source: trunk/kernel/doc/40bits @ 367

1-Event Manager:
The even manager has been changed in the fallowing ways :
        - the API has been changed to avoid given a pointer to the second argumern to event_send we give only the 
        the gpid of the target processor.
        - local event handling is gloablly the same
        - remote event has been changed. Must be carrefully handled with physical access:
                - the kfifo of the remote cpu is physically accessed
                - the event_notifier must notify after copying the event structure (on the stack ?). 
                        !!! This does not aplying that the pointer arguments are copied !!! 
                        => modify event handler

3-Processes/Thread Management

        3.1) Task Management

                There's one gloabal table that index all process by pid.
                Replicaing the kernel will also replicate this structure...
                The fatest way to corrdinate the replica : 
                        - Consiste of using the table of cluster 0 as the main table. Each table will contain
                        a pointer to the task structure if the task is local, other ways the table entry is 
                        either marked and contains the cluster id wich handle the task or a NULL value. 
                        A NULL value at the main table mean that the entry has not been allocated.
                        A NULL value in other table must be confirmed by cheching the main table.

                        This imply a change in pid allocate, wich must allocate from the main table by setting
                        the cluster id and marking the entry with a remote flag (0x1). // TODO more on synch
                
                        The lookup need to check for the task in local, if we find a task address we return' it. 
                        If we find a remote entry we go to the indicate cluster. If the remote cluster does not contain 
                        the address of the task ...


                Excuse me!  the fastest way is to directly use the main table as the only table and ignore other tables!!!
                This will require putting a physical addresse on the table to point to task structures and synchronised 
                with a lock.

        3.2)Task descriptor

        This structure contain the following structures that are shared between threads:
                - vmm (Virtual Memory Manager) : see 2-.
                - pointer to root/cwd file, fdinfo and binary file: use physical addresses
                - pointer to task parent: use a parent cluster id field, or directly a physical addresse ?
                - 2 list : one for the children and one for the siblings
                - a signal manager (NULL par default ?)
                - thread handling structures : 
                        - a table containing a reference to all the threads (and a pointer to it's page struct)
                        - a list entry root that englobe all thread: not really used, can be deleted
                
        How to manipulate these fields ? => by message passing ?
        for thread's pointer we directly placed the ppn: thread's structures are aligned on a page.
        Other pointers fd_info, bin are to locals strutures.
        vmm handling is local if a single threaded task, otherwise the handling is done by message passing.

        3.3) Thread Management
        In the second stage...
        
                
                
2-Memory Management:
        vmm: message passing to the processor that contain the main structure. 
        There is no need to acess the thread data except for statistique pupose!
        There is a part that could be done by the thread with passing a message... look in the page table part

        To think : page fault at the same addresses how to avoid them (merge the messages of page fault) ?

        kmem: all remote allocation are forbiden except for user pages (at least for now).
        (user remote allocation have been done in the second stage.)


3-Cluster Manager:
        One cluster manager for each cluster: no need for the global table...no need for the replication of cluster addresses
        ...no needs for procs base addresses: the cpu_s are directly put into the cluster structures with the cluster base 
        address the compiler will offset to the good addresse.


3-DQDT:
        initialisation has been modified to not require the remote allocation API. We set a table of two dqdt in the 



4-Drivers




5-Boot:
        A) The main proc of the boot cluster enter the boot loader, initiliase the stack bases for
        all other procs in the boot_tbl and wait for all other proc to copy the needed info (kernel, boot,
        boot info) from the boot cluster.

        B) All other procs enter the boot_loader allocate a stack a the end of the cluster ram and
        wait for their cluster to be initilised by their main proc. (except for those of the boot cluster ?)

        C) The main procs of other clusters enter the boot loader, set the sp register to point to
        the stack at the end of the current cluster ram. 
        (Enter the C code and) Copy the kernel/boot/boot_info from the main cluster to their cluster
        increment a value  and sleep  back (wait instruction).

        D) All procs sleep after incrementing a counter in the in the original BI (boot cluster).

        E) Once the counter reach the number of proc, the main proc of the boot cluster can then
        copy the kernel/boot/boot_info in the same cluster but at the correct adresses. This will erase the
        code of the preloader. (the sp pointer of the boot cluster is good, the preloader should set it to
        point

        !!! moving the boot code is dangerous, since we are currently executing it ... and the
        addreses in it are fixed ?!!! If we really want this space, we need to give up at this stage on the
        boot code, BI can be kept.
        
        ***********************************************************************************************************
        Initital memory content at boot (TODO):

        ------------------ 0x0000 0000
        Preloader
        ------------------ ENDPRELOADER
        Kernel
        ------------------ END kernel paddr (important here, the charging addresse (paddr) is differente than vaddr)
        BI
        ------------------
        Boot code
        ------------------ End of boot
        .....
        ------------------ end of cluster space - 4 boot stacks
        Boot stacks (x nbproc)
        ----------------- End of cluster space

        
        Boot code role is to set:
                1 - the kernel and boot info: deplace it to the address 0
                2 - the stacks for each proc
                3 - initialise the info structure (allocated on the stack): don't forget to put the BI info address
                        that resulted from displacing it.
                4 - initialise the page table and the mmu if we are booting in virtual mode

        ***** Phase 1: boot proc of the boot cluster
        - after the preloader, there's only the boot_proc who has a stack and all other proc are waiting (an WTI).
        - jump to C code
        - wakeup all proc (possible improvement: set a distributed wakeup)

        ***** Phase 2: all proc (the boot cluster proc)

        - start by setting a stack at the end of current cluster
        - jump to C code
        - set info structure
        - copy the kernel and BI info (only one proc per cluster) at the start of the current cluster 
        (- set the page table and the mmu)
        - jump to the kernel (of the cuurrent cluster) and wait in it ? yes wait for the
          kernel image of cluster 0 to be set (since we will probably synchronise with it)
          

        ? how to signal to the current cluster proc to jump to kernel ?

        --> the boot code is now used by only the boot cluster
        --> the preloader code is used by only the boot cluster other procs

        ***** Phase 3:
        - put a barrier to be sure that all procs have entred the boot code
        - all proc set their stack: to put after the boot proc stack
        (- set the page table and the mmu)

        --> the preloader code is no longer in use

        - copy the kernel and BI of the boot cluster (by the boot proc)
        - all proc of the boot cluster can now jump to the kernel and wait
        - the boot proc enters the kernels and wake up a proc per cluster (except for the current
          boot cluster)

        --> the boot code is no longer in use
                

        !! be careful the stack of all the procs are still in use !!
        put the in the reserved zone ?

        N.B.: all addresses of the boot cluster are found using compiled in variables


        ****** It's up to the kernel now...
        
Kernel boot:
        ***** Phase 1: only the boot proc
        wake up one proc per cluster
        !!! Must be sure that they are sleeping (waiting) ? !!!

        ***** Phase 2: A proc per cluster
        Tasks: '->' indicate dependance

        1) initialise task bootstap and task manager and set idle arg
                reported: devfs/sysfs roots

        2) initialise arch_memory : (initialise boot_tbl: done in the boot phase); initialise
                clusters_tbl ?
        initialise arch: initialise current cluster: initialise ppm (physical page memory) (can
                                alloc now!), cpu_init : event listener!
                         initialise devices (printk) (put all in cluster devices lists, root all
                        interruptions locally (to the proc handling them)); 
                         and register them (!!!to be reported!!!) -> depend on message api


        3) --Barrier--: here we should wait that everybody has set there event listener.

        only main proc of I/O cluster: initialise fs cache locally(?); 
                                        initialise devfs/sysfs roots and set devices.
        ----Barier ?------


        4) task_bootstarp_finalize :  should not be needed all should be done in task_init (TODO)
           kdmsg_init : a set of lock (isr, exception, printk)_lock  and printk_sync
        
        5) DQDT init: arch_dqdt_init and dqdt_init (TODO)

        6) Forget about bootstrap_replicate_task and cluster_init_table 
                a local cluster_init_cores is needed ?

        register devices in fs ?
        
        **** Phase 3: all procs (thread idle)

        set thread_idle for each proc (parrallèle ?)

        load thread_idle: free reserved page (the boot loader reserved pages...), create the event
                                thread event manager, kvfsd only in the cluster_IO.


        *** Phase 4: kvfsd (only main cpu of cluster IO)
        
        enable all irq.

        initialise fs -> need __sys_blk
        task load init require devfs file system
        if failed set kminishell
        








Kernel Runing Modifications:

        *** DQDT (Ressource handling): Must be set using physical addresses ?

        *** FS support: Open, read, write ... all physical access or just when reaching the device ?

        *** Memory support: when there is no multithreaded application (only kthread) the memory
                management is all done localy. Cross cluster messaging for remote creation.

        *** Task support: a replicated taskmanager, pids are now composed of XYN, X and Y are the
                cluster offset, N is the offset in the current cluster. Remote creation can be done
                only using message passing.

        *** Device support:

        *** Exception, Interruptions (not syscalls): all device interruptions are rooted and handled
                        locally, as configured at boot time.

        Add cluster span to the arch config or in the cluster_entry. Yes it must be defined by the
        boot code since this the code that decide the offset when using virtual memory.

        *** User space support:
                - Syscall: switch adressing mode in case we come or go from a syscall; save the coproc2 extend register 
                - copy from/to user must be modified


Developement process:
        ********** Phase one *******************
        Consiste of boting the kernel with no user space support and with a dummy dqdt suport

        *** Stage 0:
                Rewrite the kernel in the view of passing to 40 bits:
                        Handle gloabal variables:
                        1) remove inused gloabl variables
                        2) simplifie the message passing API
                        3) do a simplified DQDT : replace dqdt files by a dummy : next_cluster++
                        4) task and thread creation only using message passing
                   5) do a true copy_from_usr and copy_to_usr !
                   6) delete the assumption that all address superior to KERNEL_OFFSET belongs 
                                         to the kernel(genarally used to verify if an address belongd to user space)
                                         replace by checking that the address is between USR_OFFSET and _LIMIT 
                                         (see libk/elf.c) for an example.
                        

        *** Stage 1: 
                Develloping boot code: require setting the disk, develloping virtual memory for
                        the perspective of a dooble mode of booting, replicate the kernel and put it
                        at address zero.
                Develloping kernel boot code (reaching kern_init and printing a message ?), setting boot_dmsg 

        *** Stage 2: (TODO: synchronise with stage 0)
                Rewrite of most of the kernel:
                        ******* GOALS:
                        1) delete (replace) reference to (unused) global variables: 
                                - cluster_tbl : no such a thing as cluster init...
                                        reference to cluster struct should be removed to keep only cids: this will affect allot of code!
                                        No more a access to remote listener ptr (event.c), nor cpu struct, not cluster, nor ppm of remote cluster ...

                                - devfs_db replaced by the root dentry/inode of the devfs file system and since
                                there is only one kernel that use it => allocate it dynamically
                                - similar for sysfs_root_entry ?
                                - devfs_n_op/devfs_f_op : are const 
                                - (dma,fb,icu,iopic,sda,timer,xicu,tty)_count are per cluster the
                                  new naming scheme is device_cidn
                                - rt_timer: Not used => removet it
                                - (exception, isr, printk, boot)_lock: are all used to lock the printing
                                - kexcept_tty, klog_tty, tty_tbl, kisr_tty, kboot_tty ???

                                - soclib_(any device)_driver are RO: can we mark them as const ?
                                - boot_stage_done: not used

                        2) Event Manager to use physical adresses.
                        3) fs
                        4) all interruption (produced by device and WTI) are handled locally
                        5) all exception are handled locally (there is no remote page fault, since
                                there is no user space => no multithreaded user space application)
                        6) Complete the kernel boot phase.
                        7) Context switch: adding the saving of the coproc registers: DATA_extend register and Virtual
                                mode register
        ?cpu_gid2ptr: should be modified to fetch only local cluster, if a remote cluster is called, it should return an error

        *** Stage N: user space support
                $27 is now used by the kerntry code! user space should no longer used it
                Switch to user space need to modify to set off the MMU and reset when going back...
                set in init context the mmu register value (0xF)

***Structures modif (with study only struct fields)****

** NOTATIONS:
-PB: problem, 
-HD: handled in the corresponding structure. This is for structure that are 
directly embeded in the structure and wich contain a non local pointer
-NPB: no problem
-LPTR: local pointer
-RPTR: remote pointer
-RPTR-S: RPTR  that is static => point to a structure that cannot be moved across clusters  
-RPTR-D: RPTR that is dynmic=> point to a structure that can be moved across clusters 
(which mean that the  ptr need to be updated!)


