At boot we print the ranges we've mapped for the linear mapping and
what page size we've used. Also track whether the range is mapped
executable or not and display that as well.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If we look closely at the logic in create_physical_mapping(), when
we're doing STRICT_KERNEL_RWX, we do the following steps:
- determine the gap from where we are to the end of the range
- choose an appropriate mapping_size based on the gap
- check if that mapping_size would overlap the __init_begin
boundary, and if not choose an appropriate mapping_size
We can simplify the logic by taking the __init_begin boundary into
account when we calculate the initial gap.
So add a next_boundary() function which tells us what the next
boundary is, either the __init_begin boundary or end. In future we can
add more boundaries.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we have CONFIG_STRICT_KERNEL_RWX enabled, we want to split the
linear mapping at the text/data boundary so we can map the kernel
text read only.
The current logic uses a goto inside the for loop, which works, but is
hard to reason about.
When we hit the goto retry case we set max_mapping_size to PMD_SIZE
and go back to the start.
Setting max_mapping_size means we skip the PUD case and go to the PMD
case.
We know we will pass the alignment and gap checks because the only
reason we are there is we hit the goto retry, and that is guarded by
mapping_size == PUD_SIZE, which means addr is PUD aligned and gap is
greater or equal to PUD_SIZE.
So the only part of the check that can fail is the mmu_psize_defs
check for the 2M page size.
If we just duplicate that check we can avoid the goto, and we get the
same result.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we have CONFIG_STRICT_KERNEL_RWX enabled, we want to split the
linear mapping at the text/data boundary so we can map the kernel
text read only.
Currently we always use a small page at the text/data boundary, even
when that's not necessary:
Mapped 0x0000000000000000-0x0000000000e00000 with 2.00 MiB pages
Mapped 0x0000000000e00000-0x0000000001000000 with 64.0 KiB pages
Mapped 0x0000000001000000-0x0000000040000000 with 2.00 MiB pages
This is because the check that the mapping crosses the __init_begin
boundary is too strict, it also returns true when we map exactly up to
the boundary.
So fix it to check that the mapping would actually map past
__init_begin, and with that we see:
Mapped 0x0000000000000000-0x0000000040000000 with 2.00 MiB pages
Mapped 0x0000000040000000-0x0000000100000000 with 1.00 GiB pages
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we have CONFIG_STRICT_KERNEL_RWX enabled, we want to split the
linear mapping at the text/data boundary so we can map the kernel text
read only.
But the current logic uses small pages for the entire text section,
regardless of whether a larger page size would fit. eg. with the
boundary at 16M we could use 2M pages, but instead we use 64K pages up
to the 16M boundary:
Mapped 0x0000000000000000-0x0000000001000000 with 64.0 KiB pages
Mapped 0x0000000001000000-0x0000000040000000 with 2.00 MiB pages
Mapped 0x0000000040000000-0x0000000100000000 with 1.00 GiB pages
This is because the test is checking if addr is < __init_begin
and addr + mapping_size is >= _stext. But that is true for all pages
between _stext and __init_begin.
Instead what we want to check is if we are crossing the text/data
boundary, which is at __init_begin. With that fixed we see:
Mapped 0x0000000000000000-0x0000000000e00000 with 2.00 MiB pages
Mapped 0x0000000000e00000-0x0000000001000000 with 64.0 KiB pages
Mapped 0x0000000001000000-0x0000000040000000 with 2.00 MiB pages
Mapped 0x0000000040000000-0x0000000100000000 with 1.00 GiB pages
ie. we're correctly using 2MB pages below __init_begin, but we still
drop down to 64K pages unnecessarily at the boundary.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we have CONFIG_STRICT_KERNEL_RWX enabled, we try to split the
kernel linear (1:1) mapping so that the kernel text is in a separate
page to kernel data, so we can mark the former read-only.
We could achieve that just by always using 64K pages for the linear
mapping, but we try to be smarter. Instead we use huge pages when
possible, and only switch to smaller pages when necessary.
