Re: running mm/ksft_hmm.sh on arm64 results in a kernel panic

[Balbir Singh <balbirs@nvidia.com>]

On 3/19/26 12:49, Alistair Popple wrote:
> On 2026-03-19 at 02:05 +1100, "Lorenzo Stoakes (Oracle)" <ljs@kernel.org> wrote...
>> On Wed, Mar 18, 2026 at 01:26:39PM +0800, Zenghui Yu wrote:
>>> Hi all,
>>>
>>> When running mm/ksft_hmm.sh in my arm64 virtual machine, I ran into the
>>> following kernel panic:
>>>
>>> [root@localhost mm]# ./ksft_hmm.sh
>>> TAP version 13
>>> # --------------------------------
>>> # running bash ./test_hmm.sh smoke
>>> # --------------------------------
>>> # Running smoke test. Note, this test provides basic coverage.
>>> # TAP version 13
>>> # 1..74
>>> # # Starting 74 tests from 4 test cases.
>>> # #  RUN           hmm.hmm_device_private.benchmark_thp_migration ...
>>> #
>>> # HMM THP Migration Benchmark
>>> # ---------------------------
>>> # System page size: 16384 bytes
>>> #
>>> # === Small Buffer (512KB) (0.5 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 0.423 ms        | 0.182 ms        | -133.0%
>>> # Dev->Sys Migration   | 0.027 ms        | 0.025 ms        | -7.0%
>>> # S->D Throughput      | 1.15 GB/s      | 2.69 GB/s      | -57.1%
>>> # D->S Throughput      | 18.12 GB/s      | 19.38 GB/s      | -6.5%
>>> #
>>> # === Half THP Size (1MB) (1.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 0.367 ms        | 1.187 ms        | 69.0%
>>> # Dev->Sys Migration   | 0.048 ms        | 0.049 ms        | 2.2%
>>> # S->D Throughput      | 2.66 GB/s      | 0.82 GB/s      | 222.9%
>>> # D->S Throughput      | 20.53 GB/s      | 20.08 GB/s      | 2.3%
>>> #
>>> # === Single THP Size (2MB) (2.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 0.817 ms        | 0.782 ms        | -4.4%
>>> # Dev->Sys Migration   | 0.089 ms        | 0.096 ms        | 7.1%
>>> # S->D Throughput      | 2.39 GB/s      | 2.50 GB/s      | -4.2%
>>> # D->S Throughput      | 22.00 GB/s      | 20.44 GB/s      | 7.6%
>>> #
>>> # === Two THP Size (4MB) (4.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 3.419 ms        | 2.337 ms        | -46.3%
>>> # Dev->Sys Migration   | 0.321 ms        | 0.225 ms        | -42.6%
>>> # S->D Throughput      | 1.14 GB/s      | 1.67 GB/s      | -31.6%
>>> # D->S Throughput      | 12.17 GB/s      | 17.36 GB/s      | -29.9%
>>> #
>>> # === Four THP Size (8MB) (8.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 4.535 ms        | 4.563 ms        | 0.6%
>>> # Dev->Sys Migration   | 0.583 ms        | 0.582 ms        | -0.2%
>>> # S->D Throughput      | 1.72 GB/s      | 1.71 GB/s      | 0.6%
>>> # D->S Throughput      | 13.39 GB/s      | 13.43 GB/s      | -0.2%
>>> #
>>> # === Eight THP Size (16MB) (16.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 10.190 ms        | 9.805 ms        | -3.9%
>>> # Dev->Sys Migration   | 1.130 ms        | 1.195 ms        | 5.5%
>>> # S->D Throughput      | 1.53 GB/s      | 1.59 GB/s      | -3.8%
>>> # D->S Throughput      | 13.83 GB/s      | 13.07 GB/s      | 5.8%
>>> #
