The fast path of get_user_pages_fast() disables IRQs and then does:
- gup_pud_range()
- gup_pmd_range()
- gup_pte_range()
- flush_dcache_page()
However, flush_dcache_page() makes a smp_call_function(), and using
smp_call_function() when IRQs are disabled is not allowed.
In order to work around this problem, this commit prevents the fast
path from get_user_pages_fast() from being used on SMP, and directly
uses the slow path, which doesn't disable interrupts.
Fixes the following warning when get_futex_key() is called:
WARNING: CPU: 1 PID: 410 at kernel/smp.c:416 smp_call_function_many+0x1fc/0x2ac
Modules linked in:
CPU: 1 PID: 410 Comm: wdc-pfmu-test Tainted: G O 4.13.0-00035-g0d3023c733a0-dirty #363
task: 97d9a120 task.stack: 97d2e000
PC is at smp_call_function_many+0x1fc/0x2ac
PR is at smp_call_function_many+0x1e2/0x2ac
PC : 88071fd4 SP : 97d2fcd4 SR : 400080f1
TEA : c06d6840
R0 : 000000f0 R1 : 00000002 R2 : 00000000 R3 : 8849c89c
R4 : 886d62c4 R5 : 886d62c4 R6 : 886db928 R7 : 884a9dec
R8 : 00000000 R9 : ffffffff R10 : 8844667c R11 : 29ee9000
R12 : 8844667c R13 : 00000001 R14 : 97d2fcd4
MACH: 000177e7 MACL: 0000018c GBR : 295742d0 PR : 88071fba
Call trace:
[<8800fe2c>] sh4_flush_dcache_page+0x0/0xa4
[<880720a2>] smp_call_function+0x1e/0x50
[<8800fe2c>] sh4_flush_dcache_page+0x0/0xa4
[<8800ecc0>] flush_dcache_page+0x34/0x74
[<88010d1c>] gup_pte_range+0xd0/0x140
[<88010e04>] gup_pmd_range+0x78/0xa8
[<88010c4c>] gup_pte_range+0x0/0x140
[<88011006>] get_user_pages_fast+0xda/0x130
[<8806f376>] get_futex_key+0x72/0x3a0
[<8806f6a4>] futex_wait_setup+0x0/0x10c
[<88055a08>] __rcu_read_unlock+0x0/0x54
[<880559c8>] __rcu_read_lock+0x0/0x1c
[<88010f2c>] get_user_pages_fast+0x0/0x130
[<8806f6de>] futex_wait_setup+0x3a/0x10c
[<8806f6a4>] futex_wait_setup+0x0/0x10c
[<8806f852>] futex_wait+0xa2/0x1b8
[<8806f6a4>] futex_wait_setup+0x0/0x10c
[<88071c00>] smp_call_function_single+0xac/0x108
[<8800e908>] cache_noop+0x0/0xc
[<8800e908>] cache_noop+0x0/0xc
[<88071018>] do_futex+0x108/0x8f8
[<880bdfe4>] finish_fault+0x38/0x70
[<880bdff2>] finish_fault+0x46/0x70
[<880bf212>] handle_mm_fault+0x2f6/0x954
[<880bf21a>] handle_mm_fault+0x2fe/0x954
[<88030e64>] resched_curr+0x54/0x6c
[<8832be9e>] _raw_spin_lock_irqsave+0xa/0x18
[<880718dc>] SyS_futex+0xd4/0x124
[<8801488e>] SyS_clone+0x16/0x24
[<8800727e>] syscall_call+0x18/0x1e
[<88071808>] SyS_futex+0x0/0x124
Signed-off-by: Thomas Petazzoni <[email protected]>
---
Note: this is really a work around. Perhaps SuperH should use the
generic GUP implementation from mm/gup.c ? Any other solution ?
arch/sh/mm/gup.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/arch/sh/mm/gup.c b/arch/sh/mm/gup.c
index 56c86ca98ecf..075c8efe6c1d 100644
--- a/arch/sh/mm/gup.c
+++ b/arch/sh/mm/gup.c
@@ -234,6 +234,9 @@ int get_user_pages_fast(unsigned long start, int nr_pages, int write,
if (end < start)
goto slow_irqon;
+ if (IS_ENABLED(CONFIG_SMP))
+ goto slow_irqon;
+
local_irq_disable();
pgdp = pgd_offset(mm, addr);
do {
--
2.14.4
Hello!
On 06/15/2018 06:52 PM, Thomas Petazzoni wrote:
> The fast path of get_user_pages_fast() disables IRQs and then does:
>
> - gup_pud_range()
> - gup_pmd_range()
> - gup_pte_range()
> - flush_dcache_page()
>
> However, flush_dcache_page() makes a smp_call_function(), and using
> smp_call_function() when IRQs are disabled is not allowed.
>
> In order to work around this problem, this commit prevents the fast
> path from get_user_pages_fast() from being used on SMP, and directly
^^^^ of?
Else it sounds a bit tautological...
> uses the slow path, which doesn't disable interrupts.