However we have an off-by-one bug in that logic, which causes us to
calculate the wrong boundary between text and data.
For example with the end of the kernel text at 16M we see:
radix-mmu: Mapped 0x0000000000000000-0x0000000001200000 with 64.0 KiB pages
radix-mmu: Mapped 0x0000000001200000-0x0000000040000000 with 2.00 MiB pages
radix-mmu: Mapped 0x0000000040000000-0x0000000100000000 with 1.00 GiB pages
ie. we mapped from 0 to 18M with 64K pages, even though the boundary
between text and data is at 16M.
With the fix we see we're correctly hitting the 16M boundary:
radix-mmu: Mapped 0x0000000000000000-0x0000000001000000 with 64.0 KiB pages
radix-mmu: Mapped 0x0000000001000000-0x0000000040000000 with 2.00 MiB pages
radix-mmu: Mapped 0x0000000040000000-0x0000000100000000 with 1.00 GiB pages
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch fixes the following warnings (obtained with make W=1).
arch/powerpc/mm/slice.c: In function 'slice_range_to_mask':
arch/powerpc/mm/slice.c:73:12: error: comparison is always true due to limited range of data type [-Werror=type-limits]
if (start < SLICE_LOW_TOP) {
^
arch/powerpc/mm/slice.c:81:20: error: comparison is always false due to limited range of data type [-Werror=type-limits]
if ((start + len) > SLICE_LOW_TOP) {
^
arch/powerpc/mm/slice.c: In function 'slice_mask_for_free':
arch/powerpc/mm/slice.c:136:17: error: comparison is always true due to limited range of data type [-Werror=type-limits]
if (high_limit <= SLICE_LOW_TOP)
^
arch/powerpc/mm/slice.c: In function 'slice_check_range_fits':
arch/powerpc/mm/slice.c:185:12: error: comparison is always true due to limited range of data type [-Werror=type-limits]
if (start < SLICE_LOW_TOP) {
^
arch/powerpc/mm/slice.c:195:39: error: comparison is always false due to limited range of data type [-Werror=type-limits]
if (SLICE_NUM_HIGH && ((start + len) > SLICE_LOW_TOP)) {
^
arch/powerpc/mm/slice.c: In function 'slice_scan_available':
arch/powerpc/mm/slice.c:306:11: error: comparison is always true due to limited range of data type [-Werror=type-limits]
if (addr < SLICE_LOW_TOP) {
^
arch/powerpc/mm/slice.c: In function 'get_slice_psize':
arch/powerpc/mm/slice.c:709:11: error: comparison is always true due to limited range of data type [-Werror=type-limits]
if (addr < SLICE_LOW_TOP) {
^
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch fixes the following warnings (obtained with make W=1).
arch/powerpc/mm/slice.c: At top level:
arch/powerpc/mm/slice.c:682:15: error: no previous prototype for 'arch_get_unmapped_area' [-Werror=missing-prototypes]
unsigned long arch_get_unmapped_area(struct file *filp,
^
arch/powerpc/mm/slice.c:692:15: error: no previous prototype for 'arch_get_unmapped_area_topdown' [-Werror=missing-prototypes]
unsigned long arch_get_unmapped_area_topdown(struct file *filp,
^
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Add a trace point for tlbia (Translation Lookaside Buffer Invalidate
All) instruction.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Since commit bd0dbb73e0 ("powerpc/mm/books3s: Add new pte bit to
mark pte temporarily invalid."), _PAGE_PRESENT doesn't mean exactly
that a page is present. A page is also considered preset when
_PAGE_INVALID is set.
This patch changes the meaning of "present" and adds a status "valid"
associated to the _PAGE_PRESENT flag.
Fixes: bd0dbb73e0 ("powerpc/mm/books3s: Add new pte bit to mark pte temporarily invalid.")
