A recent change to the checksum code removed usage of some extra
arguments, alongside with storage on the stack for those, and the stack
pointer no longer needed to be adjusted in the function prologue.
But a left over subtraction wasn't removed in the function epilogue,
causing the function to return with the stack pointer moved 16 bytes
away from where it should have. This corrupted local state and lead to
weird crashes.
This simply removes the leftover instruction from the epilogue.
Fixes: 70d65cd555 ("ppc: propagate the calling conventions change down to csum_partial_copy_generic()")
Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Merge misc updates from Andrew Morton:
"181 patches.
Subsystems affected by this patch series: kbuild, scripts, ntfs,
ocfs2, vfs, mm (slab, slub, kmemleak, dax, debug, pagecache, fadvise,
gup, swap, memremap, memcg, selftests, pagemap, mincore, hmm, dma,
memory-failure, vmallo and migration)"
* emailed patches from Andrew Morton <akpm@linux-foundation.org>: (181 commits)
mm/migrate: remove obsolete comment about device public
mm/migrate: remove cpages-- in migrate_vma_finalize()
mm, oom_adj: don't loop through tasks in __set_oom_adj when not necessary
memblock: use separate iterators for memory and reserved regions
memblock: implement for_each_reserved_mem_region() using __next_mem_region()
memblock: remove unused memblock_mem_size()
x86/setup: simplify reserve_crashkernel()
x86/setup: simplify initrd relocation and reservation
arch, drivers: replace for_each_membock() with for_each_mem_range()
arch, mm: replace for_each_memblock() with for_each_mem_pfn_range()
memblock: reduce number of parameters in for_each_mem_range()
memblock: make memblock_debug and related functionality private
memblock: make for_each_memblock_type() iterator private
mircoblaze: drop unneeded NUMA and sparsemem initializations
riscv: drop unneeded node initialization
h8300, nds32, openrisc: simplify detection of memory extents
arm64: numa: simplify dummy_numa_init()
arm, xtensa: simplify initialization of high memory pages
dma-contiguous: simplify cma_early_percent_memory()
KVM: PPC: Book3S HV: simplify kvm_cma_reserve()
...
Pull seccomp updates from Kees Cook:
"The bulk of the changes are with the seccomp selftests to accommodate
some powerpc-specific behavioral characteristics. Additional cleanups,
fixes, and improvements are also included:
- heavily refactor seccomp selftests (and clone3 selftests
dependency) to fix powerpc (Kees Cook, Thadeu Lima de Souza
Cascardo)
- fix style issue in selftests (Zou Wei)
- upgrade "unknown action" from KILL_THREAD to KILL_PROCESS (Rich
Felker)
- replace task_pt_regs(current) with current_pt_regs() (Denis
Efremov)
- fix corner-case race in USER_NOTIF (Jann Horn)
- make CONFIG_SECCOMP no longer per-arch (YiFei Zhu)"
* tag 'seccomp-v5.10-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/kees/linux: (23 commits)
seccomp: Make duplicate listener detection non-racy
seccomp: Move config option SECCOMP to arch/Kconfig
selftests/clone3: Avoid OS-defined clone_args
selftests/seccomp: powerpc: Set syscall return during ptrace syscall exit
selftests/seccomp: Allow syscall nr and ret value to be set separately
selftests/seccomp: Record syscall during ptrace entry
selftests/seccomp: powerpc: Fix seccomp return value testing
selftests/seccomp: Remove SYSCALL_NUM_RET_SHARE_REG in favor of SYSCALL_RET_SET
selftests/seccomp: Avoid redundant register flushes
selftests/seccomp: Convert REGSET calls into ARCH_GETREG/ARCH_SETREG
selftests/seccomp: Convert HAVE_GETREG into ARCH_GETREG/ARCH_SETREG
selftests/seccomp: Remove syscall setting #ifdefs
selftests/seccomp: mips: Remove O32-specific macro
selftests/seccomp: arm64: Define SYSCALL_NUM_SET macro
selftests/seccomp: arm: Define SYSCALL_NUM_SET macro
selftests/seccomp: mips: Define SYSCALL_NUM_SET macro
selftests/seccomp: Provide generic syscall setting macro
selftests/seccomp: Refactor arch register macros to avoid xtensa special case
selftests/seccomp: Use __NR_mknodat instead of __NR_mknod
selftests/seccomp: Use bitwise instead of arithmetic operator for flags
...
