On Tue, 28 Jul 2020 08:11:42 +0300
Mike Rapoport <rppt@xxxxxxxxxx> wrote:
> From: Mike Rapoport <rppt@xxxxxxxxxxxxx>
>
> dummy_numa_init() loops over memblock.memory and passes nid=0 to
> numa_add_memblk() which essentially wraps memblock_set_node(). However,
> memblock_set_node() can cope with entire memory span itself, so the loop
> over memblock.memory regions is redundant.
>
> Replace the loop with a single call to memblock_set_node() to the entire
> memory.
Hi Mike,
I had a similar patch I was going to post shortly so can add a bit more
on the advantages of this one.
Beyond cleaning up, it also fixes an issue with a buggy ACPI firmware in which the SRAT
table covers some but not all of the memory in the EFI memory map. Stealing bits
from the draft cover letter I had for that...
> This issue can be easily triggered by having an SRAT table which fails
> to cover all elements of the EFI memory map.
>
> This firmware error is detected and a warning printed. e.g.
> "NUMA: Warning: invalid memblk node 64 [mem 0x240000000-0x27fffffff]"
> At that point we fall back to dummy_numa_init().
>
> However, the failed ACPI init has left us with our memblocks all broken
> up as we split them when trying to assign them to NUMA nodes.
>
> We then iterate over the memblocks and add them to node 0.
>
> for_each_memblock(memory, mblk) {
> ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size);
> if (!ret)
> continue;
> pr_err("NUMA init failed\n");
> return ret;
> }
>
> numa_add_memblk() calls memblock_set_node() which merges regions that
> were previously split up during the earlier attempt to add them to different
> nodes during parsing of SRAT.
>
> This means elements are moved in the memblock array and we can end up
> in a different memblock after the call to numa_add_memblk().
> Result is:
>
> Unable to handle kernel paging request at virtual address 0000000000003a40
> Mem abort info:
> ESR = 0x96000004
> EC = 0x25: DABT (current EL), IL = 32 bits
> SET = 0, FnV = 0
> EA = 0, S1PTW = 0
> Data abort info:
> ISV = 0, ISS = 0x00000004
> CM = 0, WnR = 0
> [0000000000003a40] user address but active_mm is swapper
> Internal error: Oops: 96000004 [#1] PREEMPT SMP
>
> ...
>
> Call trace:
> sparse_init_nid+0x5c/0x2b0
> sparse_init+0x138/0x170
> bootmem_init+0x80/0xe0
> setup_arch+0x2a0/0x5fc
> start_kernel+0x8c/0x648
>
> As an illustrative example:
> EFI table has one block of memory.
> memblks[0] = [0...0x2f] so we start with a single memblock.
>
> SRAT has
> [0x00...0x0f] in node 0
> [0x10...0x1f] in node 1
> but no entry covering
> [0x20...0x2f].
>
> Whilst parsing SRAT the single memblock is broken into 3.
> memblks[0] = [0x00...0x0f] in node 0
> memblks[1] = [0x10...0x1f] in node 1
> memblks[2] = [0x20...0x2f] in node MAX_NUM_NODES (invalid value)
>
> A sanity check parse then detects the invalid section and acpi_numa_init
> fails. We then fall back to the dummy path.
>
> That iterates over the memblocks. We'll use i an index in the array of memblocks
>
> i = 0;
> memblks[0] = [0x00...0x0f] set to node0.
> merge doesn't do anything because the neighbouring memblock is still in node1.
>
> i = 1
> memblks[1] = [0x10...0x1f] set to node 0.
> merge combines memblock 0 and 1 to give a new set of memblocks.
>
> memblks[0] = [0x00..0x1f] in node 0
> memblks[1] = [0x20..0x2f] in node MAX_NUM_NODES.
>
> i = 2 off the end of the now reduced array of memblocks, so exit the loop.
> (if we restart the loop here everything will be fine).
>
> Later sparse_init_nid tries to use the node of the second memblock to index
> somethings and boom.
>
> Signed-off-by: Mike Rapoport <rppt@xxxxxxxxxxxxx>
Acked-by: Jonathan Cameron <Jonathan.Cameron@xxxxxxxxxx>
> ---
> arch/arm64/mm/numa.c | 13 +++++--------
> 1 file changed, 5 insertions(+), 8 deletions(-)
>
> diff --git a/arch/arm64/mm/numa.c b/arch/arm64/mm/numa.c
> index aafcee3e3f7e..0cbdbcc885fb 100644
> --- a/arch/arm64/mm/numa.c
> +++ b/arch/arm64/mm/numa.c
> @@ -423,19 +423,16 @@ static int __init numa_init(int (*init_func)(void))
> */
> static int __init dummy_numa_init(void)
> {
> + phys_addr_t start = memblock_start_of_DRAM();
> + phys_addr_t end = memblock_end_of_DRAM();
> int ret;
> - struct memblock_region *mblk;
>
> if (numa_off)
> pr_info("NUMA disabled\n"); /* Forced off on command line. */
> - pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n",
> - memblock_start_of_DRAM(), memblock_end_of_DRAM() - 1);
> -
> - for_each_memblock(memory, mblk) {
> - ret = numa_add_memblk(0, mblk->base, mblk->base + mblk->size);
> - if (!ret)
> - continue;
> + pr_info("Faking a node at [mem %#018Lx-%#018Lx]\n", start, end - 1);
>
> + ret = numa_add_memblk(0, start, end);
> + if (ret) {
> pr_err("NUMA init failed\n");
> return ret;
> }