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Back when I added -Werror in commit ba55bd7436 ("powerpc: Add
configurable -Werror for arch/powerpc") I did it by adding it to most
of the arch Makefiles.
At the time we excluded math-emu, because apparently it didn't build
cleanly. But that seems to have been fixed somewhere in the interim.
So move the -Werror addition to the top-level of the arch, this saves
us from repeating it in every Makefile and means we won't forget to
add it to any new sub-dirs.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Currently we limit the max addressable memory to 128TB. This patch increase the
limit to 2PB. We can have devices like nvdimm which adds memory above 512TB
limit.
We still don't support regular system ram above 512TB. One of the challenge with
that is the percpu allocator, that allocates per node memory and use the max
distance between them as the percpu offsets. This means with large gap in
address space ( system ram above 1PB) we will run out of vmalloc space to map
the percpu allocation.
In order to support addressable memory above 512TB, kernel should be able to
linear map this range. To do that with hash translation we now add 4 context
to kernel linear map region. Our per context addressable range is 512TB. We
still keep VMALLOC and VMEMMAP region to old size. SLB miss handlers is updated
to validate these limit.
We also limit this update to SPARSEMEM_VMEMMAP and SPARSEMEM_EXTREME
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We will be adding get_kernel_context later. Update function name to indicate
this handle context allocation user space address.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This adds CONFIG_DEBUG_VM checks to ensure:
- The kernel stack is in the SLB after it's flushed and bolted.
- We don't insert an SLB for an address that is aleady in the SLB.
- The kernel SLB miss handler does not take an SLB miss.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
slb_flush_and_rebolt() is misleading, it is called in virtual mode, so
it can not possibly change the stack, so it should not be touching the
shadow area. And since vmalloc is no longer bolted, it should not
change any bolted mappings at all.
Change the name to slb_flush_and_restore_bolted(), and have it just
load the kernel stack from what's currently in the shadow SLB area.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When switching processes, currently all user SLBEs are cleared, and a
few (exec_base, pc, and stack) are preloaded. In trivial testing with
small apps, this tends to miss the heap and low 256MB segments, and it
will also miss commonly accessed segments on large memory workloads.
Add a simple round-robin preload cache that just inserts the last SLB
miss into the head of the cache and preloads those at context switch
time. Every 256 context switches, the oldest entry is removed from the
cache to shrink the cache and require fewer slbmte if they are unused.
Much more could go into this, including into the SLB entry reclaim
side to track some LRU information etc, which would require a study of
large memory workloads. But this is a simple thing we can do now that
is an obvious win for common workloads.
With the full series, process switching speed on the context_switch
benchmark on POWER9/hash (with kernel speculation security masures
disabled) increases from 140K/s to 178K/s (27%).
POWER8 does not change much (within 1%), it's unclear why it does not
see a big gain like POWER9.
Booting to busybox init with 256MB segments has SLB misses go down
from 945 to 69, and with 1T segments 900 to 21. These could almost all
be eliminated by preloading a bit more carefully with ELF binary
loading.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This will be used by the SLB code in the next patch, but for now this
sets the slb_addr_limit to the correct size for 32-bit tasks.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Add 32-entry bitmaps to track the allocation status of the first 32
SLB entries, and whether they are user or kernel entries. These are
used to allocate free SLB entries first, before resorting to the round
robin allocator.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch moves SLB miss handlers completely to C, using the standard
exception handler macros to set up the stack and branch to C.
This can be done because the segment containing the kernel stack is
always bolted, so accessing it with relocation on will not cause an
SLB exception.
Arbitrary kernel memory must not be accessed when handling kernel
space SLB misses, so care should be taken there. However user SLB
misses can access any kernel memory, which can be used to move some
fields out of the paca (in later patches).
User SLB misses could quite easily reconcile IRQs and set up a first
class kernel environment and exit via ret_from_except, however that
doesn't seem to be necessary at the moment, so we only do that if a
bad fault is encountered.