Pull crypto updates from Herbert Xu:
"API:
- Allow DRBG testing through user-space af_alg
- Add tcrypt speed testing support for keyed hashes
- Add type-safe init/exit hooks for ahash
Algorithms:
- Mark arc4 as obsolete and pending for future removal
- Mark anubis, khazad, sead and tea as obsolete
- Improve boot-time xor benchmark
- Add OSCCA SM2 asymmetric cipher algorithm and use it for integrity
Drivers:
- Fixes and enhancement for XTS in caam
- Add support for XIP8001B hwrng in xiphera-trng
- Add RNG and hash support in sun8i-ce/sun8i-ss
- Allow imx-rngc to be used by kernel entropy pool
- Use crypto engine in omap-sham
- Add support for Ingenic X1830 with ingenic"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6: (205 commits)
X.509: Fix modular build of public_key_sm2
crypto: xor - Remove unused variable count in do_xor_speed
X.509: fix error return value on the failed path
crypto: bcm - Verify GCM/CCM key length in setkey
crypto: qat - drop input parameter from adf_enable_aer()
crypto: qat - fix function parameters descriptions
crypto: atmel-tdes - use semicolons rather than commas to separate statements
crypto: drivers - use semicolons rather than commas to separate statements
hwrng: mxc-rnga - use semicolons rather than commas to separate statements
hwrng: iproc-rng200 - use semicolons rather than commas to separate statements
hwrng: stm32 - use semicolons rather than commas to separate statements
crypto: xor - use ktime for template benchmarking
crypto: xor - defer load time benchmark to a later time
crypto: hisilicon/zip - fix the uninitalized 'curr_qm_qp_num'
crypto: hisilicon/zip - fix the return value when device is busy
crypto: hisilicon/zip - fix zero length input in GZIP decompress
crypto: hisilicon/zip - fix the uncleared debug registers
lib/mpi: Fix unused variable warnings
crypto: x86/poly1305 - Remove assignments with no effect
hwrng: npcm - modify readl to readb
...
Pull compat mount cleanups from Al Viro:
"The last remnants of mount(2) compat buried by Christoph.
Buried into NFS, that is.
Generally I'm less enthusiastic about "let's use in_compat_syscall()
deep in call chain" kind of approach than Christoph seems to be, but
in this case it's warranted - that had been an NFS-specific wart,
hopefully not to be repeated in any other filesystems (read: any new
filesystem introducing non-text mount options will get NAKed even if
it doesn't mess the layout up).
IOW, not worth trying to grow an infrastructure that would avoid that
use of in_compat_syscall()..."
