Add a complete priority inheritance implementation to the scheduler (all mutex ownership and queue_send calls are inheritable). Priorities are differential so that dispatch depends on the runnable range of priorities. Codec priority can therefore be raised in small steps (pcmbuf updated to enable). Simplify the kernel functions to ease implementation and use the same kernel.c for both sim and target (I'm tired of maintaining two ;_). 1) Not sure if a minor audio break at first buffering issue will exist on large-sector disks (the main mutex speed issue was genuinely resolved earlier). At this point it's best dealt with at the buffering level. It seems a larger filechunk could be used again. 2) Perhaps 64-bit sims will have some minor issues (finicky) but a backroll of the code of concern there is a 5-minute job. All kernel objects become incompatible so a full rebuild and update is needed.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@16791 a1c6a512-1295-4272-9138-f99709370657

1 files changed, 167 insertions, 204 deletions

diff --git a/firmware/export/thread.h b/firmware/export/thread.h
index dd97ab1e83..bb1cb7cd17 100644
--- a/firmware/export/thread.h
+++ b/firmware/export/thread.h
@@ -26,21 +26,35 @@
 
 /* Priority scheduling (when enabled with HAVE_PRIORITY_SCHEDULING) works
  * by giving high priority threads more CPU time than less priority threads
- * when they need it.
- * 
+ * when they need it. Priority is differential such that the priority
+ * difference between a lower priority runnable thread and the highest priority
+ * runnable thread determines the amount of aging nescessary for the lower
+ * priority thread to be scheduled in order to prevent starvation.
+ *
  * If software playback codec pcm buffer is going down to critical, codec
- * can change it own priority to REALTIME to override user interface and
+ * can gradually raise its own priority to override user interface and
  * prevent playback skipping.
  */
+#define PRIORITY_RESERVED_HIGH   0   /* Reserved */
+#define PRIORITY_RESERVED_LOW    32  /* Reserved */
 #define HIGHEST_PRIORITY         1   /* The highest possible thread priority */
-#define LOWEST_PRIORITY          100 /* The lowest possible thread priority */
-#define PRIORITY_REALTIME        1
-#define PRIORITY_USER_INTERFACE  4   /* The main thread */
-#define PRIORITY_RECORDING       4   /* Recording thread */
-#define PRIORITY_PLAYBACK        4   /* or REALTIME when needed */
-#define PRIORITY_BUFFERING       4   /* Codec buffering thread */
-#define PRIORITY_SYSTEM          6   /* All other firmware threads */
-#define PRIORITY_BACKGROUND      8   /* Normal application threads */
+#define LOWEST_PRIORITY          31  /* The lowest possible thread priority */
+/* Realtime range reserved for threads that will not allow threads of lower
+ * priority to age and run (future expansion) */
+#define PRIORITY_REALTIME_1      1
+#define PRIORITY_REALTIME_2      2
+#define PRIORITY_REALTIME_3      3
+#define PRIORITY_REALTIME_4      4
+#define PRIORITY_REALTIME        4   /* Lowest realtime range */
+#define PRIORITY_USER_INTERFACE  16  /* The main thread */
+#define PRIORITY_RECORDING       16  /* Recording thread */
+#define PRIORITY_PLAYBACK        16  /* Variable between this and MAX */
+#define PRIORITY_PLAYBACK_MAX    5   /* Maximum allowable playback priority */
+#define PRIORITY_BUFFERING       16  /* Codec buffering thread */
+#define PRIORITY_SYSTEM          18  /* All other firmware threads */
+#define PRIORITY_BACKGROUND      20  /* Normal application threads */
+#define NUM_PRIORITIES           32
+#define PRIORITY_IDLE            32  /* Priority representative of no tasks */
 
 /* TODO: Only a minor tweak to create_thread would be needed to let
  * thread slots be caller allocated - no essential threading functionality
@@ -59,80 +73,40 @@
 