>>> # === One twenty eight THP Size (256MB) (256.0 MB) ===
>>> #                      | With THP        | Without THP     | Improvement
>>> # ---------------------------------------------------------------------
>>> # Sys->Dev Migration   | 80.464 ms        | 92.764 ms        | 13.3%
>>> # Dev->Sys Migration   | 9.528 ms        | 18.166 ms        | 47.6%
>>> # S->D Throughput      | 3.11 GB/s      | 2.70 GB/s      | 15.3%
>>> # D->S Throughput      | 26.24 GB/s      | 13.76 GB/s      | 90.7%
>>> # #            OK  hmm.hmm_device_private.benchmark_thp_migration
>>> # ok 1 hmm.hmm_device_private.benchmark_thp_migration
>>> # #  RUN           hmm.hmm_device_private.migrate_anon_huge_zero_err ...
>>> # # hmm-tests.c:2622:migrate_anon_huge_zero_err:Expected ret (-2) == 0 (0)
>>>
>>> [  154.077143] Unable to handle kernel paging request at virtual address
>>> 0000000000005268
>>> [  154.077179] Mem abort info:
>>> [  154.077203]   ESR = 0x0000000096000007
>>> [  154.077219]   EC = 0x25: DABT (current EL), IL = 32 bits
>>> [  154.078433]   SET = 0, FnV = 0
>>> [  154.078434]   EA = 0, S1PTW = 0
>>> [  154.078435]   FSC = 0x07: level 3 translation fault
>>> [  154.078435] Data abort info:
>>> [  154.078436]   ISV = 0, ISS = 0x00000007, ISS2 = 0x00000000
>>> [  154.078459]   CM = 0, WnR = 0, TnD = 0, TagAccess = 0
>>> [  154.078479]   GCS = 0, Overlay = 0, DirtyBit = 0, Xs = 0
>>> [  154.078484] user pgtable: 16k pages, 47-bit VAs, pgdp=000000010b920000
>>> [  154.078487] [0000000000005268] pgd=0800000101b4c403,
>>> p4d=0800000101b4c403, pud=0800000101b4c403, pmd=0800000108cd8403,
>>> pte=0000000000000000
>>> [  154.078520] Internal error: Oops: 0000000096000007 [#1]  SMP
>>> [  154.098664] Modules linked in: test_hmm rfkill drm fuse backlight ipv6
>>> [  154.100468] CPU: 7 UID: 0 PID: 1357 Comm: hmm-tests Kdump: loaded Not
>>> tainted 7.0.0-rc4-00029-ga989fde763f4-dirty #260 PREEMPT
>>> [  154.103855] Hardware name: QEMU QEMU Virtual Machine, BIOS
>>> edk2-stable202408-prebuilt.qemu.org 08/13/2024
>>> [  154.104409] pstate: 61400005 (nZCv daif +PAN -UAO -TCO +DIT -SSBS
>>> BTYPE=--)
>>> [  154.104847] pc : dmirror_devmem_fault+0xe4/0x1c0 [test_hmm]
>>> [  154.105758] lr : dmirror_devmem_fault+0xcc/0x1c0 [test_hmm]
>>> [  154.109465] sp : ffffc000855ab430
>>> [  154.109677] x29: ffffc000855ab430 x28: ffff8000c9f73e40 x27:
>>> ffff8000c9f73e40
>>> [  154.110091] x26: ffff8000cb920000 x25: ffffc000812e0000 x24:
>>> 0000000000000000
>>> [  154.110540] x23: ffff8000c9f73e40 x22: 0000000000000000 x21:
>>> 0000000000000008
>>> [  154.110888] x20: ffff8000c07e1980 x19: ffffc000855ab618 x18:
>>> ffffc000855abc40
>>> [  154.111223] x17: 0000000000000000 x16: 0000000000000000 x15:
>>> 0000000000000000
>>> [  154.111563] x14: 0000000000000000 x13: 0000000000000000 x12:
>>> ffffc00080fedd68
>>> [  154.111903] x11: 00007fffa3bf7fff x10: 0000000000000000 x9 :
>>> 1ffff00019166a41
>>> [  154.112244] x8 : ffff8000c132df20 x7 : 0000000000000000 x6 :
>>> ffff8000c53bfe88