* 'compat.mount' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
fs: remove compat_sys_mount
fs,nfs: lift compat nfs4 mount data handling into the nfs code
nfs: simplify nfs4_parse_monolithic
Pull compat quotactl cleanups from Al Viro:
"More Christoph's compat cleanups: quotactl(2)"
* 'work.quota-compat' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
quota: simplify the quotactl compat handling
compat: add a compat_need_64bit_alignment_fixup() helper
compat: lift compat_s64 and compat_u64 to <asm-generic/compat.h>
Pull compat iovec cleanups from Al Viro:
"Christoph's series around import_iovec() and compat variant thereof"
* 'work.iov_iter' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
security/keys: remove compat_keyctl_instantiate_key_iov
mm: remove compat_process_vm_{readv,writev}
fs: remove compat_sys_vmsplice
fs: remove the compat readv/writev syscalls
fs: remove various compat readv/writev helpers
iov_iter: transparently handle compat iovecs in import_iovec
iov_iter: refactor rw_copy_check_uvector and import_iovec
iov_iter: move rw_copy_check_uvector() into lib/iov_iter.c
compat.h: fix a spelling error in <linux/compat.h>
Pull copy_and_csum cleanups from Al Viro:
"Saner calling conventions for csum_and_copy_..._user() and friends"
[ Removing 800+ lines of code and cleaning stuff up is good - Linus ]
* 'work.csum_and_copy' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs:
ppc: propagate the calling conventions change down to csum_partial_copy_generic()
amd64: switch csum_partial_copy_generic() to new calling conventions
sparc64: propagate the calling convention changes down to __csum_partial_copy_...()
xtensa: propagate the calling conventions change down into csum_partial_copy_generic()
mips: propagate the calling convention change down into __csum_partial_copy_..._user()
mips: __csum_partial_copy_kernel() has no users left
mips: csum_and_copy_{to,from}_user() are never called under KERNEL_DS
sparc32: propagate the calling conventions change down to __csum_partial_copy_sparc_generic()
i386: propagate the calling conventions change down to csum_partial_copy_generic()
sh: propage the calling conventions change down to csum_partial_copy_generic()
m68k: get rid of zeroing destination on error in csum_and_copy_from_user()
arm: propagate the calling convention changes down to csum_partial_copy_from_user()
alpha: propagate the calling convention changes down to csum_partial_copy.c helpers
saner calling conventions for csum_and_copy_..._user()
csum_and_copy_..._user(): pass 0xffffffff instead of 0 as initial sum
csum_partial_copy_nocheck(): drop the last argument
unify generic instances of csum_partial_copy_nocheck()
icmp_push_reply(): reorder adding the checksum up
skb_copy_and_csum_bits(): don't bother with the last argument
Pull orphan section checking from Ingo Molnar:
"Orphan link sections were a long-standing source of obscure bugs,
because the heuristics that various linkers & compilers use to handle
them (include these bits into the output image vs discarding them
silently) are both highly idiosyncratic and also version dependent.
Instead of this historically problematic mess, this tree by Kees Cook
(et al) adds build time asserts and build time warnings if there's any
orphan section in the kernel or if a section is not sized as expected.
And because we relied on so many silent assumptions in this area, fix
a metric ton of dependencies and some outright bugs related to this,
before we can finally enable the checks on the x86, ARM and ARM64
platforms"
* tag 'core-build-2020-10-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (36 commits)
x86/boot/compressed: Warn on orphan section placement
x86/build: Warn on orphan section placement
arm/boot: Warn on orphan section placement
arm/build: Warn on orphan section placement
arm64/build: Warn on orphan section placement
x86/boot/compressed: Add missing debugging sections to output
x86/boot/compressed: Remove, discard, or assert for unwanted sections
x86/boot/compressed: Reorganize zero-size section asserts
x86/build: Add asserts for unwanted sections
x86/build: Enforce an empty .got.plt section
x86/asm: Avoid generating unused kprobe sections
arm/boot: Handle all sections explicitly
arm/build: Assert for unwanted sections
arm/build: Add missing sections
arm/build: Explicitly keep .ARM.attributes sections
arm/build: Refactor linker script headers
arm64/build: Assert for unwanted sections
arm64/build: Add missing DWARF sections
arm64/build: Use common DISCARDS in linker script
arm64/build: Remove .eh_frame* sections due to unwind tables
...
Pull x86 irq updates from Thomas Gleixner:
"Surgery of the MSI interrupt handling to prepare the support of
upcoming devices which require non-PCI based MSI handling:
- Cleanup historical leftovers all over the place
- Rework the code to utilize more core functionality
- Wrap XEN PCI/MSI interrupts into an irqdomain to make irqdomain
assignment to PCI devices possible.
- Assign irqdomains to PCI devices at initialization time which
allows to utilize the full functionality of hierarchical
irqdomains.
- Remove arch_.*_msi_irq() functions from X86 and utilize the
irqdomain which is assigned to the device for interrupt management.
