Google Git
Sign in
nest-open-source / manifest_repos / mali-driver / refs/heads/main / . / bifrost / r32p1 / kernel / drivers / gpu / arm / midgard / mali_kbase_js.c
blob: 3682486674c0d191c4fb774a691565242e1cb387 [file] [log] [blame]
// SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note
/*
*
* (C) COPYRIGHT 2011-2021 ARM Limited. All rights reserved.
*
* This program is free software and is provided to you under the terms of the
* GNU General Public License version 2 as published by the Free Software
* Foundation, and any use by you of this program is subject to the terms
* of such GNU license.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, you can access it online at
* http://www.gnu.org/licenses/gpl-2.0.html.
*
*/
/*
* Job Scheduler Implementation
*/
#include <mali_kbase.h>
#include <mali_kbase_js.h>
#include <tl/mali_kbase_tracepoints.h>
#include <mali_linux_trace.h>
#include <mali_kbase_hw.h>
#include <mali_kbase_ctx_sched.h>
#include <mali_kbase_defs.h>
#include <mali_kbase_config_defaults.h>
#include "mali_kbase_jm.h"
#include "mali_kbase_hwaccess_jm.h"
#include <linux/priority_control_manager.h>
/*
* Private types
*/
/* Bitpattern indicating the result of releasing a context */
enum {
/* The context was descheduled - caller should try scheduling in a new
* one to keep the runpool full
*/
KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED = (1u << 0),
/* Ctx attributes were changed - caller should try scheduling all
* contexts
*/
KBASEP_JS_RELEASE_RESULT_SCHED_ALL = (1u << 1)
};
typedef u32 kbasep_js_release_result;
const int kbasep_js_atom_priority_to_relative[BASE_JD_NR_PRIO_LEVELS] = {
KBASE_JS_ATOM_SCHED_PRIO_MED, /* BASE_JD_PRIO_MEDIUM */
KBASE_JS_ATOM_SCHED_PRIO_HIGH, /* BASE_JD_PRIO_HIGH */
KBASE_JS_ATOM_SCHED_PRIO_LOW, /* BASE_JD_PRIO_LOW */
KBASE_JS_ATOM_SCHED_PRIO_REALTIME /* BASE_JD_PRIO_REALTIME */
};
const base_jd_prio
kbasep_js_relative_priority_to_atom[KBASE_JS_ATOM_SCHED_PRIO_COUNT] = {
BASE_JD_PRIO_REALTIME, /* KBASE_JS_ATOM_SCHED_PRIO_REALTIME */
BASE_JD_PRIO_HIGH, /* KBASE_JS_ATOM_SCHED_PRIO_HIGH */
BASE_JD_PRIO_MEDIUM, /* KBASE_JS_ATOM_SCHED_PRIO_MED */
BASE_JD_PRIO_LOW /* KBASE_JS_ATOM_SCHED_PRIO_LOW */
};
/*
* Private function prototypes
*/
static kbasep_js_release_result kbasep_js_runpool_release_ctx_internal(
struct kbase_device *kbdev, struct kbase_context *kctx,
struct kbasep_js_atom_retained_state *katom_retained_state);
static int kbase_js_get_slot(struct kbase_device *kbdev,
struct kbase_jd_atom *katom);
static void kbase_js_foreach_ctx_job(struct kbase_context *kctx,
kbasep_js_ctx_job_cb *callback);
/* Helper for ktrace */
#if KBASE_KTRACE_ENABLE
static int kbase_ktrace_get_ctx_refcnt(struct kbase_context *kctx)
{
return atomic_read(&kctx->refcount);
}
#else /* KBASE_KTRACE_ENABLE */
static int kbase_ktrace_get_ctx_refcnt(struct kbase_context *kctx)
{
CSTD_UNUSED(kctx);
return 0;
}
#endif /* KBASE_KTRACE_ENABLE */
/*
* Private functions
*/
/**
* core_reqs_from_jsn_features - Convert JSn_FEATURES to core requirements
* @features: JSn_FEATURE register value
*
* Given a JSn_FEATURE register value returns the core requirements that match
*
* Return: Core requirement bit mask
*/
static base_jd_core_req core_reqs_from_jsn_features(u16 features)
{
base_jd_core_req core_req = 0u;
if ((features & JS_FEATURE_SET_VALUE_JOB) != 0)
core_req |= BASE_JD_REQ_V;
if ((features & JS_FEATURE_CACHE_FLUSH_JOB) != 0)
core_req |= BASE_JD_REQ_CF;
if ((features & JS_FEATURE_COMPUTE_JOB) != 0)
core_req |= BASE_JD_REQ_CS;
if ((features & JS_FEATURE_TILER_JOB) != 0)
core_req |= BASE_JD_REQ_T;
if ((features & JS_FEATURE_FRAGMENT_JOB) != 0)
core_req |= BASE_JD_REQ_FS;
return core_req;
}
static void kbase_js_sync_timers(struct kbase_device *kbdev)
{
mutex_lock(&kbdev->js_data.runpool_mutex);
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&kbdev->js_data.runpool_mutex);
}
/**
* jsctx_rb_none_to_pull_prio(): - Check if there are no pullable atoms
* @kctx: Pointer to kbase context with ring buffer.
* @js: Job slot id to check.
* @prio: Priority to check.
*
* Return true if there are no atoms to pull. There may be running atoms in the
* ring buffer even if there are no atoms to pull. It is also possible for the
* ring buffer to be full (with running atoms) when this functions returns
* true.
*
* Return: true if there are no atoms to pull, false otherwise.
*/
static inline bool
jsctx_rb_none_to_pull_prio(struct kbase_context *kctx, int js, int prio)
{
bool none_to_pull;
struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
none_to_pull = RB_EMPTY_ROOT(&rb->runnable_tree);
dev_dbg(kctx->kbdev->dev,
"Slot %d (prio %d) is %spullable in kctx %pK\n",
js, prio, none_to_pull ? "not " : "", kctx);
return none_to_pull;
}
/**
* jsctx_rb_none_to_pull(): - Check if all priority ring buffers have no
* pullable atoms
* @kctx: Pointer to kbase context with ring buffer.
* @js: Job slot id to check.
*
* Caller must hold hwaccess_lock
*
* Return: true if the ring buffers for all priorities have no pullable atoms,
* false otherwise.
*/
static inline bool
jsctx_rb_none_to_pull(struct kbase_context *kctx, int js)
{
int prio;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
for (prio = KBASE_JS_ATOM_SCHED_PRIO_FIRST;
prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
if (!jsctx_rb_none_to_pull_prio(kctx, js, prio))
return false;
}
return true;
}
/**
* jsctx_queue_foreach_prio(): - Execute callback for each entry in the queue.
* @kctx: Pointer to kbase context with the queue.
* @js: Job slot id to iterate.
* @prio: Priority id to iterate.
* @callback: Function pointer to callback.
*
* Iterate over a queue and invoke @callback for each entry in the queue, and
* remove the entry from the queue.
*
* If entries are added to the queue while this is running those entries may, or
* may not be covered. To ensure that all entries in the buffer have been
* enumerated when this function returns jsctx->lock must be held when calling
* this function.
*
* The HW access lock must always be held when calling this function.
*/
static void jsctx_queue_foreach_prio(struct kbase_context *kctx, int js,
int prio, kbasep_js_ctx_job_cb *callback)
{
struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
while (!RB_EMPTY_ROOT(&queue->runnable_tree)) {
struct rb_node *node = rb_first(&queue->runnable_tree);
struct kbase_jd_atom *entry = rb_entry(node,
struct kbase_jd_atom, runnable_tree_node);
rb_erase(node, &queue->runnable_tree);
callback(kctx->kbdev, entry);
/* Runnable end-of-renderpass atoms can also be in the linked
* list of atoms blocked on cross-slot dependencies. Remove them
* to avoid calling the callback twice.
*/
if (entry->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST) {
WARN_ON(!(entry->core_req &
BASE_JD_REQ_END_RENDERPASS));
dev_dbg(kctx->kbdev->dev,
"Del runnable atom %pK from X_DEP list\n",
(void *)entry);
list_del(&entry->queue);
entry->atom_flags &=
~KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST;
}
}
while (!list_empty(&queue->x_dep_head)) {
struct kbase_jd_atom *entry = list_entry(queue->x_dep_head.next,
struct kbase_jd_atom, queue);
WARN_ON(!(entry->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST));
dev_dbg(kctx->kbdev->dev,
"Del blocked atom %pK from X_DEP list\n",
(void *)entry);
list_del(queue->x_dep_head.next);
entry->atom_flags &=
~KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST;
callback(kctx->kbdev, entry);
}
}
/**
* jsctx_queue_foreach(): - Execute callback for each entry in every queue
* @kctx: Pointer to kbase context with queue.
* @js: Job slot id to iterate.
* @callback: Function pointer to callback.
*
* Iterate over all the different priorities, and for each call
* jsctx_queue_foreach_prio() to iterate over the queue and invoke @callback
* for each entry, and remove the entry from the queue.
*/
static inline void jsctx_queue_foreach(struct kbase_context *kctx, int js,
kbasep_js_ctx_job_cb *callback)
{
int prio;
for (prio = KBASE_JS_ATOM_SCHED_PRIO_FIRST;
prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++)
jsctx_queue_foreach_prio(kctx, js, prio, callback);
}
/**
* jsctx_rb_peek_prio(): - Check buffer and get next atom
* @kctx: Pointer to kbase context with ring buffer.
* @js: Job slot id to check.
* @prio: Priority id to check.
*
* Check the ring buffer for the specified @js and @prio and return a pointer to
* the next atom, unless the ring buffer is empty.
*
* Return: Pointer to next atom in buffer, or NULL if there is no atom.
*/
static inline struct kbase_jd_atom *
jsctx_rb_peek_prio(struct kbase_context *kctx, int js, int prio)
{
struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
struct rb_node *node;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
dev_dbg(kctx->kbdev->dev,
"Peeking runnable tree of kctx %pK for prio %d (s:%d)\n",
(void *)kctx, prio, js);
node = rb_first(&rb->runnable_tree);
if (!node) {
dev_dbg(kctx->kbdev->dev, "Tree is empty\n");
return NULL;
}
return rb_entry(node, struct kbase_jd_atom, runnable_tree_node);
}
/**
* jsctx_rb_peek(): - Check all priority buffers and get next atom
* @kctx: Pointer to kbase context with ring buffer.
* @js: Job slot id to check.
*
* Check the ring buffers for all priorities, starting from
* KBASE_JS_ATOM_SCHED_PRIO_REALTIME, for the specified @js and @prio and return a
* pointer to the next atom, unless all the priority's ring buffers are empty.
*
* Caller must hold the hwaccess_lock.
*
* Return: Pointer to next atom in buffer, or NULL if there is no atom.
*/
static inline struct kbase_jd_atom *
jsctx_rb_peek(struct kbase_context *kctx, int js)
{
int prio;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
for (prio = KBASE_JS_ATOM_SCHED_PRIO_FIRST;
prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
struct kbase_jd_atom *katom;
katom = jsctx_rb_peek_prio(kctx, js, prio);
if (katom)
return katom;
}
return NULL;
}
/**
* jsctx_rb_pull(): - Mark atom in list as running
* @kctx: Pointer to kbase context with ring buffer.
* @katom: Pointer to katom to pull.
*
* Mark an atom previously obtained from jsctx_rb_peek() as running.
*
* @katom must currently be at the head of the ring buffer.
*/
static inline void
jsctx_rb_pull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
{
int prio = katom->sched_priority;
int js = katom->slot_nr;
struct jsctx_queue *rb = &kctx->jsctx_queue[prio][js];
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
dev_dbg(kctx->kbdev->dev, "Erasing atom %pK from runnable tree of kctx %pK\n",
(void *)katom, (void *)kctx);
/* Atoms must be pulled in the correct order. */
WARN_ON(katom != jsctx_rb_peek_prio(kctx, js, prio));
rb_erase(&katom->runnable_tree_node, &rb->runnable_tree);
}
#define LESS_THAN_WRAP(a, b) ((s32)(a - b) < 0)
static void
jsctx_tree_add(struct kbase_context *kctx, struct kbase_jd_atom *katom)
{
struct kbase_device *kbdev = kctx->kbdev;
int prio = katom->sched_priority;
int js = katom->slot_nr;
struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
struct rb_node **new = &(queue->runnable_tree.rb_node), *parent = NULL;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
dev_dbg(kbdev->dev, "Adding atom %pK to runnable tree of kctx %pK (s:%d)\n",
(void *)katom, (void *)kctx, js);
while (*new) {
struct kbase_jd_atom *entry = container_of(*new,
struct kbase_jd_atom, runnable_tree_node);
parent = *new;
if (LESS_THAN_WRAP(katom->age, entry->age))
new = &((*new)->rb_left);
else
new = &((*new)->rb_right);
}
/* Add new node and rebalance tree. */
rb_link_node(&katom->runnable_tree_node, parent, new);
rb_insert_color(&katom->runnable_tree_node, &queue->runnable_tree);
KBASE_TLSTREAM_TL_ATTRIB_ATOM_STATE(kbdev, katom, TL_ATOM_STATE_READY);
}
/**
* jsctx_rb_unpull(): - Undo marking of atom in list as running
* @kctx: Pointer to kbase context with ring buffer.
* @katom: Pointer to katom to unpull.
*
* Undo jsctx_rb_pull() and put @katom back in the queue.
*
* jsctx_rb_unpull() must be called on atoms in the same order the atoms were
* pulled.