>>> [  154.112581] x5 : 0000000000000009 x4 : ffffc000855ab3d0 x3 :
>>> 0000000000000004
>>> [  154.112921] x2 : 0000000000000004 x1 : ffff8000c132df18 x0 :
>>> 0000000000005200
>>> [  154.113254] Call trace:
>>> [  154.113370]  dmirror_devmem_fault+0xe4/0x1c0 [test_hmm] (P)
>>> [  154.113679]  do_swap_page+0x132c/0x17b0
>>> [  154.113912]  __handle_mm_fault+0x7e4/0x1af4
>>> [  154.114124]  handle_mm_fault+0xb4/0x294
>>> [  154.114398]  __get_user_pages+0x210/0xbfc
>>> [  154.114607]  get_dump_page+0xd8/0x144
>>> [  154.114795]  dump_user_range+0x70/0x2e8
>>> [  154.115020]  elf_core_dump+0xb64/0xe40
>>> [  154.115212]  vfs_coredump+0xfb4/0x1ce8
>>> [  154.115397]  get_signal+0x6cc/0x844
>>> [  154.115582]  arch_do_signal_or_restart+0x7c/0x33c
>>> [  154.115805]  exit_to_user_mode_loop+0x104/0x16c
>>> [  154.116030]  el0_svc+0x174/0x178
>>> [  154.116216]  el0t_64_sync_handler+0xa0/0xe4
>>> [  154.116414]  el0t_64_sync+0x198/0x19c
>>> [  154.116594] Code: d2800083 f9400280 f9003be0 2a0303e2 (b9406800)
>>> [  154.116891] ---[ end trace 0000000000000000 ]---
>>> [  158.741771] Kernel panic - not syncing: Oops: Fatal exception
>>> [  158.742164] SMP: stopping secondary CPUs
>>> [  158.742970] Kernel Offset: disabled
>>> [  158.743162] CPU features: 0x0000000,00060005,11210501,94067723
>>> [  158.743440] Memory Limit: none
>>> [  164.002089] Starting crashdump kernel...
>>> [  164.002867] Bye!
>>
>> That 'Bye!' is delightful :)
>>
>>>
>>> [root@localhost linux]# ./scripts/faddr2line lib/test_hmm.ko
>>> dmirror_devmem_fault+0xe4/0x1c0
>>> dmirror_devmem_fault+0xe4/0x1c0:
>>> dmirror_select_device at /root/code/linux/lib/test_hmm.c:153
>>> (inlined by) dmirror_devmem_fault at /root/code/linux/lib/test_hmm.c:1659
>>>
>>> The kernel is built with arm64's virt.config plus
>>>
>>> +CONFIG_ARM64_16K_PAGES=y
>>> +CONFIG_ZONE_DEVICE=y
>>> +CONFIG_DEVICE_PRIVATE=y
>>> +CONFIG_TEST_HMM=m
>>>
>>> I *guess* the problem is that migrate_anon_huge_zero_err() has chosen an
>>> incorrect THP size (which should be 32M in a system with 16k page size),
>>
>> Yeah, it hardcodes to 2mb:
>>
>> TEST_F(hmm, migrate_anon_huge_zero_err)
>> {
>> 	...
>>
>> 	size = TWOMEG;
>> }
>>
>> Which isn't correct obviously and needs to be fixed.
>>
>> We should read /sys/kernel/mm/transparent_hugepage/hpage_pmd_size instead.
>>
>> vm_utils.h has read_pmd_pagesize() So this can be fixed with:
>>
>> 	size = read_pmd_pagesize();
>>
>> We then madvise(.., MADV_HUGEPAGE) region of size, which is now too small.:
>>
>> TEST_F(hmm, migrate_anon_huge_zero_err)
>> {
>> 	...
>>
>> 	size = TWOMEG;
>>
>> 	...
>>
>> 	ret = madvise(map, size, MADV_HUGEPAGE);
>> 	ASSERT_EQ(ret, 0); <-- but should succeed anyway, just won't do anything
>>
>> 	...
>>
>> 	ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_FLAGS, buffer,
>> 			      HMM_DMIRROR_FLAG_FAIL_ALLOC);
>> }
>>
>> Then we switch into lib/test_hmm.c:
>>
>> static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
>> 					   struct dmirror *dmirror)
>> {
>> 	...
>>