- Make the arch_.*_msi_irq() support conditional on a config switch
and let the last few users select it"
* tag 'x86-irq-2020-10-12' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (40 commits)
PCI: MSI: Fix Kconfig dependencies for PCI_MSI_ARCH_FALLBACKS
x86/apic/msi: Unbreak DMAR and HPET MSI
iommu/amd: Remove domain search for PCI/MSI
iommu/vt-d: Remove domain search for PCI/MSI[X]
x86/irq: Make most MSI ops XEN private
x86/irq: Cleanup the arch_*_msi_irqs() leftovers
PCI/MSI: Make arch_.*_msi_irq[s] fallbacks selectable
x86/pci: Set default irq domain in pcibios_add_device()
iommm/amd: Store irq domain in struct device
iommm/vt-d: Store irq domain in struct device
x86/xen: Wrap XEN MSI management into irqdomain
irqdomain/msi: Allow to override msi_domain_alloc/free_irqs()
x86/xen: Consolidate XEN-MSI init
x86/xen: Rework MSI teardown
x86/xen: Make xen_msi_init() static and rename it to xen_hvm_msi_init()
PCI/MSI: Provide pci_dev_has_special_msi_domain() helper
PCI_vmd_Mark_VMD_irqdomain_with_DOMAIN_BUS_VMD_MSI
irqdomain/msi: Provide DOMAIN_BUS_VMD_MSI
x86/irq: Initialize PCI/MSI domain at PCI init time
x86/pci: Reducde #ifdeffery in PCI init code
...
Pull RAS updates from Borislav Petkov:
- Extend the recovery from MCE in kernel space also to processes which
encounter an MCE in kernel space but while copying from user memory
by sending them a SIGBUS on return to user space and umapping the
faulty memory, by Tony Luck and Youquan Song.
- memcpy_mcsafe() rework by splitting the functionality into
copy_mc_to_user() and copy_mc_to_kernel(). This, as a result, enables
support for new hardware which can recover from a machine check
encountered during a fast string copy and makes that the default and
lets the older hardware which does not support that advance recovery,
opt in to use the old, fragile, slow variant, by Dan Williams.
- New AMD hw enablement, by Yazen Ghannam and Akshay Gupta.
- Do not use MSR-tracing accessors in #MC context and flag any fault
while accessing MCA architectural MSRs as an architectural violation
with the hope that such hw/fw misdesigns are caught early during the
hw eval phase and they don't make it into production.
- Misc fixes, improvements and cleanups, as always.
* tag 'ras_updates_for_v5.10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
x86/mce: Allow for copy_mc_fragile symbol checksum to be generated
x86/mce: Decode a kernel instruction to determine if it is copying from user
x86/mce: Recover from poison found while copying from user space
x86/mce: Avoid tail copy when machine check terminated a copy from user
x86/mce: Add _ASM_EXTABLE_CPY for copy user access
x86/mce: Provide method to find out the type of an exception handler
x86/mce: Pass pointer to saved pt_regs to severity calculation routines
x86/copy_mc: Introduce copy_mc_enhanced_fast_string()
x86, powerpc: Rename memcpy_mcsafe() to copy_mc_to_{user, kernel}()
x86/mce: Drop AMD-specific "DEFERRED" case from Intel severity rule list
x86/mce: Add Skylake quirk for patrol scrub reported errors
RAS/CEC: Convert to DEFINE_SHOW_ATTRIBUTE()
x86/mce: Annotate mce_rd/wrmsrl() with noinstr
x86/mce/dev-mcelog: Do not update kflags on AMD systems
x86/mce: Stop mce_reign() from re-computing severity for every CPU
x86/mce: Make mce_rdmsrl() panic on an inaccessible MSR
x86/mce: Increase maximum number of banks to 64
x86/mce: Delay clearing IA32_MCG_STATUS to the end of do_machine_check()
x86/MCE/AMD, EDAC/mce_amd: Remove struct smca_hwid.xec_bitmap
RAS/CEC: Fix cec_init() prototype
In order to make adding configurable features into seccomp easier,
it's better to have the options at one single location, considering
especially that the bulk of seccomp code is arch-independent. An quick
look also show that many SECCOMP descriptions are outdated; they talk
about /proc rather than prctl.
