arch/ia64: Restore arch-specific pgd_offset_k implementation
Subsystems (2)
Mike Rapoport
reviewerPATCH v5 1/1 arch/ia64: Restore arch-specific pgd_offset_k implementation
From: Jessica Clarke <[email protected]>
IA-64 is special and treats pgd_offset_k() differently to pgd_offset(), using different formulae to calculate the indices into the kernel and user PGDs. The index into the user PGDs takes into account the region number, but the index into the kernel (init_mm) PGD always assumes a predefined kernel region number. Commit 974b9b2c68 ("mm: consolidate pte_index() and pte_offset_*() definitions") made IA-64 use a generic pgd_offset_k() which incorrectly used pgd_index() for kernel page tables. As a result, the index into the kernel PGD was going out of bounds and the kernel hung during early boot.
Allow overrides of pgd_offset_k() and override it on IA-64 with the old implementation that will correctly index the kernel PGD.
---
Changes in v4:
* Correct the comment that I missed in v4 :(
Changes since v3:
* Correct commit message as Jessica suggested
Changes since v2:
* Rephrase commit message and comment about pgd_offset_k()
Changes since v1:
* Fixed typo in commit message
* Slightly reworded commit message to sound less weird
* Included Adrian's Tested-by
arch/ia64/include/asm/pgtable.h | 9 +++++++++ include/linux/pgtable.h | 2 ++ 2 files changed, 11 insertions(+)
arch/ia64/include/asm/pgtable.h +9 -0
@@ -366,6 +366,15 @@ pgd_index (unsigned long address) }#define pgd_index pgd_index +/* +* In the kernel's mapped region we know everything is in region number 5, so +* as an optimisation its PGD already points to the area for that region. +* However, this also means that we cannot use pgd_index() and we must +* never add the region here. +*/ +#define pgd_offset_k (addr) \ + (init_mm.pgd + (((addr) >> PGDIR_SHIFT) & (PTRS_PER_PGD - 1))) +/* Look up a pgd entry in the gate area. On IA-64, the gate-area resides in the kernel-mapped segment, hence we use pgd_offset_k() here. */
include/linux/pgtable.h +2 -0
@@ -117,7 +117,9 @@ static inline pgd_t *pgd_offset_pgd(pgd_t *pgd, unsigned long address)* a shortcut which implies the use of the kernel's pgd, instead * of a process's */ +#ifndef pgd_offset_k#define pgd_offset_k (address) pgd_offset(&init_mm, (address)) +#endif /* * In many cases it is known that a virtual address is mapped at PMD or PTE
Signature · 1 line
2.26.2
Quoted 3 lines · Mike Rapoport
+/*
+ * In the kernel's mapped region we know everything is in region number 5, so
+ * as an optimisation its PGD already points to the area for that region.
Is it actually an optimisation? Are there any benchmarks where this makes any difference whatsoever? Or should the comment be closer to
/* ia64 is gratuitously different. Cope with it. */
I might suggest that pgd be changed to be the same as every other pgd in the system, and every other architecture.
Quoted 6 lines · Matthew Wilcox
Quoted 3 lines · Mike Rapoport
+/*
+ * In the kernel's mapped region we know everything is in region number 5, so
+ * as an optimisation its PGD already points to the area for that region.
Quoted 7 lines
Is it actually an optimisation? Are there any benchmarks where this makes any difference whatsoever? Or should the comment be closer to
/* ia64 is gratuitously different. Cope with it. */
I might suggest that pgd be changed to be the same as every other pgd in the system, and every other architecture.
IA-64 is just weird. There's an optional hardware page table walker, and two different options for how to use it within that, short-format and long-format. My understanding is that Linux uses the short-format, and that means that the VHPTs (Virtual Hashed Page Table) are per-region. My assumption therefore is that the kernel's PGD is a completely separate allocation from the userspace ones, and it's not just a case of "moving the pointer back to the start" but that "that is all there is", so unifying the two would be a significant undertaking that avoids one special case at the cost of invasive refactoring and a performance hit by duplication and/or unused space. My knowledge of IA-64 translation is pretty limited, that's about all I know (and this likely has mistakes in it), but I know enough to avoid finding out too much more. Never make the mistake of assuming IA-64 can be made to look like any other architecture :)
Jess
Mike Rapoport
reviewerQuoted 9 lines · Matthew Wilcox
Quoted 3 lines · Mike Rapoport
+/*
+ * In the kernel's mapped region we know everything is in region number 5, so
+ * as an optimisation its PGD already points to the area for that region.Is it actually an optimisation? Are there any benchmarks where this makes any difference whatsoever? Or should the comment be closer to
/* ia64 is gratuitously different. Cope with it. */
It better suits arch/ia64/Kconfig ;-)
arch/ia64/Kconfig +1 -0
@@ -62,6 +62,7 @@ config IA64 the 32-bit X86 line. The IA-64 Linux project has a home page at <http://www.linuxia64.org/> and a mailing list at <[email protected]>. + IA-64 is gratuitously different. Cope with it.
Quoted 2 lines
I might suggest that pgd be changed to be the same as every other pgd in the system, and every other architecture.
I did some grepping and got lost in IA-64 assembly, so I agree with Jessica that it would open a can of worms.
Signature · 2 lines
Sincerely yours, Mike.
Quoted 30 lines · Mike Rapoport
From: Jessica Clarke <[email protected]>
IA-64 is special and treats
pgd_offset_k()differently topgd_offset(), using different formulae to calculate the indices into the kernel and user PGDs. The index into the user PGDs takes into account the region number, but the index into the kernel (init_mm) PGD always assumes a predefined kernel region number. Commit974b9b2c68("mm: consolidatepte_index()and pte_offset_*() definitions") made IA-64 use a genericpgd_offset_k()which incorrectly usedpgd_index()for kernel page tables. As a result, the index into the kernel PGD was going out of bounds and the kernel hung during early boot.Allow overrides of
pgd_offset_k()and override it on IA-64 with the old implementation that will correctly index the kernel PGD.Fixes 974b9b2c68 ("mm: consolidate pte_index() and pte_offset_*() definitions")Reported-by John Paul Adrian Glaubitz <[email protected]>Signed-off-by Jessica Clarke <[email protected]>Signed-off-by Mike Rapoport <[email protected]>---
Changes in v4:
* Correct the comment that I missed in v4 :(Changes since v3:
* Correct commit message as Jessica suggestedChanges since v2:
* Rephrase commit message and comment aboutpgd_offset_k()
Quoted 7 lines
Changes since v1:
* Fixed typo in commit message
* Slightly reworded commit message to sound less weird
* Included Adrian's Tested-byarch/ia64/include/asm/pgtable.h | 9 +++++++++ include/linux/pgtable.h | 2 ++ 2 files changed, 11 insertions(+)
I can take this via my tree if there are no objections.
Tony?
Signature · 2 lines
Sincerely yours, Mike.
Quoted 1 line
I can take this via my tree if there are no objections.
Sure. Go ahead. (You can count that as Acked-by)
-Tony