*/
static inline void
jsctx_rb_unpull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
{
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
jsctx_tree_add(kctx, katom);
}
static bool kbase_js_ctx_pullable(struct kbase_context *kctx,
int js,
bool is_scheduled);
static bool kbase_js_ctx_list_add_pullable_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js);
static bool kbase_js_ctx_list_add_unpullable_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js);
/*
* Functions private to KBase ('Protected' functions)
*/
int kbasep_js_devdata_init(struct kbase_device * const kbdev)
{
struct kbasep_js_device_data *jsdd;
int i, j;
KBASE_DEBUG_ASSERT(kbdev != NULL);
jsdd = &kbdev->js_data;
#ifdef CONFIG_MALI_DEBUG
/* Soft-stop will be disabled on a single context by default unless
* softstop_always is set
*/
jsdd->softstop_always = false;
#endif /* CONFIG_MALI_DEBUG */
jsdd->nr_all_contexts_running = 0;
jsdd->nr_user_contexts_running = 0;
jsdd->nr_contexts_pullable = 0;
atomic_set(&jsdd->nr_contexts_runnable, 0);
/* No ctx allowed to submit */
jsdd->runpool_irq.submit_allowed = 0u;
memset(jsdd->runpool_irq.ctx_attr_ref_count, 0,
sizeof(jsdd->runpool_irq.ctx_attr_ref_count));
memset(jsdd->runpool_irq.slot_affinities, 0,
sizeof(jsdd->runpool_irq.slot_affinities));
memset(jsdd->runpool_irq.slot_affinity_refcount, 0,
sizeof(jsdd->runpool_irq.slot_affinity_refcount));
INIT_LIST_HEAD(&jsdd->suspended_soft_jobs_list);
/* Config attributes */
jsdd->scheduling_period_ns = DEFAULT_JS_SCHEDULING_PERIOD_NS;
jsdd->soft_stop_ticks = DEFAULT_JS_SOFT_STOP_TICKS;
jsdd->soft_stop_ticks_cl = DEFAULT_JS_SOFT_STOP_TICKS_CL;
jsdd->hard_stop_ticks_ss = DEFAULT_JS_HARD_STOP_TICKS_SS;
jsdd->hard_stop_ticks_cl = DEFAULT_JS_HARD_STOP_TICKS_CL;
jsdd->hard_stop_ticks_dumping = DEFAULT_JS_HARD_STOP_TICKS_DUMPING;
jsdd->gpu_reset_ticks_ss = DEFAULT_JS_RESET_TICKS_SS;
jsdd->gpu_reset_ticks_cl = DEFAULT_JS_RESET_TICKS_CL;
jsdd->gpu_reset_ticks_dumping = DEFAULT_JS_RESET_TICKS_DUMPING;
jsdd->ctx_timeslice_ns = DEFAULT_JS_CTX_TIMESLICE_NS;
atomic_set(&jsdd->soft_job_timeout_ms, DEFAULT_JS_SOFT_JOB_TIMEOUT);
dev_dbg(kbdev->dev, "JS Config Attribs: ");
dev_dbg(kbdev->dev, "\tscheduling_period_ns:%u",
jsdd->scheduling_period_ns);
dev_dbg(kbdev->dev, "\tsoft_stop_ticks:%u",
jsdd->soft_stop_ticks);
dev_dbg(kbdev->dev, "\tsoft_stop_ticks_cl:%u",
jsdd->soft_stop_ticks_cl);
dev_dbg(kbdev->dev, "\thard_stop_ticks_ss:%u",
jsdd->hard_stop_ticks_ss);
dev_dbg(kbdev->dev, "\thard_stop_ticks_cl:%u",
jsdd->hard_stop_ticks_cl);
dev_dbg(kbdev->dev, "\thard_stop_ticks_dumping:%u",
jsdd->hard_stop_ticks_dumping);
dev_dbg(kbdev->dev, "\tgpu_reset_ticks_ss:%u",
jsdd->gpu_reset_ticks_ss);
dev_dbg(kbdev->dev, "\tgpu_reset_ticks_cl:%u",
jsdd->gpu_reset_ticks_cl);
dev_dbg(kbdev->dev, "\tgpu_reset_ticks_dumping:%u",
jsdd->gpu_reset_ticks_dumping);
dev_dbg(kbdev->dev, "\tctx_timeslice_ns:%u",
jsdd->ctx_timeslice_ns);
dev_dbg(kbdev->dev, "\tsoft_job_timeout:%i",
atomic_read(&jsdd->soft_job_timeout_ms));
if (!(jsdd->soft_stop_ticks < jsdd->hard_stop_ticks_ss &&
jsdd->hard_stop_ticks_ss < jsdd->gpu_reset_ticks_ss &&
jsdd->soft_stop_ticks < jsdd->hard_stop_ticks_dumping &&
jsdd->hard_stop_ticks_dumping <
jsdd->gpu_reset_ticks_dumping)) {
dev_err(kbdev->dev, "Job scheduler timeouts invalid; soft/hard/reset tick counts should be in increasing order\n");
return -EINVAL;
}
#if KBASE_DISABLE_SCHEDULING_SOFT_STOPS
dev_dbg(kbdev->dev, "Job Scheduling Soft-stops disabled, ignoring value for soft_stop_ticks==%u at %uns per tick. Other soft-stops may still occur.",
jsdd->soft_stop_ticks,
jsdd->scheduling_period_ns);
#endif
#if KBASE_DISABLE_SCHEDULING_HARD_STOPS
dev_dbg(kbdev->dev, "Job Scheduling Hard-stops disabled, ignoring values for hard_stop_ticks_ss==%d and hard_stop_ticks_dumping==%u at %uns per tick. Other hard-stops may still occur.",
jsdd->hard_stop_ticks_ss,
jsdd->hard_stop_ticks_dumping,
jsdd->scheduling_period_ns);
#endif
#if KBASE_DISABLE_SCHEDULING_SOFT_STOPS && KBASE_DISABLE_SCHEDULING_HARD_STOPS
dev_dbg(kbdev->dev, "Note: The JS tick timer (if coded) will still be run, but do nothing.");
#endif
for (i = 0; i < kbdev->gpu_props.num_job_slots; ++i)
jsdd->js_reqs[i] = core_reqs_from_jsn_features(
kbdev->gpu_props.props.raw_props.js_features[i]);
/* On error, we could continue on: providing none of the below resources
* rely on the ones above
*/
mutex_init(&jsdd->runpool_mutex);
mutex_init(&jsdd->queue_mutex);
sema_init(&jsdd->schedule_sem, 1);
for (i = 0; i < kbdev->gpu_props.num_job_slots; ++i) {
for (j = KBASE_JS_ATOM_SCHED_PRIO_FIRST; j < KBASE_JS_ATOM_SCHED_PRIO_COUNT; ++j) {
INIT_LIST_HEAD(&jsdd->ctx_list_pullable[i][j]);
INIT_LIST_HEAD(&jsdd->ctx_list_unpullable[i][j]);
}
}
return 0;
}
void kbasep_js_devdata_halt(struct kbase_device *kbdev)
{
CSTD_UNUSED(kbdev);
}
void kbasep_js_devdata_term(struct kbase_device *kbdev)
{
struct kbasep_js_device_data *js_devdata;
s8 zero_ctx_attr_ref_count[KBASEP_JS_CTX_ATTR_COUNT] = { 0, };
KBASE_DEBUG_ASSERT(kbdev != NULL);
js_devdata = &kbdev->js_data;
/* The caller must de-register all contexts before calling this
*/
KBASE_DEBUG_ASSERT(js_devdata->nr_all_contexts_running == 0);
KBASE_DEBUG_ASSERT(memcmp(
js_devdata->runpool_irq.ctx_attr_ref_count,
zero_ctx_attr_ref_count,
sizeof(zero_ctx_attr_ref_count)) == 0);
CSTD_UNUSED(zero_ctx_attr_ref_count);
}
int kbasep_js_kctx_init(struct kbase_context *const kctx)
{
struct kbase_device *kbdev;
struct kbasep_js_kctx_info *js_kctx_info;
int i, j;
KBASE_DEBUG_ASSERT(kctx != NULL);
kbdev = kctx->kbdev;
KBASE_DEBUG_ASSERT(kbdev != NULL);
for (i = 0; i < BASE_JM_MAX_NR_SLOTS; ++i)
INIT_LIST_HEAD(&kctx->jctx.sched_info.ctx.ctx_list_entry[i]);
js_kctx_info = &kctx->jctx.sched_info;
kctx->slots_pullable = 0;
js_kctx_info->ctx.nr_jobs = 0;
kbase_ctx_flag_clear(kctx, KCTX_SCHEDULED);
kbase_ctx_flag_clear(kctx, KCTX_DYING);
memset(js_kctx_info->ctx.ctx_attr_ref_count, 0,
sizeof(js_kctx_info->ctx.ctx_attr_ref_count));
/* Initially, the context is disabled from submission until the create
* flags are set
*/
kbase_ctx_flag_set(kctx, KCTX_SUBMIT_DISABLED);
/* On error, we could continue on: providing none of the below resources
* rely on the ones above
*/
mutex_init(&js_kctx_info->ctx.jsctx_mutex);
init_waitqueue_head(&js_kctx_info->ctx.is_scheduled_wait);
for (i = KBASE_JS_ATOM_SCHED_PRIO_FIRST; i < KBASE_JS_ATOM_SCHED_PRIO_COUNT; i++) {
for (j = 0; j < BASE_JM_MAX_NR_SLOTS; j++) {
INIT_LIST_HEAD(&kctx->jsctx_queue[i][j].x_dep_head);
kctx->jsctx_queue[i][j].runnable_tree = RB_ROOT;
}
}
return 0;
}
void kbasep_js_kctx_term(struct kbase_context *kctx)
{
struct kbase_device *kbdev;
struct kbasep_js_kctx_info *js_kctx_info;
int js;
bool update_ctx_count = false;
unsigned long flags;
KBASE_DEBUG_ASSERT(kctx != NULL);
kbdev = kctx->kbdev;
KBASE_DEBUG_ASSERT(kbdev != NULL);
js_kctx_info = &kctx->jctx.sched_info;
/* The caller must de-register all jobs before calling this */
KBASE_DEBUG_ASSERT(!kbase_ctx_flag(kctx, KCTX_SCHEDULED));
KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs == 0);
mutex_lock(&kbdev->js_data.queue_mutex);
mutex_lock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++)
list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF)) {
WARN_ON(atomic_read(&kbdev->js_data.nr_contexts_runnable) <= 0);
atomic_dec(&kbdev->js_data.nr_contexts_runnable);
update_ctx_count = true;
kbase_ctx_flag_clear(kctx, KCTX_RUNNABLE_REF);
}
mutex_unlock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
mutex_unlock(&kbdev->js_data.queue_mutex);
if (update_ctx_count) {
mutex_lock(&kbdev->js_data.runpool_mutex);
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&kbdev->js_data.runpool_mutex);
}
}
/**
* kbase_js_ctx_list_add_pullable_nolock - Variant of
* kbase_jd_ctx_list_add_pullable()
* where the caller must hold
* hwaccess_lock
* @kbdev: Device pointer
* @kctx: Context to add to queue
* @js: Job slot to use
*
* Caller must hold hwaccess_lock
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool kbase_js_ctx_list_add_pullable_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
bool ret = false;
lockdep_assert_held(&kbdev->hwaccess_lock);
dev_dbg(kbdev->dev, "Add pullable tail kctx %pK (s:%d)\n",
(void *)kctx, js);
if (!list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
list_add_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
&kbdev->js_data.ctx_list_pullable[js][kctx->priority]);
if (!kctx->slots_pullable) {
kbdev->js_data.nr_contexts_pullable++;
ret = true;
if (!atomic_read(&kctx->atoms_pulled)) {
WARN_ON(kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_set(kctx, KCTX_RUNNABLE_REF);
atomic_inc(&kbdev->js_data.nr_contexts_runnable);
}
}
kctx->slots_pullable |= (1 << js);
return ret;
}
/**
* kbase_js_ctx_list_add_pullable_head_nolock - Variant of
* kbase_js_ctx_list_add_pullable_head()
* where the caller must hold
* hwaccess_lock
* @kbdev: Device pointer
* @kctx: Context to add to queue
* @js: Job slot to use
*
* Caller must hold hwaccess_lock
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool kbase_js_ctx_list_add_pullable_head_nolock(
struct kbase_device *kbdev, struct kbase_context *kctx, int js)
{
bool ret = false;
lockdep_assert_held(&kbdev->hwaccess_lock);
dev_dbg(kbdev->dev, "Add pullable head kctx %pK (s:%d)\n",
(void *)kctx, js);
if (!list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
list_add(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
&kbdev->js_data.ctx_list_pullable[js][kctx->priority]);
if (!kctx->slots_pullable) {
kbdev->js_data.nr_contexts_pullable++;
ret = true;
if (!atomic_read(&kctx->atoms_pulled)) {
WARN_ON(kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_set(kctx, KCTX_RUNNABLE_REF);
atomic_inc(&kbdev->js_data.nr_contexts_runnable);
}
}
kctx->slots_pullable |= (1 << js);
return ret;
}
/**
* kbase_js_ctx_list_add_pullable_head - Add context to the head of the
* per-slot pullable context queue
* @kbdev: Device pointer
* @kctx: Context to add to queue
* @js: Job slot to use
*
* If the context is on either the pullable or unpullable queues, then it is
* removed before being added to the head.
*
* This function should be used when a context has been scheduled, but no jobs
* can currently be pulled from it.
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool kbase_js_ctx_list_add_pullable_head(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
bool ret;
unsigned long flags;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
ret = kbase_js_ctx_list_add_pullable_head_nolock(kbdev, kctx, js);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return ret;
}
/**
* kbase_js_ctx_list_add_unpullable_nolock - Add context to the tail of the
* per-slot unpullable context queue
* @kbdev: Device pointer
* @kctx: Context to add to queue
* @js: Job slot to use
*
* The context must already be on the per-slot pullable queue. It will be
* removed from the pullable queue before being added to the unpullable queue.
*
* This function should be used when a context has been pulled from, and there
* are no jobs remaining on the specified slot.
*
* Caller must hold hwaccess_lock
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool kbase_js_ctx_list_add_unpullable_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
bool ret = false;
lockdep_assert_held(&kbdev->hwaccess_lock);
dev_dbg(kbdev->dev, "Add unpullable tail kctx %pK (s:%d)\n",
(void *)kctx, js);
list_move_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
&kbdev->js_data.ctx_list_unpullable[js][kctx->priority]);
if (kctx->slots_pullable == (1 << js)) {
kbdev->js_data.nr_contexts_pullable--;
ret = true;
if (!atomic_read(&kctx->atoms_pulled)) {
WARN_ON(!kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_clear(kctx, KCTX_RUNNABLE_REF);
atomic_dec(&kbdev->js_data.nr_contexts_runnable);
}
}
kctx->slots_pullable &= ~(1 << js);
return ret;
}
/**
* kbase_js_ctx_list_remove_nolock - Remove context from the per-slot pullable
* or unpullable context queues
* @kbdev: Device pointer
* @kctx: Context to remove from queue
* @js: Job slot to use
*
* The context must already be on one of the queues.
*
* This function should be used when a context has no jobs on the GPU, and no
* jobs remaining for the specified slot.
*
* Caller must hold hwaccess_lock
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool kbase_js_ctx_list_remove_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
bool ret = false;
lockdep_assert_held(&kbdev->hwaccess_lock);
WARN_ON(list_empty(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]));
list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
if (kctx->slots_pullable == (1 << js)) {
kbdev->js_data.nr_contexts_pullable--;
ret = true;
if (!atomic_read(&kctx->atoms_pulled)) {
WARN_ON(!kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_clear(kctx, KCTX_RUNNABLE_REF);
atomic_dec(&kbdev->js_data.nr_contexts_runnable);
}
}
kctx->slots_pullable &= ~(1 << js);
return ret;
}
/**
* kbase_js_ctx_list_pop_head_nolock - Variant of kbase_js_ctx_list_pop_head()
* where the caller must hold
* hwaccess_lock
* @kbdev: Device pointer
* @js: Job slot to use
*
* Caller must hold hwaccess_lock
*
* Return: Context to use for specified slot.
* NULL if no contexts present for specified slot
*/
static struct kbase_context *kbase_js_ctx_list_pop_head_nolock(
struct kbase_device *kbdev,
int js)
{
struct kbase_context *kctx;
int i;
lockdep_assert_held(&kbdev->hwaccess_lock);
for (i = KBASE_JS_ATOM_SCHED_PRIO_FIRST; i < KBASE_JS_ATOM_SCHED_PRIO_COUNT; i++) {
if (list_empty(&kbdev->js_data.ctx_list_pullable[js][i]))
continue;
kctx = list_entry(kbdev->js_data.ctx_list_pullable[js][i].next,
struct kbase_context,
jctx.sched_info.ctx.ctx_list_entry[js]);
list_del_init(&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
dev_dbg(kbdev->dev,
"Popped %pK from the pullable queue (s:%d)\n",
(void *)kctx, js);
return kctx;
}
return NULL;
}
/**
* kbase_js_ctx_list_pop_head - Pop the head context off the per-slot pullable
* queue.
* @kbdev: Device pointer
* @js: Job slot to use
*
* Return: Context to use for specified slot.
