On 12.02.25 18:45, Claudio Imbrenda wrote:
On Wed, 12 Feb 2025 17:55:18 +0100
David Hildenbrand <david@xxxxxxxxxx> wrote:
On 31.01.25 12:24, Claudio Imbrenda wrote:
Move gmap related functions from kernel/uv into kvm.
Create a new file to collect gmap-related functions.
Reviewed-by: Janosch Frank <frankja@xxxxxxxxxxxxx>
Reviewed-by: Christoph Schlameuss <schlameuss@xxxxxxxxxxxxx>
[fixed unpack_one(), thanks mhartmay@xxxxxxxxxxxxx]
Link: https://lore.kernel.org/r/20250123144627.312456-6-imbrenda@xxxxxxxxxxxxx
Signed-off-by: Claudio Imbrenda <imbrenda@xxxxxxxxxxxxx>
Message-ID: <20250123144627.312456-6-imbrenda@xxxxxxxxxxxxx>
---
This patch breaks large folio splitting because you end up un-refing
the wrong folios after a split; I tried to make it work, but either
because of other changes in this patch (or in others), I
cannot get it to work and have to give up for today.
yes, I had also noticed that and I already have a fix ready. In fact my
fix was exactly like yours, except that I did not pass the struct folio
anymore to kvm_s390_wiggle_split_folio(), but instead I only pass a
page and use page_folio() at the beginning, and I use
split_huge_page_to_list_to_order() directly instead of split_folio()
unfortunately the fix does not fix the issue I'm seeing....
but putting printks everywhere seems to solve the issue, so it seems to
be a race somewhere
It also doesn't work with a single vCPU for me. The VM is stuck in
With a two vCPUs (so one can report the lockup), I get:
[ 62.645168] rcu: INFO: rcu_sched self-detected stall on CPU
[ 62.645181] rcu: 0-....: (5999 ticks this GP) idle=0104/1/0x4000000000000002 softirq=2/2 fqs=2997
[ 62.645186] rcu: (t=6000 jiffies g=-1199 q=62 ncpus=2)
[ 62.645191] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.14.0-427.33.1.el9_4.s390x #1
[ 62.645194] Hardware name: IBM 3931 LA1 400 (KVM/Linux)
[ 62.645195] Krnl PSW : 0704c00180000000 0000000024b3e776 (set_memory_decrypted+0x66/0xa0)
[ 62.645206] R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:0 PM:0 RI:0 EA:3
[ 62.645208] Krnl GPRS: 00000000ca004000 0000037f00000001 000000008092f000 0000000000000000
[ 62.645210] 0000037fffb1bbc0 0000000000000001 0000000025e75208 000000008092f000
[ 62.645211] 0000000080873808 0000037fffb1bcd8 0000000000001000 0000000025e75220
[ 62.645213] 0000000080281500 00000000258aa480 0000000024c0b17a 0000037fffb1bb20
[ 62.645220] Krnl Code: 0000000024b3e76a: a784000f brc 8,0000000024b3e788
[ 62.645220] 0000000024b3e76e: a7210fff tmll %r2,4095
[ 62.645220] #0000000024b3e772: a7740017 brc 7,0000000024b3e7a0
[ 62.645220] >0000000024b3e776: b9a40034 uvc %r3,%r4,0
[ 62.645220] 0000000024b3e77a: b2220010 ipm %r1
[ 62.645220] 0000000024b3e77e: 8810001c srl %r1,28
[ 62.645220] 0000000024b3e782: ec12fffa017e cij %r1,1,2,0000000024b3e776
[ 62.645220] 0000000024b3e788: a72b1000 aghi %r2,4096
[ 62.645232] Call Trace:
[ 62.645234] [<0000000024b3e776>] set_memory_decrypted+0x66/0xa0
[ 62.645238] [<0000000024c0b17a>] dma_direct_alloc+0x16a/0x2d0
[ 62.645242] [<0000000024c09b92>] dma_alloc_attrs+0x62/0x80
[ 62.645243] [<000000002546c950>] cio_gp_dma_create+0x60/0xa0
[ 62.645248] [<0000000025ebb712>] css_bus_init+0x102/0x1b8
[ 62.645252] [<0000000025ebb7ea>] channel_subsystem_init+0x22/0xf8
[ 62.645254] [<0000000024b149ac>] do_one_initcall+0x3c/0x200
[ 62.645256] [<0000000025e777be>] do_initcalls+0x11e/0x148
[ 62.645260] [<0000000025e77a34>] kernel_init_freeable+0x1cc/0x208
[ 62.645262] [<00000000254ad01e>] kernel_init+0x2e/0x170
[ 62.645264] [<0000000024b16fdc>] __ret_from_fork+0x3c/0x60
[ 62.645266] [<00000000254bb07a>] ret_from_fork+0xa/0x40
The removed PTE lock would only explain it if we would have a concurrent GUP etc.
from QEMU I/O ? Not sure.
To fix the wrong refcount freezing, doing exactly what folio splitting does
(migration PTEs, locking the pagecache etc., freezing->converting,
removing migration ptes) should work, but requires a bit of work.
--
Cheers,
David / dhildenb
