Newshot

Linux 7.4 Kernel to Support Parallel Deletion of NVMe Namespaces, Boosting Processing Speed

AI summary

Linux 7.4-rc4 is expected to be released today with multiple fixes and performance improvements, including a notable NVMe namespace performance optimization. The kernel will support parallel deletion of NVMe namespaces; previously, namespace deletion code used serial logic, deleting one namespace at a time, which could become a performance bottleneck in complex modern Linux server deployments. Software engineer Qiliang Yuan submitted the patch, noting that each namespace deletion must wait for an RCU grace period in del_gendisk() for both the namespace and its multipath header, totaling about 20 milliseconds per deletion, so removing a controller scales with namespace count—over 1 second for 64 namespaces—and the new approach deletes namespaces in parallel using asynchronous domains, with concurrent waits sharing the system-wide RCU grace period.

Translation(model translation, not the source's own text)

Linux 7.4 Kernel to Support Parallel Deletion of NVMe Namespaces, Improving Processing Speed

IT Home, November 22 news: According to a report by tech media Phoronix yesterday, the Linux 7.4-rc4 kernel is expected to be released today. This update will bring multiple fixes and performance improvements, as well as a notable NVMe namespace performance optimization.

IT Home learned from the original report that the Linux 7.4-rc4 kernel will support parallel deletion of NVMe namespaces (NVMe Namespaces). For reference, this feature can divide NVMe storage devices into multiple logically isolated storage volumes, mainly suitable for scenarios such as multi-tenant environments, workload isolation, and security management.

Previously, the NVMe namespace deletion code in the Linux kernel used serial logic, deleting only one namespace at a time. In some complex modern Linux server deployment environments, serial deletion may become a performance bottleneck.

Software engineer Qiliang Yuan submitted this patch, and described the implementation of parallel deletion in the comments: "nvme_remove_namespaces () deletes each namespace under the controller one by one. Each deletion operation needs to wait for an RCU grace period in del_gendisk (), separately for that namespace and its multipath head, taking about 20 milliseconds in total."

"Therefore, the time required to delete a controller (for example, disconnecting from an NVMe-oF target) is proportional to the number of its namespaces. When a controller has 64 namespaces, the entire deletion process may take more than 1 second. Before commit 4e893ca81170 ('nvme_core: scan namespaces asynchronously'), the scanning operation had the same problem. That commit implemented parallel addition of namespaces asynchronously."

"Now we can use the same approach to delete these namespaces in parallel with async domains, and wait for all operations to complete before returning. The deletion operation for each namespace still needs to wait about 20 milliseconds, but these waits can now overlap: since the RCU grace period is system-wide, deletion operations waiting at the same time can share the same grace period instead of waiting separately."

Related Articles

  • South African developer claims to have single-handedly developed the Impossible OS operating system: compatible with Win32 and Linux binaries, to be seen in two years
  • Debian Linux maintainers form a new team, planning to improve platform game compatibility and launch a dedicated distribution
  • TUXEDO InfinityBook Pro 15 - Gen11 - AMD laptop released, equipped with HX 470
  • Nginx exposed to the NGINX2Root vulnerability chain: attackers can remotely execute commands without authentication, and combined with a Linux kernel vulnerability can obtain root privileges on the host
  • AMD releases 26.17 Linux version Radeon software upgrade: adds support for high-bandwidth HDMI, etc.
  • Linux kernel vulnerability NoisyTCP hidden for 11 years exposed, sending specially crafted TCP packets can crash servers

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