Re: [PATCH v3] sparc64: Multi-page size support

[Nitin Gupta]

On 12/11/2016 06:14 PM, David Miller wrote:
> From: David Miller <davem@xxxxxxxxxxxxx>
> Date: Sun, 11 Dec 2016 21:06:30 -0500 (EST)
> 
>> Applied.
> 
> Actually, I'm reverting.
> 
> Just doing a simply "make -s -j128" kernel build on a T4-2 I'm
> getting kernel log warnings:
> 
> [2024810.925975] IPv6: ADDRCONF(NETDEV_CHANGE): eth0: link becomes ready
> [2024909.011397] random: crng init done
> [2024970.860642] ------------[ cut here ]------------
> [2024970.869970] WARNING: CPU: 85 PID: 19335 at arch/sparc/mm/tsb.c:99 __flush_huge_tsb_one_entry.constprop.0+0x30/0x74
> [2024970.890947] Modules linked in: ipv6 usb_storage loop ehci_pci sg sr_mod igb ptp pps_core ehci_hcd n2_rng rng_core
> [2024970.911785] CPU: 85 PID: 19335 Comm: ld Not tainted 4.9.0+ #9
> [2024970.923588] Call Trace:
> [2024970.928807]  [0000000000463a2c] __warn+0xb0/0xc8
> [2024970.938349]  [000000000044efa4] __flush_huge_tsb_one_entry.constprop.0+0x30/0x74
> [2024970.953463]  [000000000044f224] flush_tsb_user_page+0x88/0x9c
> [2024970.965268]  [000000000044eabc] tlb_batch_add_one+0x5c/0xd4
> [2024970.976722]  [000000000044ed84] set_pmd_at+0x104/0x184
> [2024970.987329]  [0000000000552594] zap_huge_pmd+0x30/0x244
> [2024970.998097]  [000000000052a7a8] unmap_page_range+0x18c/0x794
> [2024971.009721]  [000000000052b05c] unmap_vmas+0x18/0x44
> [2024971.019976]  [00000000005315f8] exit_mmap+0x94/0x114
> [2024971.030207]  [0000000000461930] mmput+0x50/0xf8
> [2024971.039593]  [00000000004674ac] do_exit+0x310/0x904
> [2024971.049651]  [0000000000467c10] do_group_exit+0x80/0xbc
> [2024971.060415]  [0000000000467c60] SyS_exit_group+0x14/0x20
> [2024971.071363]  [0000000000406194] linux_sparc_syscall32+0x34/0x60
> 
> which is:
> 
> #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
> static void __flush_huge_tsb_one_entry(unsigned long tsb, unsigned long v,
>                                        unsigned long hash_shift,
>                                        unsigned long nentries,
>                                        unsigned int hugepage_shift)
> {
>         unsigned int hpage_entries;
>         unsigned int i;
> 
>         hpage_entries = 1 << (hugepage_shift - REAL_HPAGE_SHIFT);
>         WARN_ON(v & ((1ul << hugepage_shift) - 1));
>         ^^^^^^^


The warning is getting triggered since we do 'vaddr |= 1'
in case the addr is executable in tlb_batch_add_one().

I originally tested the patch with stream which allocates
pages only for heap. Somehow, I cannot not reproduce it (on a Sonoma)
even when I back text segments with huagepages using:

hugectl --force-preload --heap=8m --text=8m make -s -j128


I added this WARN_ON during debugging and can simply be removed.
Do you want to see a v4 with this warning removed or can you reapply
with just this change?

Thanks,
Nitin