On 3/13/20 12:12 PM, Srikar Dronamraju wrote: > * Michael Ellerman <mpe@xxxxxxxxxxxxxx> [2020-03-13 21:48:06]: > >> Sachin Sant <sachinp@xxxxxxxxxxxxxxxxxx> writes: >> >> The patch below might work. Sachin can you test this? I tried faking up >> >> a system with a memoryless node zero but couldn't get it to even start >> >> booting. >> >> >> > The patch did not help. The kernel crashed during >> > the boot with the same call trace. >> > >> > BUG_ON() introduced with the patch was not triggered. >> >> OK, that's weird. >> >> I eventually managed to get a memoryless node going in sim, and it >> appears to work there. >> >> eg in dmesg: >> >> [ 0.000000][ T0] numa: NODE_DATA [mem 0x2000fffa2f80-0x2000fffa7fff] >> [ 0.000000][ T0] numa: NODE_DATA(0) on node 1 >> [ 0.000000][ T0] numa: NODE_DATA [mem 0x2000fff9df00-0x2000fffa2f7f] >> ... >> [ 0.000000][ T0] Early memory node ranges >> [ 0.000000][ T0] node 1: [mem 0x0000000000000000-0x00000000ffffffff] >> [ 0.000000][ T0] node 1: [mem 0x0000200000000000-0x00002000ffffffff] >> [ 0.000000][ T0] Could not find start_pfn for node 0 >> [ 0.000000][ T0] Initmem setup node 0 [mem 0x0000000000000000-0x0000000000000000] >> [ 0.000000][ T0] On node 0 totalpages: 0 >> [ 0.000000][ T0] Initmem setup node 1 [mem 0x0000000000000000-0x00002000ffffffff] >> [ 0.000000][ T0] On node 1 totalpages: 131072 >> >> # dmesg | grep set_numa >> [ 0.000000][ T0] set_numa_mem: mem node for 0 = 1 >> [ 0.005654][ T0] set_numa_mem: mem node for 1 = 1 >> >> So is the problem more than just node zero having no memory? >> > > The problem would happen with possible nodes which are not yet present. i.e > no cpus, no memory attached to those nodes. > > Please look at > http://lore.kernel.org/lkml/20200312131438.GB3277@xxxxxxxxxxxxxxxxxx/t/#u > for more details. > > The summary being: pgdat/Node_Data for such nodes is not allocated. Hence Michael's log shows that his pgdat is still allocated. But perhaps Sachin had also your 3 patches from the other thread applied, in addition to Michael's patch. So in his case pgdat for node 0 would indeed be no longer allocated, and thus SLUB code was crashing in node_present_pages() instead. > the node_present_pages(nid) called where nid is a possible but not yet > present node fails. Currently node_present_pages(nid) and node_to_mem_node > don't seem to be equipped to handle possible but not present nodes. > >> cheers >
