Force 4K pages for IO addresses.

PATCH Superseded Tony Breeds 4 participants 7 messages Mar 17, 2008 – Mar 18, 2008 PowerPC

lore

Share & tools

lore

Tony Breeds

author

2008-03-17 03:47 UTC #0 lore raw

PATCH 1/1 Force 4K pages for IO addresses.

Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.

Signed-off-by Tony Breeds <[email protected]>

---
Paul, please consider this for 2.6.25

 arch/powerpc/mm/hash_utils_64.c |    4 +---
 1 files changed, 1 insertions(+), 3 deletions(-)

Fixes the following ooops:

Unable to handle kernel paging request for data at address 0xd00008008006c040
Faulting instruction address: 0xc000000000336b08
Oops: Kernel access of bad area, sig: 7 [#1]
SMP NR_CPUS=32 pSeries
Modules linked in:
NIP: c000000000336b08 LR: c000000000336aac CTR: 8000000000138240
REGS: c0000000fcef3190 TRAP: 0300   Not tainted  (2.6.25-rc5.20080317-143135)
MSR: 8000000000009032 <EE,ME,IR,DR>  CR: 24000442  XER: 00000020
DAR: d00008008006c040, DSISR: 0000000042000000

TASK = c0000000fc979190[2781] 'ifconfig' THREAD: c0000000fcef0000 CPU: 7
GPR00: 8000000000000000 c0000000fcef3410 c0000000009cf350 0000000000000000
GPR04: 0000000000000005 0000000000000000 c0000000fc470e50 0000000000000000
GPR08: c0000000fc11fc00 d00008008006c040 f000ffffffffffff 0000000000000000
GPR12: 0000000000000000 c0000000008aed80 0000000028004422 0000000000000000
GPR16: 0000000010120588 00000000100d0000 00000000100b0000 00000000100d0000
GPR20: 0000000010010474 c0000000fea08000 c0000000fcef3cd0 0000000000008914
GPR24: 0000000000000000 0000000000000000 c0000000fe02e000 0000000000000060
GPR28: 0000000000000060 0000000000000000 c00000000094afa8 c0000000fc470e40 NIP [c000000000336b08] .ehea_create_cq+0x1c4/0x26c
LR [c000000000336aac] .ehea_create_cq+0x168/0x26c

Call Trace:
[c0000000fcef3410] [c000000000336aac] .ehea_create_cq+0x168/0x26c (unreliable)
[c0000000fcef34b0] [c00000000032efc4] .ehea_init_port_res+0xd0/0x4c0
[c0000000fcef3570] [c000000000331ba8] .ehea_up+0x134/0x754
[c0000000fcef3670] [c0000000003323c4] .ehea_open+0x84/0x104
[c0000000fcef3710] [c0000000004d55a0] .dev_open+0xb8/0x128
[c0000000fcef37a0] [c0000000004d53e0] .dev_change_flags+0x110/0x218
[c0000000fcef3840] [c000000000528db0] .devinet_ioctl+0x2d4/0x76c
[c0000000fcef3940] [c0000000005293a8] .inet_ioctl+0xe0/0x130
[c0000000fcef39c0] [c0000000004c4590] .sock_ioctl+0x2c0/0x314
[c0000000fcef3a60] [c0000000000fb970] .vfs_ioctl+0x5c/0xf0
[c0000000fcef3b00] [c0000000000fbe34] .do_vfs_ioctl+0x430/0x460
[c0000000fcef3bb0] [c0000000000fbeb0] .sys_ioctl+0x4c/0x88
[c0000000fcef3c60] [c00000000012c364] .dev_ifsioc+0x138/0x324
[c0000000fcef3d40] [c00000000012fdbc] .compat_sys_ioctl+0x398/0x428
[c0000000fcef3e30] [c000000000008748] syscall_exit+0x0/0x40

arch/powerpc/mm/hash_utils_64.c +1 -3

index 32f4161..f399b1d 100644
@@ -351,9 +351,7 @@ static void __init htab_init_page_sizes(void)
 		mmu_vmalloc_psize = MMU_PAGE_64K;
 		if (mmu_linear_psize == MMU_PAGE_4K)
 			mmu_linear_psize = MMU_PAGE_64K;
-		if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
-			mmu_io_psize = MMU_PAGE_64K;
-		else
+		if (!cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
 			mmu_ci_restrictions = 1;
 	}
 #endif /* CONFIG_PPC_64K_PAGES */

Signature · 6 lines

1.5.4.3

Yours Tony

  linux.conf.au    http://www.marchsouth.org/
  Jan 19 - 24 2009 The Australian Linux Technical Conference!