[ Credit to Aneesh for bug fixes, error checks, and improvements to
bad address handling, etc ]
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
[mpe: Disallow tracing for all of slb.c for now.]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
_PAGE_PRIVILEGED corresponds to the SH bit which doesn't protect
against user access but only disables ASID verification on kernel
accesses. User access is controlled with _PMD_USER flag.
Name it _PAGE_SH instead of _PAGE_PRIVILEGED
_PAGE_HUGE corresponds to the SPS bit which doesn't really tells
that's it is a huge page but only that it is not a 4k page.
Name it _PAGE_SPS instead of _PAGE_HUGE
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
To reduce the complexity of flag_array, and allow the removal of
default 0 value of non existing flags, lets have one flag_array
table for each platform family with only the really existing flags.
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Get rid of platform specific _PAGE_XXXX in powerpc common code and
use helpers instead.
mm/dump_linuxpagetables.c will be handled separately
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The 'access' parameter of hash_preload() is either 0 or _PAGE_EXEC.
Among the two versions of hash_preload(), only the PPC64 one is
doing something with this 'access' parameter.
In order to remove the use of _PAGE_EXEC outside platform code,
'access' parameter is replaced by 'is_exec' which will be either
true of false, and the PPC64 version of hash_preload() creates
the access flag based on 'is_exec'.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
book3s/32 doesn't define _PAGE_EXEC, so no need to use it.
All other platforms define _PAGE_EXEC so no need to check
it is not NUL when not book3s/32.
Reviewed-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
In order to avoid multiple conversions, handover directly a
pgprot_t to map_kernel_page() as already done for radix.
Do the same for __ioremap_caller() and __ioremap_at().
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Set PAGE_KERNEL directly in the caller and do not rely on a
hack adding PAGE_KERNEL flags when _PAGE_PRESENT is not set.
As already done for PPC64, use pgprot_cache() helpers instead of
_PAGE_XXX flags in PPC32 ioremap() derived functions.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Other arches have ioremap_wt() to map IO areas write-through.
Implement it on PPC as well in order to avoid drivers using
__ioremap(_PAGE_WRITETHRU)
Also implement ioremap_coherent() to avoid drivers using
__ioremap(_PAGE_COHERENT)
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
The powerpc mobility code may receive RTAS requests to perform PRRN
(Platform Resource Reassignment Notification) topology changes at any
time, including during LPAR migration operations.
In some configurations where the affinity of CPUs or memory is being
changed on that platform, the PRRN requests may apply or refer to
outdated information prior to the complete update of the device-tree.
This patch changes the duration for which topology updates are
suppressed during LPAR migrations from just the rtas_ibm_suspend_me()
/ 'ibm,suspend-me' call(s) to cover the entire migration_store()
operation to allow all changes to the device-tree to be applied prior
to accepting and applying any PRRN requests.
For tracking purposes, pr_info notices are added to the functions
start_topology_update() and stop_topology_update() of 'numa.c'.
Signed-off-by: Michael Bringmann <mwb@linux.vnet.ibm.com>
Reviewed-by: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Local radix TLB flush operations that operate on congruence classes
have explicit ERAT flushes for POWER9. The process scoped LPID flush
did not have a flush, so add it.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
PPC_INVALIDATE_ERAT is slbia IH=7 which is a new variant introduced
with POWER9, and the result is undefined on earlier CPUs.
Commits 7b9f71f974 ("powerpc/64s: POWER9 machine check handler") and
d4748276ae ("powerpc/64s: Improve local TLB flush for boot and MCE on
POWER9") caused POWER7/8 code to use this instruction. Remove it. An
ERAT flush can be made by invalidatig the SLB, but before POWER9 that
requires a flush and rebolt.
Fixes: 7b9f71f974 ("powerpc/64s: POWER9 machine check handler")
Fixes: d4748276ae ("powerpc/64s: Improve local TLB flush for boot and MCE on POWER9")
Cc: stable@vger.kernel.org # v4.11+
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When enumerating page size definitions to check hardware support,
we construct a constant which is (1U << (def->shift - 10)).
