The cpu_opv system call executes a vector of operations on behalf of
user-space on a specific CPU with interrupts disabled or within an
IPI handler. It is similar to readv() and writev() system calls which
take a "struct iovec" array as argument.
The operations available are: comparison, memcpy, memcpy_release, add,
and add_release. The system call receives a CPU number from user-space
as argument, which is the CPU on which those operations need to be
performed. All pointers in the ops must have been set up to point to
the per CPU memory of the CPU on which the operations should be
executed. The "comparison" operation can be used to check that the data
used in the preparation step did not change between preparation of
system call inputs and operation execution within the irq-off critical
section.
The reason why for requiring all pointer offsets to be calculated by
user-space beforehand is because get_user_pages() first needs to be
used to pin all pages touched by each operation. This takes care of
faulting-in the pages. Those pages are then vmap'd into the kernel
virtual address range. Then the operations are performed atomically with
respect to other thread's execution on that CPU, without generating any
page fault, by means of a IPI handler (or disabling interrupts).
An overall maximum of 4120 bytes in enforced on the sum of operation
length within an operation vector, so user-space cannot generate a
too long irq-off critical section. The maximum number of operations
supported in a vector is 4. User-space can query the maximum vector
size and the number of operation "instructions" supported with by
passing the appropriate flags as system call parameter. The cache cold
critical section duration has been measured as 4.7µs on x86-64 for 16
operations, therefore a more restrictive limit of 4 operations should
cause an even shorter irq-off latency. Each operation is also limited to
a length of 4096 bytes, meaning that an operation can touch a maximum of
4 pages (memcpy: 2 pages for source, 2 pages for destination if
addresses are not aligned on page boundaries).
**** Justification for cpu_opv ****
Here are a few reasons justifying why the cpu_opv system call is
needed in addition to rseq:
1) Allow algorithms to perform per-cpu data migration without relying on
sched_setaffinity()
The use-cases are migration of memory between per-cpu memory free-lists,
and stealing tasks from other per-cpu work queues: each require that
accesses to remote per-cpu data structures are performed.
Just rseq is not enough to cover those use-cases without additionally
relying on sched_setaffinity, which is unfortunately not
CPU-hotplug-safe.
The cpu_opv system call receives a CPU number as argument, and
performs the operation sequence in a IPI handler on the right CPU. The
IPI handler ensures any restartable sequence critical section it
interrupts will be aborted. If the requested CPU is offline, it performs
the operations from the current CPU while preventing CPU hotplug, and
with a mutex held.
2) Handling single-stepping from tools
Tools like debuggers and simulators use single-stepping to run through
existing programs. If core libraries start to use restartable sequences
for e.g. memory allocation, this means pre-existing programs cannot be
single-stepped, simply because the underlying glibc or jemalloc has
changed.
The rseq user-space does expose a __rseq_table section for the sake of
debuggers, so they can skip over the rseq critical sections if they
want. However, this requires upgrading tools, and still breaks
single-stepping in case where glibc or jemalloc is updated, but not the
tooling.
Having a performance-related library improvement break tooling is likely
to cause a big push-back against wide adoption of rseq.
3) Forward-progress guarantee
Having a piece of user-space code that stops progressing due to external
conditions is pretty bad. Developers are used to think of fast-path and
slow-path (e.g. for locking), where the contended vs uncontended cases
have different performance characteristics, but each need to provide
some level of progress guarantees.
There are concerns about using just "rseq" without the associated
slow-path (cpu_opv) that guarantees progress. It's just asking for
trouble when real-life happen: page faults, uprobes, and other
unforeseen conditions that would seldom cause a rseq fast-path to never
progress.
4) Handling page faults
It's pretty easy to come up with corner-case scenarios where rseq does
not progress without the help from cpu_opv. For instance, a system with
swap enabled which is under high memory pressure could trigger page
faults at pretty much every rseq attempt. Although this scenario
is extremely unlikely, rseq becomes the weak link of the chain.
5) Comparison with LL/SC handling of debugger single-stepping
The layman versed in the load-link/store-conditional instructions in
RISC architectures will notice the similarity between rseq and LL/SC
critical sections. The comparison can even be pushed further: since
debuggers can handle those LL/SC critical sections, they should be
able to handle rseq c.s. in the same way.
First, the way gdb recognises LL/SC c.s. patterns is very fragile:
it's limited to specific common patterns, and will miss the pattern
in all other cases. But fear not, having the rseq c.s. expose a
__rseq_table to debuggers removes that guessing part.
The main difference between LL/SC and rseq is that debuggers had
to support single-stepping through LL/SC critical sections from the
get go in order to support a given architecture. For rseq, we're
adding critical sections into pre-existing applications/libraries,
so the user expectation is that tools don't break due to a library
optimization.
6) Perform maintenance operations on per-cpu data
rseq c.s. are quite limited feature-wise: they need to end with a
*single* commit instruction that updates a memory location. On the other
hand, the cpu_opv system call can combine a sequence of operations that
need to be executed atomically with respect to concurrent execution on
a given CPU. While slower than rseq, this allows for more complex
maintenance operations to be performed on per-cpu data concurrently with
rseq fast-paths, in cases where it's not possible to map those sequences
of ops to a rseq.
7) Use cpu_opv as generic implementation for architectures not
implementing rseq assembly code
rseq critical sections require architecture-specific user-space code to
be crafted in order to port an algorithm to a given architecture. In
addition, it requires that the kernel architecture implementation adds
hooks into signal delivery and resume to user-space.
In order to facilitate integration of rseq into user-space, cpu_opv can
provide a (relatively slower) architecture-agnostic implementation of
rseq. This means that user-space code can be ported to all architectures
through use of cpu_opv initially, and have the fast-path use rseq
whenever the asm code is implemented.
8) Allow libraries with multi-part algorithms to work on same per-cpu
data without affecting the allowed cpu mask
The lttng-ust tracer presents an interesting use-case for per-cpu
buffers: the algorithm needs to update a "reserve" counter, serialize
data into the buffer, and then update a "commit" counter _on the same
per-cpu buffer_. Using rseq for both reserve and commit can bring
significant performance benefits.
Clearly, if rseq reserve fails, the algorithm can retry on a different
per-cpu buffer. However, it's not that easy for the commit. It needs to
be performed on the same per-cpu buffer as the reserve.
The cpu_opv system call solves that problem by receiving the cpu number
on which the operation needs to be performed as argument. It uses an
IPI to perform the operations on the requested CPU.
Changing the allowed cpu mask for the current thread is not an
acceptable alternative for a tracing library, because the application
being traced does not expect that mask to be changed by libraries.
9) Ensure that data structures don't need store-release/load-acquire
semantic to handle fall-back
cpu_opv performs the fall-back on the requested CPU with an IPI to that
CPU. Executing the slow-path on the right CPU ensures that
store-release/load-acquire semantic is not required neither on the
fast-path nor slow-path.
10) Allow use of rseq critical sections from signal handlers
Considering that rseq needs to be registered/unregistered from the
current thread, it means there is a window at thread creation/exit where
a signal handler can nest over the thread before rseq is registered by
glibc, or after it has been unregistered by glibc. One possibility to
handle this would be to extend clone() to have rseq registered
immediately when the thread is created, and unregistered implicitly when
the thread vanishes. Adding complexity to clone() has not been an idea
received well so far. So an alternative solution is to ensure that
signal handlers using rseq critical sections have a fallback mechanism
(cpu_opv) to work on per-cpu data structures when they are nested over
threads for which rseq is not currently registered.
11) Inability to mix rseq and non-rseq atomic operations on percpu data
A typical approach when dealing with locking fast-path and slow-path is
to fall-back on a slower/less efficient mechanism to perform what the
fast-path cannot do.
One approach that naturally comes to mind when considering rseq
fast-path abort would be to rather issue the same side-effect by
means of an atomic instruction. Arguably, before rseq, updates to
per-cpu data structures used to be done by reading the current CPU
number and by then using an atomic instruction to update the data in a
way that is safe against concurrent updates. This atomic instruction was
indeed needed to deal with migration between reading the current CPU
number and doing the update.
Unfortunately, it is a _bug_ to mix concurrent access to a per-cpu data
with both rseq (guaranteed to be on the right CPU, never migrated before
the commit) and non-rseq atomic instructions (which can be issued from
the wrong CPU), because the rseq critical section (on the right CPU)
executing concurrently with the atomic instruction (on another CPU due
to migration) on can cause data corruption of the per-cpu data.