 #define DEFAULT_STACK_SIZE 0x400 /* Bytes */
 
-/**
- * "Busy" values that can be swapped into a variable to indicate
- * that the variable or object pointed to is in use by another processor
- * core. When accessed, the busy value is swapped-in while the current
- * value is atomically returned. If the swap returns the busy value,
- * the processor should retry the operation until some other value is
- * returned. When modification is finished, the new value should be
- * written which unlocks it and updates it atomically.
- *
- * Procedure:
- * while ((curr_value = swap(&variable, BUSY_VALUE)) == BUSY_VALUE);
- *
- * Modify/examine object at mem location or variable. Create "new_value"
- * as suitable.
- *
- * variable = new_value or curr_value;
- *
- * To check a value for busy and perform an operation if not:
- * curr_value = swap(&variable, BUSY_VALUE);
- *
- * if (curr_value != BUSY_VALUE)
- * {
- *     Modify/examine object at mem location or variable. Create "new_value"
- *     as suitable.
- *     variable = new_value or curr_value;
- * }
- * else
- * {
- *     Do nothing - already busy
- * }
- *
- * Only ever restore when an actual value is returned or else it could leave
- * the variable locked permanently if another processor unlocked in the
- * meantime. The next access attempt would deadlock for all processors since
- * an abandoned busy status would be left behind.
- */
-#define STATE_BUSYuptr    ((void*)UINTPTR_MAX)
-#define STATE_BUSYu8      UINT8_MAX
-#define STATE_BUSYi       INT_MIN
-
 #ifndef SIMULATOR
 /* Need to keep structures inside the header file because debug_menu
  * needs them. */
 #ifdef CPU_COLDFIRE
 struct regs
 {
-    unsigned int macsr;  /*     0 - EMAC status register */
-    unsigned int d[6];   /*  4-24 - d2-d7 */
-    unsigned int a[5];   /* 28-44 - a2-a6 */
-    void         *sp;    /*    48 - Stack pointer (a7) */
-    void         *start; /*    52 - Thread start address, or NULL when started */
+    uint32_t macsr; /*     0 - EMAC status register */
+    uint32_t d[6];  /*  4-24 - d2-d7 */
+    uint32_t a[5];  /* 28-44 - a2-a6 */
+    uint32_t sp;    /*    48 - Stack pointer (a7) */
+    uint32_t start; /*    52 - Thread start address, or NULL when started */
 };
 #elif CONFIG_CPU == SH7034
 struct regs
 {
-    unsigned int r[7];   /*  0-24 - Registers r8 thru r14 */
-    void         *sp;    /*    28 - Stack pointer (r15) */
-    void         *pr;    /*    32 - Procedure register */
-    void         *start; /*    36 - Thread start address, or NULL when started */
+    uint32_t r[7];  /*  0-24 - Registers r8 thru r14 */
+    uint32_t sp;    /*    28 - Stack pointer (r15) */
+    uint32_t pr;    /*    32 - Procedure register */
+    uint32_t start; /*    36 - Thread start address, or NULL when started */
 };
 #elif defined(CPU_ARM)
 struct regs
 {
-    unsigned int r[8];   /*  0-28 - Registers r4-r11 */
-    void         *sp;    /*    32 - Stack pointer (r13) */
-    unsigned int lr;     /*    36 - r14 (lr) */
-    void         *start; /*    40 - Thread start address, or NULL when started */
+    uint32_t r[8];  /*  0-28 - Registers r4-r11 */
+    uint32_t sp;    /*    32 - Stack pointer (r13) */
+    uint32_t lr;    /*    36 - r14 (lr) */
+    uint32_t start; /*    40 - Thread start address, or NULL when started */
 };
 #endif /* CONFIG_CPU */
 #else
 struct regs
 {
     void *t;             /* Simulator OS thread */
-    void *c;             /* Condition for blocking and sync */
+    void *s;             /* Semaphore for blocking and wakeup */
     void (*start)(void); /* Start function */
 };
 #endif /* !SIMULATOR */
@@ -154,13 +128,13 @@ enum
                             thread_thaw is called with its ID */
     THREAD_NUM_STATES,
     TIMEOUT_STATE_FIRST = STATE_SLEEPING,
-#if NUM_CORES > 1
-    STATE_BUSY = STATE_BUSYu8, /* Thread slot is being examined */
-#endif
 };
 
 #if NUM_CORES > 1
-#define THREAD_DESTRUCT ((const char *)0x84905617)
+/* Pointer value for name field to indicate thread is being killed. Using
+ * an alternate STATE_* won't work since that would interfere with operation
+ * while the thread is still running. */
+#define THREAD_DESTRUCT ((const char *)~(intptr_t)0)
 #endif
 