>> 	for (addr = args->start; addr < args->end; ) {
>> 		...
>>
>> 		if (dmirror->flags & HMM_DMIRROR_FLAG_FAIL_ALLOC) {
>> 			dmirror->flags &= ~HMM_DMIRROR_FLAG_FAIL_ALLOC;
>> 			dpage = NULL;  <-- force failure for 1st page
>>
>> 	...
>>
>> 		if (!dpage) {
>> 			...
>>
>> 			if (!is_large) <-- isn't large, as MADV_HUGEPAGE failed
>> 				goto next;
>>
>> 	...
>> next:
>> 		src++;
>> 		dst++;
>> 		addr += PAGE_SIZE;
>> 	}
>> }
>>
>> Back to the hmm-tests.c selftest:
>>
>> TEST_F(hmm, migrate_anon_huge_zero_err)
>> {
>> 	...
>>
>> 	ret = hmm_dmirror_cmd(self->fd, HMM_DMIRROR_FLAGS, buffer,
>> 			      HMM_DMIRROR_FLAG_FAIL_ALLOC);
>> 	ASSERT_EQ(ret, 0);                  <-- succeeds but...
>> 	ASSERT_EQ(buffer->cpages, npages);  <-- cpages = npages - 1.
>> }
>>
>> So then we try to teardown which inokves:
>>
>> FIXTURE_TEARDOWN(hmm)
>> {
>> 	int ret = close(self->fd);  <-- triggers kernel dmirror_fops_release()
>> 	...
>> }
>>
>> In the kernel:
>>
>> static int dmirror_fops_release(struct inode *inode, struct file *filp)
>> {
>> 	struct dmirror *dmirror = filp->private_data;
>> 	...
>>
>> 	kfree(dmirror);  <-- frees dmirror...
>> 	return 0;
>> }
>>
>> So dmirror is fred but in dmirror_migrate_alloc_and_copy(), for all those pages
>> we DID migrate:
>>
>> static void dmirror_migrate_alloc_and_copy(struct migrate_vma *args,
>> 					   struct dmirror *dmirror)
>> {
>> 	...
>>
>> 	for (addr = args->start; addr < args->end; ) {
>> 		...
>>
>> 		if (!dpage) { <-- we will succeed allocation so don't branch.
>> 			...
>> 		}
>>
>> 		rpage = BACKING_PAGE(dpage);
>>
>> 		/*
>> 		 * Normally, a device would use the page->zone_device_data to
>> 		 * point to the mirror but here we use it to hold the page for
>> 		 * the simulated device memory and that page holds the pointer
>> 		 * to the mirror.
>> 		 */
>> 		rpage->zone_device_data = dmirror;
>>
>> 		...
>> 	}
>>
>> 	...
>> }
>>
>> So now a bunch of device private pages have a zone_device_data set to a dangling
>> dmirror pointer.
>>
>> Then on coredump, we walk the VMAs, meaning we fault in device private pages and
>> end up invoking do_swap_page() which in turn calls dmirror_devmem_fault() (via
>> the struct dev_pagemap_ops
>> dmirror_devmem_ops->migrate_to_ram=dmirror_devmem_fault callback)
>>
>> This is via get_dump_page() -> __get_user_pages_locked(..., FOLL_FORCE |
>> FOLL_DUMP | FOLL_GET) -> __get_user_pages() -> handle_mm_fault() ->
>> __handle_mm_fault() -> do_swap_page() and:
>>
>> vm_fault_t do_swap_page(struct vm_fault *vmf)
>> {
>> 	...
>> 		entry = softleaf_from_pte(vmf->orig_pte);
>> 	if (unlikely(!softleaf_is_swap(entry))) {
>> 		if (softleaf_is_migration(entry)) {
>> 			...
>> 		} else if (softleaf_is_device_private(entry)) {
>> 			...
>>
>> 			if (trylock_page(vmf->page)) {
>> 				...
>>
>> 				ret = pgmap->ops->migrate_to_ram(vmf);
>>
>> 				...
>> 			}
>>
>> 			...
>> 		}
>>
>> 		...
>> 	}
>>
>> 	...
>> }
>>
>> (BTW, we seriously need to clean this up).
> 
> What did you have in mind here?
> 