Tony Breeds

author

2008-03-17 03:54 UTC #1 lore raw

PATCH v2 1/1 Force 4K pages for IO addresses.

Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.

Signed-off-by Tony Breeds <[email protected]>

---
* Updated the comment aswell, sorry.
* Paul please consider this for 2.6.25

 arch/powerpc/mm/hash_utils_64.c |   13 +++++--------
 1 files changed, 5 insertions(+), 8 deletions(-)

Fixes the following ooops:

Unable to handle kernel paging request for data at address 0xd00008008006c040
Faulting instruction address: 0xc000000000336b08
Oops: Kernel access of bad area, sig: 7 [#1]
SMP NR_CPUS=32 pSeries
Modules linked in:
NIP: c000000000336b08 LR: c000000000336aac CTR: 8000000000138240
REGS: c0000000fcef3190 TRAP: 0300   Not tainted  (2.6.25-rc5.20080317-143135)
MSR: 8000000000009032 <EE,ME,IR,DR>  CR: 24000442  XER: 00000020
DAR: d00008008006c040, DSISR: 0000000042000000

Call Trace:
[c0000000fcef3410] [c000000000336aac] .ehea_create_cq+0x168/0x26c (unreliable)
[c0000000fcef34b0] [c00000000032efc4] .ehea_init_port_res+0xd0/0x4c0
[c0000000fcef3570] [c000000000331ba8] .ehea_up+0x134/0x754
[c0000000fcef3670] [c0000000003323c4] .ehea_open+0x84/0x104
[c0000000fcef3710] [c0000000004d55a0] .dev_open+0xb8/0x128
[c0000000fcef37a0] [c0000000004d53e0] .dev_change_flags+0x110/0x218
[c0000000fcef3840] [c000000000528db0] .devinet_ioctl+0x2d4/0x76c
[c0000000fcef3940] [c0000000005293a8] .inet_ioctl+0xe0/0x130
[c0000000fcef39c0] [c0000000004c4590] .sock_ioctl+0x2c0/0x314
[c0000000fcef3a60] [c0000000000fb970] .vfs_ioctl+0x5c/0xf0
[c0000000fcef3b00] [c0000000000fbe34] .do_vfs_ioctl+0x430/0x460
[c0000000fcef3bb0] [c0000000000fbeb0] .sys_ioctl+0x4c/0x88
[c0000000fcef3c60] [c00000000012c364] .dev_ifsioc+0x138/0x324
[c0000000fcef3d40] [c00000000012fdbc] .compat_sys_ioctl+0x398/0x428
[c0000000fcef3e30] [c000000000008748] syscall_exit+0x0/0x40

arch/powerpc/mm/hash_utils_64.c +5 -8

index 32f4161..8b2c978 100644
@@ -340,20 +340,17 @@ static void __init htab_init_page_sizes(void)
 	/*
 	 * Pick a size for the ordinary pages. Default is 4K, we support
 	 * 64K for user mappings and vmalloc if supported by the processor.
-	 * We only use 64k for ioremap if the processor
-	 * (and firmware) support cache-inhibited large pages.
-	 * If not, we use 4k and set mmu_ci_restrictions so that
-	 * hash_page knows to switch processes that use cache-inhibited
-	 * mappings to 4k pages.
+	 * We always use 4k for ioremap, however if the processor
+	 * (and firmware) support cache-inhibited large pages we set
+	 * mmu_ci_restrictions so that hash_page knows to switch processes that
+	 * use cache-inhibited mappings to 4k pages.
 	 */
 	if (mmu_psize_defs[MMU_PAGE_64K].shift) {
 		mmu_virtual_psize = MMU_PAGE_64K;
 		mmu_vmalloc_psize = MMU_PAGE_64K;
 		if (mmu_linear_psize == MMU_PAGE_4K)
 			mmu_linear_psize = MMU_PAGE_64K;
-		if (cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
-			mmu_io_psize = MMU_PAGE_64K;
-		else
+		if (!cpu_has_feature(CPU_FTR_CI_LARGE_PAGE))
 			mmu_ci_restrictions = 1;
 	}
 #endif /* CONFIG_PPC_64K_PAGES */