As a result of moving the config option and keeping it default on,
architectures arm, arm64, csky, riscv, sh, and xtensa did not have SECCOMP
on by default prior to this and SECCOMP will be default in this change.
Architectures microblaze, mips, powerpc, s390, sh, and sparc have an
outdated depend on PROC_FS and this dependency is removed in this change.
Suggested-by: Jann Horn <jannh@google.com>
Link: https://lore.kernel.org/lkml/CAG48ez1YWz9cnp08UZgeieYRhHdqh-ch7aNwc4JRBnGyrmgfMg@mail.gmail.com/
Signed-off-by: YiFei Zhu <yifeifz2@illinois.edu>
[kees: added HAVE_ARCH_SECCOMP help text, tweaked wording]
Signed-off-by: Kees Cook <keescook@chromium.org>
Link: https://lore.kernel.org/r/9ede6ef35c847e58d61e476c6a39540520066613.1600951211.git.yifeifz2@illinois.edu
On PPC64, we have mftb().
On PPC32, we have mftbl() and an #define mftb() mftbl().
mftb() and mftbl() are equivalent, their purpose is to read the
content of SPRN_TRBL, as returned by 'mftb' simplified instruction.
binutils seems to define 'mftbl' instruction as an equivalent
of 'mftb'.
However in both 32 bits and 64 bits documentation, only 'mftb' is
defined, and when performing a disassembly with objdump, the displayed
instruction is 'mftb'
No need to have two ways to do the same thing with different
names, rename mftbl() to have only mftb().
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/94dc68d3d9ef9eb549796d4b938b6ba0305a049b.1601556145.git.christophe.leroy@csgroup.eu
Unlike PPC64 which had a single head_64.S, PPC32 are multiple ones.
There is the head_32.S, selected by default based on the value of BITS
and overridden based on some CONFIG_ values. This leads to thinking
that it may be selected by different types of PPC32 platform but
indeed it ends up being selected by book3s/32 only.
Make that explicit by:
- Not doing any default selection based on BITS.
- Renaming head_32.S to head_book3s_32.S.
- Get head_book3s_32.S selected only by CONFIG_PPC_BOOK3S_32.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
[mpe: Fix head_$(BITS).o reference in arch/powerpc/Makefile]
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/319d379f696412681c66a987cc75e6abf8f958d2.1601975100.git.christophe.leroy@csgroup.eu
At the time being, an early hash table is set up when
CONFIG_KASAN is selected.
There is nothing wrong with setting such an early hash table
all the time, even if it is not used. This is a statically
allocated 256 kB table which lies in the init data section.
This makes the code simpler and may in the future allow to
setup early IO mappings with fixmap instead of hard coding BATs.
Put create_hpte() and flush_hash_pages() in the .ref.text section
in order to avoid warning for the reference to early_hash[]. This
reference is removed by MMU_init_hw_patch() before init memory is
freed.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/b8f8101c368b8a6451844a58d7bd7d83c14cf2aa.1601566529.git.christophe.leroy@csgroup.eu
Add NVDIMM_FAMILY_PAPR to the list of valid 'dimm_family_mask'
acceptable by papr_scm. This is needed as since commit
92fe2aa859 ("libnvdimm: Validate command family indices") libnvdimm
performs a validation of 'nd_cmd_pkg.nd_family' received as part of
ND_CMD_CALL processing to ensure only known command families can use
the general ND_CMD_CALL pass-through functionality.
Without this change the ND_CMD_CALL pass-through targeting
NVDIMM_FAMILY_PAPR error out with -EINVAL.
Fixes: 92fe2aa859 ("libnvdimm: Validate command family indices")
Signed-off-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200913211904.24472-1-vaibhav@linux.ibm.com
The inline execution path for the hardware assisted branch flush
instruction failed to set CTR to the correct value before bcctr,
causing a crash when the feature is enabled.
