So promote them to meminit... "Yinghai" <yinghai.lu@xxxxxxxxxx> wrote: >On 05/10/2010 12:10 AM, Benjamin Herrenschmidt wrote: >> On Sat, 2010-05-08 at 08:17 -0700, Yinghai Lu wrote: >>> They will check if the region array is big enough. >>> >>> __check_and_double_region_array will try to double the region array if that >>> array spare slots is not big enough. Old array will be copied to new array. >>> >>> Arch code should set lmb.default_alloc_limit accordingly, so the new array is in >>> accessiable address. >> >> More issues... >> >>> +static void __init __check_and_double_region_array(struct lmb_region *type, >>> + struct lmb_property *static_region) >>> +{ >>> + u64 size, mem; >>> + struct lmb_property *new, *old; >>> + unsigned long rgnsz = type->nr_regions; >>> + >>> + /* Do we have enough slots left ? */ >>> + if ((rgnsz - type->cnt) > 2) >>> + return; >>> + >>> + old = type->region; >>> + /* Double the array size */ >>> + size = sizeof(struct lmb_property) * rgnsz * 2; >>> + >>> + mem = __lmb_alloc_base(size, sizeof(struct lmb_property), lmb.default_alloc_limit); >>> + if (mem == 0) >>> + panic("can not find more space for lmb.reserved.region array"); >> >> Now, that is not right because we do memory hotplug. Thus lmb_add() must >> be able to deal with things running past LMB init. >> >> slab_is_available() will do the job for now, unless somebody has bootmem >> and tries to lmb_add() memory while bootmem is active, but screw that >> for now. See the code I'll post tonight. >> >>> + new = __va(mem); >>> + /* Copy old to new */ >>> + memcpy(&new[0], &old[0], sizeof(struct lmb_property) * rgnsz); >>> + memset(&new[rgnsz], 0, sizeof(struct lmb_property) * rgnsz); >>> + >>> + memset(&old[0], 0, sizeof(struct lmb_property) * rgnsz); >>> + type->region = new; >>> + type->nr_regions = rgnsz * 2; >>> + printk(KERN_DEBUG "lmb.reserved.region array is doubled to %ld at [%llx - %llx]\n", >>> + type->nr_regions, mem, mem + size - 1); >>> + >>> + /* Free old one ?*/ >>> + if (old != static_region) >>> + lmb_free(__pa(old), sizeof(struct lmb_property) * rgnsz); >>> +} >> >> Similar comment, don't bother if slab is available. >> >>> +void __init lmb_add_memory(u64 start, u64 end) >>> +{ >>> + lmb_add_region(&lmb.memory, start, end - start); >>> + __check_and_double_region_array(&lmb.memory, &lmb_memory_region[0]); >>> +} >> >> So you duplicate lmb_add() gratuituously ? >> >>> +void __init lmb_reserve_area(u64 start, u64 end, char *name) >>> +{ >>> + if (start == end) >>> + return; >>> + >>> + if (WARN_ONCE(start > end, "lmb_reserve_area: wrong range [%#llx, %#llx]\n", start, end)) >>> + return; >>> + >>> + lmb_add_region(&lmb.reserved, start, end - start); >>> + __check_and_double_region_array(&lmb.reserved, &lmb_reserved_region[0]); >>> +} >> >> And lmb_reserve() ? >> >> Do we want to end up with 5 copies of the same API with subtle >> differences just for fun ? > >those functions have __init markers, and only can be used on boot stage. so do need to worry about hotplug mem. > >what I do is: use current lmb code for x86, and keep the affects to original lmb users to minimum. (should be near 0) > >YH -- Sent from my Android phone with K-9 Mail. Please excuse my brevity.