Skip to contentSkip to conversation

m68k: mm: Remove check for VM_IO to fix deferred I/O

PATCH In review Geert Uytterhoeven 3 participants 11 messages Jan 28, 2022 – Feb 7, 2022

Subsystems (4)
lore
Share & tools
lore

Geert Uytterhoeven

author maintainer
#0 lore raw

PATCH 1/1 m68k: mm: Remove check for VM_IO to fix deferred I/O

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	vma = find_vma(mm, address);
 	if (!vma)
 		goto map_err;
-	if (vma->vm_flags & VM_IO)
-		goto acc_err;
 	if (vma->vm_start <= address)
 		goto good_area;
 	if (!(vma->vm_flags & VM_GROWSDOWN))
Signature · 1 line
2.25.1
#1 lore raw

Hi Geert,

for hwregs_present(), the exception fixup will handle any access error (through send_fault_sig()), so this should continue to work.

Why the special handling of VM_IO pages? Maybe hp300 had marked all IO register pages VM_IO to distinguish IO faults from VM faults...

The only other area I can imagine this might have an impact is the Mac's pseudo-DMA - FInn might want to give this some testing.

Cheers,

Michael

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 32 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	vma = find_vma(mm, address);
 	if (!vma)
 		goto map_err;
-	if (vma->vm_flags & VM_IO)
-		goto acc_err;
 	if (vma->vm_start <= address)
 		goto good_area;
 	if (!(vma->vm_flags & VM_GROWSDOWN))
#2 lore raw
Quoted 11 lines · Michael Schmitz

Hi Geert,

for hwregs_present(), the exception fixup will handle any access error (through send_fault_sig()), so this should continue to work.

Why the special handling of VM_IO pages? Maybe hp300 had marked all IO register pages VM_IO to distinguish IO faults from VM faults...

The only other area I can imagine this might have an impact is the Mac's pseudo-DMA - FInn might want to give this some testing.

mac_scsi.c and mac_esp.c don't use ioremap(). They rely on head.S:

        mmu_map_eq      #0x50000000,#0x03000000,%d3

Having said that, I will run some tests if you still think it necessary.

#3 lore raw

Hi Finn,

Am 29.01.2022 um 12:55 schrieb Finn Thain:

Quoted 14 lines · Michael Schmitz
Quoted 12 lines

Hi Geert,

for hwregs_present(), the exception fixup will handle any access error (through send_fault_sig()), so this should continue to work.

Why the special handling of VM_IO pages? Maybe hp300 had marked all IO register pages VM_IO to distinguish IO faults from VM faults...

The only other area I can imagine this might have an impact is the Mac's pseudo-DMA - FInn might want to give this some testing.

Quoted 5 lines

mac_scsi.c and mac_esp.c don't use ioremap(). They rely on head.S:

        mmu_map_eq      #0x50000000,#0x03000000,%d3

Having said that, I will run some tests if you still think it necessary.

No need for test then, thanks!

Cheers,

Michael

#4 lore raw
In reply to Geert Uytterhoeven

Hi Geert,

testing this patch on my Falcon 030, I'm seeing a weird error checking and mounting the root filesystem (pata-falcon). The system appears to sit idle, never completing the journal recovery and mount. Still investigating that.

Can't see how that would be caused by your patch, just saying I could not yet test it.

Cheers,

Michael

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 32 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	vma = find_vma(mm, address);
 	if (!vma)
 		goto map_err;
-	if (vma->vm_flags & VM_IO)
-		goto acc_err;
 	if (vma->vm_start <= address)
 		goto good_area;
 	if (!(vma->vm_flags & VM_GROWSDOWN))
#5 lore raw

Hi Geert,

Am 30.01.2022 um 13:32 schrieb Michael Schmitz:

Quoted 6 lines

Hi Geert,

testing this patch on my Falcon 030, I'm seeing a weird error checking and mounting the root filesystem (pata-falcon). The system appears to sit idle, never completing the journal recovery and mount. Still investigating that.

Belay that - not related to your patch, must be some other regression since v5.16 that I'm seeing there.

Just ignore the noise ...

Cheers,

Michael

Quoted 42 lines

Can't see how that would be caused by your patch, just saying I could not yet test it.

Cheers,

Michael

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 33 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned
 long address,
     vma = find_vma(mm, address);
     if (!vma)
         goto map_err;
-    if (vma->vm_flags & VM_IO)
-        goto acc_err;
     if (vma->vm_start <= address)
         goto good_area;
     if (!(vma->vm_flags & VM_GROWSDOWN))
#6 lore raw

Regression in 5.17-rc1 on pata-falcon (was: Re: [PATCH] m68k: mm: Remove check for VM_IO to fix deferred I/O)

Hi Jens,

commit 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 (blk-mq: fix tag_get wait task can't be awakened) does cause a regression on my m68k hardware test rig (m68k Falcon030, IDE disk attached through pata-falcon driver which does use polled IO instead of interrupts, so may be a little on the slow side).

While it usually takes 8 minutes for my system to boot to a point where the network driver is loaded, and another 10 minutes before I can ssh into the box, all the while with IO activity on the disk as seem from the disk activity LED, the boot takes a few hours to complete since v15-rc1, with IO activity only very rarely seen.

In the one case where I could log in remotely, I had to abort the attempted reboot after another few hours.

This problem occurs only on real hardware, and isn't seen on e.g. ARAnyM which is frequently used to test changes.

Bisection between v5.16 and v5.17-rc1 points to 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 as the culprit, which is corroborated by reverting that commit in v5.17-rc1 and booting as rapidly as before.