Fixes: 4d24e21cc6 ("powerpc/security: Allow for processors that flush the link stack using the special bcctr")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201007080605.64423-1-npiggin@gmail.com
Patch here adds a cpumask attr to hv_gpci pmu along with ABI documentation.
Primary use to expose the cpumask is for the perf tool which has the
capability to parse the driver sysfs folder and understand the
cpumask file. Having cpumask file will reduce the number of perf command
line parameters (will avoid "-C" option in the perf tool
command line). It can also notify the user which is
the current cpu used to retrieve the counter data.
command:# cat /sys/devices/hv_gpci/cpumask
0
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201003074943.338618-5-kjain@linux.ibm.com
Patch here adds cpu hotplug functions to hv_gpci pmu.
A new cpuhp_state "CPUHP_AP_PERF_POWERPC_HV_GPCI_ONLINE" enum
is added.
The online callback function updates the cpumask only if its
empty. As the primary intention of adding hotplug support
is to designate a CPU to make HCALL to collect the
counter data.
The offline function test and clear corresponding cpu in a cpumask
and update cpumask to any other active cpu.
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201003074943.338618-4-kjain@linux.ibm.com
Commit 9e9f601084 ("powerpc/perf/{hv-gpci, hv-common}: generate
requests with counters annotated") adds a framework for defining
gpci counters.
In this patch, they adds starting_index value as '0xffffffffffffffff'.
which is wrong as starting_index is of size 32 bits.
Because of this, incase we try to run hv-gpci event we get error.
In power9 machine:
command#: perf stat -e hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/
-C 0 -I 1000
event syntax error: '..bie_count_and_time_tlbie_instructions_issued/'
\___ value too big for format, maximum is 4294967295
This patch fix this issue and changes starting_index value to '0xffffffff'
After this patch:
command#: perf stat -e hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/ -C 0 -I 1000
1.000085786 1,024 hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/
2.000287818 1,024 hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/
2.439113909 17,408 hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/
Fixes: 9e9f601084 ("powerpc/perf/{hv-gpci, hv-common}: generate requests with counters annotated")
Signed-off-by: Kajol Jain <kjain@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201003074943.338618-1-kjain@linux.ibm.com
If the RTAS call to query the PE address for a device fails we jump the
err: label where an error message is printed along with the return code.
However, the printed return code is from the "ret" variable which isn't set
at that point since we assigned the result to "addr" instead. Fix this by
consistently using the "ret" variable for the result of the RTAS call
helpers an dropping the "addr" local variable"
Fixes: 98ba956f6a ("powerpc/pseries/eeh: Rework device EEH PE determination")
Signed-off-by: Oliver O'Halloran <oohall@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201007040903.819081-2-oohall@gmail.com
During memory hot-add, dlpar_add_lmb() calls memory_add_physaddr_to_nid()
to determine which node id (nid) to use when later calling __add_memory().
This is wasteful. On pseries, memory_add_physaddr_to_nid() finds an
appropriate nid for a given address by looking up the LMB containing the
address and then passing that LMB to of_drconf_to_nid_single() to get the
nid. In dlpar_add_lmb() we get this address from the LMB itself.
In short, we have a pointer to an LMB and then we are searching for
that LMB *again* in order to find its nid.
If we call of_drconf_to_nid_single() directly from dlpar_add_lmb() we
can skip the redundant lookup. The only error handling we need to
duplicate from memory_add_physaddr_to_nid() is the fallback to the
default nid when drconf_to_nid_single() returns -1 (NUMA_NO_NODE) or
an invalid nid.
Skipping the extra lookup makes hot-add operations faster, especially
on machines with many LMBs.
Consider an LPAR with 126976 LMBs. In one test, hot-adding 126000
LMBs on an upatched kernel took ~3.5 hours while a patched kernel
completed the same operation in ~2 hours:
Unpatched (12450 seconds):
Sep 9 04:06:31 ltc-brazos1 drmgr[810169]: drmgr: -c mem -a -q 126000
Sep 9 04:06:31 ltc-brazos1 kernel: pseries-hotplug-mem: Attempting to hot-add 126000 LMB(s)
[...]
Sep 9 07:34:01 ltc-brazos1 kernel: pseries-hotplug-mem: Memory at 20000000 (drc index 80000002) was hot-added
Patched (7065 seconds):
Sep 8 21:49:57 ltc-brazos1 drmgr[877703]: drmgr: -c mem -a -q 126000
Sep 8 21:49:57 ltc-brazos1 kernel: pseries-hotplug-mem: Attempting to hot-add 126000 LMB(s)
[...]
Sep 8 23:27:42 ltc-brazos1 kernel: pseries-hotplug-mem: Memory at 20000000 (drc index 80000002) was hot-added
It should be noted that the speedup grows more substantial when
hot-adding LMBs at the end of the drconf range. This is because we
are skipping a linear LMB search.
To see the distinction, consider smaller hot-add test on the same
LPAR. A perf-stat run with 10 iterations showed that hot-adding 4096
LMBs completed less than 1 second faster on a patched kernel:
Unpatched:
Performance counter stats for 'drmgr -c mem -a -q 4096' (10 runs):
104,753.42 msec task-clock # 0.992 CPUs utilized ( +- 0.55% )
4,708 context-switches # 0.045 K/sec ( +- 0.69% )
2,444 cpu-migrations # 0.023 K/sec ( +- 1.25% )
394 page-faults # 0.004 K/sec ( +- 0.22% )
445,902,503,057 cycles # 4.257 GHz ( +- 0.55% ) (66.67%)
8,558,376,740 stalled-cycles-frontend # 1.92% frontend cycles idle ( +- 0.88% ) (49.99%)
300,346,181,651 stalled-cycles-backend # 67.36% backend cycles idle ( +- 0.76% ) (50.01%)
258,091,488,691 instructions # 0.58 insn per cycle
# 1.16 stalled cycles per insn ( +- 0.22% ) (66.67%)
70,568,169,256 branches # 673.660 M/sec ( +- 0.17% ) (50.01%)
3,100,725,426 branch-misses # 4.39% of all branches ( +- 0.20% ) (49.99%)
105.583 +- 0.589 seconds time elapsed ( +- 0.56% )
Patched:
Performance counter stats for 'drmgr -c mem -a -q 4096' (10 runs):
104,055.69 msec task-clock # 0.993 CPUs utilized ( +- 0.32% )
4,606 context-switches # 0.044 K/sec ( +- 0.20% )
2,463 cpu-migrations # 0.024 K/sec ( +- 0.93% )
394 page-faults # 0.004 K/sec ( +- 0.25% )
442,951,129,921 cycles # 4.257 GHz ( +- 0.32% ) (66.66%)
8,710,413,329 stalled-cycles-frontend # 1.97% frontend cycles idle ( +- 0.47% ) (50.06%)
299,656,905,836 stalled-cycles-backend # 67.65% backend cycles idle ( +- 0.39% ) (50.02%)
252,731,168,193 instructions # 0.57 insn per cycle
# 1.19 stalled cycles per insn ( +- 0.20% ) (66.66%)
68,902,851,121 branches # 662.173 M/sec ( +- 0.13% ) (49.94%)
3,100,242,882 branch-misses # 4.50% of all branches ( +- 0.15% ) (49.98%)
104.829 +- 0.325 seconds time elapsed ( +- 0.31% )
This is consistent. An add-by-count hot-add operation adds LMBs
greedily, so LMBs near the start of the drconf range are considered
first. On an otherwise idle LPAR with so many LMBs we would expect to
find the LMBs we need near the start of the drconf range, hence the
smaller speedup.
Signed-off-by: Scott Cheloha <cheloha@linux.ibm.com>
Reviewed-by: Laurent Dufour <ldufour@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20200916145122.3408129-1-cheloha@linux.ibm.com