**** rseq and cpu_opv use-cases ****
1) per-cpu spinlock
A per-cpu spinlock can be implemented as a rseq consisting of a
comparison operation (== 0) on a word, and a word store (1), followed
by an acquire barrier after control dependency. The unlock path can be
performed with a simple store-release of 0 to the word, which does
not require rseq.
The cpu_opv fallback requires a single-word comparison (== 0) and a
single-word store (1).
2) per-cpu statistics counters
A per-cpu statistics counters can be implemented as a rseq consisting
of a final "add" instruction on a word as commit.
The cpu_opv fallback can be implemented as a "ADD" operation.
Besides statistics tracking, these counters can be used to implement
user-space RCU per-cpu grace period tracking for both single and
multi-process user-space RCU.
3) per-cpu LIFO linked-list (unlimited size stack)
A per-cpu LIFO linked-list has a "push" and "pop" operation,
which respectively adds an item to the list, and removes an
item from the list.
The "push" operation can be implemented as a rseq consisting of
a word comparison instruction against head followed by a word store
(commit) to head. Its cpu_opv fallback can be implemented as a
word-compare followed by word-store as well.
The "pop" operation can be implemented as a rseq consisting of
loading head, comparing it against NULL, loading the next pointer
at the right offset within the head item, and the next pointer as
a new head, returning the old head on success.
The cpu_opv fallback for "pop" differs from its rseq algorithm:
considering that cpu_opv requires to know all pointers at system
call entry so it can pin all pages, so cpu_opv cannot simply load
head and then load the head->next address within the irq-off
critical section. User-space needs to pass the head and head->next
addresses to the kernel, and the kernel needs to check that the
head address is unchanged since it has been loaded by user-space.
However, when accessing head->next in a ABA situation, it's
possible that head is unchanged, but loading head->next can
result in a page fault due to a concurrently freed head object.
This is why the "expect_fault" operation field is introduced: if a
fault is triggered by this access, "-EAGAIN" will be returned by
cpu_opv rather than -EFAULT, thus indicating the the operation
vector should be attempted again. The "pop" operation can thus be
implemented as a word comparison of head against the head loaded
by user-space, followed by a load of the head->next pointer (which
may fault), and a store of that pointer as a new head.
4) per-cpu LIFO ring buffer with pointers to objects (fixed-sized stack)
This structure is useful for passing around allocated objects
by passing pointers through per-cpu fixed-sized stack.
The "push" side can be implemented with a check of the current
offset against the maximum buffer length, followed by a rseq
consisting of a comparison of the previously loaded offset
against the current offset, a word "try store" operation into the
next ring buffer array index (it's OK to abort after a try-store,
since it's not the commit, and its side-effect can be overwritten),
then followed by a word-store to increment the current offset (commit).
The "push" cpu_opv fallback can be done with the comparison, and
two consecutive word stores, all within the irq-off section.
The "pop" side can be implemented with a check that offset is not
0 (whether the buffer is empty), a load of the "head" pointer before the
offset array index, followed by a rseq consisting of a word
comparison checking that the offset is unchanged since previously
loaded, another check ensuring that the "head" pointer is unchanged,
followed by a store decrementing the current offset.
The cpu_opv "pop" can be implemented with the same algorithm
as the rseq fast-path (compare, compare, store).
5) per-cpu LIFO ring buffer with pointers to objects (fixed-sized stack)
supporting "peek" from remote CPU
In order to implement work queues with work-stealing between CPUs, it is
useful to ensure the offset "commit" in scenario 4) "push" have a
store-release semantic, thus allowing remote CPU to load the offset
with acquire semantic, and load the top pointer, in order to check if
work-stealing should be performed. The task (work queue item) existence
should be protected by other means, e.g. RCU.
If the peek operation notices that work-stealing should indeed be
performed, a thread can use cpu_opv to move the task between per-cpu
workqueues, by first invoking cpu_opv passing the remote work queue
cpu number as argument to pop the task, and then again as "push" with
the target work queue CPU number.
6) per-cpu LIFO ring buffer with data copy (fixed-sized stack)
(with and without acquire-release)
This structure is useful for passing around data without requiring
memory allocation by copying the data content into per-cpu fixed-sized
stack.
The "push" operation is performed with an offset comparison against
the buffer size (figuring out if the buffer is full), followed by
a rseq consisting of a comparison of the offset, a try-memcpy attempting
to copy the data content into the buffer (which can be aborted and
overwritten), and a final store incrementing the offset.
The cpu_opv fallback needs to same operations, except that the memcpy
is guaranteed to complete, given that it is performed with irqs
disabled or from IPI handler. This requires a memcpy operation
supporting length up to 4kB.
The "pop" operation is similar to the "push, except that the offset
is first compared to 0 to ensure the buffer is not empty. The
copy source is the ring buffer, and the destination is an output
buffer.
7) per-cpu FIFO ring buffer (fixed-sized queue)
This structure is useful wherever a FIFO behavior (queue) is needed.
One major use-case is tracer ring buffer.
An implementation of this ring buffer has a "reserve", followed by
serialization of multiple bytes into the buffer, ended by a "commit".
The "reserve" can be implemented as a rseq consisting of a word
comparison followed by a word store. The reserve operation moves the
producer "head". The multi-byte serialization can be performed
non-atomically. Finally, the "commit" update can be performed with
a rseq "add" commit instruction with store-release semantic. The
ring buffer consumer reads the commit value with load-acquire
semantic to know whenever it is safe to read from the ring buffer.
This use-case requires that both "reserve" and "commit" operations
be performed on the same per-cpu ring buffer, even if a migration
happens between those operations. In the typical case, both operations
will happens on the same CPU and use rseq. In the unlikely event of a
migration, the cpu_opv system call will ensure the commit can be
performed on the right CPU by migrating the task to that CPU.
On the consumer side, an alternative to using store-release and
load-acquire on the commit counter would be to use cpu_opv to
ensure the commit counter load is performed on the right CPU through an
IPI.
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
CC: "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx>
CC: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
CC: Paul Turner <pjt@xxxxxxxxxx>
CC: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
CC: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
CC: Andi Kleen <andi@xxxxxxxxxxxxxx>
CC: Dave Watson <davejwatson@xxxxxx>
CC: Chris Lameter <cl@xxxxxxxxx>
CC: Ingo Molnar <mingo@xxxxxxxxxx>
CC: "H. Peter Anvin" <hpa@xxxxxxxxx>
CC: Ben Maurer <bmaurer@xxxxxx>
CC: Steven Rostedt <rostedt@xxxxxxxxxxx>
CC: Josh Triplett <josh@xxxxxxxxxxxxxxxx>
CC: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
CC: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
CC: Russell King <linux@xxxxxxxxxxxxxxxx>
CC: Catalin Marinas <catalin.marinas@xxxxxxx>
CC: Will Deacon <will.deacon@xxxxxxx>
CC: Michael Kerrisk <mtk.manpages@xxxxxxxxx>
CC: Boqun Feng <boqun.feng@xxxxxxxxx>
CC: linux-api@xxxxxxxxxxxxxxx
---
Changes since v1:
- handle CPU hotplug,
- cleanup implementation using function pointers: We can use function
pointers to implement the operations rather than duplicating all the
user-access code.
- refuse device pages: Performing cpu_opv operations on io map'd pages
with preemption disabled could generate long preempt-off critical
sections, which leads to unwanted scheduler latency. Return EFAULT if
a device page is received as parameter
- restrict op vector to 4216 bytes length sum: Restrict the operation
vector to length sum of:
- 4096 bytes (typical page size on most architectures, should be
enough for a string, or structures)
- 15 * 8 bytes (typical operations on integers or pointers).
The goal here is to keep the duration of preempt off critical section
short, so we don't add significant scheduler latency.
- Add INIT_ONSTACK macro: Introduce the
CPU_OP_FIELD_u32_u64_INIT_ONSTACK() macros to ensure that users
correctly initialize the upper bits of CPU_OP_FIELD_u32_u64() on their
stack to 0 on 32-bit architectures.
- Add CPU_MB_OP operation:
Use-cases with:
- two consecutive stores,
- a mempcy followed by a store,
require a memory barrier before the final store operation. A typical
use-case is a store-release on the final store. Given that this is a
slow path, just providing an explicit full barrier instruction should
be sufficient.
- Add expect fault field:
The use-case of list_pop brings interesting challenges. With rseq, we
can use rseq_cmpnev_storeoffp_load(), and therefore load a pointer,
compare it against NULL, add an offset, and load the target "next"
pointer from the object, all within a single req critical section.
Life is not so easy for cpu_opv in this use-case, mainly because we
need to pin all pages we are going to touch in the preempt-off
critical section beforehand. So we need to know the target object (in
which we apply an offset to fetch the next pointer) when we pin pages
before disabling preemption.
So the approach is to load the head pointer and compare it against
NULL in user-space, before doing the cpu_opv syscall. User-space can
then compute the address of the head->next field, *without loading it*.
The cpu_opv system call will first need to pin all pages associated
with input data. This includes the page backing the head->next object,
which may have been concurrently deallocated and unmapped. Therefore,
in this case, getting -EFAULT when trying to pin those pages may
happen: it just means they have been concurrently unmapped. This is
an expected situation, and should just return -EAGAIN to user-space,
to user-space can distinguish between "should retry" type of
situations and actual errors that should be handled with extreme
prejudice to the program (e.g. abort()).
Therefore, add "expect_fault" fields along with op input address
pointers, so user-space can identify whether a fault when getting a
field should return EAGAIN rather than EFAULT.
- Add compiler barrier between operations: Adding a compiler barrier
between store operations in a cpu_opv sequence can be useful when
paired with membarrier system call.
An algorithm with a paired slow path and fast path can use
sys_membarrier on the slow path to replace fast-path memory barriers
by compiler barrier.
Adding an explicit compiler barrier between operations allows
cpu_opv to be used as fallback for operations meant to match
the membarrier system call.
Changes since v2:
- Fix memory leak by introducing struct cpu_opv_pinned_pages.
Suggested by Boqun Feng.
- Cast argument 1 passed to access_ok from integer to void __user *,
fixing sparse warning.
Changes since v3:
- Fix !SMP by adding push_task_to_cpu() empty static inline.
- Add missing sys_cpu_opv() asmlinkage declaration to
include/linux/syscalls.h.
Changes since v4:
- Cleanup based on Thomas Gleixner's feedback.
- Handle retry in case where the scheduler migrates the thread away
from the target CPU after migration within the syscall rather than
returning EAGAIN to user-space.
- Move push_task_to_cpu() to its own patch.
- New scheme for touching user-space memory:
1) get_user_pages_fast() to pin/get all pages (which can sleep),
2) vm_map_ram() those pages
3) grab mmap_sem (read lock)
4) __get_user_pages_fast() (or get_user_pages() on failure)
-> Confirm that the same page pointers are returned. This
catches cases where COW mappings are changed concurrently.
-> If page pointers differ, or on gup failure, release mmap_sem,
vm_unmap_ram/put_page and retry from step (1).
-> perform put_page on the extra reference immediately for each
page.
5) preempt disable
6) Perform operations on vmap. Those operations are normal
loads/stores/memcpy.
7) preempt enable
8) release mmap_sem
9) vm_unmap_ram() all virtual addresses
10) put_page() all pages
- Handle architectures with VIVT caches along with vmap(): call
flush_kernel_vmap_range() after each "write" operation. This
ensures that the user-space mapping and vmap reach a consistent
state between each operation.
- Depend on MMU for is_zero_pfn(). e.g. Blackfin and SH architectures
don't provide the zero_pfn symbol.
Changes since v5:
- Fix handling of push_task_to_cpu() when argument is a cpu which is
not part of the task's allowed cpu mask.
- Add CPU_OP_NR_FLAG flag, which returns the number of operations
supported by the system call.
Changes since v6:
- Use __u* in public uapi header rather than uint*_t.
- Disallow cpu_opv targeting noncached vma, which requires using
get_user_pages() rather than get_user_pages_fast() to get the
vma.
- Fix handling of vm_map_ram() errors by increasing nr_vaddr only after
success.
- Issue vm_unmap_aliases() after each cpu_opv system call, thus ensuring
lazy unmapping does not exhaust vmalloc address space in stress-tests on
32-bit systems.
- Use vm_map_user_ram() and vm_unmap_user_ram() to ensure cache coherency
on virtually aliased architectures.
Changes since v7:
- Adapt to removal of types_32_64.h.
Changes since v8:
- Use IPI to interpret operation vector (or interrupt off critical
section). This is possible now that the interpreter touches a shadow
mapping (vmap) of the user-space pages, and it is simpler than trying
to migrate the current thread.
- Update documentation to reflect the change from preempt-off critical
section to IPI.
- Introduce SPDX license comments.
- Remove unused bitwise and shift operations (reduce instruction-set),
- Remove "mb" instruction,
- Introduce memcpy_release and add_release instructions,
- Allow user-space to query operation vector size supported by the kernel,
- Reduce operation vector size supported from 16 to 4.
---
Man page associated:
CPU_OPV(2) Linux Programmer's Manual CPU_OPV(2)
NAME
cpu_opv - Per-CPU-atomic operation vector system call
SYNOPSIS
#include <linux/cpu_opv.h>
int cpu_opv(struct cpu_op * cpu_opv, int cpuopcnt, int cpu, int flags);
DESCRIPTION
The cpu_opv system call executes a vector of operations on behalf
of user-space on a specific CPU atomically with respect to con‐
current execution on that CPU.
The term CPU used in this documentation refers to a hardware exe‐
cution context.
The operations available are: comparison, memcpy, add. Both mem‐
cpy and add operations have a counterpart with release semantic.
The system call receives a CPU number from user-space as argu‐
ment, which is the CPU on which those operations need to be per‐
formed. All pointers in the ops must have been set up to point
to the per CPU memory of the CPU on which the operations should
be executed. The "comparison" operation can be used to check that
the data used in the preparation step did not change between
preparation of system call inputs and operation execution by the
kernel.
An overall maximum of 4216 bytes in enforced on the sum of opera‐
tion length within an operation vector, so user-space cannot gen‐
erate a too long interrupt-off critical section or inter-proces‐
sor interrupt handler. Each operation is also limited a length of
4096 bytes. A maximum limit of 4 operations per cpu_opv syscall
invocation is enforced.
The layout of struct cpu_opv is as follows:
Fields
op Operation of type enum cpu_op_type to perform. This opera‐
tion type selects the associated "u" union field.
len
Length (in bytes) of data to consider for this operation.
u.compare_op
For a CPU_COMPARE_EQ_OP , and CPU_COMPARE_NE_OP , a and b
are pointers to data meant to be compared. The
expect_fault_a and expect_fault_b fields indicate whether
a page fault should be expected when accessing the memory
holding this data. If expect_fault_a , or expect_fault_b
is set, EAGAIN is returned on fault, else EFAULT is
returned. The len field is allowed to take values from 0
to 4096 for comparison operations.
u.memcpy_op
For a CPU_MEMCPY_OP , or CPU_MEMCPY_RELEASE_OP , contains
the dst and src pointers, expressing a copy of src into
dst. The expect_fault_dst and expect_fault_src fields
indicate whether a page fault should be expected when
accessing the memory holding both source and destination,
which starts at the pointer address, of length len . If
expect_fault_dst , or expect_fault_src is set, EAGAIN is
returned on fault, else EFAULT is returned. The len field
is allowed to take values from 0 to 4096 for memcpy opera‐
tions.
u.arithmetic_op
For a CPU_ADD_OP , or CPU_ADD_RELEASE_OP , contains the p
, count , and expect_fault_p fields, which are respec‐
tively a pointer to the memory location to increment, the
64-bit signed integer value to add, and whether a page
fault should be expected for p . If expect_fault_p is
set, EAGAIN is returned on fault, else EFAULT is returned.
The len field is allowed to take values of 1, 2, 4, 8
bytes for arithmetic operations.
The enum cpu_op_types contains the following operations:
· CPU_COMPARE_EQ_OP: Compare whether two memory locations are
equal,
· CPU_COMPARE_NE_OP: Compare whether two memory locations differ,
· CPU_MEMCPY_OP: Copy a source memory location into a destina‐
tion,
· CPU_MEMCPY_RELEASE_OP: Copy a source memory location into a
destination, with release semantic,
· CPU_ADD_OP: Increment a target memory location of a given
count,
· CPU_ADD_RELEASE_OP: Increment a target memory location of a
given count, with release semantic,
All of the operations above provide single-copy atomicity guar‐
antees for word-sized, word-aligned target pointers, for both
loads and stores.
The cpuopcnt argument is the number of elements in the cpu_opv
array. It can take values from 0 to an upper limit returned by
invoking cpu_opv() with the CPU_OP_VEC_LEN_MAX_FLAG flag set.
The cpu argument is the CPU number on which the operation
sequence needs to be executed.
The flags argument is a bitmask. When CPU_OP_NR_FLAG is set, the
cpu_opv() system call returns the number of operations available.
When CPU_OP_VEC_LEN_MAX_FLAG is set, the cpu_opv() system call
returns the maximum length of the sequence of operations that is
accepted as input argument by the system call. When flags is 0,
the sequence of operations received as parameter is performed.
RETURN VALUE
When invoked with flags set at 0, a return value of 0 indicates
success. On error, -1 is returned, and errno is set. If a com‐
parison operation fails, execution of the operation vector is
stopped, and the return value is the index after the comparison
operation (values between 1 and 4).
When flags is non-zero, on error, -1 is returned, and errno is
set. On success, the behavior is described in the DESCRIPTION
section for each flag.
ERRORS
EAGAIN cpu_opv() system call should be attempted again.
EINVAL Either flags contains an invalid value, or cpu contains an
invalid value or a value not allowed by the current
thread's allowed cpu mask, or cpuopcnt contains an invalid
value, or the cpu_opv operation vector contains an invalid
op value, or the cpu_opv operation vector contains an
invalid len value, or the cpu_opv operation vector sum of
len values is too large.
ENOSYS The cpu_opv() system call is not implemented by this ker‐
nel.
EFAULT cpu_opv is an invalid address, or a pointer contained
within an operation is invalid (and a fault is not
expected for that pointer). Pointers to device and non‐
cached memory within an operation are considered invalid.
VERSIONS
The cpu_opv() system call was added in Linux 4.X (TODO).
CONFORMING TO
cpu_opv() is Linux-specific.
SEE ALSO
membarrier(2), rseq(2)
Linux 2018-10-27 CPU_OPV(2)
---
MAINTAINERS | 7 +
include/linux/syscalls.h | 3 +
include/uapi/linux/cpu_opv.h | 69 ++++
init/Kconfig | 18 +
kernel/Makefile | 1 +
kernel/cpu_opv.c | 955 +++++++++++++++++++++++++++++++++++++++++++
kernel/sys_ni.c | 1 +
7 files changed, 1054 insertions(+)
create mode 100644 include/uapi/linux/cpu_opv.h
create mode 100644 kernel/cpu_opv.c
diff --git a/MAINTAINERS b/MAINTAINERS
index b2f710eee67a..de59c7c12c8f 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -3872,6 +3872,13 @@ B: https://bugzilla.kernel.org
F: drivers/cpuidle/*
F: include/linux/cpuidle.h
+PER-CPU-ATOMIC OPERATION VECTOR SUPPORT
+M: Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
+L: linux-kernel@xxxxxxxxxxxxxxx
+S: Supported
+F: kernel/cpu_opv.c
+F: include/uapi/linux/cpu_opv.h
+
CRAMFS FILESYSTEM
M: Nicolas Pitre <nico@xxxxxxxxxx>
S: Maintained
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index 2ff814c92f7f..c5af29eccd0e 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -68,6 +68,7 @@ struct perf_event_attr;
struct file_handle;
struct sigaltstack;
struct rseq;
+struct cpu_op;
union bpf_attr;
#include <linux/types.h>
@@ -906,6 +907,8 @@ asmlinkage long sys_statx(int dfd, const char __user *path, unsigned flags,
unsigned mask, struct statx __user *buffer);
asmlinkage long sys_rseq(struct rseq __user *rseq, uint32_t rseq_len,
int flags, uint32_t sig);
+asmlinkage long sys_cpu_opv(struct cpu_op __user *ucpuopv, int cpuopcnt,
+ int cpu, int flags);
/*
* Architecture-specific system calls
diff --git a/include/uapi/linux/cpu_opv.h b/include/uapi/linux/cpu_opv.h
new file mode 100644
index 000000000000..350e5a7a61f2
--- /dev/null
+++ b/include/uapi/linux/cpu_opv.h
@@ -0,0 +1,69 @@
+/* SPDX-License-Identifier: GPL-2.0+ WITH Linux-syscall-note */
+#ifndef _UAPI_LINUX_CPU_OPV_H
+#define _UAPI_LINUX_CPU_OPV_H
+
+/*
+ * linux/cpu_opv.h
+ *
+ * Per-CPU-atomic operation vector system call API
+ *
+ * Copyright (c) 2017-2018 Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
+ */
+
+#include <linux/types.h>
+
+/* Maximum size of operation structure within struct cpu_op. */
+#define CPU_OP_ARG_LEN_MAX 24
+/* Maximum data len for compare and memcpy operations. */
+#define CPU_OP_DATA_LEN_MAX 4096
+/* Maximum data len for arithmetic operations. */
+#define CPU_OP_ARITHMETIC_DATA_LEN_MAX 8
+
+enum cpu_op_flags {
+ CPU_OP_NR_FLAG = (1U << 0),
+ CPU_OP_VEC_LEN_MAX_FLAG = (1U << 1),
+};
+
+enum cpu_op_type {
+ /* compare */
+ CPU_COMPARE_EQ_OP,
+ CPU_COMPARE_NE_OP,
+ /* memcpy */
+ CPU_MEMCPY_OP,
+ CPU_MEMCPY_RELEASE_OP,
+ /* arithmetic */
+ CPU_ADD_OP,
+ CPU_ADD_RELEASE_OP,
+
+ NR_CPU_OPS,
+};
+
+/* Vector of operations to perform. Limited to 16. */
+struct cpu_op {
+ /* enum cpu_op_type. */
+ __s32 op;
+ /* data length, in bytes. */
+ __u32 len;
+ union {
+ struct {
+ __u64 a;
+ __u64 b;
+ __u8 expect_fault_a;
+ __u8 expect_fault_b;
+ } compare_op;
+ struct {
+ __u64 dst;
+ __u64 src;
+ __u8 expect_fault_dst;
+ __u8 expect_fault_src;
+ } memcpy_op;
+ struct {
+ __u64 p;
+ __s64 count;
+ __u8 expect_fault_p;
+ } arithmetic_op;
+ char __padding[CPU_OP_ARG_LEN_MAX];
+ } u;
+};
+
+#endif /* _UAPI_LINUX_CPU_OPV_H */
diff --git a/init/Kconfig b/init/Kconfig
index 1e234e2f1cba..e7c21a683642 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1483,6 +1483,8 @@ config RSEQ
bool "Enable rseq() system call" if EXPERT
default y
depends on HAVE_RSEQ
+ depends on MMU
+ select CPU_OPV
select MEMBARRIER
help
Enable the restartable sequences system call. It provides a
@@ -1502,6 +1504,22 @@ config DEBUG_RSEQ
If unsure, say N.
+# CPU_OPV depends on MMU for is_zero_pfn()
+config CPU_OPV
+ bool "Enable cpu_opv() system call" if EXPERT
+ default y
+ depends on MMU
+ help
+ Enable the per-CPU-atomic operation vector system call.
+ It allows user-space to perform a sequence of operations on
+ per-CPU data atomically with respect to concurrent execution on that
+ CPU. Useful as single-stepping fall-back for restartable sequences,
+ migration of data between per-CPU data structures, and for performing
+ more complex operations on per-CPU data that would not be otherwise
+ possible to do with restartable sequences.
+
+ If unsure, say Y.
+
config EMBEDDED
bool "Embedded system"
option allnoconfig_y
diff --git a/kernel/Makefile b/kernel/Makefile
index 7a63d567fdb5..507150b93521 100644
--- a/kernel/Makefile
+++ b/kernel/Makefile
@@ -116,6 +116,7 @@ obj-$(CONFIG_TORTURE_TEST) += torture.o
obj-$(CONFIG_HAS_IOMEM) += iomem.o
obj-$(CONFIG_ZONE_DEVICE) += memremap.o
obj-$(CONFIG_RSEQ) += rseq.o
+obj-$(CONFIG_CPU_OPV) += cpu_opv.o
$(obj)/configs.o: $(obj)/config_data.h
diff --git a/kernel/cpu_opv.c b/kernel/cpu_opv.c
new file mode 100644
index 000000000000..6ee7ca3376be
--- /dev/null
+++ b/kernel/cpu_opv.c
@@ -0,0 +1,955 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Per-CPU-atomic operation vector system call
+ *
+ * It allows user-space to perform a sequence of operations on per-cpu
+ * data in the user-space address space atomically with respect to concurrent
+ * accesses from the same cpu. Useful as single-stepping fall-back for
+ * restartable sequences, and for performing more complex operations on per-cpu
+ * data that would not be otherwise possible to do with restartable sequences,
+ * such as migration of per-cpu data from one cpu to another.
+ *
+ * Copyright (C) 2017-2018 EfficiOS Inc.,
+ * Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx>
+ */
+
+#include <linux/sched.h>
+#include <linux/uaccess.h>
+#include <linux/syscalls.h>
+#include <linux/cpu_opv.h>
+#include <linux/types.h>
+#include <linux/mutex.h>
+#include <linux/pagemap.h>
+#include <linux/mm.h>
+#include <linux/vmalloc.h>
+#include <linux/atomic.h>
+#include <linux/smp.h>
+#include <asm/ptrace.h>
+#include <asm/byteorder.h>
+#include <asm/cacheflush.h>
+
+#include "sched/sched.h"
+
+/*
+ * Maximum number of operations expected for a vector.
+ */
+#define CPU_OP_VEC_LEN_MAX 4
+
+/*
+ * Maximum data len for overall vector. Restrict the amount of user-space
+ * data touched by the kernel in interrupt-off or IPI handler context, so it
+ * does not introduce long interrupt latencies.
+ * cpu_opv allows one copy of up to 4096 bytes, and 3 operations touching 8
+ * bytes each.
+ * This limit is applied to the sum of length specified for all operations
+ * in a vector.
+ */
+#define CPU_OP_VEC_DATA_LEN_MAX (CPU_OP_DATA_LEN_MAX + \
+ (CPU_OP_VEC_LEN_MAX - 1) * CPU_OP_ARITHMETIC_DATA_LEN_MAX)
+
+/*
+ * Invocation of cpu_opv requires a maximum of 8 virtual address pointers.
+ * Keep those in an array on the stack of the cpu_opv system call.
+ */
+#define NR_VADDR 8
+
+/* Maximum pages per op. */
+#define CPU_OP_MAX_PAGES 4
+
+/* Maximum number of virtual addresses per op. */
+#define CPU_OP_VEC_MAX_ADDR (2 * CPU_OP_VEC_LEN_MAX)
+
+union op_fn_data {
+ uint8_t _u8;
+ uint16_t _u16;
+ uint32_t _u32;
+ uint64_t _u64;
+#if (BITS_PER_LONG < 64)
+ uint32_t _u64_split[2];
+#endif
+};
+
+struct vaddr {
+ unsigned long mem;
+ unsigned long uaddr;
+ struct page *pages[2];
+ unsigned int nr_pages;
+ int write;
+};
+
+struct cpu_opv_vaddr {
+ struct vaddr addr[NR_VADDR];
+ size_t nr_vaddr;
+};
+
+typedef int (*op_fn_t)(union op_fn_data *data, uint64_t v, uint32_t len);
+
+struct opv_ipi_args {
+ struct cpu_op *cpuop;
+ int cpuopcnt;
+ int ret;
+};
+
+/*
+ * Provide mutual exclution for threads executing a cpu_opv against an
+ * offline CPU.
+ */
+static DEFINE_MUTEX(cpu_opv_offline_lock);
+
+/*
+ * The cpu_opv system call executes a vector of operations on behalf of
+ * user-space on a specific CPU either with interrupts disabled or within
+ * a interrupt handler. It is inspired by readv() and writev() system
+ * calls which take a "struct iovec" array as argument.
+ *
+ * The operations available are: comparison, memcpy, memcpy_release, add,
+ * add_release. The system call receives a CPU number from user-space as
+ * argument, which is the CPU on which those operations need to be
+ * performed. All pointers in the ops must have been set up to point to
+ * the per CPU memory of the CPU on which the operations should be
+ * executed. The "comparison" operation can be used to check that the data
+ * used in the preparation step did not change between preparation of
+ * system call inputs and interpretation of the operation vector within
+ * the kernel.
+ *
+ * The reason why we require all pointer offsets to be calculated by
+ * user-space beforehand is because we need to use get_user_pages()
+ * to first pin all pages touched by each operation. This takes care of
+ * faulting-in the pages. Then, either interrupts are disabled if the
+ * target CPU is the current CPU, or an IPI is sent to the target CPU, and
+ * the operations are performed atomically with respect to other thread
+ * execution on that CPU, without generating any page fault. If the IPI
+ * nests over a restartable sequence critical section, it will abort that
+ * critical section.
+ *
+ * An overall maximum of 4120 bytes in enforced on the sum of operation
+ * length within an operation vector, so user-space cannot generate a
+ * too long interrupt-off critical section or IPI handler. The operation
+ * vector size is limited to 4 operations. The cache cold critical section
+ * duration has been measured as 4.7µs for 16 operations on x86-64. Each
+ * operation is also limited a length of 4096 bytes, meaning that an
+ * operation can touch a maximum of 4 pages (memcpy: 2 pages for source, 2
+ * pages for destination if addresses are not aligned on page boundaries).
+ *
+ * If the current thread running on the requested CPU, interrupts are
+ * disabled around interpretation of the operation vector. If the target
+ * CPU differs from the current CPU, an IPI is sent to the remote CPU
+ * to interpret the operation vector. If the remote CPU is offline, the
+ * operation vector is executed while holding a reference count preventing
+ * concurrent CPU hotplug changes, with cpu_opv_offline_lock mutex held.
+ */
+
+static unsigned long cpu_op_range_nr_pages(unsigned long addr,
+ unsigned long len)
+{
+ return ((addr + len - 1) >> PAGE_SHIFT) - (addr >> PAGE_SHIFT) + 1;
+}
+
+static int cpu_op_count_pages(u64 addr, unsigned long len)
+{
+ unsigned long nr_pages;
+
+ /*
+ * Validate that the address is within the process address space.
+ * This allows cast of those addresses to unsigned long throughout the
+ * rest of this system call, because it would be invalid to have an
+ * address over 4GB on a 32-bit kernel.
+ */
+ if (addr >= TASK_SIZE)
+ return -EINVAL;
+ if (!len)
+ return 0;
+ nr_pages = cpu_op_range_nr_pages(addr, len);
+ if (nr_pages > 2) {
+ WARN_ON(1);
+ return -EINVAL;
+ }
+ return nr_pages;
+}
+
+/*
+ * Check operation types and length parameters. Count number of pages.
+ */
+static int cpu_opv_check_op(struct cpu_op *op, int *nr_vaddr, uint32_t *sum)
+{
+ int ret;
+
+ *sum += op->len;
+
+ /* Validate inputs. */
+ switch (op->op) {
+ case CPU_COMPARE_EQ_OP:
+ case CPU_COMPARE_NE_OP:
+ case CPU_MEMCPY_OP:
+ case CPU_MEMCPY_RELEASE_OP:
+ if (op->len > CPU_OP_DATA_LEN_MAX)
+ return -EINVAL;
+ break;
+ case CPU_ADD_OP:
+ case CPU_ADD_RELEASE_OP:
+ switch (op->len) {
+ case 1:
+ case 2:
+ case 4:
+ case 8:
+ break;
+ default:
+ return -EINVAL;
+ }
+ break;
+ default:
+ return -EINVAL;
+ }
+
+ /* Validate pointers, count pages and virtual addresses. */
+ switch (op->op) {
+ case CPU_COMPARE_EQ_OP:
+ case CPU_COMPARE_NE_OP:
+ ret = cpu_op_count_pages(op->u.compare_op.a, op->len);
+ if (ret < 0)
+ return ret;
+ ret = cpu_op_count_pages(op->u.compare_op.b, op->len);
+ if (ret < 0)
+ return ret;
+ *nr_vaddr += 2;
+ break;
+ case CPU_MEMCPY_OP:
+ case CPU_MEMCPY_RELEASE_OP:
+ ret = cpu_op_count_pages(op->u.memcpy_op.dst, op->len);
+ if (ret < 0)
+ return ret;
+ ret = cpu_op_count_pages(op->u.memcpy_op.src, op->len);
+ if (ret < 0)
+ return ret;
+ *nr_vaddr += 2;
+ break;
+ case CPU_ADD_OP:
+ case CPU_ADD_RELEASE_OP:
+ ret = cpu_op_count_pages(op->u.arithmetic_op.p, op->len);
+ if (ret < 0)
+ return ret;
+ (*nr_vaddr)++;
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+/*
+ * Check operation types and length parameters. Count number of pages.
+ */
+static int cpu_opv_check(struct cpu_op *cpuopv, int cpuopcnt, int *nr_vaddr)
+{
+ uint32_t sum = 0;
+ int i, ret;
+
+ for (i = 0; i < cpuopcnt; i++) {
+ ret = cpu_opv_check_op(&cpuopv[i], nr_vaddr, &sum);
+ if (ret)
+ return ret;
+ }
+ if (sum > CPU_OP_VEC_DATA_LEN_MAX)
+ return -EINVAL;
+ return 0;
+}
+
+static int cpu_op_check_page(struct page *page, int write)
+{
+ struct address_space *mapping;
+
+ if (is_zone_device_page(page))
+ return -EFAULT;
+
+ /*
+ * The page lock protects many things but in this context the page
+ * lock stabilizes mapping, prevents inode freeing in the shared
+ * file-backed region case and guards against movement to swap
+ * cache.
+ *
+ * Strictly speaking the page lock is not needed in all cases being
+ * considered here and page lock forces unnecessarily serialization
+ * From this point on, mapping will be re-verified if necessary and
+ * page lock will be acquired only if it is unavoidable
+ *
+ * Mapping checks require the head page for any compound page so the
+ * head page and mapping is looked up now.
+ */
+ page = compound_head(page);
+ mapping = READ_ONCE(page->mapping);
+
+ /*
+ * If page->mapping is NULL, then it cannot be a PageAnon page;
+ * but it might be the ZERO_PAGE (which is OK to read from), or
+ * in the gate area or in a special mapping (for which this
+ * check should fail); or it may have been a good file page when
+ * get_user_pages found it, but truncated or holepunched or
+ * subjected to invalidate_complete_page2 before the page lock
+ * is acquired (also cases which should fail). Given that a
+ * reference to the page is currently held, refcount care in
+ * invalidate_complete_page's remove_mapping prevents
+ * drop_caches from setting mapping to NULL concurrently.
+ *
+ * The case to guard against is when memory pressure cause
+ * shmem_writepage to move the page from filecache to swapcache
+ * concurrently: an unlikely race, but a retry for page->mapping
+ * is required in that situation.
+ */
+ if (!mapping) {
+ int shmem_swizzled;
+
+ /*
+ * Check again with page lock held to guard against
+ * memory pressure making shmem_writepage move the page
+ * from filecache to swapcache.
+ */
+ lock_page(page);
+ shmem_swizzled = PageSwapCache(page) || page->mapping;
+ unlock_page(page);
+ if (shmem_swizzled)
+ return -EAGAIN;
+ /*
+ * It is valid to read from, but invalid to write to the
+ * ZERO_PAGE.
+ */
+ if (!(is_zero_pfn(page_to_pfn(page)) ||
+ is_huge_zero_page(page)) || write)
+ return -EFAULT;
+ }
+ return 0;
+}
+
+static int cpu_op_check_pages(struct page **pages,
+ unsigned long nr_pages,
+ int write)
+{
+ unsigned long i;
+
+ for (i = 0; i < nr_pages; i++) {
+ int ret;
+
+ ret = cpu_op_check_page(pages[i], write);
+ if (ret)
+ return ret;
+ }
+ return 0;
+}
+
+static int cpu_op_pin_pages(unsigned long addr, unsigned long len,
+ struct cpu_opv_vaddr *vaddr_ptrs,
+ unsigned long *vaddr, int write)
+{
+ struct page *pages[2];
+ struct vm_area_struct *vmas[2];
+ int ret, nr_pages, nr_put_pages, n;
+ unsigned long _vaddr;
+ struct vaddr *va;
+ struct mm_struct *mm = current->mm;
+
+ nr_pages = cpu_op_count_pages(addr, len);
+ if (nr_pages <= 0)
+ return nr_pages;
+again:
+ down_read(&mm->mmap_sem);
+ ret = get_user_pages(addr, nr_pages, write ? FOLL_WRITE : 0, pages,
+ vmas);
+ if (ret < nr_pages) {
+ if (ret >= 0) {
+ nr_put_pages = ret;
+ ret = -EFAULT;
+ } else {
+ nr_put_pages = 0;
+ }
+ up_read(&mm->mmap_sem);
+ goto error;
+ }
+ /*
+ * cpu_opv() accesses its own cached mapping of the userspace pages.
+ * Considering that concurrent noncached and cached accesses may yield
+ * to unexpected results in terms of memory consistency, explicitly
+ * disallow cpu_opv on noncached memory.
+ */
+ for (n = 0; n < nr_pages; n++) {
+ if (is_vma_noncached(vmas[n])) {
+ nr_put_pages = nr_pages;
+ ret = -EFAULT;
+ up_read(&mm->mmap_sem);
+ goto error;
+ }
+ }
+ up_read(&mm->mmap_sem);
+ ret = cpu_op_check_pages(pages, nr_pages, write);
+ if (ret) {
+ nr_put_pages = nr_pages;
+ goto error;
+ }
+ _vaddr = (unsigned long)vm_map_user_ram(pages, nr_pages, addr,
+ numa_node_id(), PAGE_KERNEL);
+ if (!_vaddr) {
+ nr_put_pages = nr_pages;
+ ret = -ENOMEM;
+ goto error;
+ }
+ va = &vaddr_ptrs->addr[vaddr_ptrs->nr_vaddr++];
+ va->mem = _vaddr;
+ va->uaddr = addr;
+ for (n = 0; n < nr_pages; n++)
+ va->pages[n] = pages[n];
+ va->nr_pages = nr_pages;
+ va->write = write;
+ *vaddr = _vaddr + (addr & ~PAGE_MASK);
+ return 0;
+
+error:
+ for (n = 0; n < nr_put_pages; n++)
+ put_page(pages[n]);
+ /*
+ * Retry if a page has been faulted in, or is being swapped in.
+ */
+ if (ret == -EAGAIN)
+ goto again;
+ return ret;
+}
+
+static int cpu_opv_pin_pages_op(struct cpu_op *op,
+ struct cpu_opv_vaddr *vaddr_ptrs,
+ bool *expect_fault)
+{
+ int ret;
+ unsigned long vaddr = 0;
+
+ switch (op->op) {
+ case CPU_COMPARE_EQ_OP:
+ case CPU_COMPARE_NE_OP:
+ ret = -EFAULT;
+ *expect_fault = op->u.compare_op.expect_fault_a;
+ if (!access_ok(VERIFY_READ,
+ (void __user *)(unsigned long)op->u.compare_op.a,
+ op->len))
+ return ret;
+ ret = cpu_op_pin_pages(op->u.compare_op.a, op->len,
+ vaddr_ptrs, &vaddr, 0);
+ if (ret)
+ return ret;
+ op->u.compare_op.a = vaddr;
+ ret = -EFAULT;
+ *expect_fault = op->u.compare_op.expect_fault_b;
+ if (!access_ok(VERIFY_READ,
+ (void __user *)(unsigned long)op->u.compare_op.b,
+ op->len))
+ return ret;
+ ret = cpu_op_pin_pages(op->u.compare_op.b, op->len,
+ vaddr_ptrs, &vaddr, 0);
+ if (ret)
+ return ret;
+ op->u.compare_op.b = vaddr;
+ break;
+ case CPU_MEMCPY_OP:
+ case CPU_MEMCPY_RELEASE_OP:
+ ret = -EFAULT;
+ *expect_fault = op->u.memcpy_op.expect_fault_dst;
+ if (!access_ok(VERIFY_WRITE,
+ (void __user *)(unsigned long)op->u.memcpy_op.dst,
+ op->len))
+ return ret;
+ ret = cpu_op_pin_pages(op->u.memcpy_op.dst, op->len,
+ vaddr_ptrs, &vaddr, 1);
+ if (ret)
+ return ret;
+ op->u.memcpy_op.dst = vaddr;
+ ret = -EFAULT;
+ *expect_fault = op->u.memcpy_op.expect_fault_src;
+ if (!access_ok(VERIFY_READ,
+ (void __user *)(unsigned long)op->u.memcpy_op.src,
+ op->len))
+ return ret;
+ ret = cpu_op_pin_pages(op->u.memcpy_op.src, op->len,
+ vaddr_ptrs, &vaddr, 0);
+ if (ret)
+ return ret;
+ op->u.memcpy_op.src = vaddr;
+ break;
+ case CPU_ADD_OP:
+ case CPU_ADD_RELEASE_OP:
+ ret = -EFAULT;
+ *expect_fault = op->u.arithmetic_op.expect_fault_p;
+ if (!access_ok(VERIFY_WRITE,
+ (void __user *)(unsigned long)op->u.arithmetic_op.p,
+ op->len))
+ return ret;
+ ret = cpu_op_pin_pages(op->u.arithmetic_op.p, op->len,
+ vaddr_ptrs, &vaddr, 1);
+ if (ret)
+ return ret;
+ op->u.arithmetic_op.p = vaddr;
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+static int cpu_opv_pin_pages(struct cpu_op *cpuop, int cpuopcnt,
+ struct cpu_opv_vaddr *vaddr_ptrs)
+{
+ int ret, i;
+ bool expect_fault = false;
+
+ /* Check access, pin pages. */
+ for (i = 0; i < cpuopcnt; i++) {
+ ret = cpu_opv_pin_pages_op(&cpuop[i], vaddr_ptrs,
+ &expect_fault);
+ if (ret)
+ goto error;
+ }
+ return 0;
+
+error:
+ /*
+ * If faulting access is expected, return EAGAIN to user-space.
+ * It allows user-space to distinguish between a fault caused by
+ * an access which is expect to fault (e.g. due to concurrent
+ * unmapping of underlying memory) from an unexpected fault from
+ * which a retry would not recover.
+ */
+ if (ret == -EFAULT && expect_fault)
+ return -EAGAIN;
+ return ret;
+}
+
+static int __op_get(union op_fn_data *data, void *p, size_t len)
+{
+ switch (len) {
+ case 1:
+ data->_u8 = READ_ONCE(*(uint8_t *)p);
+ break;
+ case 2:
+ data->_u16 = READ_ONCE(*(uint16_t *)p);
+ break;
+ case 4:
+ data->_u32 = READ_ONCE(*(uint32_t *)p);
+ break;
+ case 8:
+#if (BITS_PER_LONG == 64)
+ data->_u64 = READ_ONCE(*(uint64_t *)p);
+#else
+ {
+ data->_u64_split[0] = READ_ONCE(*(uint32_t *)p);
+ data->_u64_split[1] = READ_ONCE(*((uint32_t *)p + 1));
+ }
+#endif
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+static int __op_put(union op_fn_data *data, void *p, size_t len, bool release)
+{
+ switch (len) {
+ case 1:
+ if (release)
+ smp_store_release((uint8_t *)p, data->_u8);
+ else
+ WRITE_ONCE(*(uint8_t *)p, data->_u8);
+ break;
+ case 2:
+ if (release)
+ smp_store_release((uint16_t *)p, data->_u16);
+ else
+ WRITE_ONCE(*(uint16_t *)p, data->_u16);
+ break;
+ case 4:
+ if (release)
+ smp_store_release((uint32_t *)p, data->_u32);
+ else
+ WRITE_ONCE(*(uint32_t *)p, data->_u32);
+ break;
+ case 8:
+#if (BITS_PER_LONG == 64)
+ if (release)
+ smp_store_release((uint64_t *)p, data->_u64);
+ else
+ WRITE_ONCE(*(uint64_t *)p, data->_u64);
+#else
+ {
+ if (release)
+ smp_store_release((uint32_t *)p, data->_u64_split[0]);
+ else
+ WRITE_ONCE(*(uint32_t *)p, data->_u64_split[0]);
+ WRITE_ONCE(*((uint32_t *)p + 1), data->_u64_split[1]);
+ }
+#endif
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+/* Return 0 if same, > 0 if different, < 0 on error. */
+static int do_cpu_op_compare(unsigned long _a, unsigned long _b, uint32_t len)
+{
+ void *a = (void *)_a;
+ void *b = (void *)_b;
+ union op_fn_data tmp[2];
+ int ret;
+
+ switch (len) {
+ case 1:
+ case 2:
+ case 4:
+ case 8:
+ if (!IS_ALIGNED(_a, len) || !IS_ALIGNED(_b, len))
+ goto memcmp;
+ break;
+ default:
+ goto memcmp;
+ }
+
+ ret = __op_get(&tmp[0], a, len);
+ if (ret)
+ return ret;
+ ret = __op_get(&tmp[1], b, len);
+ if (ret)
+ return ret;
+
+ switch (len) {
+ case 1:
+ ret = !!(tmp[0]._u8 != tmp[1]._u8);
+ break;
+ case 2:
+ ret = !!(tmp[0]._u16 != tmp[1]._u16);
+ break;
+ case 4:
+ ret = !!(tmp[0]._u32 != tmp[1]._u32);
+ break;
+ case 8:
+ ret = !!(tmp[0]._u64 != tmp[1]._u64);
+ break;
+ default:
+ return -EINVAL;
+ }
+ return ret;
+
+memcmp:
+ if (memcmp(a, b, len))
+ return 1;
+ return 0;
+}
+
+/* Return 0 on success, < 0 on error. */
+static int do_cpu_op_memcpy(unsigned long _dst, unsigned long _src,
+ uint32_t len, bool release)
+{
+ void *dst = (void *)_dst;
+ void *src = (void *)_src;
+ union op_fn_data tmp;
+ int ret;
+
+ switch (len) {
+ case 1:
+ case 2:
+ case 4:
+ case 8:
+ if (!IS_ALIGNED(_dst, len) || !IS_ALIGNED(_src, len))
+ goto memcpy;
+ break;
+ default:
+ goto memcpy;
+ }
+
+ ret = __op_get(&tmp, src, len);
+ if (ret)
+ return ret;
+ return __op_put(&tmp, dst, len, release);
+
+memcpy:
+ if (release)
+ smp_mb();
+ memcpy(dst, src, len);
+ return 0;
+}
+
+static int op_add_fn(union op_fn_data *data, uint64_t count, uint32_t len)
+{
+ switch (len) {
+ case 1:
+ data->_u8 += (uint8_t)count;
+ break;
+ case 2:
+ data->_u16 += (uint16_t)count;
+ break;
+ case 4:
+ data->_u32 += (uint32_t)count;
+ break;
+ case 8:
+ data->_u64 += (uint64_t)count;
+ break;
+ default:
+ return -EINVAL;
+ }
+ return 0;
+}
+
+/* Return 0 on success, < 0 on error. */
+static int do_cpu_op_fn(op_fn_t op_fn, unsigned long _p, uint64_t v,
+ uint32_t len, bool release)
+{
+ union op_fn_data tmp;
+ void *p = (void *)_p;
+ int ret;
+
+ ret = __op_get(&tmp, p, len);
+ if (ret)
+ return ret;
+ ret = op_fn(&tmp, v, len);
+ if (ret)
+ return ret;
+ ret = __op_put(&tmp, p, len, release);
+ if (ret)
+ return ret;
+ return 0;
+}
+
+/*
+ * Return negative value on error, positive value if comparison
+ * fails, 0 on success.
+ */
+static int __do_cpu_opv_op(struct cpu_op *op)
+{
+ /* Guarantee a compiler barrier between each operation. */
+ barrier();
+
+ switch (op->op) {
+ case CPU_COMPARE_EQ_OP:
+ return do_cpu_op_compare(op->u.compare_op.a,
+ op->u.compare_op.b,
+ op->len);
+ case CPU_COMPARE_NE_OP:
+ {
+ int ret;
+
+ ret = do_cpu_op_compare(op->u.compare_op.a,
+ op->u.compare_op.b,
+ op->len);
+ if (ret < 0)
+ return ret;
+ /*
+ * Stop execution, return positive value if comparison
+ * is identical.
+ */
+ if (ret == 0)
+ return 1;
+ return 0;
+ }
+ case CPU_MEMCPY_OP:
+ return do_cpu_op_memcpy(op->u.memcpy_op.dst,
+ op->u.memcpy_op.src,
+ op->len, false);
+ case CPU_MEMCPY_RELEASE_OP:
+ return do_cpu_op_memcpy(op->u.memcpy_op.dst,
+ op->u.memcpy_op.src,
+ op->len, true);
+ case CPU_ADD_OP:
+ return do_cpu_op_fn(op_add_fn, op->u.arithmetic_op.p,
+ op->u.arithmetic_op.count, op->len, false);
+ case CPU_ADD_RELEASE_OP:
+ return do_cpu_op_fn(op_add_fn, op->u.arithmetic_op.p,
+ op->u.arithmetic_op.count, op->len, true);
+ default:
+ return -EINVAL;
+ }
+}
+
+static int __do_cpu_opv(struct cpu_op *cpuop, int cpuopcnt)
+{
+ int i, ret;
+
+ for (i = 0; i < cpuopcnt; i++) {
+ ret = __do_cpu_opv_op(&cpuop[i]);
+ /* If comparison fails, stop execution and return index + 1. */
+ if (ret > 0)
+ return i + 1;
+ /* On error, stop execution. */
+ if (ret < 0)
+ return ret;
+ }
+ return 0;
+}
+
+/*
+ * Check that the page pointers pinned by get_user_pages()
+ * are still in the page table. Invoked with mmap_sem held.
+ * Return 0 if pointers match, -EAGAIN if they don't.
+ */
+static int vaddr_check(struct vaddr *vaddr)
+{
+ struct page *pages[2];
+ int ret, n;
+
+ ret = __get_user_pages_fast(vaddr->uaddr, vaddr->nr_pages,
+ vaddr->write, pages);
+ for (n = 0; n < ret; n++)
+ put_page(pages[n]);
+ if (ret < vaddr->nr_pages) {
+ ret = get_user_pages(vaddr->uaddr, vaddr->nr_pages,
+ vaddr->write ? FOLL_WRITE : 0,
+ pages, NULL);
+ if (ret < 0)
+ return -EAGAIN;
+ for (n = 0; n < ret; n++)
+ put_page(pages[n]);
+ if (ret < vaddr->nr_pages)
+ return -EAGAIN;
+ }
+ for (n = 0; n < vaddr->nr_pages; n++) {
+ if (pages[n] != vaddr->pages[n])
+ return -EAGAIN;
+ }
+ return 0;
+}
+
+static int vaddr_ptrs_check(struct cpu_opv_vaddr *vaddr_ptrs)
+{
+ int i;
+
+ for (i = 0; i < vaddr_ptrs->nr_vaddr; i++) {
+ int ret;
+
+ ret = vaddr_check(&vaddr_ptrs->addr[i]);
+ if (ret)
+ return ret;
+ }
+ return 0;
+}
+
+static void cpu_opv_ipi(void *info)
+{
+ struct opv_ipi_args *args = info;
+
+ rseq_preempt(current);
+ args->ret = __do_cpu_opv(args->cpuop, args->cpuopcnt);
+}
+
+static int do_cpu_opv(struct cpu_op *cpuop, int cpuopcnt,
+ struct cpu_opv_vaddr *vaddr_ptrs, int cpu)
+{
+ struct mm_struct *mm = current->mm;
+ struct opv_ipi_args args = {
+ .cpuop = cpuop,
+ .cpuopcnt = cpuopcnt,
+ };
+ int ret;
+
+retry:
+ if (!cpumask_test_cpu(cpu, ¤t->cpus_allowed))
+ return -EINVAL;
+ down_read(&mm->mmap_sem);
+ ret = vaddr_ptrs_check(vaddr_ptrs);
+ if (ret)
+ goto end;
+ ret = smp_call_function_single(cpu, cpu_opv_ipi, &args, 1);
+ if (ret) {
+ up_read(&mm->mmap_sem);
+ goto check_online;
+ }
+ ret = args.ret;
+end:
+ up_read(&mm->mmap_sem);
+ return ret;
+
+check_online:
+ get_online_cpus();
+ if (cpu_online(cpu)) {
+ put_online_cpus();
+ goto retry;
+ }
+ /*
+ * CPU is offline. Perform operation from the current CPU with
+ * cpu_online read lock held, preventing that CPU from coming online,
+ * and with mutex held, providing mutual exclusion against other
+ * CPUs also finding out about an offline CPU.
+ */
+ down_read(&mm->mmap_sem);
+ ret = vaddr_ptrs_check(vaddr_ptrs);
+ if (ret)
+ goto offline_end;
+ mutex_lock(&cpu_opv_offline_lock);
+ ret = __do_cpu_opv(cpuop, cpuopcnt);
+ mutex_unlock(&cpu_opv_offline_lock);
+offline_end:
+ up_read(&mm->mmap_sem);
+ put_online_cpus();
+ return ret;
+}
+
+/*
+ * cpu_opv - execute operation vector on a given CPU in interrupt context.
+ *
+ * Userspace should pass the CPU number on which the operation vector
+ * should be executed as parameter.
+ */
+SYSCALL_DEFINE4(cpu_opv, struct cpu_op __user *, ucpuopv, int, cpuopcnt,
+ int, cpu, int, flags)
+{
+ struct cpu_op cpuopv[CPU_OP_VEC_LEN_MAX];
+ struct cpu_opv_vaddr vaddr_ptrs = {
+ .nr_vaddr = 0,
+ };
+ int ret, i, nr_vaddr = 0;
+ bool retry = false;
+
+ if (unlikely(flags & ~(CPU_OP_NR_FLAG | CPU_OP_VEC_LEN_MAX_FLAG)))
+ return -EINVAL;
+ if (flags & CPU_OP_NR_FLAG) {
+ if (flags & CPU_OP_VEC_LEN_MAX_FLAG)
+ return -EINVAL;
+ return NR_CPU_OPS;
+ }
+ if (flags & CPU_OP_VEC_LEN_MAX_FLAG)
+ return CPU_OP_VEC_LEN_MAX;
+ if (unlikely(cpu < 0))
+ return -EINVAL;
+ if (cpuopcnt < 0 || cpuopcnt > CPU_OP_VEC_LEN_MAX)
+ return -EINVAL;
+ if (copy_from_user(cpuopv, ucpuopv, cpuopcnt * sizeof(struct cpu_op)))
+ return -EFAULT;
+ ret = cpu_opv_check(cpuopv, cpuopcnt, &nr_vaddr);
+ if (ret)
+ return ret;
+ if (nr_vaddr > NR_VADDR)
+ return -EINVAL;
+again:
+ ret = cpu_opv_pin_pages(cpuopv, cpuopcnt, &vaddr_ptrs);
+ if (ret)
+ goto end;
+ ret = do_cpu_opv(cpuopv, cpuopcnt, &vaddr_ptrs, cpu);
+ if (ret == -EAGAIN)
+ retry = true;
+end:
+ for (i = 0; i < vaddr_ptrs.nr_vaddr; i++) {
+ struct vaddr *vaddr = &vaddr_ptrs.addr[i];
+ int j;
+
+ vm_unmap_user_ram((void *)vaddr->mem, vaddr->nr_pages);
+ for (j = 0; j < vaddr->nr_pages; j++) {
+ if (vaddr->write)
+ set_page_dirty(vaddr->pages[j]);
+ put_page(vaddr->pages[j]);
+ }
+ }
+ /*
+ * Force vm_map flush to ensure we don't exhaust available vmalloc
+ * address space.
+ */
+ if (vaddr_ptrs.nr_vaddr)
+ vm_unmap_aliases();
+ if (retry) {
+ retry = false;
+ vaddr_ptrs.nr_vaddr = 0;
+ goto again;
+ }
+ return ret;
+}
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index df556175be50..0a6410d77c33 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -435,3 +435,4 @@ COND_SYSCALL(setuid16);
/* restartable sequence */
COND_SYSCALL(rseq);
+COND_SYSCALL(cpu_opv);
--
2.11.0