However, the array of page size definitions is only initalised for
various MMU_PAGE_* constants, so it contains a number of 0-initialised
elements with def->shift == 0. This means we end up shifting by a
very large number, which gives the following UBSan splat:
================================================================================
UBSAN: Undefined behaviour in /home/dja/dev/linux/linux/arch/powerpc/mm/tlb_nohash.c:506:21
shift exponent 4294967286 is too large for 32-bit type 'unsigned int'
CPU: 0 PID: 0 Comm: swapper Not tainted 4.19.0-rc3-00045-ga604f927b012-dirty #6
Call Trace:
[c00000000101bc20] [c000000000a13d54] .dump_stack+0xa8/0xec (unreliable)
[c00000000101bcb0] [c0000000004f20a8] .ubsan_epilogue+0x18/0x64
[c00000000101bd30] [c0000000004f2b10] .__ubsan_handle_shift_out_of_bounds+0x110/0x1a4
[c00000000101be20] [c000000000d21760] .early_init_mmu+0x1b4/0x5a0
[c00000000101bf10] [c000000000d1ba28] .early_setup+0x100/0x130
[c00000000101bf90] [c000000000000528] start_here_multiplatform+0x68/0x80
================================================================================
Fix this by first checking if the element exists (shift != 0) before
constructing the constant.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When a process allocates a hugepage, the following leak is
reported by kmemleak. This is a false positive which is
due to the pointer to the table being stored in the PGD
as physical memory address and not virtual memory pointer.
unreferenced object 0xc30f8200 (size 512):
comm "mmap", pid 374, jiffies 4872494 (age 627.630s)
hex dump (first 32 bytes):
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ................
backtrace:
[<e32b68da>] huge_pte_alloc+0xdc/0x1f8
[<9e0df1e1>] hugetlb_fault+0x560/0x8f8
[<7938ec6c>] follow_hugetlb_page+0x14c/0x44c
[<afbdb405>] __get_user_pages+0x1c4/0x3dc
[<b8fd7cd9>] __mm_populate+0xac/0x140
[<3215421e>] vm_mmap_pgoff+0xb4/0xb8
[<c148db69>] ksys_mmap_pgoff+0xcc/0x1fc
[<4fcd760f>] ret_from_syscall+0x0/0x38
See commit a984506c54 ("powerpc/mm: Don't report PUDs as
memory leaks when using kmemleak") for detailed explanation.
To fix that, this patch tells kmemleak to ignore the allocated
hugepage table.
Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Make sure we are operating on THP and hugetlb entries in the respective hash
fault handling routines.
No functional change in this patch. If we walked the table wrongly before, we
will retry the access.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Update few code paths to check for pmd_large.
set_pmd_at:
We want to use this to store swap pte at pmd level. For swap ptes we don't want
to set H_PAGE_THP_HUGE. Hence check for pmd_large in set_pmd_at. This remove
the false WARN_ON when using this with swap pmd entry.
pmd_page:
We don't really use them on pmd migration entries. But they can also work with
migration entries and we don't differentiate at the pte level. Hence update
pmd_page to work with pmd migration entries too
__find_linux_pte:
lockless page table walk need to handle pmd migration entries. pmd_trans_huge
check will return false on them. We don't set thp = 1 for such entries, but
update hpage_shift correctly. Without this we will walk pmd migration entries
as a pte page pointer which is wrong.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This make hugetlb directory pointer similar to other page able entries. A hugepd
entry is identified by lack of _PAGE_PTE bit set and directory size stored in
HUGEPD_SHIFT_MASK. We update that to also look at _PAGE_PRESENT
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
With this patch we use 0x8000000000000000UL (_PAGE_PRESENT) to indicate a valid
pgd/pud/pmd entry. We also switch the p**_present() to look at this bit.