I don't pretend to understand the purpose of the problematic commit, and cannot spot anything glaringly obvious with the change in logic in e.g. __blk_mq_tag_idle(). If there's anything you'd like me to test that could make that commit work for my unusual set-up, I'd be happy to help.

Cheers,

Michael

Am 30.01.2022 um 19:57 schrieb Michael Schmitz:

Quoted 62 lines

Hi Geert,

Am 30.01.2022 um 13:32 schrieb Michael Schmitz:

Quoted 6 lines

Hi Geert,

testing this patch on my Falcon 030, I'm seeing a weird error checking and mounting the root filesystem (pata-falcon). The system appears to sit idle, never completing the journal recovery and mount. Still investigating that.

Belay that - not related to your patch, must be some other regression since v5.16 that I'm seeing there.

Just ignore the noise ...

Cheers,

Michael

Quoted 42 lines

Can't see how that would be caused by your patch, just saying I could not yet test it.

Cheers,

Michael

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 33 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned
 long address,
     vma = find_vma(mm, address);
     if (!vma)
         goto map_err;
-    if (vma->vm_flags & VM_IO)
-        goto acc_err;
     if (vma->vm_start <= address)
         goto good_area;
     if (!(vma->vm_flags & VM_GROWSDOWN))
author maintainer
#7 lore raw

Re: Regression in 5.17-rc1 on pata-falcon (was: Re: [PATCH] m68k: mm: Remove check for VM_IO to fix deferred I/O)

Hi Michael,

Quoted 10 lines · Michael Schmitz

commit 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 (blk-mq: fix tag_get wait task can't be awakened) does cause a regression on my m68k hardware test rig (m68k Falcon030, IDE disk attached through pata-falcon driver which does use polled IO instead of interrupts, so may be a little on the slow side).

Bisection between v5.16 and v5.17-rc1 points to 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 as the culprit, which is corroborated by reverting that commit in v5.17-rc1 and booting as rapidly as before.

Now you know the culprit, it looks like several other people ran into this. Does this fix help?
https://lore.kernel.org/all/1643040870.3bwvk3sis4.none@localhost/

It is commit 10825410b956dc1e ("blk-mq: Fix wrong wakeup batch configuration which will cause hang") in v5.17-rc2.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

#8 lore raw

Re: Regression in 5.17-rc1 on pata-falcon

Hi Geert, Jens,

thanks for the hint - applying 10825410b956dc1e on top of 5.17-rc1 does indeed fix the issue.

nr_tags == 1 on the Falcon may explain why I ran into this. Oddly enough, I have the same on ARAnyM. Adding a second 'disk' there does reproduce the issue on ARAnyM though. nr_tags == 1 && nr_disks > 1 appears to be sufficient to reproduce this.

I surmised I couldn't be the only one to run into this, but hadn't seen any other reports yet.

Cheers,

Michael

Am 05.02.2022 um 21:31 schrieb Geert Uytterhoeven:

Quoted 32 lines

Hi Michael,

Quoted 10 lines · Michael Schmitz

commit 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 (blk-mq: fix tag_get wait task can't be awakened) does cause a regression on my m68k hardware test rig (m68k Falcon030, IDE disk attached through pata-falcon driver which does use polled IO instead of interrupts, so may be a little on the slow side).

Bisection between v5.16 and v5.17-rc1 points to 180dccb0dba4f5e84a4a70c1be1d34cbb6528b32 as the culprit, which is corroborated by reverting that commit in v5.17-rc1 and booting as rapidly as before.

Now you know the culprit, it looks like several other people ran into this. Does this fix help?
https://lore.kernel.org/all/1643040870.3bwvk3sis4.none@localhost/

It is commit 10825410b956dc1e ("blk-mq: Fix wrong wakeup batch configuration which will cause hang") in v5.17-rc2.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds

Tested-by on v1 1/1
#9 lore raw
In reply to Geert Uytterhoeven

Hi Geert,

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 5 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

Works fine on my Falcon030 when applied to v5.16.

Quoted 27 lines

---
This check was never present in a fault handler on any other architecture than m68k.
Some digging revealed that it was added in v2.1.106, but I couldn't find an email with a patch adding it. That same kernel version extended the use of the hwreg_present() helper to HP9000/300, so the check might have been needed there, perhaps only during development?
The Atari kernel relies heavily on hwreg_present() (both the success and failure cases), and these still work, at least on ARAnyM.

---
 arch/m68k/mm/fault.c | 2 --
 1 file changed, 2 deletions(-)
arch/m68k/mm/fault.c +0 -2
index 1493cf5eac1e7a39..71aa9f6315dc8028 100644
@@ -93,8 +93,6 @@ int do_page_fault(struct pt_regs *regs, unsigned long address,
 	vma = find_vma(mm, address);
 	if (!vma)
 		goto map_err;
-	if (vma->vm_flags & VM_IO)
-		goto acc_err;
 	if (vma->vm_start <= address)
 		goto good_area;
 	if (!(vma->vm_flags & VM_GROWSDOWN))
author maintainer
#10 lore raw
Quoted 10 lines · Michael Schmitz

Am 29.01.2022 um 06:30 schrieb Geert Uytterhoeven:

Quoted 5 lines

When an application accesses a mapped frame buffer backed by deferred I/O, it receives a segmentation fault. Fix this by removing the check for VM_IO in do_page_fault().

Works fine on my Falcon030 when applied to v5.16.

Thanks, queued in the m68k for-v5.18 branch.

Gr{oetje,eeting}s,

Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]

In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds