zephyr/kernel/pipes.c
Andy Ross 73cb9586ce k_mem_pool: Complete rework
This patch amounts to a mostly complete rewrite of the k_mem_pool
allocator, which had been the source of historical complaints vs. the
one easily available in newlib.  The basic design of the allocator is
unchanged (it's still a 4-way buddy allocator), but the implementation
has made different choices throughout.  Major changes:

Space efficiency: The old implementation required ~2.66 bytes per
"smallest block" in overhead, plus 16 bytes per log4 "level" of the
allocation tree, plus a global tracking struct of 32 bytes and a very
surprising 12 byte overhead (in struct k_mem_block) per active
allocation on top of the returned data pointer.  This new allocator
uses a simple bit array as the only per-block storage and places the
free list into the freed blocks themselves, requiring only ~1.33 bits
per smallest block, 12 bytes per level, 32 byte globally and only 4
bytes of per-allocation bookeeping.  And it puts more of the generated
tree into BSS, slightly reducing binary sizes for non-trivial pool
sizes (even as the code size itself has increased a tiny bit).

IRQ safe: atomic operations on the store have been cut down to be at
most "4 bit sets and dlist operations" (i.e. a few dozen
instructions), reducing latency significantly and allowing us to lock
against interrupts cleanly from all APIs.  Allocations and frees can
be done from ISRs now without limitation (well, obviously you can't
sleep, so "timeout" must be K_NO_WAIT).

Deterministic performance: there is no more "defragmentation" step
that must be manually managed.  Block coalescing is done synchronously
at free time and takes constant time (strictly log4(num_levels)), as
the detection of four free "partner bits" is just a simple shift and
mask operation.

Cleaner behavior with odd sizes.  The old code assumed that the
specified maximum size would be a power of four multiple of the
minimum size, making use of non-standard buffer sizes problematic.
This implementation re-aligns the sub-blocks at each level and can
handle situations wehre alignment restrictions mean fewer than 4x will
be available.  If you want precise layout control, you can still
specify the sizes rigorously.  It just doesn't break if you don't.

More portable: the original implementation made use of GNU assembler
macros embedded inline within C __asm__ statements.  Not all
toolchains are actually backed by a GNU assembler even when the
support the GNU assembly syntax.  This is pure C, albeit with some
hairy macros to expand the compile-time-computed values.

Related changes that had to be rolled into this patch for bisectability:

* The new allocator has a firm minimum block size of 8 bytes (to store
  the dlist_node_t).  It will "work" with smaller requested min_size
  values, but obviously makes no firm promises about layout or how
  many will be available.  Unfortunately many of the tests were
  written with very small 4-byte minimum sizes and to assume exactly
  how many they could allocate.  Bump the sizes to match the allocator
  minimum.

* The mbox and pipes API made use of the internals of k_mem_block and
  had to be ported to the new scheme.  Blocks no longer store a
  backpointer to the pool that allocated them (it's an integer ID in a
  bitfield) , so if you want to "nullify" them you have to use the
  data pointer.

* test_mbox_api had a bug were it was prematurely freeing k_mem_blocks
  that it sent through the mailbox.  This worked in the old allocator
  because the memory wouldn't be touched when freed, but now we stuff
  list pointers in there and the bug was exposed.

* Remove test_mpool_options: the options (related to defragmentation
  behavior) tested no longer exist.

Signed-off-by: Andy Ross <andrew.j.ross@intel.com>
2017-05-13 14:39:41 -04:00

701 lines
19 KiB
C

/*
* Copyright (c) 2016 Wind River Systems, Inc.
*
* SPDX-License-Identifier: Apache-2.0
*/
/**
* @file
*
* @brief Pipes
*/
#include <kernel.h>
#include <kernel_structs.h>
#include <debug/object_tracing_common.h>
#include <toolchain.h>
#include <sections.h>
#include <wait_q.h>
#include <misc/dlist.h>
#include <init.h>
struct k_pipe_desc {
unsigned char *buffer; /* Position in src/dest buffer */
size_t bytes_to_xfer; /* # bytes left to transfer */
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
struct k_mem_block *block; /* Pointer to memory block */
struct k_mem_block copy_block; /* For backwards compatibility */
struct k_sem *sem; /* Semaphore to give if async */
#endif
};
struct k_pipe_async {
struct _thread_base thread; /* Dummy thread object */
struct k_pipe_desc desc; /* Pipe message descriptor */
};
extern struct k_pipe _k_pipe_list_start[];
extern struct k_pipe _k_pipe_list_end[];
struct k_pipe *_trace_list_k_pipe;
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
/* Array of asynchronous message descriptors */
static struct k_pipe_async __noinit async_msg[CONFIG_NUM_PIPE_ASYNC_MSGS];
/* stack of unused asynchronous message descriptors */
K_STACK_DEFINE(pipe_async_msgs, CONFIG_NUM_PIPE_ASYNC_MSGS);
/* Allocate an asynchronous message descriptor */
static void _pipe_async_alloc(struct k_pipe_async **async)
{
k_stack_pop(&pipe_async_msgs, (u32_t *)async, K_FOREVER);
}
/* Free an asynchronous message descriptor */
static void _pipe_async_free(struct k_pipe_async *async)
{
k_stack_push(&pipe_async_msgs, (u32_t)async);
}
/* Finish an asynchronous operation */
static void _pipe_async_finish(struct k_pipe_async *async_desc)
{
/*
* An asynchronous operation is finished with the scheduler locked
* to prevent the called routines from scheduling a new thread.
*/
k_mem_pool_free(async_desc->desc.block);
if (async_desc->desc.sem != NULL) {
k_sem_give(async_desc->desc.sem);
}
_pipe_async_free(async_desc);
}
#endif /* CONFIG_NUM_PIPE_ASYNC_MSGS > 0 */
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0) || \
defined(CONFIG_OBJECT_TRACING)
/*
* Do run-time initialization of pipe object subsystem.
*/
static int init_pipes_module(struct device *dev)
{
ARG_UNUSED(dev);
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
/*
* Create pool of asynchronous pipe message descriptors.
*
* A dummy thread requires minimal initialization, since it never gets
* to execute. The _THREAD_DUMMY flag is sufficient to distinguish a
* dummy thread from a real one. The threads are *not* added to the
* kernel's list of known threads.
*
* Once initialized, the address of each descriptor is added to a stack
* that governs access to them.
*/
for (int i = 0; i < CONFIG_NUM_PIPE_ASYNC_MSGS; i++) {
async_msg[i].thread.thread_state = _THREAD_DUMMY;
async_msg[i].thread.swap_data = &async_msg[i].desc;
k_stack_push(&pipe_async_msgs, (u32_t)&async_msg[i]);
}
#endif /* CONFIG_NUM_PIPE_ASYNC_MSGS > 0 */
/* Complete initialization of statically defined mailboxes. */
#ifdef CONFIG_OBJECT_TRACING
struct k_pipe *pipe;
for (pipe = _k_pipe_list_start; pipe < _k_pipe_list_end; pipe++) {
SYS_TRACING_OBJ_INIT(k_pipe, pipe);
}
#endif /* CONFIG_OBJECT_TRACING */
return 0;
}
SYS_INIT(init_pipes_module, PRE_KERNEL_1, CONFIG_KERNEL_INIT_PRIORITY_OBJECTS);
#endif /* CONFIG_NUM_PIPE_ASYNC_MSGS or CONFIG_OBJECT_TRACING */
void k_pipe_init(struct k_pipe *pipe, unsigned char *buffer, size_t size)
{
pipe->buffer = buffer;
pipe->size = size;
pipe->bytes_used = 0;
pipe->read_index = 0;
pipe->write_index = 0;
sys_dlist_init(&pipe->wait_q.writers);
sys_dlist_init(&pipe->wait_q.readers);
SYS_TRACING_OBJ_INIT(k_pipe, pipe);
}
/**
* @brief Copy bytes from @a src to @a dest
*
* @return Number of bytes copied
*/
static size_t _pipe_xfer(unsigned char *dest, size_t dest_size,
const unsigned char *src, size_t src_size)
{
size_t num_bytes = min(dest_size, src_size);
const unsigned char *end = src + num_bytes;
while (src != end) {
*dest = *src;
dest++;
src++;
}
return num_bytes;
}
/**
* @brief Put data from @a src into the pipe's circular buffer
*
* Modifies the following fields in @a pipe:
* buffer, bytes_used, write_index
*
* @return Number of bytes written to the pipe's circular buffer
*/
static size_t _pipe_buffer_put(struct k_pipe *pipe,
const unsigned char *src, size_t src_size)
{
size_t bytes_copied;
size_t run_length;
size_t num_bytes_written = 0;
int i;
for (i = 0; i < 2; i++) {
run_length = min(pipe->size - pipe->bytes_used,
pipe->size - pipe->write_index);
bytes_copied = _pipe_xfer(pipe->buffer + pipe->write_index,
run_length,
src + num_bytes_written,
src_size - num_bytes_written);
num_bytes_written += bytes_copied;
pipe->bytes_used += bytes_copied;
pipe->write_index += bytes_copied;
if (pipe->write_index == pipe->size) {
pipe->write_index = 0;
}
}
return num_bytes_written;
}
/**
* @brief Get data from the pipe's circular buffer
*
* Modifies the following fields in @a pipe:
* bytes_used, read_index
*
* @return Number of bytes read from the pipe's circular buffer
*/
static size_t _pipe_buffer_get(struct k_pipe *pipe,
unsigned char *dest, size_t dest_size)
{
size_t bytes_copied;
size_t run_length;
size_t num_bytes_read = 0;
int i;
for (i = 0; i < 2; i++) {
run_length = min(pipe->bytes_used,
pipe->size - pipe->read_index);
bytes_copied = _pipe_xfer(dest + num_bytes_read,
dest_size - num_bytes_read,
pipe->buffer + pipe->read_index,
run_length);
num_bytes_read += bytes_copied;
pipe->bytes_used -= bytes_copied;
pipe->read_index += bytes_copied;
if (pipe->read_index == pipe->size) {
pipe->read_index = 0;
}
}
return num_bytes_read;
}
/**
* @brief Prepare a working set of readers/writers
*
* Prepare a list of "working threads" into/from which the data
* will be directly copied. This list is useful as it is used to ...
*
* 1. avoid double copying
* 2. minimize interrupt latency as interrupts are unlocked
* while copying data
* 3. ensure a timeout can not make the request impossible to satisfy
*
* The list is populated with previously pended threads that will be ready to
* run after the pipe call is complete.
*
* Important things to remember when reading from the pipe ...
* 1. If there are writers int @a wait_q, then the pipe's buffer is full.
* 2. Conversely if the pipe's buffer is not full, there are no writers.
* 3. The amount of available data in the pipe is the sum the bytes used in
* the pipe (@a pipe_space) and all the requests from the waiting writers.
* 4. Since data is read from the pipe's buffer first, the working set must
* include writers that will (try to) re-fill the pipe's buffer afterwards.
*
* Important things to remember when writing to the pipe ...
* 1. If there are readers in @a wait_q, then the pipe's buffer is empty.
* 2. Conversely if the pipe's buffer is not empty, then there are no readers.
* 3. The amount of space available in the pipe is the sum of the bytes unused
* in the pipe (@a pipe_space) and all the requests from the waiting readers.
*
* @return false if request is unsatisfiable, otherwise true
*/
static bool _pipe_xfer_prepare(sys_dlist_t *xfer_list,
struct k_thread **waiter,
_wait_q_t *wait_q,
size_t pipe_space,
size_t bytes_to_xfer,
size_t min_xfer,
s32_t timeout)
{
sys_dnode_t *node;
struct k_thread *thread;
struct k_pipe_desc *desc;
size_t num_bytes = 0;
if (timeout == K_NO_WAIT) {
for (node = sys_dlist_peek_head(wait_q); node != NULL;
node = sys_dlist_peek_next(wait_q, node)) {
thread = (struct k_thread *)node;
desc = (struct k_pipe_desc *)thread->base.swap_data;
num_bytes += desc->bytes_to_xfer;
if (num_bytes >= bytes_to_xfer) {
break;
}
}
if (num_bytes + pipe_space < min_xfer) {
return false;
}
}
/*
* Either @a timeout is not K_NO_WAIT (so the thread may pend) or
* the entire request can be satisfied. Generate the working list.
*/
sys_dlist_init(xfer_list);
num_bytes = 0;
while ((thread = (struct k_thread *) sys_dlist_peek_head(wait_q))) {
desc = (struct k_pipe_desc *)thread->base.swap_data;
num_bytes += desc->bytes_to_xfer;
if (num_bytes > bytes_to_xfer) {
/*
* This request can not be fully satisfied.
* Do not remove it from the wait_q.
* Do not abort its timeout (if applicable).
* Do not add it to the transfer list
*/
break;
}
/*
* This request can be fully satisfied.
* Remove it from the wait_q.
* Abort its timeout.
* Add it to the transfer list.
*/
_unpend_thread(thread);
_abort_thread_timeout(thread);
sys_dlist_append(xfer_list, &thread->base.k_q_node);
}
*waiter = (num_bytes > bytes_to_xfer) ? thread : NULL;
return true;
}
/**
* @brief Determine the correct return code
*
* Bytes Xferred No Wait Wait
* >= Minimum 0 0
* < Minimum -EIO* -EAGAIN
*
* * The "-EIO No Wait" case was already checked when the "working set"
* was created in _pipe_xfer_prepare().
*
* @return See table above
*/
static int _pipe_return_code(size_t min_xfer, size_t bytes_remaining,
size_t bytes_requested)
{
if (bytes_requested - bytes_remaining >= min_xfer) {
/*
* At least the minimum number of requested
* bytes have been transferred.
*/
return 0;
}
return -EAGAIN;
}
/**
* @brief Ready a pipe thread
*
* If the pipe thread is a real thread, then add it to the ready queue.
* If it is a dummy thread, then finish the asynchronous work.
*
* @return N/A
*/
static void _pipe_thread_ready(struct k_thread *thread)
{
unsigned int key;
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
if (thread->base.thread_state & _THREAD_DUMMY) {
_pipe_async_finish((struct k_pipe_async *)thread);
return;
}
#endif
key = irq_lock();
_ready_thread(thread);
irq_unlock(key);
}
/**
* @brief Internal API used to send data to a pipe
*/
int _k_pipe_put_internal(struct k_pipe *pipe, struct k_pipe_async *async_desc,
unsigned char *data, size_t bytes_to_write,
size_t *bytes_written, size_t min_xfer,
s32_t timeout)
{
struct k_thread *reader;
struct k_pipe_desc *desc;
sys_dlist_t xfer_list;
unsigned int key;
size_t num_bytes_written = 0;
size_t bytes_copied;
#if (CONFIG_NUM_PIPE_ASYNC_MSGS == 0)
ARG_UNUSED(async_desc);
#endif
key = irq_lock();
/*
* Create a list of "working readers" into which the data will be
* directly copied.
*/
if (!_pipe_xfer_prepare(&xfer_list, &reader, &pipe->wait_q.readers,
pipe->size - pipe->bytes_used, bytes_to_write,
min_xfer, timeout)) {
irq_unlock(key);
*bytes_written = 0;
return -EIO;
}
_sched_lock();
irq_unlock(key);
/*
* 1. 'xfer_list' currently contains a list of reader threads that can
* have their read requests fulfilled by the current call.
* 2. 'reader' if not NULL points to a thread on the reader wait_q
* that can get some of its requested data.
* 3. Interrupts are unlocked but the scheduler is locked to allow
* ticks to be delivered but no scheduling to occur
* 4. If 'reader' times out while we are copying data, not only do we
* still have a pointer to it, but it can not execute until this call
* is complete so it is still safe to copy data to it.
*/
struct k_thread *thread = (struct k_thread *)
sys_dlist_get(&xfer_list);
while (thread) {
desc = (struct k_pipe_desc *)thread->base.swap_data;
bytes_copied = _pipe_xfer(desc->buffer, desc->bytes_to_xfer,
data + num_bytes_written,
bytes_to_write - num_bytes_written);
num_bytes_written += bytes_copied;
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
/* The thread's read request has been satisfied. Ready it. */
key = irq_lock();
_ready_thread(thread);
irq_unlock(key);
thread = (struct k_thread *)sys_dlist_get(&xfer_list);
}
/*
* Copy any data to the reader that we left on the wait_q.
* It is possible no data will be copied.
*/
if (reader) {
desc = (struct k_pipe_desc *)reader->base.swap_data;
bytes_copied = _pipe_xfer(desc->buffer, desc->bytes_to_xfer,
data + num_bytes_written,
bytes_to_write - num_bytes_written);
num_bytes_written += bytes_copied;
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
}
/*
* As much data as possible has been directly copied to any waiting
* readers. Add as much as possible to the pipe's circular buffer.
*/
num_bytes_written +=
_pipe_buffer_put(pipe, data + num_bytes_written,
bytes_to_write - num_bytes_written);
if (num_bytes_written == bytes_to_write) {
*bytes_written = num_bytes_written;
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
if (async_desc != NULL) {
_pipe_async_finish(async_desc);
}
#endif
k_sched_unlock();
return 0;
}
/* Not all data was copied. */
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
if (async_desc != NULL) {
/*
* Lock interrupts and unlock the scheduler before
* manipulating the writers wait_q.
*/
key = irq_lock();
_sched_unlock_no_reschedule();
_pend_thread((struct k_thread *) &async_desc->thread,
&pipe->wait_q.writers, K_FOREVER);
_reschedule_threads(key);
return 0;
}
#endif
struct k_pipe_desc pipe_desc;
pipe_desc.buffer = data + num_bytes_written;
pipe_desc.bytes_to_xfer = bytes_to_write - num_bytes_written;
if (timeout != K_NO_WAIT) {
_current->base.swap_data = &pipe_desc;
/*
* Lock interrupts and unlock the scheduler before
* manipulating the writers wait_q.
*/
key = irq_lock();
_sched_unlock_no_reschedule();
_pend_current_thread(&pipe->wait_q.writers, timeout);
_Swap(key);
} else {
k_sched_unlock();
}
*bytes_written = bytes_to_write - pipe_desc.bytes_to_xfer;
return _pipe_return_code(min_xfer, pipe_desc.bytes_to_xfer,
bytes_to_write);
}
int k_pipe_get(struct k_pipe *pipe, void *data, size_t bytes_to_read,
size_t *bytes_read, size_t min_xfer, s32_t timeout)
{
struct k_thread *writer;
struct k_pipe_desc *desc;
sys_dlist_t xfer_list;
unsigned int key;
size_t num_bytes_read = 0;
size_t bytes_copied;
__ASSERT(min_xfer <= bytes_to_read, "");
__ASSERT(bytes_read != NULL, "");
key = irq_lock();
/*
* Create a list of "working readers" into which the data will be
* directly copied.
*/
if (!_pipe_xfer_prepare(&xfer_list, &writer, &pipe->wait_q.writers,
pipe->bytes_used, bytes_to_read,
min_xfer, timeout)) {
irq_unlock(key);
*bytes_read = 0;
return -EIO;
}
_sched_lock();
irq_unlock(key);
num_bytes_read = _pipe_buffer_get(pipe, data, bytes_to_read);
/*
* 1. 'xfer_list' currently contains a list of writer threads that can
* have their write requests fulfilled by the current call.
* 2. 'writer' if not NULL points to a thread on the writer wait_q
* that can post some of its requested data.
* 3. Data will be copied from each writer's buffer to either the
* reader's buffer and/or to the pipe's circular buffer.
* 4. Interrupts are unlocked but the scheduler is locked to allow
* ticks to be delivered but no scheduling to occur
* 5. If 'writer' times out while we are copying data, not only do we
* still have a pointer to it, but it can not execute until this
* call is complete so it is still safe to copy data from it.
*/
struct k_thread *thread = (struct k_thread *)
sys_dlist_get(&xfer_list);
while (thread && (num_bytes_read < bytes_to_read)) {
desc = (struct k_pipe_desc *)thread->base.swap_data;
bytes_copied = _pipe_xfer(data + num_bytes_read,
bytes_to_read - num_bytes_read,
desc->buffer, desc->bytes_to_xfer);
num_bytes_read += bytes_copied;
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
/*
* It is expected that the write request will be satisfied.
* However, if the read request was satisfied before the
* write request was satisfied, then the write request must
* finish later when writing to the pipe's circular buffer.
*/
if (num_bytes_read == bytes_to_read) {
break;
}
_pipe_thread_ready(thread);
thread = (struct k_thread *)sys_dlist_get(&xfer_list);
}
if (writer && (num_bytes_read < bytes_to_read)) {
desc = (struct k_pipe_desc *)writer->base.swap_data;
bytes_copied = _pipe_xfer(data + num_bytes_read,
bytes_to_read - num_bytes_read,
desc->buffer, desc->bytes_to_xfer);
num_bytes_read += bytes_copied;
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
}
/*
* Copy as much data as possible from the writers (if any)
* into the pipe's circular buffer.
*/
while (thread) {
desc = (struct k_pipe_desc *)thread->base.swap_data;
bytes_copied = _pipe_buffer_put(pipe, desc->buffer,
desc->bytes_to_xfer);
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
/* Write request has been satsified */
_pipe_thread_ready(thread);
thread = (struct k_thread *)sys_dlist_get(&xfer_list);
}
if (writer) {
desc = (struct k_pipe_desc *)writer->base.swap_data;
bytes_copied = _pipe_buffer_put(pipe, desc->buffer,
desc->bytes_to_xfer);
desc->buffer += bytes_copied;
desc->bytes_to_xfer -= bytes_copied;
}
if (num_bytes_read == bytes_to_read) {
k_sched_unlock();
*bytes_read = num_bytes_read;
return 0;
}
/* Not all data was read. */
struct k_pipe_desc pipe_desc;
pipe_desc.buffer = data + num_bytes_read;
pipe_desc.bytes_to_xfer = bytes_to_read - num_bytes_read;
if (timeout != K_NO_WAIT) {
_current->base.swap_data = &pipe_desc;
key = irq_lock();
_sched_unlock_no_reschedule();
_pend_current_thread(&pipe->wait_q.readers, timeout);
_Swap(key);
} else {
k_sched_unlock();
}
*bytes_read = bytes_to_read - pipe_desc.bytes_to_xfer;
return _pipe_return_code(min_xfer, pipe_desc.bytes_to_xfer,
bytes_to_read);
}
int k_pipe_put(struct k_pipe *pipe, void *data, size_t bytes_to_write,
size_t *bytes_written, size_t min_xfer, s32_t timeout)
{
__ASSERT(min_xfer <= bytes_to_write, "");
__ASSERT(bytes_written != NULL, "");
return _k_pipe_put_internal(pipe, NULL, data,
bytes_to_write, bytes_written,
min_xfer, timeout);
}
#if (CONFIG_NUM_PIPE_ASYNC_MSGS > 0)
void k_pipe_block_put(struct k_pipe *pipe, struct k_mem_block *block,
size_t bytes_to_write, struct k_sem *sem)
{
struct k_pipe_async *async_desc;
size_t dummy_bytes_written;
/* For simplicity, always allocate an asynchronous descriptor */
_pipe_async_alloc(&async_desc);
async_desc->desc.block = &async_desc->desc.copy_block;
async_desc->desc.copy_block = *block;
async_desc->desc.sem = sem;
async_desc->thread.prio = k_thread_priority_get(_current);
(void) _k_pipe_put_internal(pipe, async_desc, block->data,
bytes_to_write, &dummy_bytes_written,
bytes_to_write, K_FOREVER);
}
#endif