* NULL if no contexts present for specified slot
*/
static struct kbase_context *kbase_js_ctx_list_pop_head(
struct kbase_device *kbdev, int js)
{
struct kbase_context *kctx;
unsigned long flags;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
kctx = kbase_js_ctx_list_pop_head_nolock(kbdev, js);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return kctx;
}
/**
* kbase_js_ctx_pullable - Return if a context can be pulled from on the
* specified slot
* @kctx: Context pointer
* @js: Job slot to use
* @is_scheduled: true if the context is currently scheduled
*
* Caller must hold hwaccess_lock
*
* Return: true if context can be pulled from on specified slot
* false otherwise
*/
static bool kbase_js_ctx_pullable(struct kbase_context *kctx, int js,
bool is_scheduled)
{
struct kbasep_js_device_data *js_devdata;
struct kbase_jd_atom *katom;
struct kbase_device *kbdev = kctx->kbdev;
lockdep_assert_held(&kbdev->hwaccess_lock);
js_devdata = &kbdev->js_data;
if (is_scheduled) {
if (!kbasep_js_is_submit_allowed(js_devdata, kctx)) {
dev_dbg(kbdev->dev, "JS: No submit allowed for kctx %pK\n",
(void *)kctx);
return false;
}
}
katom = jsctx_rb_peek(kctx, js);
if (!katom) {
dev_dbg(kbdev->dev, "JS: No pullable atom in kctx %pK (s:%d)\n",
(void *)kctx, js);
return false; /* No pullable atoms */
}
if (kctx->blocked_js[js][katom->sched_priority]) {
dev_dbg(kbdev->dev,
"JS: kctx %pK is blocked from submitting atoms at priority %d (s:%d)\n",
(void *)kctx, katom->sched_priority, js);
return false;
}
if (atomic_read(&katom->blocked)) {
dev_dbg(kbdev->dev, "JS: Atom %pK is blocked in js_ctx_pullable\n",
(void *)katom);
return false; /* next atom blocked */
}
if (kbase_js_atom_blocked_on_x_dep(katom)) {
if (katom->x_pre_dep->gpu_rb_state ==
KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB ||
katom->x_pre_dep->will_fail_event_code) {
dev_dbg(kbdev->dev,
"JS: X pre-dep %pK is not present in slot FIFO or will fail\n",
(void *)katom->x_pre_dep);
return false;
}
if ((katom->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) &&
kbase_backend_nr_atoms_on_slot(kctx->kbdev, js)) {
dev_dbg(kbdev->dev,
"JS: Atom %pK has cross-slot fail dependency and atoms on slot (s:%d)\n",
(void *)katom, js);
return false;
}
}
dev_dbg(kbdev->dev, "JS: Atom %pK is pullable in kctx %pK (s:%d)\n",
(void *)katom, (void *)kctx, js);
return true;
}
static bool kbase_js_dep_validate(struct kbase_context *kctx,
struct kbase_jd_atom *katom)
{
struct kbase_device *kbdev = kctx->kbdev;
bool ret = true;
bool has_dep = false, has_x_dep = false;
int js = kbase_js_get_slot(kbdev, katom);
int prio = katom->sched_priority;
int i;
for (i = 0; i < 2; i++) {
struct kbase_jd_atom *dep_atom = katom->dep[i].atom;
if (dep_atom) {
int dep_js = kbase_js_get_slot(kbdev, dep_atom);
int dep_prio = dep_atom->sched_priority;
dev_dbg(kbdev->dev,
"Checking dep %d of atom %pK (s:%d) on %pK (s:%d)\n",
i, (void *)katom, js, (void *)dep_atom, dep_js);
/* Dependent atom must already have been submitted */
if (!(dep_atom->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_TREE)) {
dev_dbg(kbdev->dev,
"Blocker not submitted yet\n");
ret = false;
break;
}
/* Dependencies with different priorities can't
* be represented in the ringbuffer
*/
if (prio != dep_prio) {
dev_dbg(kbdev->dev,
"Different atom priorities\n");
ret = false;
break;
}
if (js == dep_js) {
/* Only one same-slot dependency can be
* represented in the ringbuffer
*/
if (has_dep) {
dev_dbg(kbdev->dev,
"Too many same-slot deps\n");
ret = false;
break;
}
/* Each dependee atom can only have one
* same-slot dependency
*/
if (dep_atom->post_dep) {
dev_dbg(kbdev->dev,
"Too many same-slot successors\n");
ret = false;
break;
}
has_dep = true;
} else {
/* Only one cross-slot dependency can be
* represented in the ringbuffer
*/
if (has_x_dep) {
dev_dbg(kbdev->dev,
"Too many cross-slot deps\n");
ret = false;
break;
}
/* Each dependee atom can only have one
* cross-slot dependency
*/
if (dep_atom->x_post_dep) {
dev_dbg(kbdev->dev,
"Too many cross-slot successors\n");
ret = false;
break;
}
/* The dependee atom can not already be in the
* HW access ringbuffer
*/
if (dep_atom->gpu_rb_state !=
KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB) {
dev_dbg(kbdev->dev,
"Blocker already in ringbuffer (state:%d)\n",
dep_atom->gpu_rb_state);
ret = false;
break;
}
/* The dependee atom can not already have
* completed
*/
if (dep_atom->status !=
KBASE_JD_ATOM_STATE_IN_JS) {
dev_dbg(kbdev->dev,
"Blocker already completed (status:%d)\n",
dep_atom->status);
ret = false;
break;
}
has_x_dep = true;
}
/* Dependency can be represented in ringbuffers */
}
}
/* If dependencies can be represented by ringbuffer then clear them from
* atom structure
*/
if (ret) {
for (i = 0; i < 2; i++) {
struct kbase_jd_atom *dep_atom = katom->dep[i].atom;
if (dep_atom) {
int dep_js = kbase_js_get_slot(kbdev, dep_atom);
dev_dbg(kbdev->dev,
"Clearing dep %d of atom %pK (s:%d) on %pK (s:%d)\n",
i, (void *)katom, js, (void *)dep_atom,
dep_js);
if ((js != dep_js) &&
(dep_atom->status !=
KBASE_JD_ATOM_STATE_COMPLETED)
&& (dep_atom->status !=
KBASE_JD_ATOM_STATE_HW_COMPLETED)
&& (dep_atom->status !=
KBASE_JD_ATOM_STATE_UNUSED)) {
katom->atom_flags |=
KBASE_KATOM_FLAG_X_DEP_BLOCKED;
dev_dbg(kbdev->dev, "Set X_DEP flag on atom %pK\n",
(void *)katom);
katom->x_pre_dep = dep_atom;
dep_atom->x_post_dep = katom;
if (kbase_jd_katom_dep_type(
&katom->dep[i]) ==
BASE_JD_DEP_TYPE_DATA)
katom->atom_flags |=
KBASE_KATOM_FLAG_FAIL_BLOCKER;
}
if ((kbase_jd_katom_dep_type(&katom->dep[i])
== BASE_JD_DEP_TYPE_DATA) &&
(js == dep_js)) {
katom->pre_dep = dep_atom;
dep_atom->post_dep = katom;
}
list_del(&katom->dep_item[i]);
kbase_jd_katom_dep_clear(&katom->dep[i]);
}
}
} else {
dev_dbg(kbdev->dev,
"Deps of atom %pK (s:%d) could not be represented\n",
(void *)katom, js);
}
return ret;
}
void kbase_js_set_ctx_priority(struct kbase_context *kctx, int new_priority)
{
struct kbase_device *kbdev = kctx->kbdev;
int js;
lockdep_assert_held(&kbdev->hwaccess_lock);
/* Move kctx to the pullable/upullable list as per the new priority */
if (new_priority != kctx->priority) {
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
if (kctx->slots_pullable & (1 << js))
list_move_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
&kbdev->js_data.ctx_list_pullable[js][new_priority]);
else
list_move_tail(&kctx->jctx.sched_info.ctx.ctx_list_entry[js],
&kbdev->js_data.ctx_list_unpullable[js][new_priority]);
}
kctx->priority = new_priority;
}
}
void kbase_js_update_ctx_priority(struct kbase_context *kctx)
{
struct kbase_device *kbdev = kctx->kbdev;
int new_priority = KBASE_JS_ATOM_SCHED_PRIO_LOW;
int prio;
lockdep_assert_held(&kbdev->hwaccess_lock);
if (kbdev->js_ctx_scheduling_mode == KBASE_JS_SYSTEM_PRIORITY_MODE) {
/* Determine the new priority for context, as per the priority
* of currently in-use atoms.
*/
for (prio = KBASE_JS_ATOM_SCHED_PRIO_FIRST;
prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
if (kctx->atoms_count[prio]) {
new_priority = prio;
break;
}
}
}
kbase_js_set_ctx_priority(kctx, new_priority);
}
KBASE_EXPORT_TEST_API(kbase_js_update_ctx_priority);
/**
* js_add_start_rp() - Add an atom that starts a renderpass to the job scheduler
* @start_katom: Pointer to the atom to be added.
* Return: 0 if successful or a negative value on failure.
*/
static int js_add_start_rp(struct kbase_jd_atom *const start_katom)
{
struct kbase_context *const kctx = start_katom->kctx;
struct kbase_jd_renderpass *rp;
struct kbase_device *const kbdev = kctx->kbdev;
unsigned long flags;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(start_katom->core_req & BASE_JD_REQ_START_RENDERPASS)))
return -EINVAL;
if (start_katom->core_req & BASE_JD_REQ_END_RENDERPASS)
return -EINVAL;
compiletime_assert((1ull << (sizeof(start_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[start_katom->renderpass_id];
if (rp->state != KBASE_JD_RP_COMPLETE)
return -EINVAL;
dev_dbg(kctx->kbdev->dev, "JS add start atom %pK of RP %d\n",
(void *)start_katom, start_katom->renderpass_id);
/* The following members are read when updating the job slot
* ringbuffer/fifo therefore they require additional locking.
*/
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
rp->state = KBASE_JD_RP_START;
rp->start_katom = start_katom;
rp->end_katom = NULL;
INIT_LIST_HEAD(&rp->oom_reg_list);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return 0;
}
/**
* js_add_end_rp() - Add an atom that ends a renderpass to the job scheduler
* @end_katom: Pointer to the atom to be added.
* Return: 0 if successful or a negative value on failure.
*/
static int js_add_end_rp(struct kbase_jd_atom *const end_katom)
{
struct kbase_context *const kctx = end_katom->kctx;
struct kbase_jd_renderpass *rp;
struct kbase_device *const kbdev = kctx->kbdev;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(end_katom->core_req & BASE_JD_REQ_END_RENDERPASS)))
return -EINVAL;
if (end_katom->core_req & BASE_JD_REQ_START_RENDERPASS)
return -EINVAL;
compiletime_assert((1ull << (sizeof(end_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[end_katom->renderpass_id];
dev_dbg(kbdev->dev, "JS add end atom %pK in state %d of RP %d\n",
(void *)end_katom, (int)rp->state, end_katom->renderpass_id);
if (rp->state == KBASE_JD_RP_COMPLETE)
return -EINVAL;
if (rp->end_katom == NULL) {
/* We can't be in a retry state until the fragment job chain
* has completed.
*/
unsigned long flags;
WARN_ON(rp->state == KBASE_JD_RP_RETRY);
WARN_ON(rp->state == KBASE_JD_RP_RETRY_PEND_OOM);
WARN_ON(rp->state == KBASE_JD_RP_RETRY_OOM);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
rp->end_katom = end_katom;
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
} else
WARN_ON(rp->end_katom != end_katom);
return 0;
}
bool kbasep_js_add_job(struct kbase_context *kctx,
struct kbase_jd_atom *atom)
{
unsigned long flags;
struct kbasep_js_kctx_info *js_kctx_info;
struct kbase_device *kbdev;
struct kbasep_js_device_data *js_devdata;
int err = 0;
bool enqueue_required = false;
bool timer_sync = false;
KBASE_DEBUG_ASSERT(kctx != NULL);
KBASE_DEBUG_ASSERT(atom != NULL);
lockdep_assert_held(&kctx->jctx.lock);
kbdev = kctx->kbdev;
js_devdata = &kbdev->js_data;
js_kctx_info = &kctx->jctx.sched_info;
mutex_lock(&js_devdata->queue_mutex);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
if (atom->core_req & BASE_JD_REQ_START_RENDERPASS)
err = js_add_start_rp(atom);
else if (atom->core_req & BASE_JD_REQ_END_RENDERPASS)
err = js_add_end_rp(atom);
if (err < 0) {
atom->event_code = BASE_JD_EVENT_JOB_INVALID;
atom->status = KBASE_JD_ATOM_STATE_COMPLETED;
goto out_unlock;
}
/*
* Begin Runpool transaction
*/
mutex_lock(&js_devdata->runpool_mutex);
/* Refcount ctx.nr_jobs */
KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs < U32_MAX);
++(js_kctx_info->ctx.nr_jobs);
dev_dbg(kbdev->dev, "Add atom %pK to kctx %pK; now %d in ctx\n",
(void *)atom, (void *)kctx, js_kctx_info->ctx.nr_jobs);
/* Lock for state available during IRQ */
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (++kctx->atoms_count[atom->sched_priority] == 1)
kbase_js_update_ctx_priority(kctx);
if (!kbase_js_dep_validate(kctx, atom)) {
/* Dependencies could not be represented */
--(js_kctx_info->ctx.nr_jobs);
dev_dbg(kbdev->dev,
"Remove atom %pK from kctx %pK; now %d in ctx\n",
(void *)atom, (void *)kctx, js_kctx_info->ctx.nr_jobs);
/* Setting atom status back to queued as it still has unresolved
* dependencies
*/
atom->status = KBASE_JD_ATOM_STATE_QUEUED;
dev_dbg(kbdev->dev, "Atom %pK status to queued\n", (void *)atom);
/* Undo the count, as the atom will get added again later but
* leave the context priority adjusted or boosted, in case if
* this was the first higher priority atom received for this
* context.
* This will prevent the scenario of priority inversion, where
* another context having medium priority atoms keeps getting
* scheduled over this context, which is having both lower and
* higher priority atoms, but higher priority atoms are blocked
* due to dependency on lower priority atoms. With priority
* boost the high priority atom will get to run at earliest.
*/
kctx->atoms_count[atom->sched_priority]--;
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&js_devdata->runpool_mutex);
goto out_unlock;
}
enqueue_required = kbase_js_dep_resolved_submit(kctx, atom);
KBASE_KTRACE_ADD_JM_REFCOUNT(kbdev, JS_ADD_JOB, kctx, atom, atom->jc,
kbase_ktrace_get_ctx_refcnt(kctx));
/* Context Attribute Refcounting */
kbasep_js_ctx_attr_ctx_retain_atom(kbdev, kctx, atom);
if (enqueue_required) {
if (kbase_js_ctx_pullable(kctx, atom->slot_nr, false))
timer_sync = kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, atom->slot_nr);
else
timer_sync = kbase_js_ctx_list_add_unpullable_nolock(
kbdev, kctx, atom->slot_nr);
}
/* If this context is active and the atom is the first on its slot,
* kick the job manager to attempt to fast-start the atom
*/
if (enqueue_required && kctx ==
kbdev->hwaccess.active_kctx[atom->slot_nr])
kbase_jm_try_kick(kbdev, 1 << atom->slot_nr);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (timer_sync)
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&js_devdata->runpool_mutex);
/* End runpool transaction */
if (!kbase_ctx_flag(kctx, KCTX_SCHEDULED)) {
if (kbase_ctx_flag(kctx, KCTX_DYING)) {
/* A job got added while/after kbase_job_zap_context()
* was called on a non-scheduled context. Kill that job
* by killing the context.
*/
kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx,
false);
} else if (js_kctx_info->ctx.nr_jobs == 1) {
/* Handle Refcount going from 0 to 1: schedule the
* context on the Queue
*/
KBASE_DEBUG_ASSERT(!kbase_ctx_flag(kctx, KCTX_SCHEDULED));
dev_dbg(kbdev->dev, "JS: Enqueue Context %pK", kctx);
/* Queue was updated - caller must try to schedule the
* head context
*/
WARN_ON(!enqueue_required);
}
}
out_unlock:
dev_dbg(kbdev->dev, "Enqueue of kctx %pK is %srequired\n",
kctx, enqueue_required ? "" : "not ");
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
return enqueue_required;
}
void kbasep_js_remove_job(struct kbase_device *kbdev,
struct kbase_context *kctx, struct kbase_jd_atom *atom)
{
struct kbasep_js_kctx_info *js_kctx_info;
unsigned long flags;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
KBASE_DEBUG_ASSERT(atom != NULL);
js_kctx_info = &kctx->jctx.sched_info;
KBASE_KTRACE_ADD_JM_REFCOUNT(kbdev, JS_REMOVE_JOB, kctx, atom, atom->jc,
kbase_ktrace_get_ctx_refcnt(kctx));
/* De-refcount ctx.nr_jobs */
KBASE_DEBUG_ASSERT(js_kctx_info->ctx.nr_jobs > 0);
--(js_kctx_info->ctx.nr_jobs);
dev_dbg(kbdev->dev,
"Remove atom %pK from kctx %pK; now %d in ctx\n",
(void *)atom, (void *)kctx, js_kctx_info->ctx.nr_jobs);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (--kctx->atoms_count[atom->sched_priority] == 0)
kbase_js_update_ctx_priority(kctx);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
}
bool kbasep_js_remove_cancelled_job(struct kbase_device *kbdev,
struct kbase_context *kctx, struct kbase_jd_atom *katom)
{
unsigned long flags;
struct kbasep_js_atom_retained_state katom_retained_state;
bool attr_state_changed;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
KBASE_DEBUG_ASSERT(katom != NULL);
kbasep_js_atom_retained_state_copy(&katom_retained_state, katom);
kbasep_js_remove_job(kbdev, kctx, katom);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
/* The atom has 'finished' (will not be re-run), so no need to call
* kbasep_js_has_atom_finished().
*
* This is because it returns false for soft-stopped atoms, but we
* want to override that, because we're cancelling an atom regardless of
* whether it was soft-stopped or not
*/
attr_state_changed = kbasep_js_ctx_attr_ctx_release_atom(kbdev, kctx,
&katom_retained_state);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return attr_state_changed;
}
/**
* kbasep_js_run_jobs_after_ctx_and_atom_release - Try running more jobs after
* releasing a context and/or atom
* @kbdev: The kbase_device to operate on
* @kctx: The kbase_context to operate on
* @katom_retained_state: Retained state from the atom
* @runpool_ctx_attr_change: True if the runpool context attributes have changed
*
* This collates a set of actions that must happen whilst hwaccess_lock is held.
*
* This includes running more jobs when:
* - The previously released kctx caused a ctx attribute change,
* - The released atom caused a ctx attribute change,
* - Slots were previously blocked due to affinity restrictions,
* - Submission during IRQ handling failed.
*
* Return: %KBASEP_JS_RELEASE_RESULT_SCHED_ALL if context attributes were
* changed. The caller should try scheduling all contexts
*/
static kbasep_js_release_result kbasep_js_run_jobs_after_ctx_and_atom_release(
struct kbase_device *kbdev,
struct kbase_context *kctx,
struct kbasep_js_atom_retained_state *katom_retained_state,
bool runpool_ctx_attr_change)
{
struct kbasep_js_device_data *js_devdata;
kbasep_js_release_result result = 0;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
KBASE_DEBUG_ASSERT(katom_retained_state != NULL);
js_devdata = &kbdev->js_data;
lockdep_assert_held(&kctx->jctx.sched_info.ctx.jsctx_mutex);
lockdep_assert_held(&js_devdata->runpool_mutex);
lockdep_assert_held(&kbdev->hwaccess_lock);
if (js_devdata->nr_user_contexts_running != 0 && runpool_ctx_attr_change) {
/* A change in runpool ctx attributes might mean we can
* run more jobs than before
*/
result = KBASEP_JS_RELEASE_RESULT_SCHED_ALL;
KBASE_KTRACE_ADD_JM_SLOT(kbdev, JD_DONE_TRY_RUN_NEXT_JOB,
kctx, NULL, 0u, 0);
}
return result;
}
/**
* kbasep_js_runpool_release_ctx_internal - Internal function to release the reference
* on a ctx and an atom's "retained state", only
* taking the runpool and as transaction mutexes
* @kbdev: The kbase_device to operate on
* @kctx: The kbase_context to operate on
* @katom_retained_state: Retained state from the atom
*
* This also starts more jobs running in the case of an ctx-attribute state change
*
* This does none of the followup actions for scheduling:
* - It does not schedule in a new context
* - It does not requeue or handle dying contexts
*
* For those tasks, just call kbasep_js_runpool_release_ctx() instead
*
* Has following requirements
* - Context is scheduled in, and kctx->as_nr matches kctx_as_nr
* - Context has a non-zero refcount
* - Caller holds js_kctx_info->ctx.jsctx_mutex
* - Caller holds js_devdata->runpool_mutex
*
* Return: A bitpattern, containing KBASEP_JS_RELEASE_RESULT_* flags, indicating
* the result of releasing a context that whether the caller should try
* scheduling a new context or should try scheduling all contexts.
*/
static kbasep_js_release_result kbasep_js_runpool_release_ctx_internal(
struct kbase_device *kbdev,
struct kbase_context *kctx,
struct kbasep_js_atom_retained_state *katom_retained_state)
{
unsigned long flags;
struct kbasep_js_device_data *js_devdata;
struct kbasep_js_kctx_info *js_kctx_info;
kbasep_js_release_result release_result = 0u;
bool runpool_ctx_attr_change = false;
int kctx_as_nr;
int new_ref_count;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
js_kctx_info = &kctx->jctx.sched_info;
js_devdata = &kbdev->js_data;
/* Ensure context really is scheduled in */
KBASE_DEBUG_ASSERT(kbase_ctx_flag(kctx, KCTX_SCHEDULED));
kctx_as_nr = kctx->as_nr;
KBASE_DEBUG_ASSERT(kctx_as_nr != KBASEP_AS_NR_INVALID);
KBASE_DEBUG_ASSERT(atomic_read(&kctx->refcount) > 0);
/*
* Transaction begins on AS and runpool_irq
*
* Assert about out calling contract
*/
mutex_lock(&kbdev->pm.lock);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
KBASE_DEBUG_ASSERT(kctx_as_nr == kctx->as_nr);
KBASE_DEBUG_ASSERT(atomic_read(&kctx->refcount) > 0);
/* Update refcount */
kbase_ctx_sched_release_ctx(kctx);
new_ref_count = atomic_read(&kctx->refcount);
/* Release the atom if it finished (i.e. wasn't soft-stopped) */
if (kbasep_js_has_atom_finished(katom_retained_state))
runpool_ctx_attr_change |= kbasep_js_ctx_attr_ctx_release_atom(
kbdev, kctx, katom_retained_state);
if (new_ref_count == 2 && kbase_ctx_flag(kctx, KCTX_PRIVILEGED) &&
#ifdef CONFIG_MALI_ARBITER_SUPPORT
!kbase_pm_is_gpu_lost(kbdev) &&
#endif
!kbase_pm_is_suspending(kbdev)) {
/* Context is kept scheduled into an address space even when
* there are no jobs, in this case we have to handle the
* situation where all jobs have been evicted from the GPU and
* submission is disabled.
*
* At this point we re-enable submission to allow further jobs
* to be executed
*/
kbasep_js_set_submit_allowed(js_devdata, kctx);
}
/* Make a set of checks to see if the context should be scheduled out.
* Note that there'll always be at least 1 reference to the context
* which was previously acquired by kbasep_js_schedule_ctx().
*/
if (new_ref_count == 1 &&
(!kbasep_js_is_submit_allowed(js_devdata, kctx) ||
#ifdef CONFIG_MALI_ARBITER_SUPPORT
kbase_pm_is_gpu_lost(kbdev) ||
#endif
kbase_pm_is_suspending(kbdev))) {
int num_slots = kbdev->gpu_props.num_job_slots;
int slot;
/* Last reference, and we've been told to remove this context
* from the Run Pool
*/
dev_dbg(kbdev->dev, "JS: RunPool Remove Context %pK because refcount=%d, jobs=%d, allowed=%d",
kctx, new_ref_count, js_kctx_info->ctx.nr_jobs,
kbasep_js_is_submit_allowed(js_devdata, kctx));
KBASE_TLSTREAM_TL_NRET_AS_CTX(kbdev, &kbdev->as[kctx->as_nr], kctx);
kbase_backend_release_ctx_irq(kbdev, kctx);
for (slot = 0; slot < num_slots; slot++) {
if (kbdev->hwaccess.active_kctx[slot] == kctx) {
dev_dbg(kbdev->dev, "Marking kctx %pK as inactive (s:%d)\n",
(void *)kctx, slot);
kbdev->hwaccess.active_kctx[slot] = NULL;
}
}
/* Ctx Attribute handling
*
* Releasing atoms attributes must either happen before this, or
* after the KCTX_SHEDULED flag is changed, otherwise we
* double-decount the attributes
*/
runpool_ctx_attr_change |=
kbasep_js_ctx_attr_runpool_release_ctx(kbdev, kctx);
/* Releasing the context and katom retained state can allow
* more jobs to run
*/
release_result |=
kbasep_js_run_jobs_after_ctx_and_atom_release(kbdev,
kctx, katom_retained_state,
runpool_ctx_attr_change);
/*
* Transaction ends on AS and runpool_irq:
*
* By this point, the AS-related data is now clear and ready
* for re-use.
*
* Since releases only occur once for each previous successful
* retain, and no more retains are allowed on this context, no
* other thread will be operating in this
* code whilst we are
*/
/* Recalculate pullable status for all slots */
for (slot = 0; slot < num_slots; slot++) {
if (kbase_js_ctx_pullable(kctx, slot, false))
kbase_js_ctx_list_add_pullable_nolock(kbdev,
kctx, slot);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
kbase_backend_release_ctx_noirq(kbdev, kctx);
mutex_unlock(&kbdev->pm.lock);
/* Note: Don't reuse kctx_as_nr now */
/* Synchronize with any timers */
kbase_backend_ctx_count_changed(kbdev);
/* update book-keeping info */
kbase_ctx_flag_clear(kctx, KCTX_SCHEDULED);
/* Signal any waiter that the context is not scheduled, so is
* safe for termination - once the jsctx_mutex is also dropped,
* and jobs have finished.
*/
wake_up(&js_kctx_info->ctx.is_scheduled_wait);
/* Queue an action to occur after we've dropped the lock */
release_result |= KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED |
KBASEP_JS_RELEASE_RESULT_SCHED_ALL;
} else {
kbasep_js_run_jobs_after_ctx_and_atom_release(kbdev, kctx,
katom_retained_state, runpool_ctx_attr_change);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->pm.lock);
}
return release_result;
}
void kbasep_js_runpool_release_ctx_nolock(struct kbase_device *kbdev,
struct kbase_context *kctx)
{
struct kbasep_js_atom_retained_state katom_retained_state;
/* Setup a dummy katom_retained_state */
kbasep_js_atom_retained_state_init_invalid(&katom_retained_state);
kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
&katom_retained_state);
}
void kbasep_js_runpool_requeue_or_kill_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx, bool has_pm_ref)
{
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
/* This is called if and only if you've you've detached the context from
* the Runpool Queue, and not added it back to the Runpool
*/
KBASE_DEBUG_ASSERT(!kbase_ctx_flag(kctx, KCTX_SCHEDULED));
if (kbase_ctx_flag(kctx, KCTX_DYING)) {
/* Dying: don't requeue, but kill all jobs on the context. This
* happens asynchronously
*/
dev_dbg(kbdev->dev,
"JS: ** Killing Context %pK on RunPool Remove **", kctx);
kbase_js_foreach_ctx_job(kctx, &kbase_jd_cancel);
}
}
void kbasep_js_runpool_release_ctx_and_katom_retained_state(
struct kbase_device *kbdev, struct kbase_context *kctx,
struct kbasep_js_atom_retained_state *katom_retained_state)
{
struct kbasep_js_device_data *js_devdata;
struct kbasep_js_kctx_info *js_kctx_info;
kbasep_js_release_result release_result;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
js_kctx_info = &kctx->jctx.sched_info;
js_devdata = &kbdev->js_data;
mutex_lock(&js_devdata->queue_mutex);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
mutex_lock(&js_devdata->runpool_mutex);
release_result = kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
katom_retained_state);
/* Drop the runpool mutex to allow requeing kctx */
mutex_unlock(&js_devdata->runpool_mutex);
if ((release_result & KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED) != 0u)
kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, true);
/* Drop the jsctx_mutex to allow scheduling in a new context */
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
if (release_result & KBASEP_JS_RELEASE_RESULT_SCHED_ALL)
kbase_js_sched_all(kbdev);
}
void kbasep_js_runpool_release_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx)
{
struct kbasep_js_atom_retained_state katom_retained_state;
kbasep_js_atom_retained_state_init_invalid(&katom_retained_state);
kbasep_js_runpool_release_ctx_and_katom_retained_state(kbdev, kctx,
&katom_retained_state);
}
/* Variant of kbasep_js_runpool_release_ctx() that doesn't call into
* kbase_js_sched_all()
*/
static void kbasep_js_runpool_release_ctx_no_schedule(
struct kbase_device *kbdev, struct kbase_context *kctx)
{
struct kbasep_js_device_data *js_devdata;
struct kbasep_js_kctx_info *js_kctx_info;
kbasep_js_release_result release_result;
struct kbasep_js_atom_retained_state katom_retained_state_struct;
struct kbasep_js_atom_retained_state *katom_retained_state =
&katom_retained_state_struct;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
js_kctx_info = &kctx->jctx.sched_info;
js_devdata = &kbdev->js_data;
kbasep_js_atom_retained_state_init_invalid(katom_retained_state);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
mutex_lock(&js_devdata->runpool_mutex);
release_result = kbasep_js_runpool_release_ctx_internal(kbdev, kctx,
katom_retained_state);
/* Drop the runpool mutex to allow requeing kctx */
mutex_unlock(&js_devdata->runpool_mutex);
if ((release_result & KBASEP_JS_RELEASE_RESULT_WAS_DESCHEDULED) != 0u)
kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, true);
/* Drop the jsctx_mutex to allow scheduling in a new context */
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
/* NOTE: could return release_result if the caller would like to know
* whether it should schedule a new context, but currently no callers do
*/
}
void kbase_js_set_timeouts(struct kbase_device *kbdev)
{
lockdep_assert_held(&kbdev->hwaccess_lock);
kbase_backend_timeouts_changed(kbdev);
}
static bool kbasep_js_schedule_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
struct kbasep_js_device_data *js_devdata;
struct kbasep_js_kctx_info *js_kctx_info;
unsigned long flags;
bool kctx_suspended = false;
int as_nr;
dev_dbg(kbdev->dev, "Scheduling kctx %pK (s:%d)\n", kctx, js);
js_devdata = &kbdev->js_data;
js_kctx_info = &kctx->jctx.sched_info;
/* Pick available address space for this context */
mutex_lock(&kbdev->mmu_hw_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
as_nr = kbase_ctx_sched_retain_ctx(kctx);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->mmu_hw_mutex);
if (as_nr == KBASEP_AS_NR_INVALID) {
as_nr = kbase_backend_find_and_release_free_address_space(
kbdev, kctx);
if (as_nr != KBASEP_AS_NR_INVALID) {
/* Attempt to retain the context again, this should
* succeed
*/
mutex_lock(&kbdev->mmu_hw_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
as_nr = kbase_ctx_sched_retain_ctx(kctx);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->mmu_hw_mutex);
WARN_ON(as_nr == KBASEP_AS_NR_INVALID);
}
}
if (as_nr == KBASEP_AS_NR_INVALID)
return false; /* No address spaces currently available */
/*
* Atomic transaction on the Context and Run Pool begins
*/
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
mutex_lock(&js_devdata->runpool_mutex);
mutex_lock(&kbdev->mmu_hw_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
/* Check to see if context is dying due to kbase_job_zap_context() */
if (kbase_ctx_flag(kctx, KCTX_DYING)) {
/* Roll back the transaction so far and return */
kbase_ctx_sched_release_ctx(kctx);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->mmu_hw_mutex);
mutex_unlock(&js_devdata->runpool_mutex);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
return false;
}
KBASE_KTRACE_ADD_JM_REFCOUNT(kbdev, JS_TRY_SCHEDULE_HEAD_CTX, kctx, NULL,
0u,
kbase_ktrace_get_ctx_refcnt(kctx));
kbase_ctx_flag_set(kctx, KCTX_SCHEDULED);
/* Assign context to previously chosen address space */
if (!kbase_backend_use_ctx(kbdev, kctx, as_nr)) {
/* Roll back the transaction so far and return */
kbase_ctx_sched_release_ctx(kctx);
kbase_ctx_flag_clear(kctx, KCTX_SCHEDULED);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->mmu_hw_mutex);
mutex_unlock(&js_devdata->runpool_mutex);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
return false;
}
kbdev->hwaccess.active_kctx[js] = kctx;
KBASE_TLSTREAM_TL_RET_AS_CTX(kbdev, &kbdev->as[kctx->as_nr], kctx);
/* Cause any future waiter-on-termination to wait until the context is
* descheduled
*/
wake_up(&js_kctx_info->ctx.is_scheduled_wait);
/* Re-check for suspending: a suspend could've occurred, and all the
* contexts could've been removed from the runpool before we took this
* lock. In this case, we don't want to allow this context to run jobs,
* we just want it out immediately.