 /* Link information for lists thread is in */
@@ -188,7 +162,7 @@ void corelock_unlock(struct corelock *cl);
 /* Use native atomic swap/exchange instruction */
 struct corelock
 {
-    unsigned char locked;
+    volatile unsigned char locked;
 } __attribute__((packed));
 
 #define corelock_init(cl) \
@@ -207,15 +181,36 @@ struct corelock
 #define corelock_unlock(cl)
 #endif /* core locking selection */
 
-struct thread_queue
+#ifdef HAVE_PRIORITY_SCHEDULING
+struct blocker
 {
-    struct thread_entry *queue; /* list of threads waiting -
-                                   _must_ be first member */
-#if CONFIG_CORELOCK == SW_CORELOCK
-    struct corelock cl;         /* lock for atomic list operations */
-#endif
+    struct thread_entry *thread;   /* thread blocking other threads
+                                      (aka. object owner) */
+    int priority;                  /* highest priority waiter */
+    struct thread_entry * (*wakeup_protocol)(struct thread_entry *thread);
+};
+
+/* Choices of wakeup protocol */
+
+/* For transfer of object ownership by one thread to another thread by
+ * the owning thread itself (mutexes) */
+struct thread_entry *
+    wakeup_priority_protocol_transfer(struct thread_entry *thread);
+
+/* For release by owner where ownership doesn't change - other threads,
+ * interrupts, timeouts, etc. (mutex timeout, queues) */
+struct thread_entry *
+    wakeup_priority_protocol_release(struct thread_entry *thread);
+
+
+struct priority_distribution
+{
+    uint8_t  hist[NUM_PRIORITIES]; /* Histogram: Frequency for each priority */
+    uint32_t mask;                 /* Bitmask of hist entries that are not zero */
 };
 
+#endif /* HAVE_PRIORITY_SCHEDULING */
+
 /* Information kept in each thread slot
  * members are arranged according to size - largest first - in order
  * to ensure both alignment and packing at the same time.
@@ -224,88 +219,83 @@ struct thread_entry
 {
     struct regs context;       /* Register context at switch -
                                   _must_ be first member */
-    void *stack;               /* Pointer to top of stack */
+    uintptr_t *stack;          /* Pointer to top of stack */
     const char *name;          /* Thread name */
     long tmo_tick;             /* Tick when thread should be woken from
-                                  timeout */
+                                  timeout -
+                                  states: STATE_SLEEPING/STATE_BLOCKED_W_TMO */
     struct thread_list l;      /* Links for blocked/waking/running -
                                   circular linkage in both directions */
     struct thread_list tmo;    /* Links for timeout list -
-                                  Self-pointer-terminated in reverse direction,
-                                  NULL-terminated in forward direction */
-    struct thread_queue *bqp;  /* Pointer to list variable in kernel
+                                  Circular in reverse direction, NULL-terminated in
+                                  forward direction -
+                                  states: STATE_SLEEPING/STATE_BLOCKED_W_TMO */
+    struct thread_entry **bqp; /* Pointer to list variable in kernel
                                   object where thread is blocked - used
-                                  for implicit unblock and explicit wake */
-#if CONFIG_CORELOCK == SW_CORELOCK
-    struct thread_entry **bqnlp; /* Pointer to list variable in kernel
-                                    object where thread is blocked - non-locked
-                                    operations will be used */
+                                  for implicit unblock and explicit wake
+                                  states: STATE_BLOCKED/STATE_BLOCKED_W_TMO  */
+#if NUM_CORES > 1
+    struct corelock *obj_cl;   /* Object corelock where thead is blocked -
+                                  states: STATE_BLOCKED/STATE_BLOCKED_W_TMO */
 #endif
     struct thread_entry *queue; /* List of threads waiting for thread to be
                                   removed */
 #ifdef HAVE_EXTENDED_MESSAGING_AND_NAME
-    intptr_t retval;           /* Return value from a blocked operation */
+    #define HAVE_WAKEUP_EXT_CB
+    void (*wakeup_ext_cb)(struct thread_entry *thread); /* Callback that
+                                  performs special steps needed when being
+                                  forced off of an object's wait queue that
+                                  go beyond the standard wait queue removal
+                                  and priority disinheritance */
+    /* Only enabled when using queue_send for now */
+#endif
+#if defined(HAVE_EXTENDED_MESSAGING_AND_NAME) || NUM_CORES > 1
+    intptr_t retval;           /* Return value from a blocked operation/
+                                  misc. use */
 #endif
 #ifdef HAVE_PRIORITY_SCHEDULING
-    long last_run;             /* Last tick when started */
+    /* Priority summary of owned objects that support inheritance */
+    struct blocker *blocker;   /* Pointer to blocker when this thread is blocked
+                                  on an object that supports PIP -
+                                  states: STATE_BLOCKED/STATE_BLOCKED_W_TMO  */
+    struct priority_distribution pdist; /* Priority summary of owned objects
+                                  that have blocked threads and thread's own
+                                  base priority */
+    int skip_count;            /* Number of times skipped if higher priority
+                                  thread was running */
 #endif
     unsigned short stack_size; /* Size of stack in bytes */
 #ifdef HAVE_PRIORITY_SCHEDULING
-    unsigned char priority;    /* Current priority */
-    unsigned char priority_x;  /* Inherited priority - right now just a
-                                  runtime guarantee flag */
+    unsigned char base_priority; /* Base priority (set explicitly during
+                                  creation or thread_set_priority) */
+    unsigned char priority;    /* Scheduled priority (higher of base or
+                                  all threads blocked by this one) */
 #endif
     unsigned char state;       /* Thread slot state (STATE_*) */
-#if NUM_CORES > 1
-    unsigned char core;        /* The core to which thread belongs */
-#endif
 #ifdef HAVE_SCHEDULER_BOOSTCTRL
-    unsigned char boosted;     /* CPU frequency boost flag */
+    unsigned char cpu_boost;   /* CPU frequency boost flag */
 #endif
-#if CONFIG_CORELOCK == SW_CORELOCK
-    struct corelock cl;        /* Corelock to lock thread slot */
+#if NUM_CORES > 1
+    unsigned char core;        /* The core to which thread belongs */
+    struct corelock waiter_cl; /* Corelock for thread_wait */
+    struct corelock slot_cl;   /* Corelock to lock thread slot */
 #endif
 };
 