>
> Fixes the following warning when get_futex_key() is called:
>
> WARNING: CPU: 1 PID: 410 at kernel/smp.c:416 smp_call_function_many+0x1fc/0x2ac
> Modules linked in:
>
> CPU: 1 PID: 410 Comm: wdc-pfmu-test Tainted: G O 4.13.0-00035-g0d3023c733a0-dirty #363
> task: 97d9a120 task.stack: 97d2e000
> PC is at smp_call_function_many+0x1fc/0x2ac
> PR is at smp_call_function_many+0x1e2/0x2ac
> PC : 88071fd4 SP : 97d2fcd4 SR : 400080f1
> TEA : c06d6840
> R0 : 000000f0 R1 : 00000002 R2 : 00000000 R3 : 8849c89c
> R4 : 886d62c4 R5 : 886d62c4 R6 : 886db928 R7 : 884a9dec
> R8 : 00000000 R9 : ffffffff R10 : 8844667c R11 : 29ee9000
> R12 : 8844667c R13 : 00000001 R14 : 97d2fcd4
> MACH: 000177e7 MACL: 0000018c GBR : 295742d0 PR : 88071fba
>
> Call trace:
> [<8800fe2c>] sh4_flush_dcache_page+0x0/0xa4
> [<880720a2>] smp_call_function+0x1e/0x50
> [<8800fe2c>] sh4_flush_dcache_page+0x0/0xa4
> [<8800ecc0>] flush_dcache_page+0x34/0x74
> [<88010d1c>] gup_pte_range+0xd0/0x140
> [<88010e04>] gup_pmd_range+0x78/0xa8
> [<88010c4c>] gup_pte_range+0x0/0x140
> [<88011006>] get_user_pages_fast+0xda/0x130
> [<8806f376>] get_futex_key+0x72/0x3a0
> [<8806f6a4>] futex_wait_setup+0x0/0x10c
> [<88055a08>] __rcu_read_unlock+0x0/0x54
> [<880559c8>] __rcu_read_lock+0x0/0x1c
> [<88010f2c>] get_user_pages_fast+0x0/0x130
> [<8806f6de>] futex_wait_setup+0x3a/0x10c
> [<8806f6a4>] futex_wait_setup+0x0/0x10c
> [<8806f852>] futex_wait+0xa2/0x1b8
> [<8806f6a4>] futex_wait_setup+0x0/0x10c
> [<88071c00>] smp_call_function_single+0xac/0x108
> [<8800e908>] cache_noop+0x0/0xc
> [<8800e908>] cache_noop+0x0/0xc
> [<88071018>] do_futex+0x108/0x8f8
> [<880bdfe4>] finish_fault+0x38/0x70
> [<880bdff2>] finish_fault+0x46/0x70
> [<880bf212>] handle_mm_fault+0x2f6/0x954
> [<880bf21a>] handle_mm_fault+0x2fe/0x954
> [<88030e64>] resched_curr+0x54/0x6c
> [<8832be9e>] _raw_spin_lock_irqsave+0xa/0x18
> [<880718dc>] SyS_futex+0xd4/0x124
> [<8801488e>] SyS_clone+0x16/0x24
> [<8800727e>] syscall_call+0x18/0x1e
> [<88071808>] SyS_futex+0x0/0x124
>
> Signed-off-by: Thomas Petazzoni <[email protected]>
[...]
MBR, Sergei
Hello,
On Fri, 15 Jun 2018 19:04:45 +0300, Sergei Shtylyov wrote:
> Hello!
>
> On 06/15/2018 06:52 PM, Thomas Petazzoni wrote:
>
> > The fast path of get_user_pages_fast() disables IRQs and then does:
> >
> > - gup_pud_range()
> > - gup_pmd_range()
> > - gup_pte_range()
> > - flush_dcache_page()
> >
> > However, flush_dcache_page() makes a smp_call_function(), and using
> > smp_call_function() when IRQs are disabled is not allowed.
> >
> > In order to work around this problem, this commit prevents the fast
> > path from get_user_pages_fast() from being used on SMP, and directly
> ^^^^ of?
> Else it sounds a bit tautological...
I re-read again my sentence, and I believe it's correct and it has the
meaning I intended to give. Of course, I'm not a native english
speaker, so I might be wrong on this :-)
I'm actually more interested in hearing whether there is a better
solution to this problem. This patch submission is really a very
proposal, and I'm hoping to get some better suggestions on how to
approach the problem.
Best regards,
Thomas
--
Thomas Petazzoni, CTO, Bootlin (formerly Free Electrons)
Embedded Linux and Kernel engineering
https://bootlin.com
Hello,
On Fri, 15 Jun 2018 19:04:45 +0300, Sergei Shtylyov wrote:
> Hello!
>
> On 06/15/2018 06:52 PM, Thomas Petazzoni wrote:
>
> > The fast path of get_user_pages_fast() disables IRQs and then does:
> >
> > - gup_pud_range()
> > - gup_pmd_range()
> > - gup_pte_range()
> > - flush_dcache_page()
> >
> > However, flush_dcache_page() makes a smp_call_function(), and using
> > smp_call_function() when IRQs are disabled is not allowed.
> >
> > In order to work around this problem, this commit prevents the fast
> > path from get_user_pages_fast() from being used on SMP, and directly
> ^^^^ of?
> Else it sounds a bit tautological...
Yes, agreed.
I can fix that up for a v2, or perhaps such a small typo can be fixed
while applying the patches.
Could someone review and/or merge those patches ?
I sent four patches, which are still pending:
[PATCH] arch/sh: kernel: use KERN_CONT in print_sh_insn()
[PATCH] arch/sh: kernel: use KERN_CONT in dump_mem()
[PATCH] arch/sh: fix SH4 flush_dcache_all() for SMP
[PATCH] arch/sh: use slow path for get_user_pages_fast() on SMP
Note that this last one is more a workaround than a proper fix. It does
the job, but there is very likely a better solution, and I was hoping
to get some insights/suggestions.
Thanks a lot,
Thomas
--
Thomas Petazzoni, CTO, Bootlin (formerly Free Electrons)
Embedded Linux and Kernel engineering
https://bootlin.com