*
* The DMB required to read the suspend flag was issued recently as part
* of the hwaccess_lock locking. If a suspend occurs *after* that lock
* was taken (i.e. this condition doesn't execute), then the
* kbasep_js_suspend() code will cleanup this context instead (by virtue
* of it being called strictly after the suspend flag is set, and will
* wait for this lock to drop)
*/
#ifdef CONFIG_MALI_ARBITER_SUPPORT
if (kbase_pm_is_suspending(kbdev) || kbase_pm_is_gpu_lost(kbdev)) {
#else
if (kbase_pm_is_suspending(kbdev)) {
#endif
/* Cause it to leave at some later point */
bool retained;
retained = kbase_ctx_sched_inc_refcount_nolock(kctx);
KBASE_DEBUG_ASSERT(retained);
kbasep_js_clear_submit_allowed(js_devdata, kctx);
kctx_suspended = true;
}
kbase_ctx_flag_clear(kctx, KCTX_PULLED_SINCE_ACTIVE_JS0 << js);
/* Transaction complete */
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kbdev->mmu_hw_mutex);
/* Synchronize with any timers */
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&js_devdata->runpool_mutex);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
/* Note: after this point, the context could potentially get scheduled
* out immediately
*/
if (kctx_suspended) {
/* Finishing forcing out the context due to a suspend. Use a
* variant of kbasep_js_runpool_release_ctx() that doesn't
* schedule a new context, to prevent a risk of recursion back
* into this function
*/
kbasep_js_runpool_release_ctx_no_schedule(kbdev, kctx);
return false;
}
return true;
}
static bool kbase_js_use_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx,
int js)
{
unsigned long flags;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (kbase_ctx_flag(kctx, KCTX_SCHEDULED) &&
kbase_backend_use_ctx_sched(kbdev, kctx, js)) {
dev_dbg(kbdev->dev,
"kctx %pK already has ASID - mark as active (s:%d)\n",
(void *)kctx, js);
if (kbdev->hwaccess.active_kctx[js] != kctx) {
kbdev->hwaccess.active_kctx[js] = kctx;
kbase_ctx_flag_clear(kctx,
KCTX_PULLED_SINCE_ACTIVE_JS0 << js);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return true; /* Context already scheduled */
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
return kbasep_js_schedule_ctx(kbdev, kctx, js);
}
void kbasep_js_schedule_privileged_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx)
{
struct kbasep_js_kctx_info *js_kctx_info;
struct kbasep_js_device_data *js_devdata;
bool is_scheduled;
KBASE_DEBUG_ASSERT(kbdev != NULL);
KBASE_DEBUG_ASSERT(kctx != NULL);
js_devdata = &kbdev->js_data;
js_kctx_info = &kctx->jctx.sched_info;
#ifdef CONFIG_MALI_ARBITER_SUPPORT
/* This should only happen in response to a system call
* from a user-space thread.
* In a non-arbitrated environment this can never happen
* whilst suspending.
*
* In an arbitrated environment, user-space threads can run
* while we are suspended (for example GPU not available
* to this VM), however in that case we will block on
* the wait event for KCTX_SCHEDULED, since no context
* can be scheduled until we have the GPU again.
*/
if (kbdev->arb.arb_if == NULL)
if (WARN_ON(kbase_pm_is_suspending(kbdev)))
return;
#else
/* This should only happen in response to a system call
* from a user-space thread.
* In a non-arbitrated environment this can never happen
* whilst suspending.
*/
if (WARN_ON(kbase_pm_is_suspending(kbdev)))
return;
#endif
mutex_lock(&js_devdata->queue_mutex);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
/* Mark the context as privileged */
kbase_ctx_flag_set(kctx, KCTX_PRIVILEGED);
is_scheduled = kbase_ctx_flag(kctx, KCTX_SCHEDULED);
if (!is_scheduled) {
/* Add the context to the pullable list */
if (kbase_js_ctx_list_add_pullable_head(kbdev, kctx, 0))
kbase_js_sync_timers(kbdev);
/* Fast-starting requires the jsctx_mutex to be dropped,
* because it works on multiple ctxs
*/
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
/* Try to schedule the context in */
kbase_js_sched_all(kbdev);
/* Wait for the context to be scheduled in */
wait_event(kctx->jctx.sched_info.ctx.is_scheduled_wait,
kbase_ctx_flag(kctx, KCTX_SCHEDULED));
} else {
/* Already scheduled in - We need to retain it to keep the
* corresponding address space
*/
WARN_ON(!kbase_ctx_sched_inc_refcount(kctx));
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
}
}
KBASE_EXPORT_TEST_API(kbasep_js_schedule_privileged_ctx);
void kbasep_js_release_privileged_ctx(struct kbase_device *kbdev,
struct kbase_context *kctx)
{
struct kbasep_js_kctx_info *js_kctx_info;
KBASE_DEBUG_ASSERT(kctx != NULL);
js_kctx_info = &kctx->jctx.sched_info;
/* We don't need to use the address space anymore */
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
kbase_ctx_flag_clear(kctx, KCTX_PRIVILEGED);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
/* Release the context - it will be scheduled out */
kbasep_js_runpool_release_ctx(kbdev, kctx);
kbase_js_sched_all(kbdev);
}
KBASE_EXPORT_TEST_API(kbasep_js_release_privileged_ctx);
void kbasep_js_suspend(struct kbase_device *kbdev)
{
unsigned long flags;
struct kbasep_js_device_data *js_devdata;
int i;
u16 retained = 0u;
KBASE_DEBUG_ASSERT(kbdev);
KBASE_DEBUG_ASSERT(kbase_pm_is_suspending(kbdev));
js_devdata = &kbdev->js_data;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
/* Prevent all contexts from submitting */
js_devdata->runpool_irq.submit_allowed = 0;
/* Retain each of the contexts, so we can cause it to leave even if it
* had no refcount to begin with
*/
for (i = BASE_MAX_NR_AS - 1; i >= 0; --i) {
struct kbase_context *kctx = kbdev->as_to_kctx[i];
retained = retained << 1;
if (kctx && !(kbdev->as_free & (1u << i))) {
kbase_ctx_sched_retain_ctx_refcount(kctx);
retained |= 1u;
/* This loop will not have an effect on the privileged
* contexts as they would have an extra ref count
* compared to the normal contexts, so they will hold
* on to their address spaces. MMU will re-enabled for
* them on resume.
*/
}
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
/* De-ref the previous retain to ensure each context gets pulled out
* sometime later.
*/
for (i = 0;
i < BASE_MAX_NR_AS;
++i, retained = retained >> 1) {
struct kbase_context *kctx = kbdev->as_to_kctx[i];
if (retained & 1u)
kbasep_js_runpool_release_ctx(kbdev, kctx);
}
/* Caller must wait for all Power Manager active references to be
* dropped
*/
}
void kbasep_js_resume(struct kbase_device *kbdev)
{
struct kbasep_js_device_data *js_devdata;
int js, prio;
KBASE_DEBUG_ASSERT(kbdev);
js_devdata = &kbdev->js_data;
KBASE_DEBUG_ASSERT(!kbase_pm_is_suspending(kbdev));
mutex_lock(&js_devdata->queue_mutex);
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
for (prio = KBASE_JS_ATOM_SCHED_PRIO_FIRST;
prio < KBASE_JS_ATOM_SCHED_PRIO_COUNT; prio++) {
struct kbase_context *kctx, *n;
unsigned long flags;
#ifndef CONFIG_MALI_ARBITER_SUPPORT
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
list_for_each_entry_safe(kctx, n,
&kbdev->js_data.ctx_list_unpullable[js][prio],
jctx.sched_info.ctx.ctx_list_entry[js]) {
struct kbasep_js_kctx_info *js_kctx_info;
bool timer_sync = false;
/* Drop lock so we can take kctx mutexes */
spin_unlock_irqrestore(&kbdev->hwaccess_lock,
flags);
js_kctx_info = &kctx->jctx.sched_info;
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
mutex_lock(&js_devdata->runpool_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (!kbase_ctx_flag(kctx, KCTX_SCHEDULED) &&
kbase_js_ctx_pullable(kctx, js, false))
timer_sync =
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, js);
spin_unlock_irqrestore(&kbdev->hwaccess_lock,
flags);
if (timer_sync)
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&js_devdata->runpool_mutex);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
/* Take lock before accessing list again */
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
#else
bool timer_sync = false;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
list_for_each_entry_safe(kctx, n,
&kbdev->js_data.ctx_list_unpullable[js][prio],
jctx.sched_info.ctx.ctx_list_entry[js]) {
if (!kbase_ctx_flag(kctx, KCTX_SCHEDULED) &&
kbase_js_ctx_pullable(kctx, js, false))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, js);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (timer_sync) {
mutex_lock(&js_devdata->runpool_mutex);
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&js_devdata->runpool_mutex);
}
#endif
}
}
mutex_unlock(&js_devdata->queue_mutex);
/* Restart atom processing */
kbase_js_sched_all(kbdev);
/* JS Resume complete */
}
bool kbase_js_is_atom_valid(struct kbase_device *kbdev,
struct kbase_jd_atom *katom)
{
if ((katom->core_req & BASE_JD_REQ_FS) &&
(katom->core_req & (BASE_JD_REQ_CS | BASE_JD_REQ_ONLY_COMPUTE |
BASE_JD_REQ_T)))
return false;
if ((katom->core_req & BASE_JD_REQ_JOB_SLOT) &&
(katom->jobslot >= BASE_JM_MAX_NR_SLOTS))
return false;
return true;
}
static int kbase_js_get_slot(struct kbase_device *kbdev,
struct kbase_jd_atom *katom)
{
if (katom->core_req & BASE_JD_REQ_JOB_SLOT)
return katom->jobslot;
if (katom->core_req & BASE_JD_REQ_FS)
return 0;
if (katom->core_req & BASE_JD_REQ_ONLY_COMPUTE) {
if (katom->device_nr == 1 &&
kbdev->gpu_props.num_core_groups == 2)
return 2;
}
return 1;
}
bool kbase_js_dep_resolved_submit(struct kbase_context *kctx,
struct kbase_jd_atom *katom)
{
bool enqueue_required, add_required = true;
katom->slot_nr = kbase_js_get_slot(kctx->kbdev, katom);
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
lockdep_assert_held(&kctx->jctx.lock);
/* If slot will transition from unpullable to pullable then add to
* pullable list
*/
if (jsctx_rb_none_to_pull(kctx, katom->slot_nr)) {
enqueue_required = true;
} else {
enqueue_required = false;
}
if ((katom->atom_flags & KBASE_KATOM_FLAG_X_DEP_BLOCKED) ||
(katom->pre_dep && (katom->pre_dep->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST))) {
int prio = katom->sched_priority;
int js = katom->slot_nr;
struct jsctx_queue *queue = &kctx->jsctx_queue[prio][js];
dev_dbg(kctx->kbdev->dev, "Add atom %pK to X_DEP list (s:%d)\n",
(void *)katom, js);
list_add_tail(&katom->queue, &queue->x_dep_head);
katom->atom_flags |= KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST;
if (kbase_js_atom_blocked_on_x_dep(katom)) {
enqueue_required = false;
add_required = false;
}
} else {
dev_dbg(kctx->kbdev->dev, "Atom %pK not added to X_DEP list\n",
(void *)katom);
}
if (add_required) {
/* Check if there are lower priority jobs to soft stop */
kbase_job_slot_ctx_priority_check_locked(kctx, katom);
/* Add atom to ring buffer. */
jsctx_tree_add(kctx, katom);
katom->atom_flags |= KBASE_KATOM_FLAG_JSCTX_IN_TREE;
}
dev_dbg(kctx->kbdev->dev,
"Enqueue of kctx %pK is %srequired to submit atom %pK\n",
kctx, enqueue_required ? "" : "not ", katom);
return enqueue_required;
}
/**
* kbase_js_move_to_tree - Move atom (and any dependent atoms) to the
* runnable_tree, ready for execution
* @katom: Atom to submit
*
* It is assumed that @katom does not have KBASE_KATOM_FLAG_X_DEP_BLOCKED set,
* but is still present in the x_dep list. If @katom has a same-slot dependent
* atom then that atom (and any dependents) will also be moved.
*/
static void kbase_js_move_to_tree(struct kbase_jd_atom *katom)
{
struct kbase_context *const kctx = katom->kctx;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
while (katom) {
WARN_ON(!(katom->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST));
if (!kbase_js_atom_blocked_on_x_dep(katom)) {
dev_dbg(kctx->kbdev->dev,
"Del atom %pK from X_DEP list in js_move_to_tree\n",
(void *)katom);
list_del(&katom->queue);
katom->atom_flags &=
~KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST;
/* For incremental rendering, an end-of-renderpass atom
* may have had its dependency on start-of-renderpass
* ignored and may therefore already be in the tree.
*/
if (!(katom->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_TREE)) {
jsctx_tree_add(kctx, katom);
katom->atom_flags |=
KBASE_KATOM_FLAG_JSCTX_IN_TREE;
}
} else {
dev_dbg(kctx->kbdev->dev,
"Atom %pK blocked on x-dep in js_move_to_tree\n",
(void *)katom);
break;
}
katom = katom->post_dep;
}
}
/**
* kbase_js_evict_deps - Evict dependencies of a failed atom.
* @kctx: Context pointer
* @katom: Pointer to the atom that has failed.
* @js: The job slot the katom was run on.
* @prio: Priority of the katom.
*
* Remove all post dependencies of an atom from the context ringbuffers.
*
* The original atom's event_code will be propogated to all dependent atoms.
*
* Context: Caller must hold the HW access lock
*/
static void kbase_js_evict_deps(struct kbase_context *kctx,
struct kbase_jd_atom *katom, int js, int prio)
{
struct kbase_jd_atom *x_dep = katom->x_post_dep;
struct kbase_jd_atom *next_katom = katom->post_dep;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
if (next_katom) {
KBASE_DEBUG_ASSERT(next_katom->status !=
KBASE_JD_ATOM_STATE_HW_COMPLETED);
next_katom->will_fail_event_code = katom->event_code;
}
/* Has cross slot depenency. */
if (x_dep && (x_dep->atom_flags & (KBASE_KATOM_FLAG_JSCTX_IN_TREE |
KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST))) {
/* Remove dependency.*/
x_dep->atom_flags &= ~KBASE_KATOM_FLAG_X_DEP_BLOCKED;
dev_dbg(kctx->kbdev->dev, "Cleared X_DEP flag on atom %pK\n",
(void *)x_dep);
/* Fail if it had a data dependency. */
if (x_dep->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) {
x_dep->will_fail_event_code = katom->event_code;
}
if (x_dep->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST)
kbase_js_move_to_tree(x_dep);
}
}
struct kbase_jd_atom *kbase_js_pull(struct kbase_context *kctx, int js)
{
struct kbase_jd_atom *katom;
struct kbasep_js_device_data *js_devdata;
struct kbase_device *kbdev;
int pulled;
KBASE_DEBUG_ASSERT(kctx);
kbdev = kctx->kbdev;
dev_dbg(kbdev->dev, "JS: pulling an atom from kctx %pK (s:%d)\n",
(void *)kctx, js);
js_devdata = &kbdev->js_data;
lockdep_assert_held(&kbdev->hwaccess_lock);
if (!kbasep_js_is_submit_allowed(js_devdata, kctx)) {
dev_dbg(kbdev->dev, "JS: No submit allowed for kctx %pK\n",
(void *)kctx);
return NULL;
}
#ifdef CONFIG_MALI_ARBITER_SUPPORT
if (kbase_pm_is_suspending(kbdev) || kbase_pm_is_gpu_lost(kbdev))
#else
if (kbase_pm_is_suspending(kbdev))
#endif
return NULL;
katom = jsctx_rb_peek(kctx, js);
if (!katom) {
dev_dbg(kbdev->dev, "JS: No pullable atom in kctx %pK (s:%d)\n",
(void *)kctx, js);
return NULL;
}
if (kctx->blocked_js[js][katom->sched_priority]) {
dev_dbg(kbdev->dev,
"JS: kctx %pK is blocked from submitting atoms at priority %d (s:%d)\n",
(void *)kctx, katom->sched_priority, js);
return NULL;
}
if (atomic_read(&katom->blocked)) {
dev_dbg(kbdev->dev, "JS: Atom %pK is blocked in js_pull\n",
(void *)katom);
return NULL;
}
/* Due to ordering restrictions when unpulling atoms on failure, we do
* not allow multiple runs of fail-dep atoms from the same context to be
* present on the same slot
*/
if (katom->pre_dep && atomic_read(&kctx->atoms_pulled_slot[js])) {
struct kbase_jd_atom *prev_atom =
kbase_backend_inspect_tail(kbdev, js);
if (prev_atom && prev_atom->kctx != kctx)
return NULL;
}
if (kbase_js_atom_blocked_on_x_dep(katom)) {
if (katom->x_pre_dep->gpu_rb_state ==
KBASE_ATOM_GPU_RB_NOT_IN_SLOT_RB ||
katom->x_pre_dep->will_fail_event_code) {
dev_dbg(kbdev->dev,
"JS: X pre-dep %pK is not present in slot FIFO or will fail\n",
(void *)katom->x_pre_dep);
return NULL;
}
if ((katom->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER) &&
kbase_backend_nr_atoms_on_slot(kbdev, js)) {
dev_dbg(kbdev->dev,
"JS: Atom %pK has cross-slot fail dependency and atoms on slot (s:%d)\n",
(void *)katom, js);
return NULL;
}
}
kbase_ctx_flag_set(kctx, KCTX_PULLED);
kbase_ctx_flag_set(kctx, (KCTX_PULLED_SINCE_ACTIVE_JS0 << js));
pulled = atomic_inc_return(&kctx->atoms_pulled);
if (pulled == 1 && !kctx->slots_pullable) {
WARN_ON(kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_set(kctx, KCTX_RUNNABLE_REF);
atomic_inc(&kbdev->js_data.nr_contexts_runnable);
}
atomic_inc(&kctx->atoms_pulled_slot[katom->slot_nr]);
kctx->atoms_pulled_slot_pri[katom->slot_nr][katom->sched_priority]++;
jsctx_rb_pull(kctx, katom);
kbase_ctx_sched_retain_ctx_refcount(kctx);
katom->atom_flags |= KBASE_KATOM_FLAG_HOLDING_CTX_REF;
katom->ticks = 0;
dev_dbg(kbdev->dev, "JS: successfully pulled atom %pK from kctx %pK (s:%d)\n",
(void *)katom, (void *)kctx, js);
return katom;
}
/**
* js_return_of_start_rp() - Handle soft-stop of an atom that starts a
* renderpass
* @start_katom: Pointer to the start-of-renderpass atom that was soft-stopped
*
* This function is called to switch to incremental rendering if the tiler job
* chain at the start of a renderpass has used too much memory. It prevents the
* tiler job being pulled for execution in the job scheduler again until the
* next phase of incremental rendering is complete.
*
* If the end-of-renderpass atom is already in the job scheduler (because a
* previous attempt at tiling used too much memory during the same renderpass)
* then it is unblocked; otherwise, it is run by handing it to the scheduler.
*/
static void js_return_of_start_rp(struct kbase_jd_atom *const start_katom)
{
struct kbase_context *const kctx = start_katom->kctx;
struct kbase_device *const kbdev = kctx->kbdev;
struct kbase_jd_renderpass *rp;
struct kbase_jd_atom *end_katom;
unsigned long flags;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(start_katom->core_req & BASE_JD_REQ_START_RENDERPASS)))
return;
compiletime_assert((1ull << (sizeof(start_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[start_katom->renderpass_id];
if (WARN_ON(rp->start_katom != start_katom))
return;
dev_dbg(kctx->kbdev->dev,
"JS return start atom %pK in state %d of RP %d\n",
(void *)start_katom, (int)rp->state,
start_katom->renderpass_id);
if (WARN_ON(rp->state == KBASE_JD_RP_COMPLETE))
return;
/* The tiler job might have been soft-stopped for some reason other
* than running out of memory.
*/
if (rp->state == KBASE_JD_RP_START || rp->state == KBASE_JD_RP_RETRY) {
dev_dbg(kctx->kbdev->dev,
"JS return isn't OOM in state %d of RP %d\n",
(int)rp->state, start_katom->renderpass_id);
return;
}
dev_dbg(kctx->kbdev->dev,
"JS return confirm OOM in state %d of RP %d\n",
(int)rp->state, start_katom->renderpass_id);
if (WARN_ON(rp->state != KBASE_JD_RP_PEND_OOM &&
rp->state != KBASE_JD_RP_RETRY_PEND_OOM))
return;
/* Prevent the tiler job being pulled for execution in the
* job scheduler again.
*/
dev_dbg(kbdev->dev, "Blocking start atom %pK\n",
(void *)start_katom);
atomic_inc(&start_katom->blocked);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
rp->state = (rp->state == KBASE_JD_RP_PEND_OOM) ?
KBASE_JD_RP_OOM : KBASE_JD_RP_RETRY_OOM;
/* Was the fragment job chain submitted to kbase yet? */
end_katom = rp->end_katom;
if (end_katom) {
dev_dbg(kctx->kbdev->dev, "JS return add end atom %pK\n",
(void *)end_katom);
if (rp->state == KBASE_JD_RP_RETRY_OOM) {
/* Allow the end of the renderpass to be pulled for
* execution again to continue incremental rendering.
*/
dev_dbg(kbdev->dev, "Unblocking end atom %pK\n",
(void *)end_katom);
atomic_dec(&end_katom->blocked);
WARN_ON(!(end_katom->atom_flags &
KBASE_KATOM_FLAG_JSCTX_IN_TREE));
WARN_ON(end_katom->status != KBASE_JD_ATOM_STATE_IN_JS);
kbase_js_ctx_list_add_pullable_nolock(kbdev, kctx,
end_katom->slot_nr);
/* Expect the fragment job chain to be scheduled without
* further action because this function is called when
* returning an atom to the job scheduler ringbuffer.
*/
end_katom = NULL;
} else {
WARN_ON(end_katom->status !=
KBASE_JD_ATOM_STATE_QUEUED &&
end_katom->status != KBASE_JD_ATOM_STATE_IN_JS);
}
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (end_katom)
kbase_jd_dep_clear_locked(end_katom);
}
/**
* js_return_of_end_rp() - Handle completion of an atom that ends a renderpass
* @end_katom: Pointer to the end-of-renderpass atom that was completed
*
* This function is called to continue incremental rendering if the tiler job
* chain at the start of a renderpass used too much memory. It resets the
* mechanism for detecting excessive memory usage then allows the soft-stopped
* tiler job chain to be pulled for execution again.
*
* The start-of-renderpass atom must already been submitted to kbase.
*/
static void js_return_of_end_rp(struct kbase_jd_atom *const end_katom)
{
struct kbase_context *const kctx = end_katom->kctx;
struct kbase_device *const kbdev = kctx->kbdev;
struct kbase_jd_renderpass *rp;
struct kbase_jd_atom *start_katom;
unsigned long flags;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(end_katom->core_req & BASE_JD_REQ_END_RENDERPASS)))
return;
compiletime_assert((1ull << (sizeof(end_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[end_katom->renderpass_id];
if (WARN_ON(rp->end_katom != end_katom))
return;
dev_dbg(kctx->kbdev->dev,
"JS return end atom %pK in state %d of RP %d\n",
(void *)end_katom, (int)rp->state, end_katom->renderpass_id);
if (WARN_ON(rp->state != KBASE_JD_RP_OOM &&
rp->state != KBASE_JD_RP_RETRY_OOM))
return;
/* Reduce the number of mapped pages in the memory regions that
* triggered out-of-memory last time so that we can detect excessive
* memory usage again.
*/
kbase_gpu_vm_lock(kctx);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
while (!list_empty(&rp->oom_reg_list)) {
struct kbase_va_region *reg =
list_first_entry(&rp->oom_reg_list,
struct kbase_va_region, link);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
dev_dbg(kbdev->dev,
"Reset backing to %zu pages for region %pK\n",
reg->threshold_pages, (void *)reg);
if (!WARN_ON(reg->flags & KBASE_REG_VA_FREED))
kbase_mem_shrink(kctx, reg, reg->threshold_pages);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
dev_dbg(kbdev->dev, "Deleting region %pK from list\n",
(void *)reg);
list_del_init(&reg->link);
kbase_va_region_alloc_put(kctx, reg);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
kbase_gpu_vm_unlock(kctx);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
rp->state = KBASE_JD_RP_RETRY;
dev_dbg(kbdev->dev, "Changed state to %d for retry\n", rp->state);
/* Allow the start of the renderpass to be pulled for execution again
* to begin/continue incremental rendering.
*/
start_katom = rp->start_katom;
if (!WARN_ON(!start_katom)) {
dev_dbg(kbdev->dev, "Unblocking start atom %pK\n",
(void *)start_katom);
atomic_dec(&start_katom->blocked);
(void)kbase_js_ctx_list_add_pullable_head_nolock(kbdev, kctx,
start_katom->slot_nr);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
}
static void js_return_worker(struct work_struct *data)
{
struct kbase_jd_atom *katom = container_of(data, struct kbase_jd_atom,
work);
struct kbase_context *kctx = katom->kctx;
struct kbase_device *kbdev = kctx->kbdev;
struct kbasep_js_device_data *js_devdata = &kbdev->js_data;
struct kbasep_js_kctx_info *js_kctx_info = &kctx->jctx.sched_info;
struct kbasep_js_atom_retained_state retained_state;
int js = katom->slot_nr;
int prio = katom->sched_priority;
bool timer_sync = false;
bool context_idle = false;
unsigned long flags;
base_jd_core_req core_req = katom->core_req;
dev_dbg(kbdev->dev, "%s for atom %pK with event code 0x%x\n",
__func__, (void *)katom, katom->event_code);
if (katom->event_code != BASE_JD_EVENT_END_RP_DONE)
KBASE_TLSTREAM_TL_EVENT_ATOM_SOFTSTOP_EX(kbdev, katom);
kbase_backend_complete_wq(kbdev, katom);
kbasep_js_atom_retained_state_copy(&retained_state, katom);
mutex_lock(&js_devdata->queue_mutex);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
atomic_dec(&kctx->atoms_pulled);
atomic_dec(&kctx->atoms_pulled_slot[js]);
if (katom->event_code != BASE_JD_EVENT_END_RP_DONE)
atomic_dec(&katom->blocked);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
kctx->atoms_pulled_slot_pri[js][katom->sched_priority]--;
if (!atomic_read(&kctx->atoms_pulled_slot[js]) &&
jsctx_rb_none_to_pull(kctx, js))
timer_sync |= kbase_js_ctx_list_remove_nolock(kbdev, kctx, js);
/* If this slot has been blocked due to soft-stopped atoms, and all
* atoms have now been processed, then unblock the slot
*/
if (!kctx->atoms_pulled_slot_pri[js][prio] &&
kctx->blocked_js[js][prio]) {
kctx->blocked_js[js][prio] = false;
/* Only mark the slot as pullable if the context is not idle -
* that case is handled below
*/
if (atomic_read(&kctx->atoms_pulled) &&
kbase_js_ctx_pullable(kctx, js, true))
timer_sync |= kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, js);
}
if (!atomic_read(&kctx->atoms_pulled)) {
dev_dbg(kbdev->dev,
"No atoms currently pulled from context %pK\n",
(void *)kctx);
if (!kctx->slots_pullable) {
dev_dbg(kbdev->dev,
"Context %pK %s counted as runnable\n",
(void *)kctx,
kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF) ?
"is" : "isn't");
WARN_ON(!kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_clear(kctx, KCTX_RUNNABLE_REF);
atomic_dec(&kbdev->js_data.nr_contexts_runnable);
timer_sync = true;
}
if (kctx->as_nr != KBASEP_AS_NR_INVALID &&
!kbase_ctx_flag(kctx, KCTX_DYING)) {
int num_slots = kbdev->gpu_props.num_job_slots;
int slot;
if (!kbasep_js_is_submit_allowed(js_devdata, kctx))
kbasep_js_set_submit_allowed(js_devdata, kctx);
for (slot = 0; slot < num_slots; slot++) {
if (kbase_js_ctx_pullable(kctx, slot, true))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, slot);
}
}
kbase_jm_idle_ctx(kbdev, kctx);
context_idle = true;
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (context_idle) {
dev_dbg(kbdev->dev,
"Context %pK %s counted as active\n",
(void *)kctx,
kbase_ctx_flag(kctx, KCTX_ACTIVE) ?
"is" : "isn't");
WARN_ON(!kbase_ctx_flag(kctx, KCTX_ACTIVE));
kbase_ctx_flag_clear(kctx, KCTX_ACTIVE);
kbase_pm_context_idle(kbdev);
}
if (timer_sync)
kbase_js_sync_timers(kbdev);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
if (katom->core_req & BASE_JD_REQ_START_RENDERPASS) {
mutex_lock(&kctx->jctx.lock);
js_return_of_start_rp(katom);
mutex_unlock(&kctx->jctx.lock);
} else if (katom->event_code == BASE_JD_EVENT_END_RP_DONE) {
mutex_lock(&kctx->jctx.lock);
js_return_of_end_rp(katom);
mutex_unlock(&kctx->jctx.lock);
}
katom->atom_flags &= ~KBASE_KATOM_FLAG_HOLDING_CTX_REF;
dev_dbg(kbdev->dev, "JS: retained state %s finished",
kbasep_js_has_atom_finished(&retained_state) ?
"has" : "hasn't");
WARN_ON(kbasep_js_has_atom_finished(&retained_state));
kbasep_js_runpool_release_ctx_and_katom_retained_state(kbdev, kctx,
&retained_state);
kbase_js_sched_all(kbdev);
kbase_backend_complete_wq_post_sched(kbdev, core_req);
dev_dbg(kbdev->dev, "Leaving %s for atom %pK\n",
__func__, (void *)katom);
}
void kbase_js_unpull(struct kbase_context *kctx, struct kbase_jd_atom *katom)
{
dev_dbg(kctx->kbdev->dev, "Unpulling atom %pK in kctx %pK\n",
(void *)katom, (void *)kctx);
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
jsctx_rb_unpull(kctx, katom);
WARN_ON(work_pending(&katom->work));
/* Block re-submission until workqueue has run */
atomic_inc(&katom->blocked);
kbase_job_check_leave_disjoint(kctx->kbdev, katom);
INIT_WORK(&katom->work, js_return_worker);
queue_work(kctx->jctx.job_done_wq, &katom->work);
}
/**
* js_complete_start_rp() - Handle completion of atom that starts a renderpass
* @kctx: Context pointer
* @start_katom: Pointer to the atom that completed
*
* Put any references to virtual memory regions that might have been added by
* kbase_job_slot_softstop_start_rp() because the tiler job chain completed
* despite any pending soft-stop request.
*
* If the atom that just completed was soft-stopped during a previous attempt to
* run it then there should be a blocked end-of-renderpass atom waiting for it,
* which we must unblock to process the output of the tiler job chain.
*
* Return: true if caller should call kbase_backend_ctx_count_changed()
*/
static bool js_complete_start_rp(struct kbase_context *kctx,
struct kbase_jd_atom *const start_katom)
{
struct kbase_device *const kbdev = kctx->kbdev;
struct kbase_jd_renderpass *rp;
bool timer_sync = false;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(start_katom->core_req & BASE_JD_REQ_START_RENDERPASS)))
return false;
compiletime_assert((1ull << (sizeof(start_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[start_katom->renderpass_id];
if (WARN_ON(rp->start_katom != start_katom))
return false;
dev_dbg(kctx->kbdev->dev,
"Start atom %pK is done in state %d of RP %d\n",
(void *)start_katom, (int)rp->state,
start_katom->renderpass_id);
if (WARN_ON(rp->state == KBASE_JD_RP_COMPLETE))
return false;
if (rp->state == KBASE_JD_RP_PEND_OOM ||
rp->state == KBASE_JD_RP_RETRY_PEND_OOM) {
unsigned long flags;
dev_dbg(kctx->kbdev->dev,
"Start atom %pK completed before soft-stop\n",
(void *)start_katom);
kbase_gpu_vm_lock(kctx);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
while (!list_empty(&rp->oom_reg_list)) {
struct kbase_va_region *reg =
list_first_entry(&rp->oom_reg_list,
struct kbase_va_region, link);
WARN_ON(reg->flags & KBASE_REG_VA_FREED);
dev_dbg(kctx->kbdev->dev, "Deleting region %pK from list\n",
(void *)reg);
list_del_init(&reg->link);
kbase_va_region_alloc_put(kctx, reg);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
kbase_gpu_vm_unlock(kctx);
} else {
dev_dbg(kctx->kbdev->dev,
"Start atom %pK did not exceed memory threshold\n",
(void *)start_katom);
WARN_ON(rp->state != KBASE_JD_RP_START &&
rp->state != KBASE_JD_RP_RETRY);
}
if (rp->state == KBASE_JD_RP_RETRY ||
rp->state == KBASE_JD_RP_RETRY_PEND_OOM) {
struct kbase_jd_atom *const end_katom = rp->end_katom;
if (!WARN_ON(!end_katom)) {
unsigned long flags;
/* Allow the end of the renderpass to be pulled for
* execution again to continue incremental rendering.
*/
dev_dbg(kbdev->dev, "Unblocking end atom %pK!\n",
(void *)end_katom);
atomic_dec(&end_katom->blocked);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
timer_sync = kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, end_katom->slot_nr);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
}
}
return timer_sync;
}
/**
* js_complete_end_rp() - Handle final completion of atom that ends a renderpass
* @kctx: Context pointer
* @end_katom: Pointer to the atom that completed for the last time
*
* This function must only be called if the renderpass actually completed
* without the tiler job chain at the start using too much memory; otherwise
* completion of the end-of-renderpass atom is handled similarly to a soft-stop.
*/
static void js_complete_end_rp(struct kbase_context *kctx,
struct kbase_jd_atom *const end_katom)
{
struct kbase_device *const kbdev = kctx->kbdev;
unsigned long flags;
struct kbase_jd_renderpass *rp;
lockdep_assert_held(&kctx->jctx.lock);
if (WARN_ON(!(end_katom->core_req & BASE_JD_REQ_END_RENDERPASS)))
return;
compiletime_assert((1ull << (sizeof(end_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[end_katom->renderpass_id];
if (WARN_ON(rp->end_katom != end_katom))
return;
dev_dbg(kbdev->dev, "End atom %pK is done in state %d of RP %d\n",
(void *)end_katom, (int)rp->state, end_katom->renderpass_id);
if (WARN_ON(rp->state == KBASE_JD_RP_COMPLETE) ||
WARN_ON(rp->state == KBASE_JD_RP_OOM) ||
WARN_ON(rp->state == KBASE_JD_RP_RETRY_OOM))
return;
/* Rendering completed without running out of memory.
*/
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
WARN_ON(!list_empty(&rp->oom_reg_list));
rp->state = KBASE_JD_RP_COMPLETE;
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
dev_dbg(kbdev->dev, "Renderpass %d is complete\n",
end_katom->renderpass_id);
}
bool kbase_js_complete_atom_wq(struct kbase_context *kctx,
struct kbase_jd_atom *katom)
{
struct kbasep_js_kctx_info *js_kctx_info;
struct kbasep_js_device_data *js_devdata;
struct kbase_device *kbdev;
unsigned long flags;
bool timer_sync = false;
int atom_slot;
bool context_idle = false;
int prio = katom->sched_priority;
kbdev = kctx->kbdev;
atom_slot = katom->slot_nr;
dev_dbg(kbdev->dev, "%s for atom %pK (s:%d)\n",
__func__, (void *)katom, atom_slot);
/* Update the incremental rendering state machine.
*/
if (katom->core_req & BASE_JD_REQ_START_RENDERPASS)
timer_sync |= js_complete_start_rp(kctx, katom);
else if (katom->core_req & BASE_JD_REQ_END_RENDERPASS)
js_complete_end_rp(kctx, katom);
js_kctx_info = &kctx->jctx.sched_info;
js_devdata = &kbdev->js_data;
lockdep_assert_held(&js_kctx_info->ctx.jsctx_mutex);
mutex_lock(&js_devdata->runpool_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (katom->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_TREE) {
dev_dbg(kbdev->dev, "Atom %pK is in runnable_tree\n",
(void *)katom);
context_idle = !atomic_dec_return(&kctx->atoms_pulled);
atomic_dec(&kctx->atoms_pulled_slot[atom_slot]);
kctx->atoms_pulled_slot_pri[atom_slot][prio]--;
if (!atomic_read(&kctx->atoms_pulled) &&
!kctx->slots_pullable) {
WARN_ON(!kbase_ctx_flag(kctx, KCTX_RUNNABLE_REF));
kbase_ctx_flag_clear(kctx, KCTX_RUNNABLE_REF);
atomic_dec(&kbdev->js_data.nr_contexts_runnable);
timer_sync = true;
}
/* If this slot has been blocked due to soft-stopped atoms, and
* all atoms have now been processed, then unblock the slot
*/
if (!kctx->atoms_pulled_slot_pri[atom_slot][prio]
&& kctx->blocked_js[atom_slot][prio]) {
dev_dbg(kbdev->dev,
"kctx %pK is no longer blocked from submitting on slot %d at priority %d\n",
(void *)kctx, atom_slot, prio);
kctx->blocked_js[atom_slot][prio] = false;
if (kbase_js_ctx_pullable(kctx, atom_slot, true))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, atom_slot);
}
}
WARN_ON(!(katom->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_TREE));
if (!atomic_read(&kctx->atoms_pulled_slot[atom_slot]) &&
jsctx_rb_none_to_pull(kctx, atom_slot)) {
if (!list_empty(
&kctx->jctx.sched_info.ctx.ctx_list_entry[atom_slot]))
timer_sync |= kbase_js_ctx_list_remove_nolock(
kctx->kbdev, kctx, atom_slot);
}
/*
* If submission is disabled on this context (most likely due to an
* atom failure) and there are now no atoms left in the system then
* re-enable submission so that context can be scheduled again.
*/
if (!kbasep_js_is_submit_allowed(js_devdata, kctx) &&
!atomic_read(&kctx->atoms_pulled) &&
!kbase_ctx_flag(kctx, KCTX_DYING)) {
int js;
kbasep_js_set_submit_allowed(js_devdata, kctx);
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
if (kbase_js_ctx_pullable(kctx, js, true))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, js);
}
} else if (katom->x_post_dep &&
kbasep_js_is_submit_allowed(js_devdata, kctx)) {
int js;
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
if (kbase_js_ctx_pullable(kctx, js, true))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kbdev, kctx, js);
}
}
/* Mark context as inactive. The pm reference will be dropped later in
* jd_done_worker().
*/
if (context_idle) {
dev_dbg(kbdev->dev, "kctx %pK is no longer active\n",
(void *)kctx);
kbase_ctx_flag_clear(kctx, KCTX_ACTIVE);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
if (timer_sync)
kbase_backend_ctx_count_changed(kbdev);
mutex_unlock(&js_devdata->runpool_mutex);
dev_dbg(kbdev->dev, "Leaving %s\n", __func__);
return context_idle;
}
/**
* js_end_rp_is_complete() - Check whether an atom that ends a renderpass has
* completed for the last time.
*
* @end_katom: Pointer to the atom that completed on the hardware.
*
* An atom that ends a renderpass may be run on the hardware several times
* before notifying userspace or allowing dependent atoms to be executed.
*
* This function is used to decide whether or not to allow end-of-renderpass
* atom completion. It only returns false if the atom at the start of the
* renderpass was soft-stopped because it used too much memory during the most
* recent attempt at tiling.
*
* Return: True if the atom completed for the last time.
*/
static bool js_end_rp_is_complete(struct kbase_jd_atom *const end_katom)
{
struct kbase_context *const kctx = end_katom->kctx;
struct kbase_device *const kbdev = kctx->kbdev;
struct kbase_jd_renderpass *rp;
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
if (WARN_ON(!(end_katom->core_req & BASE_JD_REQ_END_RENDERPASS)))
return true;
compiletime_assert((1ull << (sizeof(end_katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[end_katom->renderpass_id];
if (WARN_ON(rp->end_katom != end_katom))
return true;
dev_dbg(kbdev->dev,
"JS complete end atom %pK in state %d of RP %d\n",
(void *)end_katom, (int)rp->state,
end_katom->renderpass_id);
if (WARN_ON(rp->state == KBASE_JD_RP_COMPLETE))
return true;
/* Failure of end-of-renderpass atoms must not return to the
* start of the renderpass.
*/
if (end_katom->event_code != BASE_JD_EVENT_DONE)
return true;
if (rp->state != KBASE_JD_RP_OOM &&
rp->state != KBASE_JD_RP_RETRY_OOM)
return true;
dev_dbg(kbdev->dev, "Suppressing end atom completion\n");
return false;
}
struct kbase_jd_atom *kbase_js_complete_atom(struct kbase_jd_atom *katom,
ktime_t *end_timestamp)
{
struct kbase_device *kbdev;
struct kbase_context *kctx = katom->kctx;
struct kbase_jd_atom *x_dep = katom->x_post_dep;
kbdev = kctx->kbdev;
dev_dbg(kbdev->dev, "Atom %pK complete in kctx %pK (post-dep %pK)\n",
(void *)katom, (void *)kctx, (void *)x_dep);
lockdep_assert_held(&kctx->kbdev->hwaccess_lock);
if ((katom->core_req & BASE_JD_REQ_END_RENDERPASS) &&
!js_end_rp_is_complete(katom)) {
katom->event_code = BASE_JD_EVENT_END_RP_DONE;
kbase_js_unpull(kctx, katom);
return NULL;
}
if (katom->will_fail_event_code)
katom->event_code = katom->will_fail_event_code;
katom->status = KBASE_JD_ATOM_STATE_HW_COMPLETED;
dev_dbg(kbdev->dev, "Atom %pK status to HW completed\n", (void *)katom);
if (katom->event_code != BASE_JD_EVENT_DONE) {
kbase_js_evict_deps(kctx, katom, katom->slot_nr,
katom->sched_priority);
}
KBASE_TLSTREAM_AUX_EVENT_JOB_SLOT(kbdev, NULL,
katom->slot_nr, 0, TL_JS_EVENT_STOP);
trace_sysgraph_gpu(SGR_COMPLETE, kctx->id,
kbase_jd_atom_id(katom->kctx, katom), katom->slot_nr);
kbase_jd_done(katom, katom->slot_nr, end_timestamp, 0);
/* Unblock cross dependency if present */
if (x_dep && (katom->event_code == BASE_JD_EVENT_DONE ||
!(x_dep->atom_flags & KBASE_KATOM_FLAG_FAIL_BLOCKER)) &&
(x_dep->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_X_DEP_LIST)) {
bool was_pullable = kbase_js_ctx_pullable(kctx, x_dep->slot_nr,
false);
x_dep->atom_flags &= ~KBASE_KATOM_FLAG_X_DEP_BLOCKED;
dev_dbg(kbdev->dev, "Cleared X_DEP flag on atom %pK\n",
(void *)x_dep);
kbase_js_move_to_tree(x_dep);
if (!was_pullable && kbase_js_ctx_pullable(kctx, x_dep->slot_nr,
false))
kbase_js_ctx_list_add_pullable_nolock(kbdev, kctx,
x_dep->slot_nr);
if (x_dep->atom_flags & KBASE_KATOM_FLAG_JSCTX_IN_TREE) {
dev_dbg(kbdev->dev, "Atom %pK is in runnable tree\n",
(void *)x_dep);
return x_dep;
}
} else {
dev_dbg(kbdev->dev,
"No cross-slot dep to unblock for atom %pK\n",
(void *)katom);
}
return NULL;
}
/**
* kbase_js_atom_blocked_on_x_dep - Decide whether to ignore a cross-slot
* dependency
* @katom: Pointer to an atom in the slot ringbuffer
*
* A cross-slot dependency is ignored if necessary to unblock incremental
* rendering. If the atom at the start of a renderpass used too much memory
* and was soft-stopped then the atom at the end of a renderpass is submitted
* to hardware regardless of its dependency on the start-of-renderpass atom.
* This can happen multiple times for the same pair of atoms.
*
* Return: true to block the atom or false to allow it to be submitted to
* hardware
*/
bool kbase_js_atom_blocked_on_x_dep(struct kbase_jd_atom *const katom)
{
struct kbase_context *const kctx = katom->kctx;
struct kbase_device *kbdev = kctx->kbdev;
struct kbase_jd_renderpass *rp;
lockdep_assert_held(&kbdev->hwaccess_lock);
if (!(katom->atom_flags &
KBASE_KATOM_FLAG_X_DEP_BLOCKED)) {
dev_dbg(kbdev->dev, "Atom %pK is not blocked on a cross-slot dependency",
(void *)katom);
return false;
}
if (!(katom->core_req & BASE_JD_REQ_END_RENDERPASS)) {
dev_dbg(kbdev->dev, "Atom %pK is blocked on a cross-slot dependency",
(void *)katom);
return true;
}
compiletime_assert((1ull << (sizeof(katom->renderpass_id) * 8)) <=
ARRAY_SIZE(kctx->jctx.renderpasses),
"Should check invalid access to renderpasses");
rp = &kctx->jctx.renderpasses[katom->renderpass_id];
/* We can read a subset of renderpass state without holding
* higher-level locks (but not end_katom, for example).