 #if NUM_CORES > 1
 /* Operations to be performed just before stopping a thread and starting
    a new one if specified before calling switch_thread */
-#define TBOP_UNLOCK_LIST     0x01 /* Set a pointer variable address var_ptrp */
-#if CONFIG_CORELOCK == CORELOCK_SWAP
-#define TBOP_SET_VARi        0x02 /* Set an int at address var_ip */
-#define TBOP_SET_VARu8       0x03 /* Set an unsigned char at address var_u8p */
-#define TBOP_VAR_TYPE_MASK   0x03 /* Mask for variable type*/
-#endif /* CONFIG_CORELOCK */
-#define TBOP_UNLOCK_CORELOCK 0x04
-#define TBOP_UNLOCK_THREAD   0x08 /* Unlock a thread's slot */
-#define TBOP_UNLOCK_CURRENT  0x10 /* Unlock the current thread's slot */
-#define TBOP_SWITCH_CORE     0x20 /* Call the core switch preparation routine */
+enum
+{
+    TBOP_CLEAR = 0,       /* No operation to do */
+    TBOP_UNLOCK_CORELOCK, /* Unlock a corelock variable */
+    TBOP_SWITCH_CORE,     /* Call the core switch preparation routine */
+};
 
 struct thread_blk_ops
 {
-#if CONFIG_CORELOCK != SW_CORELOCK
-    union
-    {
-        int var_iv;                   /* int variable value to set */
-        uint8_t var_u8v;              /* unsigned char valur to set */
-        struct thread_entry *list_v;  /* list pointer queue value to set */
-    };
-#endif
-    union
-    {
-#if CONFIG_CORELOCK != SW_CORELOCK
-        int *var_ip;                  /* pointer to int variable */
-        uint8_t *var_u8p;             /* pointer to unsigned char varuable */
-#endif
-        struct thread_queue *list_p;  /* pointer to list variable */
-    };
-#if CONFIG_CORELOCK == SW_CORELOCK
-    struct corelock *cl_p;            /* corelock to unlock */
-    struct thread_entry *thread;      /* thread to unlock */
-#elif CONFIG_CORELOCK == CORELOCK_SWAP
-    unsigned char state;              /* new thread state (performs unlock) */
-#endif /* SOFTWARE_CORELOCK */
-    unsigned char flags;    /* TBOP_* flags */
+    struct corelock *cl_p;    /* pointer to corelock */
+    unsigned char    flags;   /* TBOP_* flags */
 };
 #endif /* NUM_CORES > 1 */
 
@@ -316,28 +306,30 @@ struct core_entry
 {
     /* "Active" lists - core is constantly active on these and are never
        locked and interrupts do not access them */
-    struct thread_entry *running;  /* threads that are running */
+    struct thread_entry *running;  /* threads that are running (RTR) */
     struct thread_entry *timeout;  /* threads that are on a timeout before
                                       running again */
-    /* "Shared" lists - cores interact in a synchronized manner - access
-       is locked between cores and interrupts */
-    struct thread_queue  waking;   /* intermediate locked list that
-                                      hold threads other core should wake up
-                                      on next task switch */
+    struct thread_entry *block_task; /* Task going off running list */
+#ifdef HAVE_PRIORITY_SCHEDULING
+    struct priority_distribution rtr; /* Summary of running and ready-to-run
+                                         threads */
+#endif
     long next_tmo_check;           /* soonest time to check tmo threads */
 #if NUM_CORES > 1
     struct thread_blk_ops blk_ops; /* operations to perform when
                                       blocking a thread */
-#endif /* NUM_CORES */
 #ifdef HAVE_PRIORITY_SCHEDULING
-    unsigned char highest_priority;
+    struct corelock rtr_cl;        /* Lock for rtr list */
 #endif
+#endif /* NUM_CORES */
 };
 
 #ifdef HAVE_PRIORITY_SCHEDULING
 #define IF_PRIO(...)    __VA_ARGS__
+#define IFN_PRIO(...)
 #else
 #define IF_PRIO(...)
+#define IFN_PRIO(...)   __VA_ARGS__
 #endif
 
 /* Macros generate better code than an inline function is this case */
@@ -464,13 +456,18 @@ struct core_entry
 void core_idle(void);
 void core_wake(IF_COP_VOID(unsigned int core));
 
+/* Initialize the scheduler */
+void init_threads(void);
+
+/* Allocate a thread in the scheduler */
 #define CREATE_THREAD_FROZEN   0x00000001 /* Thread is frozen at create time */
 struct thread_entry*
-    create_thread(void (*function)(void), void* stack, int stack_size,
+    create_thread(void (*function)(void), void* stack, size_t stack_size,
                   unsigned flags, const char *name
                   IF_PRIO(, int priority)
                           IF_COP(, unsigned int core));
 
+/* Set and clear the CPU frequency boost flag for the calling thread */
 #ifdef HAVE_SCHEDULER_BOOSTCTRL
 void trigger_cpu_boost(void);
 void cancel_cpu_boost(void);
@@ -478,86 +475,52 @@ void cancel_cpu_boost(void);
 #define trigger_cpu_boost()
 #define cancel_cpu_boost()
 #endif
+/* Make a frozed thread runnable (when started with CREATE_THREAD_FROZEN).
+ * Has no effect on a thread not frozen. */
 void thread_thaw(struct thread_entry *thread);
+/* Wait for a thread to exit */
 void thread_wait(struct thread_entry *thread);
+/* Exit the current thread */
+void thread_exit(void);
+#if defined(DEBUG) || defined(ROCKBOX_HAS_LOGF)
+#define ALLOW_REMOVE_THREAD
+/* Remove a thread from the scheduler */
 void remove_thread(struct thread_entry *thread);
-void switch_thread(struct thread_entry *old);
-void sleep_thread(int ticks);
+#endif
 
-/**
- * Setup to allow using thread queues as locked or non-locked without speed
- * sacrifices in both core locking types.
- *
- * The blocking/waking function inline two different version of the real
- * function into the stubs when a software or other separate core locking
- * mechanism is employed.
- *
- * When a simple test-and-set or similar instruction is available, locking
- * has no cost and so one version is used and the internal worker is called
- * directly.
- *
- * CORELOCK_NONE is treated the same as when an atomic instruction can be
- * used.
- */
+/* Switch to next runnable thread */
+void switch_thread(void);
+/* Blocks a thread for at least the specified number of ticks (0 = wait until
+ * next tick) */
+void sleep_thread(int ticks);
+/* Indefinitely blocks the current thread on a thread queue */
+void block_thread(struct thread_entry *current);
+/* Blocks the current thread on a thread queue until explicitely woken or
+ * the timeout is reached */
+void block_thread_w_tmo(struct thread_entry *current, int timeout);
+
+/* Return bit flags for thread wakeup */
+#define THREAD_NONE     0x0 /* No thread woken up (exclusive) */
+#define THREAD_OK       0x1 /* A thread was woken up */
+#define THREAD_SWITCH   0x2 /* Task switch recommended (one or more of
+                               higher priority than current were woken) */
 
-/* Blocks the current thread on a thread queue */
-#if CONFIG_CORELOCK == SW_CORELOCK
-void block_thread(struct thread_queue *tq);
-void block_thread_no_listlock(struct thread_entry **list);
-#else
-void _block_thread(struct thread_queue *tq);
-static inline void block_thread(struct thread_queue *tq)
-    { _block_thread(tq); }
-static inline void block_thread_no_listlock(struct thread_entry **list)
-    { _block_thread((struct thread_queue *)list); }
-#endif /* CONFIG_CORELOCK */
-
-/* Blocks the current thread on a thread queue for a max amount of time
- * There is no "_no_listlock" version because timeout blocks without sync on
- * the blocking queues is not permitted since either core could access the
- * list at any time to do an implicit wake. In other words, objects with
- * timeout support require lockable queues. */
-void block_thread_w_tmo(struct thread_queue *tq, int timeout);
-
-/* Wakes up the thread at the head of the queue */
-#define THREAD_WAKEUP_NONE    ((struct thread_entry *)NULL)
-#define THREAD_WAKEUP_MISSING ((struct thread_entry *)(NULL+1))
-#if CONFIG_CORELOCK == SW_CORELOCK
-struct thread_entry * wakeup_thread(struct thread_queue *tq);
-struct thread_entry * wakeup_thread_no_listlock(struct thread_entry **list);
-#else
-struct thread_entry * _wakeup_thread(struct thread_queue *list);
-static inline struct thread_entry * wakeup_thread(struct thread_queue *tq)
-    { return _wakeup_thread(tq); }
-static inline struct thread_entry * wakeup_thread_no_listlock(struct thread_entry **list)
-    { return _wakeup_thread((struct thread_queue *)list); }
-#endif /* CONFIG_CORELOCK */
-
-/* Initialize a thread_queue object. */
-static inline void thread_queue_init(struct thread_queue *tq)
-    { tq->queue = NULL; IF_SWCL(corelock_init(&tq->cl);) }
 /* A convenience function for waking an entire queue of threads. */
-static inline void thread_queue_wake(struct thread_queue *tq)
-    { while (wakeup_thread(tq) != NULL); }
-/* The no-listlock version of thread_queue_wake() */
-static inline void thread_queue_wake_no_listlock(struct thread_entry **list)
-    { while (wakeup_thread_no_listlock(list) != NULL); }
+unsigned int thread_queue_wake(struct thread_entry **list);
+
+/* Wakeup a thread at the head of a list */
+unsigned int wakeup_thread(struct thread_entry **list);
 
 #ifdef HAVE_PRIORITY_SCHEDULING
 int thread_set_priority(struct thread_entry *thread, int priority);
 int thread_get_priority(struct thread_entry *thread);
-/* Yield that guarantees thread execution once per round regardless of
-   thread's scheduler priority - basically a transient realtime boost
-   without altering the scheduler's thread precedence. */
-void priority_yield(void);
-#else
-#define priority_yield  yield
 #endif /* HAVE_PRIORITY_SCHEDULING */
 #if NUM_CORES > 1
 unsigned int switch_core(unsigned int new_core);
 #endif
 struct thread_entry * thread_get_current(void);
-void init_threads(void);
+
+/* Debugging info - only! */
 int thread_stack_usage(const struct thread_entry *thread);
 #if NUM_CORES > 1
 int idle_stack_usage(unsigned int core);