Signature · 1 line

1.5.4.3

Olof Johansson

2008-03-18 00:27 UTC #2 lore raw

Quoted 2 lines · Tony Breeds

Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.

Well, that's too bad. Why penalize all platforms for it?

I.e.: Nack, we use 64K iopages on pa6t and it works well. No need to waste tlb and erat space.

-Olof

Benjamin Herrenschmidt

2008-03-18 02:44 UTC #3 lore raw

Quoted 8 lines · Olof Johansson

Quoted 2 lines · Tony Breeds
Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.
Well, that's too bad. Why penalize all platforms for it?
I.e.: Nack, we use 64K iopages on pa6t and it works well. No need to waste tlb and erat space.

We would have to make that pSeries specific for now I suppose...

We don't have a way to know that there "can" be an EHEA right ? It may not be in the device-tree at boot and dynamically added to the partition... And we can't dynamically "demote" the ioremap area like we do for userspace as we have no tracking of bolted entries hashed in early during boot (I plan to fix that but that will take a while).

Cheers,
Ben.

Nathan Lynch

2008-03-18 14:45 UTC #4 lore raw

Quoted 8 lines · Benjamin Herrenschmidt

Quoted 6 lines · Olof Johansson
Quoted 2 lines · Tony Breeds
Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.
Well, that's too bad. Why penalize all platforms for it?

Quoted 8 lines

Quoted 2 lines
I.e.: Nack, we use 64K iopages on pa6t and it works well. No need to waste tlb and erat space.
We would have to make that pSeries specific for now I suppose...
We don't have a way to know that there "can" be an EHEA right ? It may not be in the device-tree at boot and dynamically added to the partition...

The ibm,drc-names property of the root node should have "HEA" strings in it on systems where EHEA can potentially be present.

Benjamin Herrenschmidt

2008-03-18 20:32 UTC #5 lore raw

Quoted 12 lines · Nathan Lynch

Quoted 11 lines · Benjamin Herrenschmidt
Quoted 8 lines · Olof Johansson
Quoted 2 lines · Tony Breeds
Currently HEA requires 4K pages for IO resources. Just set the pages size to IO to 4K.
Well, that's too bad. Why penalize all platforms for it?
I.e.: Nack, we use 64K iopages on pa6t and it works well. No need to waste tlb and erat space.

Quoted 8 lines

Quoted 5 lines
We would have to make that pSeries specific for now I suppose...
We don't have a way to know that there "can" be an EHEA right ? It may not be in the device-tree at boot and dynamically added to the partition...
The ibm,drc-names property of the root node should have "HEA" strings in it on systems where EHEA can potentially be present.

Ah good, I didn't know. That will force us to stick an ugly wart somewhere to test it though, maybe we can just force mmu_ci_restrictions to 1 ? Paul what do you think ?

Ben.

Tony Breeds

author

2008-03-18 21:52 UTC #6 lore raw

Quoted 7 lines · Benjamin Herrenschmidt

Quoted 2 lines · Nathan Lynch
The ibm,drc-names property of the root node should have "HEA" strings in it on systems where EHEA can potentially be present.
Ah good, I didn't know. That will force us to stick an ugly wart somewhere to test it though, maybe we can just force mmu_ci_restrictions to 1 ? Paul what do you think ?

That's not enough, mmu_io_psize also needs to be 4K. I'll poke at doing something platform specific today.

Yours Tony

linux.conf.au http://www.marchsouth.org/
Jan 19 - 24 2009 The Australian Linux Technical Conference!