| Message ID | 20170824230850.1810408-1-arnd@arndb.de |
|---|---|
| State | New |
| Headers | show |
| Series | mm: HMM always needs MMU_NOTIFIER | expand |
On Thu, 24 Aug 2017 20:42:26 -0400 jglisse@redhat.com wrote: > The struct hmm is only use if the HMM mirror functionality is enabled > move associated code behind CONFIG_HMM_MIRROR to avoid build error if > one enable some of the HMM memory configuration without the mirror > feature. It's unclear whether this patch is in addition to Arnd's "mm: HMM always needs MMU_NOTIFIER" or is a replacement for it. (If the latter, it should have Arnd's Reported-by). But I cannot get this patch to apply cleanly in either situation. So I'll skip both patches. Please send something which applies on top of hmm-heterogeneous-memory-management-documentation-v3.patch mm-hmm-heterogeneous-memory-management-hmm-for-short-v5.patch mm-hmm-mirror-mirror-process-address-space-on-device-with-hmm-helpers-v3.patch mm-hmm-mirror-helper-to-snapshot-cpu-page-table-v4.patch mm-hmm-mirror-device-page-fault-handler.patch mm-memory_hotplug-introduce-add_pages.patch mm-zone_device-new-type-of-zone_device-for-unaddressable-memory-v5.patch mm-zone_device-new-type-of-zone_device-for-unaddressable-memory-fix.patch mm-zone_device-special-case-put_page-for-device-private-pages-v4.patch mm-memcontrol-allow-to-uncharge-page-without-using-page-lru-field.patch mm-memcontrol-support-memory_device_private-v4.patch mm-hmm-devmem-device-memory-hotplug-using-zone_device-v7.patch mm-hmm-devmem-dummy-hmm-device-for-zone_device-memory-v3.patch mm-migrate-new-migrate-mode-migrate_sync_no_copy.patch mm-migrate-new-memory-migration-helper-for-use-with-device-memory-v5.patch mm-migrate-migrate_vma-unmap-page-from-vma-while-collecting-pages.patch mm-migrate-support-un-addressable-zone_device-page-in-migration-v3.patch mm-migrate-allow-migrate_vma-to-alloc-new-page-on-empty-entry-v4.patch mm-device-public-memory-device-memory-cache-coherent-with-cpu-v5.patch mm-hmm-add-new-helper-to-hotplug-cdm-memory-region-v3.patch # mm-hmm-avoid-bloating-arch-that-do-not-make-use-of-hmm.patch thanks.
diff --git a/mm/hmm.c b/mm/hmm.c index 4a179a16ab10..b9e9f14e7454 100644 --- a/mm/hmm.c +++ b/mm/hmm.c @@ -41,11 +41,16 @@ */ DEFINE_STATIC_KEY_FALSE(device_private_key); EXPORT_SYMBOL(device_private_key); -static const struct mmu_notifier_ops hmm_mmu_notifier_ops; #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ +void hmm_mm_destroy(struct mm_struct *mm) +{ + kfree(mm->hmm); +} + +#if IS_ENABLED(CONFIG_HMM_MIRROR) +static const struct mmu_notifier_ops hmm_mmu_notifier_ops; -#ifdef CONFIG_HMM /* * struct hmm - HMM per mm struct * @@ -124,13 +129,6 @@ static struct hmm *hmm_register(struct mm_struct *mm) return mm->hmm; } -void hmm_mm_destroy(struct mm_struct *mm) -{ - kfree(mm->hmm); -} -#endif /* CONFIG_HMM */ - -#if IS_ENABLED(CONFIG_HMM_MIRROR) static void hmm_invalidate_range(struct hmm *hmm, enum hmm_update_type action, unsigned long start,
When building a kernel with HMM enabled but without MMU_NOTIFIER, we run into a build error: mm/hmm.c:66:22: error: field 'mmu_notifier' has incomplete type struct mmu_notifier mmu_notifier; If I read this right, the dependency is correct, but the #ifdef annotations in the mm/hmm.c are not. This changes them in a way to make it all build cleanly. Fixes: e4e0061ea15c ("mm/device-public-memory: device memory cache coherent with CPU") Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- I did not try very hard to understand what the code is supposed to do, please check if this makes sense beyond fixing the build before applying. --- mm/hmm.c | 16 +++++++--------- 1 file changed, 7 insertions(+), 9 deletions(-) -- 2.9.0