I have the same question, cc's would be helpful as well. (+)Jordan has been running 
this test for his patchset, not sure if he ran into this.

>> And in dmirror_devmem_fault callback():
>>
>> static vm_fault_t dmirror_devmem_fault(struct vm_fault *vmf)
>> {
>> 	...
>>
>> 	/*
>> 	 * Normally, a device would use the page->zone_device_data to point to
>> 	 * the mirror but here we use it to hold the page for the simulated
>> 	 * device memory and that page holds the pointer to the mirror.
>> 	 */
>> 	rpage = folio_zone_device_data(page_folio(vmf->page));
>> 	dmirror = rpage->zone_device_data;
>>
>> 	...
>>
>> 	args.pgmap_owner = dmirror->mdevice; <-- oops
>>
>> 	...
>> }
>>
>> So in terms of fixing:
>>
>> 1. Fix the test (trivial)
>>
>> Use
>>
>> 	size = read_pmd_pagesize();
>>
>> Instead of:
>>
>> 	size = TWOMEG;
> 
> Adding Balbir as this would have come in with his hugepage changes.
> 

Yes I did, agree with 1

>> 2. Have dmirror_fops_release() migrate all the device private pages back to ram
>>    before freeing dmirror or something like this
> 
> Oh yeah that's bad. We definitely need to do that migration once the file is
> closed.
> 

Agreed, it's been that way for a while, it does need cleanup.

>> You'd want to abstract code from dmirror_migrate_to_system() to be shared
>> between the two functions I think.
>>
>> But I leave that as an exercise for the reader :)
> 
> Good thing I can't read :) I can try and put something together but that won't
> happen before next week, so I won't complain if someone beats me to it. Thanks
> for the detailed analysis and report though!
> 

I'll try that at my end as well and see I can reproduce it, but I don't think
I'll win the race with Al or come close at this point.

>>> leading to the failure of the first hmm_migrate_sys_to_dev(). The test
>>> program received a SIGABRT signal and initiated vfs_coredump(). And
>>> something in the test_hmm module doesn't play well with the coredump
>>> process, which ends up with a panic. I'm not familiar with that.
>>>
>>> Note that I can also reproduce the panic by aborting the test manually
>>> with following diff (and skipping migrate_anon_huge{,_zero}_err()):
>>>
>>> diff --git a/tools/testing/selftests/mm/hmm-tests.c
>>> b/tools/testing/selftests/mm/hmm-tests.c
>>> index e8328c89d855..8d8ea8063a73 100644
>>> --- a/tools/testing/selftests/mm/hmm-tests.c
>>> +++ b/tools/testing/selftests/mm/hmm-tests.c
>>> @@ -1027,6 +1027,8 @@ TEST_F(hmm, migrate)
>>>  	ASSERT_EQ(ret, 0);
>>>  	ASSERT_EQ(buffer->cpages, npages);
>>>
>>> +	ASSERT_TRUE(0);
>>
>> This makes sense as the same dangling dmirror pointer issue arises.
>>
>>> +
>>>  	/* Check what the device read. */
>>>  	for (i = 0, ptr = buffer->mirror; i < size / sizeof(*ptr); ++i)
>>>  		ASSERT_EQ(ptr[i], i);
>>>
>>> Please have a look!
>>
>> Hopefully did so usefully here :)
>>
>>>
>>> Thanks,
>>> Zenghui
>>
>> Cheers, Lorenzo
>>

Thanks for the bug report

Balbir