** Structures that need to accessed remotly (and does they need to be dynamic ?)
struct task_s;
struct thread_s;


** Structures details

        + thread structure (thread.h)
        struct thread_s
        {
                struct cpu_uzone_s uzone;//NOPB
                spinlock_t lock;//PB:HD
                thread_state_t state;//NOPB
                ...//NOPB
                struct cpu_s *lcpu;//NOPB: LPTR   /*! pointer to the local CPU description structure */
                struct sched_s *local_sched; //NOPB LPTR /*! pointer to the local scheduler structure */

*               struct task_s *task; //PB: RPTR

                thread_attr_t type;//NOPB           /*! 3 types : usr (PTHREAD), kernel (KTHREAD) or idle (TH_IDLE) */
                struct cpu_context_s pws;//NOPB     /*! processor work state (register saved zone) */

*               struct list_entry list;//PB       /*! next/pred threads at the same state */
*               struct list_entry rope;//OB       /*! next/pred threads in the __rope list of thread */
                struct thread_info info;//HD      /*! (exit value, statistics, ...) */

                uint_t signature;//NOPB
        };
        struct thread_info
        {
                ...//NOPB
                struct cpu_s *ocpu;//PB: RPTR-S
                uint_t wakeup_date;//NOPB           /*! wakeup date in seconds */
                void *exit_value;//NOPB                  /*! exit value returned by thread or joined one */
*               struct thread_s *join;//PB:RPTR-D              /*! points to waiting thread in join case */
*               struct thread_s *waker;//RPTR-D
                struct wait_queue_s wait_queue;//PB:HD
*               struct wait_queue_s *queue;//PB
                struct cpu_context_s pss;//NPB      /*! Processor Saved State */
                pthread_attr_t attr;//HD
                void  *kstack_addr;//NPB:LPTR (always moved with the thread)
                uint_t kstack_size;//OK
                struct event_s *e_info;//never used!!!!!!!!!!!
                struct page_s *page;//LTPR ?
        };
        typedef struct
        {
                ...//NOPB
                void *stack_addr;//user space ?
                size_t stack_size;
                void *entry_func;//USP                  /* mandatory */
                void *exit_func;//USP                   /* mandatory */
                void *arg1;//?
                void *arg2;//?
                void *sigreturn_func;//?
                void *sigstack_addr;//USP todo
                size_t sigstack_size;//?
                struct sched_param  sched_param;//PPB
                ...//NOPB                        /* mandatory */
        } pthread_attr_t;
        
        + task structure
        struct task_s
        {
                /* Various Locks */
                mcs_lock_t block;//HD
                spinlock_t lock;//HD
                spinlock_t th_lock;//HD
                struct rwlock_s cwd_lock;//HD
                spinlock_t tm_lock;//local:HD

                /* Memory Management */
                struct vmm_s vmm;//HD

                /* Placement Info */
                struct cluster_s *cluster;//local
                struct cpu_s *cpu;//local

                /* File system *
                //for cwd and root they should always point to the same 
                //file struct (not node) to avoid ensurring the coherence
                //if one of them change
                struct vfs_inode_s *vfs_root;//
                struct vfs_inode_s *vfs_cwd;//
                struct fd_info_s  *fd_info;//embed it
                struct vfs_file_s *bin;//Static: embed it
          
                /* Task management */
                pid_t pid;//NPB
                uid_t uid;//NPB
                gid_t gid;//NPB
                uint_t state;//NPB
                atomic_t childs_nr;//NPB
                uint16_t childs_limit;//NPB

                //make them of type faddr
                struct task_s *parent;//PB
                struct list_entry children;//PB 
                struct list_entry list;//PB

                /* Threads */
                uint_t threads_count;
                uint_t threads_nr;
                uint_t threads_limit;
                uint_t next_order;
                uint_t max_order;

                //add a function to register after we creat it ?
                BITMAP_DECLARE(bitmap, (CONFIG_PTHREAD_THREADS_MAX >> 3));
                struct thread_s **th_tbl;//PB: TODO: local table of pointer to avoid the page below?
                struct list_entry th_root;
                struct page_s *th_tbl_pg;//NPB: local (make it local)

                /* Signal management */
                struct sig_mgr_s sig_mgr;

        #if CONFIG_FORK_LOCAL_ALLOC
                struct cluster_s *current_clstr;
        #endif
        };

                
                
        
        
File per file modif:
        *** kern dir files:
*PKB            - kern_init.c (see up)
                - atomic.c: handle atomic_t and refcount_t objects (try to keep them local ? It will
                  be hard (fs, event handler?))
                - barrier.c: used only by sys_barrier.c
                - blkio.c: this layer allow an fs to send request to block device, all request pass
                  here.(manipulate a device struct. With this it's best to replicate the notion
                  of device ? NO). Use blkio async to send an IPI blokio request.
                - cluster.c : some init functions, cid2ptr (to delete ?), manager, key_op (never used)
                - cond_var.c: used only by user space (at least for now)
                - cpu.c : the function cpu_gid2ptr access the cluster structure! see the structure for more
                - do_exec : use copy_uspace!
                - do_interupt : remove thread migration handling ?
                - do_syscall : (N.B. the context save is for the fork syscall) remove thread migration ?
                - event : modify to support the ...
                - (keysdb/kfifo/mwmr/mcs_sync/radix/rwlock/semaphore/spinlock).c: their modification depend on the scope in which their are used
                - kthread_create: local
                - rr-sched: should be local
                - scheduler: (thread_local_cpu) should be local
                - task.c: should get simplified => removing replicate calls (they should may be used to see how the migration goes)
                - thread_create.c: use of cpu_gid2ptr, should be local
                - destroy|dup|idle.c: should be local
                - thread migrate.c: need to be modified
        SYSCALLS
        Miscelanious:
                - sys_alarm.c: manipulte local data (And thread migration is initiated by each thread indepndantly...)
*PKB            - sys_barier.c: not local, to complexe for no reason except performance, we should pospone this functionnality?
                                I think we need such a struc for the boot. This mean that posponing is not possible, but we can simplify it
                - sys_clock.c: local
                - sys_cond_var.c:
                - sys_dma_memcpy.c:
                - sys_rwlock.c
                - sys_sem.c
        Task:
                - sys_exec.c
                - sys_fork.c
                - sys_getpid.c
                - sys_ps.c
                - sys_signal.c
                - sys_thread_create.c
                - sys_thread_detach.c
                - sys_thread_exit.c
                - sys_thread_getattr.c
                - sys_thread_join.c
                - sys_thread_migrate.c
                - sys_thread_sleep.c
                - sys_thread_wakeup.c
                - sys_thread_yield.c
        Task Mem:
                - sys_madvise.c
                - sys_mcntl.c
                - sys_mmap.c
                - sys_sbrk.c
                - sys_utls.c
        FS:
                - sys_chdir.c:
                - sys_close.c:
                - sys_closedir.c:
                - sys_creat.c:
                - sys_getcwd.c
                - sys_lseek.c
                - sys_mkdir.c
                - sys_mkfifo.c
                - sys_open.c
                - sys_opendir.c
                - sys_pipe.c
                - sys_read.c
                - sys_readdir.c
                - sys_stat.c
                - sys_unlink.c
                - sys_write.c
        mm/:
                Task related:
                vmm.c: munmap (region:detach, split, resize); mmap (shared, private | file, anomymous); 
                        skbrk(resize heap region TODO: use the resize func), madvise_migrate, vmm_auto_migrate,
                        (modify page table attribute and set migrate flag), madvise_will_need;
                vm_region.c:...
                Other:
                ppm.c: handle page descriptors. We find some use of cluster ptr that should be removed ?
                page.c: to make local
                
                

Question:
1) How is the time handled in almos. Does we use a unique timer to keep the time seamleese across node ?
2) Copy to/from userspace
3) Cross cluster lists: thread lists
4) Only the kentry is mapped when entring the kernel from user space... This is going to be a complexe point.
Three new things need to be done:
1- if from user space (or simply if we were in the vitual mode was actif), switch to physical mode of local cluster
2- if from kernel save DATA_EXT register. Also, if we do cpy_uspace by switchin to the virtulle space, we need to 
do what we do in one.


Remark:
cross cluster access also keep in mind that other access are possible, like page based access (temporarely map a page).




Task and thread subsystems:
Per cluster vmm or per thread ?

Per cluster: 1) interest: less physical space cosummed and less coherence to handle
             2) Problems: "resharing" of the stack zone could complex (make it impossible to migrate the stack of thread and thus the thread) ?

Per thread:  1) interest: virtualise the virtual mapping between thread become easy  (for stack and TLS)
                                and temporary mappping are allot easier...
                                (and we could free the $27 register!)
             2) Problems: to much use of physical space (cache) + require more coherence message



Files need to accessible across cluster, since they could be share by tasks!



TODO:
Check that we have mapped enough of the kernel, for the kentry ?



VMM replicated region (mapper) require that the insertion to the page cache be atomic!


DEveloppement results:
event_fifo.c
remote operation.c
arch_init.c:ongoing for devices
cluster_init.c
kentry.c:k1 is now used by the kernel! No, it's that we k1 only when we were in
kernel mode... Otherwise when coming in usert mode we don't need it and save it?