With pmd_present, we have a special case. We need to make sure we consider a
pmd marked invalid during THP split as present. Right now we clear the
_PAGE_PRESENT bit during a pmdp_invalidate. Inorder to consider this special
case we add a new pte bit _PAGE_INVALID (mapped to _RPAGE_SW0). This bit is
only used with _PAGE_PRESENT cleared. Hence we are not really losing a pte bit
for this special case. pmd_present is also updated to look at _PAGE_INVALID.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This reverts commits:
5e46e29e6a ("powerpc/64s/hash: convert SLB miss handlers to C")
8fed04d0f6 ("powerpc/64s/hash: remove user SLB data from the paca")
655deecf67 ("powerpc/64s/hash: SLB allocation status bitmaps")
2e1626744e ("powerpc/64s/hash: provide arch_setup_exec hooks for hash slice setup")
89ca4e126a ("powerpc/64s/hash: Add a SLB preload cache")
This series had a few bugs, and the fixes are not all trivial. So
revert most of it for now.
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
After migration of a powerpc LPAR, the kernel executes code to
update the system state to reflect new platform characteristics.
Such changes include modifications to device tree properties provided
to the system by PHYP. Property notifications received by the
post_mobility_fixup() code are passed along to the kernel in general
through a call to of_update_property() which in turn passes such
events back to all modules through entries like the '.notifier_call'
function within the NUMA module.
When the NUMA module updates its state, it resets its event timer. If
this occurs after a previous call to stop_topology_update() or on a
system without VPHN enabled, the code runs into an unitialized timer
structure and crashes. This patch adds a safety check along this path
toward the problem code.
An example crash log is as follows.
ibmvscsi 30000081: Re-enabling adapter!
------------[ cut here ]------------
kernel BUG at kernel/time/timer.c:958!
Oops: Exception in kernel mode, sig: 5 [#1]
LE SMP NR_CPUS=2048 NUMA pSeries
Modules linked in: nfsv3 nfs_acl nfs tcp_diag udp_diag inet_diag lockd unix_diag af_packet_diag netlink_diag grace fscache sunrpc xts vmx_crypto pseries_rng sg binfmt_misc ip_tables xfs libcrc32c sd_mod ibmvscsi ibmveth scsi_transport_srp dm_mirror dm_region_hash dm_log dm_mod
CPU: 11 PID: 3067 Comm: drmgr Not tainted 4.17.0+ #179
...
NIP mod_timer+0x4c/0x400
LR reset_topology_timer+0x40/0x60
Call Trace:
0xc0000003f9407830 (unreliable)
reset_topology_timer+0x40/0x60
dt_update_callback+0x100/0x120
notifier_call_chain+0x90/0x100
__blocking_notifier_call_chain+0x60/0x90
of_property_notify+0x90/0xd0
of_update_property+0x104/0x150
update_dt_property+0xdc/0x1f0
pseries_devicetree_update+0x2d0/0x510
post_mobility_fixup+0x7c/0xf0
migration_store+0xa4/0xc0
kobj_attr_store+0x30/0x60
sysfs_kf_write+0x64/0xa0
kernfs_fop_write+0x16c/0x240
__vfs_write+0x40/0x200
vfs_write+0xc8/0x240
ksys_write+0x5c/0x100
system_call+0x58/0x6c
Fixes: 5d88aa85c0 ("powerpc/pseries: Update CPU maps when device tree is updated")
Cc: stable@vger.kernel.org # v3.10+
Signed-off-by: Michael Bringmann <mwb@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
scan_pkey_feature() uses of_property_read_u32_array() to read the
ibm,processor-storage-keys property and calls be32_to_cpu() on the
value it gets. The problem is that of_property_read_u32_array() already
returns the value converted to the CPU byte order.
The value of pkeys_total ends up more or less sane because there's a min()
call in pkey_initialize() which reduces pkeys_total to 32. So in practice
the kernel ignores the fact that the hypervisor reserved one key for
itself (the device tree advertises 31 keys in my test VM).
