liblitmus.git/src/migration.c, branch wip-shared-mem [ARCHIVE] The userspace library for LITMUS^RT 2010. Now on GitHub. read_mapping() breaks on NR_CPUS not div by 32 2014-06-06T17:23:37+00:00 Glenn Elliott gelliott@cs.unc.edu 2014-06-06T17:23:37+00:00 aabfa10d86f2be6416eabedd03f52b548379f2d6 This patch fixes a bug in read_mapping(), which is used for reading CPU and cluster (domain) mappings. read_mapping() did not account for the case when NR_CPUS is not evenly divisible by 32 (quite common when NR_CPUS < 32). 
This patch fixes a bug in read_mapping(), which is used for
reading CPU and cluster (domain) mappings. read_mapping()
did not account for the case when NR_CPUS is not evenly
divisible by 32 (quite common when NR_CPUS < 32).
Remove unneeded code from domain migration. 2014-06-06T16:16:36+00:00 Glenn Elliott gelliott@cs.unc.edu 2014-06-06T16:16:36+00:00 963d692458fdcd87dabe599d92b70f67a4c15f0b This patch removes unneeded code from be_migrate_thread_to_domain(). This function no longer needs to check the number of online CPUs since the CPU mask for a given domain is now returned directly from read_mapping(). 
This patch removes unneeded code from be_migrate_thread_to_domain().
This function no longer needs to check the number of online
CPUs since the CPU mask for a given domain is now returned directly
from read_mapping().
Migration: Support systems with more than 32 CPUs. 2014-05-20T01:06:08+00:00 Glenn Elliott gelliott@cs.unc.edu 2014-05-20T01:06:08+00:00 5f2866d43d9a2e33bc2961edf9966cad5708cc4d This patch improves the cpu bit map parsing used to determine CPU/cluster mapping. Namely, it supports clusters with more than 32 CPUs. Prior to this patch, mapping would fail due to unexpected commas in bitfield data read from /proc (groups of 32 CPUs are separated by commas). NOTE: Bug could still be encountered if NR_CPUS > 32. New CPU limit: 4096. 
This patch improves the cpu bit map parsing used to determine
CPU/cluster mapping. Namely, it supports clusters with more
than 32 CPUs. Prior to this patch, mapping would fail due to
unexpected commas in bitfield data read from /proc (groups
of 32 CPUs are separated by commas).

NOTE: Bug could still be encountered if NR_CPUS > 32.

New CPU limit: 4096.
Use /proc/litmus/domains to set up affinity masks 2014-02-24T19:01:11+00:00 Glenn Elliott gelliott@cs.unc.edu 2014-02-05T05:31:08+00:00 1f323b00f95c4704cdeea7918853f1e9e8d26c2e This patch replaces the algorithm used to compute CPU affinity masks. The old algorithm was fragile and would break on systems with multiple levels of shared caches. (As found on some older Intel chips, and as one would probably find on a system with hyper-threading enabled.) The bug stems from fact that Linux enumerates consecutive CPUs in a way that distributes them across the physical system, thereby reducing contention resources (e.g., cache). Cluster size no longer needs to be provided now that cluster/CPU mappings are explicit in /proc/litmus/domains/. This keeps the following migration functions, but marks them as deprecated: - be_migrate_to_cluster() - cluster_to_first_cpu() - partition_to_cpu() Although the deprecated interfaces are supported, the implementations for these functions call the new be_migrate_to_domain() and domain_to_first_cpu() functions. [bbb: resolved several merge conflicts] 
This patch replaces the algorithm used to compute CPU affinity masks.
The old algorithm was fragile and would break on systems with multiple
levels of shared caches. (As found on some older Intel chips, and as
one would probably find on a system with hyper-threading enabled.) The
bug stems from fact that Linux enumerates consecutive CPUs in a way
that distributes them across the physical system, thereby reducing
contention resources (e.g., cache).

Cluster size no longer needs to be provided now that cluster/CPU
mappings are explicit in /proc/litmus/domains/.

This keeps the following migration functions, but marks them as
deprecated:
  - be_migrate_to_cluster()
  - cluster_to_first_cpu()
  - partition_to_cpu()

Although the deprecated interfaces are supported, the implementations
for these functions call the new be_migrate_to_domain() and
domain_to_first_cpu() functions.

[bbb: resolved several merge conflicts]
User large enough buffer to read /proc/litmus/release_master 2013-07-02T07:03:05+00:00 Hiroyuki Chishiro chishiro@cs.unc.edu 2013-07-02T07:03:05+00:00 5e1cee01ba46cffef72bae9c98d025a1c6cf7011 The string 'NO_CPU' needs to fit into the buffer. 
The string 'NO_CPU' needs to fit into the buffer.
Auto-CPU-affinity from part./cluster assignment. 2013-03-12T16:02:20+00:00 Glenn Elliott gelliott@cs.unc.edu 2013-03-12T15:57:42+00:00 93ffe5be8fd1954bbfe5a04e55b81ac1d83d2de7 This patch replaces be_migrate_to() with several be_migrate_*() APIs to automatically assign CPU affinity masks from a task's partition (or cluster) assignment. Routines are release-master-aware such that the release master (if one exists) will NOT be included in the task's affinity mask. (Note that release-master avoidance may be overridden by calling __be_migrate_thread_to_cluster().) 
This patch replaces be_migrate_to() with several
be_migrate_*() APIs to automatically assign CPU affinity
masks from a task's partition (or cluster) assignment.
Routines are release-master-aware such that the release
master (if one exists) will NOT be included in the
task's affinity mask. (Note that release-master
avoidance may be overridden by calling
__be_migrate_thread_to_cluster().)