Message ID | 20200604133938.GA1513@cosmos (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | mm/memory_hotplug: fix default_zone_for_pfn() to include highmem zone range | expand |
On 04.06.20 15:39, Vamshi K Sthambamkadi wrote: > On x86_32, while onlining highmem sections, the func default_zone_for_pfn() > defaults target zone to ZONE_NORMAL (movable_node_enabled = 0). Onlining of > pages is successful, and these highmem pages are moved into zone_normal. > > As a consequence, these pages are treated as low mem, and page addresses > are calculated using lowmem_page_address() which effectively overflows the > 32 bit virtual addresses, leading to kernel panics and system becomes > unusable. > > Change default_kernel_zone_for_pfn() to intersect highmem pfn range, and > calculate the default zone accordingly. We discussed this recently [1], and decided that we don't really care about memory hotplug on 32-bit anymore (especially, user space could still configure a different zone and make things crash). There was a patch from Michal in [1], looks like it has not been picked up yet. @Andrew, can we queue Michals patch? [1] https://lkml.kernel.org/r/20200218100532.GA4151@dhcp22.suse.cz > > Signed-off-by: Vamshi K Sthambamkadi <vamshi.k.sthambamkadi@gmail.com> > --- > mm/memory_hotplug.c | 7 ++++++- > 1 file changed, 6 insertions(+), 1 deletion(-) > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c > index c4d5c45..30f101a 100644 > --- a/mm/memory_hotplug.c > +++ b/mm/memory_hotplug.c > @@ -725,8 +725,13 @@ static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn > { > struct pglist_data *pgdat = NODE_DATA(nid); > int zid; > + int nr_zones = ZONE_NORMAL; > > - for (zid = 0; zid <= ZONE_NORMAL; zid++) { > +#ifdef CONFIG_HIGHMEM > + nr_zones = ZONE_HIGHMEM; > +#endif > + > + for (zid = 0; zid <= nr_zones; zid++) { > struct zone *zone = &pgdat->node_zones[zid]; > > if (zone_intersects(zone, start_pfn, nr_pages)) >
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c index c4d5c45..30f101a 100644 --- a/mm/memory_hotplug.c +++ b/mm/memory_hotplug.c @@ -725,8 +725,13 @@ static struct zone *default_kernel_zone_for_pfn(int nid, unsigned long start_pfn { struct pglist_data *pgdat = NODE_DATA(nid); int zid; + int nr_zones = ZONE_NORMAL; - for (zid = 0; zid <= ZONE_NORMAL; zid++) { +#ifdef CONFIG_HIGHMEM + nr_zones = ZONE_HIGHMEM; +#endif + + for (zid = 0; zid <= nr_zones; zid++) { struct zone *zone = &pgdat->node_zones[zid]; if (zone_intersects(zone, start_pfn, nr_pages))
On x86_32, while onlining highmem sections, the func default_zone_for_pfn() defaults target zone to ZONE_NORMAL (movable_node_enabled = 0). Onlining of pages is successful, and these highmem pages are moved into zone_normal. As a consequence, these pages are treated as low mem, and page addresses are calculated using lowmem_page_address() which effectively overflows the 32 bit virtual addresses, leading to kernel panics and system becomes unusable. Change default_kernel_zone_for_pfn() to intersect highmem pfn range, and calculate the default zone accordingly. Signed-off-by: Vamshi K Sthambamkadi <vamshi.k.sthambamkadi@gmail.com> --- mm/memory_hotplug.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)