| Message ID | 20210209133854.17399-5-osalvador@suse.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Allocate memmap from hotadded memory (per device) | expand |
On 09.02.21 14:38, Oscar Salvador wrote: > When struct page's size is not multiple of PMD, these do not get > fully populated when adding sections, hence two sections will > intersect the same the PMD. This goes against the vmemmap-per-device > premise, so reject it if that is the case. > > Signed-off-by: Oscar Salvador <osalvador@suse.de> > --- > mm/memory_hotplug.c | 5 +++-- > 1 file changed, 3 insertions(+), 2 deletions(-) > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c > index 95695483a622..d3fb036d33fd 100644 > --- a/mm/memory_hotplug.c > +++ b/mm/memory_hotplug.c > @@ -57,10 +57,11 @@ static __meminit int memmap_on_memory_store(const char *val, > const struct kernel_param *kp) > { > /* > - * Fail silently in case we cannot enable it due to platform constraints. > + * Fail silently in case we cannot enable it due to system constraints. > * User can always check whether it is enabled or not via /sys/module. > */ > - if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE)) > + if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE) || > + (PMD_SIZE % sizeof(struct page))) > return 0; > > return param_set_bool(val, kp); > Squash that into #1 - it's part of basic operation.
On Thu, Feb 25, 2021 at 07:26:50PM +0100, David Hildenbrand wrote:
> Squash that into #1 - it's part of basic operation.
Ok, will do.
On Thu, Feb 25, 2021 at 07:26:50PM +0100, David Hildenbrand wrote: > On 09.02.21 14:38, Oscar Salvador wrote: > > When struct page's size is not multiple of PMD, these do not get > > fully populated when adding sections, hence two sections will > > intersect the same the PMD. This goes against the vmemmap-per-device > > premise, so reject it if that is the case. > > > > Signed-off-by: Oscar Salvador <osalvador@suse.de> > > --- > > mm/memory_hotplug.c | 5 +++-- > > 1 file changed, 3 insertions(+), 2 deletions(-) > > > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c > > index 95695483a622..d3fb036d33fd 100644 > > --- a/mm/memory_hotplug.c > > +++ b/mm/memory_hotplug.c > > @@ -57,10 +57,11 @@ static __meminit int memmap_on_memory_store(const char *val, > > const struct kernel_param *kp) > > { > > /* > > - * Fail silently in case we cannot enable it due to platform constraints. > > + * Fail silently in case we cannot enable it due to system constraints. > > * User can always check whether it is enabled or not via /sys/module. > > */ > > - if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE)) > > + if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE) || > > + (PMD_SIZE % sizeof(struct page))) > > return 0; > > return param_set_bool(val, kp); > > > > Squash that into #1 - it's part of basic operation. Just to be sure we are on the same page. You mean to move this patch and patch#5 into the runtime check in mhp_supports_memmap_on_memory(), right?
On 01.03.21 11:06, Oscar Salvador wrote: > On Thu, Feb 25, 2021 at 07:26:50PM +0100, David Hildenbrand wrote: >> On 09.02.21 14:38, Oscar Salvador wrote: >>> When struct page's size is not multiple of PMD, these do not get >>> fully populated when adding sections, hence two sections will >>> intersect the same the PMD. This goes against the vmemmap-per-device >>> premise, so reject it if that is the case. >>> >>> Signed-off-by: Oscar Salvador <osalvador@suse.de> >>> --- >>> mm/memory_hotplug.c | 5 +++-- >>> 1 file changed, 3 insertions(+), 2 deletions(-) >>> >>> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c >>> index 95695483a622..d3fb036d33fd 100644 >>> --- a/mm/memory_hotplug.c >>> +++ b/mm/memory_hotplug.c >>> @@ -57,10 +57,11 @@ static __meminit int memmap_on_memory_store(const char *val, >>> const struct kernel_param *kp) >>> { >>> /* >>> - * Fail silently in case we cannot enable it due to platform constraints. >>> + * Fail silently in case we cannot enable it due to system constraints. >>> * User can always check whether it is enabled or not via /sys/module. >>> */ >>> - if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE)) >>> + if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE) || >>> + (PMD_SIZE % sizeof(struct page))) >>> return 0; >>> return param_set_bool(val, kp); >>> >> >> Squash that into #1 - it's part of basic operation. > > Just to be sure we are on the same page. > You mean to move this patch and patch#5 into the runtime check in > mhp_supports_memmap_on_memory(), right? Yes, then the compiler might even be able to figure out that what follows the check is dead code.
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c index 95695483a622..d3fb036d33fd 100644 --- a/mm/memory_hotplug.c +++ b/mm/memory_hotplug.c @@ -57,10 +57,11 @@ static __meminit int memmap_on_memory_store(const char *val, const struct kernel_param *kp) { /* - * Fail silently in case we cannot enable it due to platform constraints. + * Fail silently in case we cannot enable it due to system constraints. * User can always check whether it is enabled or not via /sys/module. */ - if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE)) + if (!IS_ENABLED(CONFIG_ARCH_MHP_MEMMAP_ON_MEMORY_ENABLE) || + (PMD_SIZE % sizeof(struct page))) return 0; return param_set_bool(val, kp);
When struct page's size is not multiple of PMD, these do not get fully populated when adding sections, hence two sections will intersect the same the PMD. This goes against the vmemmap-per-device premise, so reject it if that is the case. Signed-off-by: Oscar Salvador <osalvador@suse.de> --- mm/memory_hotplug.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)