Bug 104111
| Summary: | If CONFIG_NO_HZ_FULL_ALL=y, all user space processes have 0x1 affinity since 4.2.0 | ||
|---|---|---|---|
| Product: | Process Management | Reporter: | Petr Pisar (petr.pisar) |
| Component: | Scheduler | Assignee: | Ingo Molnar (mingo) |
| Status: | NEW --- | ||
| Severity: | normal | CC: | ansla80, epopevad, eugene.shalygin, fweisbec, kernel, keruspe+bugs-kernel, szg00000 |
| Priority: | P1 | ||
| Hardware: | x86-64 | ||
| OS: | Linux | ||
| Kernel Version: | 4.2.0 | Subsystem: | |
| Regression: | Yes | Bisected commit-id: | |
| Attachments: |
4.2.0 kernel log
4.2.0 .config failing config Working config |
||
Created attachment 186921 [details]
4.2.0 .config
*** Bug 104151 has been marked as a duplicate of this bug. *** I also have the same problem. Had to make defconfig for 4.2.0 and then tune options manually to get a working kernel. Created attachment 186971 [details]
failing config
The buggy config created via make oldconfig from 4.1.6
Created attachment 186981 [details]
Working config
The working one, tuned after make defconfig
Ok, After some digging, I found the exact config option causing it. I'll have a look tomorrow at fixing it. Here is the .config diff which makes it work: diff --git a/.config b/.config index 4391737..87eea22 100644 --- a/.config +++ b/.config @@ -106,7 +106,7 @@ CONFIG_NO_HZ_COMMON=y # CONFIG_HZ_PERIODIC is not set # CONFIG_NO_HZ_IDLE is not set CONFIG_NO_HZ_FULL=y -CONFIG_NO_HZ_FULL_ALL=y +# CONFIG_NO_HZ_FULL_ALL is not set CONFIG_NO_HZ_FULL_SYSIDLE=y CONFIG_NO_HZ_FULL_SYSIDLE_SMALL=8 # CONFIG_NO_HZ is not set I can confirm that unsetting the CONFIG_NO_HZ_FULL_ALL makes my scheduler more usable. It's expected behaviour since we include all nohz full CPUs to cpu_isolated_map. But we are starting to think that we should change that behaviour... |
Created attachment 186911 [details] 4.2.0 kernel log I have a 4-core AMD A10-5700 CPU. I noticed a regression between 4.1.6 and 4.2.0 kernel: All user space processes stick to first core no matter of CPU utilization. (Kernel threads use more CPUs.) "taskset -p $$" shows "1". If I set affinity to more CPUs, processes do not move. If I exclude currently occupied CPU from the affinity, the process moves to the lowest allowed CPU number. If I reboot to 4.1.6, "taskset -p $$" shows "f" and processes are distributed evenly across all CPUs. This is a difference of kernel configurations as created more or less by "make oldconfig": --- 4.1.6 2015-09-06 19:00:51.240979625 +0200 +++ 4.2.0 2015-09-06 19:00:59.123486618 +0200 -CONFIG_X86_HT=y -CONFIG_GENERIC_IRQ_LEGACY_ALLOC_HWIRQ=y +CONFIG_IRQ_DOMAIN_HIERARCHY=y +CONFIG_GENERIC_MSI_IRQ_DOMAIN=y -CONFIG_RCU_FANOUT=64 -CONFIG_RCU_FANOUT_LEAF=16 -CONFIG_RCU_KTHREAD_PRIO=0 +CONFIG_ARCH_WANTS_DYNAMIC_TASK_STRUCT=y +CONFIG_HAVE_COPY_THREAD_TLS=y +CONFIG_MODULES_TREE_LOOKUP=y -CONFIG_ARCH_USE_QUEUE_RWLOCK=y -CONFIG_QUEUE_RWLOCK=y +CONFIG_ARCH_USE_QUEUED_SPINLOCKS=y +CONFIG_QUEUED_SPINLOCKS=y +CONFIG_ARCH_USE_QUEUED_RWLOCKS=y +CONFIG_QUEUED_RWLOCKS=y +CONFIG_ARCH_SUPPORTS_DEFERRED_STRUCT_PAGE_INIT=y +CONFIG_ACPI_REV_OVERRIDE_POSSIBLE=y +CONFIG_PCI_MSI_IRQ_DOMAIN=y +CONFIG_MAC80211_STA_HASH_MAX_SIZE=0 +CONFIG_INPUT_LEDS=m +CONFIG_SND_JACK=y +CONFIG_SND_PROC_FS=y -CONFIG_SND_KCTL_JACK=y +CONFIG_EDAC_ATOMIC_SCRUB=y +CONFIG_EDAC_SUPPORT=y +CONFIG_EFI_ESRT=y -CONFIG_KASAN_SHADOW_OFFSET=0xdffffc0000000000 +CONFIG_SCHED_INFO=y +CONFIG_CRYPTO_RNG_DEFAULT=m +CONFIG_CRYPTO_AKCIPHER2=y +CONFIG_CRYPTO_NULL=m +CONFIG_CRYPTO_ECHAINIV=m +CONFIG_CRYPTO_DRBG_MENU=m +CONFIG_CRYPTO_DRBG_HMAC=y +CONFIG_CRYPTO_DRBG=m +CONFIG_CRYPTO_JITTERENTROPY=m +CONFIG_ARCH_HAS_PMEM_API=y