improve gdr_copy_to_mapping and gdr_copy_from_mapping performance with runtime-selected CPU copy implementations for AVX2, AVX-512, MOVDIR64B, NEON, and LS64-capable systems.
  • Add -M to gdrcopy_apiperf, gdrcopy_copybw, gdrcopy_copylat, and gdrcopy_pplat to request default, WC, cache, or device mappings.
  • Change gdrcopy_copylat and gdrcopy_pplat to report P0 and P50 latency metrics instead of average latency.
  • Fix build and install errors on Ubuntu 25.10 by using numeric install modes and dropping deprecated depmod -r usage.

v2.5.2

Choose a tag to compare

@pakmarkthub pakmarkthub released this 12 Mar 15:27
c91ad9f
  • Fix build errors on recent kernels (v6.15+): mark test-dummy.ko with Dual MIT/GPL so it may use GPL-only symbol __vma_start_write (vm_flags_set).

v2.5.1

Choose a tag to compare

@pakmarkthub pakmarkthub released this 19 Aug 23:52
  • Support RHEL 10.0.
  • Support CUDA 13.0.
  • Support the CDMM mode.
  • Support Blackwell.
  • Fix errors when compiling with CUDA 12.6.
  • Fix build-deb-packages.sh to always create .deb with deb versions in the filenames.

v2.5

Choose a tag to compare

@pakmarkthub pakmarkthub released this 26 Mar 20:45
  • Introduce gdr_pin_buffer_v2 API, GDR_PIN_FLAG_FORCE_PCIE pin flag and the GDR_ATTR_SUPPORT_PIN_FLAG_FORCE_PCIE attribute. Extend gdrcopy_sanity coverage to those new APIs.
  • Waive some unit tests in gdrcopy_sanity if the GPU compute mode is not set as default.
  • Introduce gdr_get_attribute API and GDR_ATTR_USE_PERSISTENT_MAPPING.
  • Add persistent mapping coverage to gdrcopy_sanity.
  • Support setting GPU ID in gdrcopy_sanity.
  • Fix nvidia_p2p_get_pages_persistent leakage in gdrdrv.
  • Introduce leakage_pin_pages_fork_cumemalloc unit test.
  • Add support for mix mappings.
  • Introduce /proc/driver/gdrdrv/params and /proc/driver/gdrdrv/nv_get_pages_refcount.
  • Fix silent failure in gdrcopy_copybw and gdrcopy_apiperf.
  • Add -l to shmoo sizes in gdrcopy_copybw.
  • Set the default NVCC_FLAGS to generate PTX and SASS for all supported architectures.
  • Improve the error reporting in gdrcopy_pplat due to PTX loading errors.
  • Add cold cache testing to gdrcopy_copylat.
  • Fix errors in gdr_map due to PAT.
  • Improve the gdrcopy_pplat algorithm and add more test cases.
  • Improve vm_flags_set detection mechanism.
  • Change the default to set use_persistent_mapping=1.

v2.4.4

Choose a tag to compare

@drossetti drossetti released this 16 Dec 18:06

This is a quick hotfix release.

Users are strongly encouraged to upgrade to this release in order to avoid potential CPU and GPU memory leaks, i.e. possibly associated to the ungraceful termination of multi-process applications.

The changes below have been fixed in the master branch and are being backported to the R2.4 branch:

  • Fix a use-after-free bug of mr objects in gdrdv_vma_close(), see #314.
  • Fix a resource leak in gdrdrv_release, see #313.

v2.4.3

Choose a tag to compare

@pakmarkthub pakmarkthub released this 10 Dec 01:42
  • Fix NVIDIA_IS_OPENSOURCE detection when compile with NVIDIA driver version 545 or newer.
  • Fix compile error in gdrdrv when compile on RHEL9.5.

v2.4.2

Choose a tag to compare

@pakmarkthub pakmarkthub released this 01 Nov 04:46
  • Fix the size alignment bug in gdrdrv.
  • Fix memory leak in gdr_pin_buffer.
  • Add support for another flavor of BF3.

v2.4.1

Choose a tag to compare

@pakmarkthub pakmarkthub released this 20 Dec 09:05
  • Add support for persistent mapping.
  • Fix bug in src/gdrdrv/Makefile.
  • Fix compile-time bug when check.h is not found.

v2.4

Choose a tag to compare

@pakmarkthub pakmarkthub released this 19 Sep 01:18
  • Various bug fixes in the test and benchmark applications.
  • Prefix all applications with "gdrcopy_".
  • Introduce more unit tests in gdrcopy_sanity.
  • Introduce gdrcopy_pplat benchmark application.
  • Remove dependency on libcheck and libsubunit
  • Introduce gdr_get_info_v2.
  • Introduce new copy algorithm for device mappings.
  • Add support for NVIDIA BLUEFIELD-3.
  • Add support for Linux kernel >= 6.3.
  • Add support for SLES and OpenSUSE.
  • Add support for systemd service on RHEL9.
  • Relicense gdrdrv to Dual MIT/GPL.
  • Fix bugs in gdrdrv when pinning two small buffers back-to-back.
  • Add support for coherent platforms such as Grace-Hopper.
  • Add support for Confidential Computing (CC).

v2.3.1

Choose a tag to compare

@pakmarkthub pakmarkthub released this 12 May 14:48
  • Add a workaround for the GPL-compatibility issue when compile with CONFIG_ARCH_HAS_CC_PLATFORM on Linux kernel 5.18+.
  • Fix error in init.d/gdrcopy due to missing /etc/rc.d/init.d/functions.