*/
WARN_ON(rp->state == KBASE_JD_RP_COMPLETE);
dev_dbg(kbdev->dev, "End atom has cross-slot dep in state %d\n",
(int)rp->state);
if (rp->state != KBASE_JD_RP_OOM && rp->state != KBASE_JD_RP_RETRY_OOM)
return true;
/* Tiler ran out of memory so allow the fragment job chain to run
* if it only depends on the tiler job chain.
*/
if (katom->x_pre_dep != rp->start_katom) {
dev_dbg(kbdev->dev, "Dependency is on %pK not start atom %pK\n",
(void *)katom->x_pre_dep, (void *)rp->start_katom);
return true;
}
dev_dbg(kbdev->dev, "Ignoring cross-slot dep on atom %pK\n",
(void *)katom->x_pre_dep);
return false;
}
void kbase_js_sched(struct kbase_device *kbdev, int js_mask)
{
struct kbasep_js_device_data *js_devdata;
struct kbase_context *last_active[BASE_JM_MAX_NR_SLOTS];
bool timer_sync = false;
bool ctx_waiting[BASE_JM_MAX_NR_SLOTS];
int js;
dev_dbg(kbdev->dev, "%s kbdev %pK mask 0x%x\n",
__func__, (void *)kbdev, (unsigned int)js_mask);
js_devdata = &kbdev->js_data;
down(&js_devdata->schedule_sem);
mutex_lock(&js_devdata->queue_mutex);
for (js = 0; js < BASE_JM_MAX_NR_SLOTS; js++) {
last_active[js] = kbdev->hwaccess.active_kctx[js];
ctx_waiting[js] = false;
}
while (js_mask) {
js = ffs(js_mask) - 1;
while (1) {
struct kbase_context *kctx;
unsigned long flags;
bool context_idle = false;
kctx = kbase_js_ctx_list_pop_head(kbdev, js);
if (!kctx) {
js_mask &= ~(1 << js);
dev_dbg(kbdev->dev,
"No kctx on pullable list (s:%d)\n",
js);
break;
}
if (!kbase_ctx_flag(kctx, KCTX_ACTIVE)) {
context_idle = true;
dev_dbg(kbdev->dev,
"kctx %pK is not active (s:%d)\n",
(void *)kctx, js);
if (kbase_pm_context_active_handle_suspend(
kbdev,
KBASE_PM_SUSPEND_HANDLER_DONT_INCREASE)) {
dev_dbg(kbdev->dev,
"Suspend pending (s:%d)\n", js);
/* Suspend pending - return context to
* queue and stop scheduling
*/
mutex_lock(
&kctx->jctx.sched_info.ctx.jsctx_mutex);
if (kbase_js_ctx_list_add_pullable_head(
kctx->kbdev, kctx, js))
kbase_js_sync_timers(kbdev);
mutex_unlock(
&kctx->jctx.sched_info.ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
up(&js_devdata->schedule_sem);
return;
}
kbase_ctx_flag_set(kctx, KCTX_ACTIVE);
}
if (!kbase_js_use_ctx(kbdev, kctx, js)) {
mutex_lock(
&kctx->jctx.sched_info.ctx.jsctx_mutex);
dev_dbg(kbdev->dev,
"kctx %pK cannot be used at this time\n",
kctx);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
if (kbase_js_ctx_pullable(kctx, js, false)
|| kbase_ctx_flag(kctx, KCTX_PRIVILEGED))
timer_sync |=
kbase_js_ctx_list_add_pullable_head_nolock(
kctx->kbdev, kctx, js);
else
timer_sync |=
kbase_js_ctx_list_add_unpullable_nolock(
kctx->kbdev, kctx, js);
spin_unlock_irqrestore(&kbdev->hwaccess_lock,
flags);
mutex_unlock(
&kctx->jctx.sched_info.ctx.jsctx_mutex);
if (context_idle) {
WARN_ON(!kbase_ctx_flag(kctx, KCTX_ACTIVE));
kbase_ctx_flag_clear(kctx, KCTX_ACTIVE);
kbase_pm_context_idle(kbdev);
}
/* No more jobs can be submitted on this slot */
js_mask &= ~(1 << js);
break;
}
mutex_lock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
kbase_ctx_flag_clear(kctx, KCTX_PULLED);
if (!kbase_jm_kick(kbdev, 1 << js)) {
dev_dbg(kbdev->dev,
"No more jobs can be submitted (s:%d)\n",
js);
js_mask &= ~(1 << js);
}
if (!kbase_ctx_flag(kctx, KCTX_PULLED)) {
bool pullable;
dev_dbg(kbdev->dev,
"No atoms pulled from kctx %pK (s:%d)\n",
(void *)kctx, js);
pullable = kbase_js_ctx_pullable(kctx, js,
true);
/* Failed to pull jobs - push to head of list.
* Unless this context is already 'active', in
* which case it's effectively already scheduled
* so push it to the back of the list.
*/
if (pullable && kctx == last_active[js] &&
kbase_ctx_flag(kctx,
(KCTX_PULLED_SINCE_ACTIVE_JS0 <<
js)))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kctx->kbdev,
kctx, js);
else if (pullable)
timer_sync |=
kbase_js_ctx_list_add_pullable_head_nolock(
kctx->kbdev,
kctx, js);
else
timer_sync |=
kbase_js_ctx_list_add_unpullable_nolock(
kctx->kbdev,
kctx, js);
/* If this context is not the active context,
* but the active context is pullable on this
* slot, then we need to remove the active
* marker to prevent it from submitting atoms in
* the IRQ handler, which would prevent this
* context from making progress.
*/
if (last_active[js] && kctx != last_active[js]
&& kbase_js_ctx_pullable(
last_active[js], js, true))
ctx_waiting[js] = true;
if (context_idle) {
kbase_jm_idle_ctx(kbdev, kctx);
spin_unlock_irqrestore(
&kbdev->hwaccess_lock,
flags);
WARN_ON(!kbase_ctx_flag(kctx, KCTX_ACTIVE));
kbase_ctx_flag_clear(kctx, KCTX_ACTIVE);
kbase_pm_context_idle(kbdev);
} else {
spin_unlock_irqrestore(
&kbdev->hwaccess_lock,
flags);
}
mutex_unlock(
&kctx->jctx.sched_info.ctx.jsctx_mutex);
js_mask &= ~(1 << js);
break; /* Could not run atoms on this slot */
}
dev_dbg(kbdev->dev, "Push kctx %pK to back of list\n",
(void *)kctx);
if (kbase_js_ctx_pullable(kctx, js, true))
timer_sync |=
kbase_js_ctx_list_add_pullable_nolock(
kctx->kbdev, kctx, js);
else
timer_sync |=
kbase_js_ctx_list_add_unpullable_nolock(
kctx->kbdev, kctx, js);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&kctx->jctx.sched_info.ctx.jsctx_mutex);
}
}
if (timer_sync)
kbase_js_sync_timers(kbdev);
for (js = 0; js < BASE_JM_MAX_NR_SLOTS; js++) {
if (kbdev->hwaccess.active_kctx[js] == last_active[js] &&
ctx_waiting[js]) {
dev_dbg(kbdev->dev, "Marking kctx %pK as inactive (s:%d)\n",
(void *)last_active[js], js);
kbdev->hwaccess.active_kctx[js] = NULL;
}
}
mutex_unlock(&js_devdata->queue_mutex);
up(&js_devdata->schedule_sem);
}
void kbase_js_zap_context(struct kbase_context *kctx)
{
struct kbase_device *kbdev = kctx->kbdev;
struct kbasep_js_device_data *js_devdata = &kbdev->js_data;
struct kbasep_js_kctx_info *js_kctx_info = &kctx->jctx.sched_info;
/*
* Critical assumption: No more submission is possible outside of the
* workqueue. This is because the OS *must* prevent U/K calls (IOCTLs)
* whilst the struct kbase_context is terminating.
*/
/* First, atomically do the following:
* - mark the context as dying
* - try to evict it from the queue
*/
mutex_lock(&kctx->jctx.lock);
mutex_lock(&js_devdata->queue_mutex);
mutex_lock(&js_kctx_info->ctx.jsctx_mutex);
kbase_ctx_flag_set(kctx, KCTX_DYING);
dev_dbg(kbdev->dev, "Zap: Try Evict Ctx %pK", kctx);
/*
* At this point we know:
* - If eviction succeeded, it was in the queue, but now no
* longer is
* - We must cancel the jobs here. No Power Manager active reference to
* release.
* - This happens asynchronously - kbase_jd_zap_context() will wait for
* those jobs to be killed.
* - If eviction failed, then it wasn't in the queue. It is one
* of the following:
* - a. it didn't have any jobs, and so is not in the Queue or
* the Run Pool (not scheduled)
* - Hence, no more work required to cancel jobs. No Power Manager
* active reference to release.
* - b. it was in the middle of a scheduling transaction (and thus must
* have at least 1 job). This can happen from a syscall or a
* kernel thread. We still hold the jsctx_mutex, and so the thread
* must be waiting inside kbasep_js_try_schedule_head_ctx(),
* before checking whether the runpool is full. That thread will
* continue after we drop the mutex, and will notice the context
* is dying. It will rollback the transaction, killing all jobs at
* the same time. kbase_jd_zap_context() will wait for those jobs
* to be killed.
* - Hence, no more work required to cancel jobs, or to release the
* Power Manager active reference.
* - c. it is scheduled, and may or may not be running jobs
* - We must cause it to leave the runpool by stopping it from
* submitting any more jobs. When it finally does leave,
* kbasep_js_runpool_requeue_or_kill_ctx() will kill all remaining jobs
* (because it is dying), release the Power Manager active reference,
* and will not requeue the context in the queue.
* kbase_jd_zap_context() will wait for those jobs to be killed.
* - Hence, work required just to make it leave the runpool. Cancelling
* jobs and releasing the Power manager active reference will be
* handled when it leaves the runpool.
*/
if (!kbase_ctx_flag(kctx, KCTX_SCHEDULED)) {
unsigned long flags;
int js;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++) {
if (!list_empty(
&kctx->jctx.sched_info.ctx.ctx_list_entry[js]))
list_del_init(
&kctx->jctx.sched_info.ctx.ctx_list_entry[js]);
}
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
/* The following events require us to kill off remaining jobs
* and update PM book-keeping:
* - we evicted it correctly (it must have jobs to be in the
* Queue)
*
* These events need no action, but take this path anyway:
* - Case a: it didn't have any jobs, and was never in the Queue
* - Case b: scheduling transaction will be partially rolled-
* back (this already cancels the jobs)
*/
KBASE_KTRACE_ADD_JM(kbdev, JM_ZAP_NON_SCHEDULED, kctx, NULL, 0u, kbase_ctx_flag(kctx, KCTX_SCHEDULED));
dev_dbg(kbdev->dev, "Zap: Ctx %pK scheduled=0", kctx);
/* Only cancel jobs when we evicted from the
* queue. No Power Manager active reference was held.
*
* Having is_dying set ensures that this kills, and doesn't
* requeue
*/
kbasep_js_runpool_requeue_or_kill_ctx(kbdev, kctx, false);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
mutex_unlock(&kctx->jctx.lock);
} else {
unsigned long flags;
bool was_retained;
/* Case c: didn't evict, but it is scheduled - it's in the Run
* Pool
*/
KBASE_KTRACE_ADD_JM(kbdev, JM_ZAP_SCHEDULED, kctx, NULL, 0u, kbase_ctx_flag(kctx, KCTX_SCHEDULED));
dev_dbg(kbdev->dev, "Zap: Ctx %pK is in RunPool", kctx);
/* Disable the ctx from submitting any more jobs */
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
kbasep_js_clear_submit_allowed(js_devdata, kctx);
/* Retain and (later) release the context whilst it is is now
* disallowed from submitting jobs - ensures that someone
* somewhere will be removing the context later on
*/
was_retained = kbase_ctx_sched_inc_refcount_nolock(kctx);
/* Since it's scheduled and we have the jsctx_mutex, it must be
* retained successfully
*/
KBASE_DEBUG_ASSERT(was_retained);
dev_dbg(kbdev->dev, "Zap: Ctx %pK Kill Any Running jobs", kctx);
/* Cancel any remaining running jobs for this kctx - if any.
* Submit is disallowed which takes effect immediately, so no
* more new jobs will appear after we do this.
*/
kbase_backend_jm_kill_running_jobs_from_kctx(kctx);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
mutex_unlock(&js_kctx_info->ctx.jsctx_mutex);
mutex_unlock(&js_devdata->queue_mutex);
mutex_unlock(&kctx->jctx.lock);
dev_dbg(kbdev->dev, "Zap: Ctx %pK Release (may or may not schedule out immediately)",
kctx);
kbasep_js_runpool_release_ctx(kbdev, kctx);
}
KBASE_KTRACE_ADD_JM(kbdev, JM_ZAP_DONE, kctx, NULL, 0u, 0u);
/* After this, you must wait on both the
* kbase_jd_context::zero_jobs_wait and the
* kbasep_js_kctx_info::ctx::is_scheduled_waitq - to wait for the jobs
* to be destroyed, and the context to be de-scheduled (if it was on the
* runpool).
*
* kbase_jd_zap_context() will do this.
*/
}
static inline int trace_get_refcnt(struct kbase_device *kbdev,
struct kbase_context *kctx)
{
return atomic_read(&kctx->refcount);
}
/**
* kbase_js_foreach_ctx_job(): - Call a function on all jobs in context
* @kctx: Pointer to context.
* @callback: Pointer to function to call for each job.
*
* Call a function on all jobs belonging to a non-queued, non-running
* context, and detach the jobs from the context as it goes.
*
* Due to the locks that might be held at the time of the call, the callback
* may need to defer work on a workqueue to complete its actions (e.g. when
* cancelling jobs)
*
* Atoms will be removed from the queue, so this must only be called when
* cancelling jobs (which occurs as part of context destruction).
*
* The locking conditions on the caller are as follows:
* - it will be holding kbasep_js_kctx_info::ctx::jsctx_mutex.
*/
static void kbase_js_foreach_ctx_job(struct kbase_context *kctx,
kbasep_js_ctx_job_cb *callback)
{
struct kbase_device *kbdev;
unsigned long flags;
u32 js;
kbdev = kctx->kbdev;
spin_lock_irqsave(&kbdev->hwaccess_lock, flags);
KBASE_KTRACE_ADD_JM_REFCOUNT(kbdev, JS_POLICY_FOREACH_CTX_JOBS, kctx, NULL,
0u, trace_get_refcnt(kbdev, kctx));
/* Invoke callback on jobs on each slot in turn */
for (js = 0; js < kbdev->gpu_props.num_job_slots; js++)
jsctx_queue_foreach(kctx, js, callback);
spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags);
}
base_jd_prio kbase_js_priority_check(struct kbase_device *kbdev, base_jd_prio priority)
{
struct priority_control_manager_device *pcm_device = kbdev->pcm_dev;
int req_priority, out_priority;
base_jd_prio out_jd_priority = priority;
if (pcm_device) {
req_priority = kbasep_js_atom_prio_to_sched_prio(priority);
out_priority = pcm_device->ops.pcm_scheduler_priority_check(pcm_device, current, req_priority);
out_jd_priority = kbasep_js_sched_prio_to_atom_prio(out_priority);
}
return out_jd_priority;
}