This is wrong, but the effect in practice is that when a process tries to
allocate the 32nd key, it gets an -EINVAL error instead of -ENOSPC which
would indicate that there aren't any keys available
Fixes: cf43d3b264 ("powerpc: Enable pkey subsystem")
Cc: stable@vger.kernel.org # v4.16+
Signed-off-by: Thiago Jung Bauermann <bauerman@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When switching processes, currently all user SLBEs are cleared, and a
few (exec_base, pc, and stack) are preloaded. In trivial testing with
small apps, this tends to miss the heap and low 256MB segments, and it
will also miss commonly accessed segments on large memory workloads.
Add a simple round-robin preload cache that just inserts the last SLB
miss into the head of the cache and preloads those at context switch
time. Every 256 context switches, the oldest entry is removed from the
cache to shrink the cache and require fewer slbmte if they are unused.
Much more could go into this, including into the SLB entry reclaim
side to track some LRU information etc, which would require a study of
large memory workloads. But this is a simple thing we can do now that
is an obvious win for common workloads.
With the full series, process switching speed on the context_switch
benchmark on POWER9/hash (with kernel speculation security masures
disabled) increases from 140K/s to 178K/s (27%).
POWER8 does not change much (within 1%), it's unclear why it does not
see a big gain like POWER9.
Booting to busybox init with 256MB segments has SLB misses go down
from 945 to 69, and with 1T segments 900 to 21. These could almost all
be eliminated by preloading a bit more carefully with ELF binary
loading.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This will be used by the SLB code in the next patch, but for now this
sets the slb_addr_limit to the correct size for 32-bit tasks.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Add 32-entry bitmaps to track the allocation status of the first 32
SLB entries, and whether they are user or kernel entries. These are
used to allocate free SLB entries first, before resorting to the round
robin allocator.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
User SLB mappig data is copied into the PACA from the mm->context so
it can be accessed by the SLB miss handlers.
After the C conversion, SLB miss handlers now run with relocation on,
and user SLB misses are able to take recursive kernel SLB misses, so
the user SLB mapping data can be removed from the paca and accessed
directly.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch moves SLB miss handlers completely to C, using the standard
exception handler macros to set up the stack and branch to C.
This can be done because the segment containing the kernel stack is
always bolted, so accessing it with relocation on will not cause an
SLB exception.
Arbitrary kernel memory may not be accessed when handling kernel space
SLB misses, so care should be taken there. However user SLB misses can
access any kernel memory, which can be used to move some fields out of
the paca (in later patches).
User SLB misses could quite easily reconcile IRQs and set up a first
class kernel environment and exit via ret_from_except, however that
doesn't seem to be necessary at the moment, so we only do that if a
bad fault is encountered.
[ Credit to Aneesh for bug fixes, error checks, and improvements to bad
address handling, etc ]
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Since RFC:
- Added MSR[RI] handling
- Fixed up a register loss bug exposed by irq tracing (Aneesh)
- Reject misses outside the defined kernel regions (Aneesh)
- Added several more sanity checks and error handling (Aneesh), we may
look at consolidating these tests and tightenig up the code but for
a first pass we decided it's better to check carefully.
Since v1:
- Fixed SLB cache corruption (Aneesh)
- Fixed untidy SLBE allocation "leak" in get_vsid error case
- Now survives some stress testing on real hardware
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
POWER9 introduces SLBIA IH=3, which invalidates all SLB entries and
associated lookaside information that have a class value of 1, which
Linux assigns to user addresses. This matches what switch_slb wants,
and allows a simple fast implementation that avoids the slb_cache
complexity.
As a side-effect, the POWER5 < DD2.1 SLB invalidation workaround is
also avoided on POWER9.
Process context switching rate is improved about 2.2% for a small
process that hits the slb cache which is the best case for the current
code.
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
