source: trunk/kernel/mm/vmm.h @ 688

/*
 * vmm.h - virtual memory management related operations
 *
 * Authors   Ghassan Almaless (2008,2009,2010,2011, 2012)
 *           Alain Greiner    (2016,2017,2018,2019,2020))
 *
 * Copyright (c) UPMC Sorbonne Universites
 *
 * This file is part of ALMOS-MKH.
 *
 * ALMOS-MKH is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License as published by
 * the Free Software Foundation; version 2.0 of the License.
 *
 * ALMOS-MKH 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 ALMOS-MKH; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
 */

#ifndef _VMM_H_
#define _VMM_H_

#include <hal_kernel_types.h>
#include <bits.h>
#include <list.h>
#include <queuelock.h>
#include <remote_queuelock.h>
#include <hal_gpt.h>
#include <vseg.h>
#include <page.h>

/****  Forward declarations  ****/

struct process_s;
struct vseg_s;

/*********************************************************************************************
 * This structure defines the STACK allocator used by the VMM to dynamically handle
 * vseg allocation or release requests for an user thread.
 * This allocator handles a fixed size array of fixed size slots in STACK zone of user space.
 * The stack size and the number of slots are defined by the CONFIG_VMM_STACK_SIZE, and
 * CONFIG_VMM_STACK_BASE parameters.
 * Each slot can contain one user stack vseg. The first 4 Kbytes page in the slot is not
 * mapped to detect stack overflow.
 * In this implementation, the slot index is defined by the user thead LTID.
 * All allocated stacks are registered in a bitmap defining the STACK zone state: 
 * - The allocator checks that the requested slot has not been already allocated, and set the
 *   corresponding bit in the bitmap.
 * - The de-allocator reset the corresponding bit in the bitmap.
 *
 * An  architecture dependant hal_vmm_display() function is defined in <hal_vmm.h> file.
 ********************************************************************************************/

typedef struct stack_mgr_s
{
    busylock_t     lock;               /*! lock protecting STACK allocator                  */
    vpn_t          vpn_base;           /*! first page of STACK zone                         */
    bitmap_t       bitmap;             /*! bit vector of allocated stacks                   */
}
stack_mgr_t;

/*********************************************************************************************
 * This structure defines the MMAP allocator used by the VMM to dynamically handle MMAP vsegs
 * requested or released by an user process. It must be called in the reference cluster.
 * - allocation policy : 
 *   This allocator implements the buddy algorithm. All allocated vsegs occupy an integer
 *   number of pages, that is power of 2, and are aligned (vpn_base is multiple of vpn_size).
 *   The requested number of pages is rounded if required. The global allocator state is
 *   completely defined by the free_pages_root[] array indexed by the vseg order.
 *   These free lists are local, but are implemented as xlist because we use the existing
 *   vseg.xlist to register a free vseg in its free list.
 * - release policy : 
 *   A released vseg is recursively merged with the "buddy" vseg when it is free, in
 *   order to build the largest possible aligned free vsegs. The resulting vseg.vpn_size
 *   field is updated.
 * Implementation note:
 * The only significant (and documented) fiels in the vsegs registered in the MMAP allocator
 * free lists are "xlist", "vpn_base", and "vpn_size".
 ********************************************************************************************/

typedef struct mmap_mgr_s
{
    busylock_t     lock;               /*! lock protecting MMAP allocator                   */
    vpn_t          vpn_base;           /*! first page of MMAP zone                          */
    vpn_t          vpn_size;           /*! number of pages in MMAP zone                     */
    xlist_entry_t  free_list_root[CONFIG_VMM_HEAP_MAX_ORDER + 1];  /* roots of free lists   */
}
mmap_mgr_t;

/*********************************************************************************************
 * This structure defines the Virtual Memory Manager for a given process in a given cluster.
 * This VMM implements four main services:
 * 1) It contains the local copy of the vseg list (VSL), only complete in reference cluster.
 * 2) It contains the local copy of the generic page table (GPT), only complete in reference.
 * 3) The stack manager dynamically allocates virtual memory space for the STACK vsegs.
 * 4) The mmap manager dynamically allocates virtual memory for the (FILE/ANON/REMOTE) vsegs.
 ******************************************************a**************************************
 * Implementation notes:
 * 1. In most clusters, the VSL and GPT are only partial copies of the reference VSL and GPT
 *    structures, stored in the reference cluster.
 * 2. The VSL contains only local vsegs, but it is implemented as an xlist, and protected by
 *    a remote_rwlock, because it can be accessed by a thread running in a remote cluster.
 *    An example is the vmm_fork_copy() function.
 * 3. The GPT in the reference cluster can be directly accessed by remote threads to handle
 *    false page-fault (page is mapped in the reference GPT, but the PTE copy is missing
 *    in the local GPT). As each PTE can be protected by a specific GPT_LOCKED attribute
 *    for exclusive access, it is NOT protected by a global lock. 
 ********************************************************************************************/

typedef struct vmm_s
{
        remote_queuelock_t vsl_lock;            /*! lock protecting the local VSL               */
        xlist_entry_t      vsegs_root;          /*! Virtual Segment List root                   */
        uint32_t           vsegs_nr;            /*! total number of local vsegs                 */

    gpt_t              gpt;                 /*! Generic Page Table descriptor               */

    stack_mgr_t        stack_mgr;           /*! embedded STACK vsegs allocator              */

    mmap_mgr_t         mmap_mgr;            /*! embedded MMAP vsegs allocator               */

        uint32_t           false_pgfault_nr;    /*! false page fault counter (for all threads)  */
        uint32_t           local_pgfault_nr;    /*! false page fault counter (for all threads)  */
        uint32_t           global_pgfault_nr;   /*! false page fault counter (for all threads)  */
    uint32_t           false_pgfault_cost;  /*! cumulated cost (for all threads)            */
    uint32_t           local_pgfault_cost;  /*! cumulated cost (for all threads)            */
    uint32_t           global_pgfault_cost; /*! cumulated cost (for all threads)            */

    vpn_t              args_vpn_base;       /*! args vseg first page                        */
    vpn_t              envs_vpn_base;       /*! envs vseg first page                        */
        vpn_t              code_vpn_base;       /*! code vseg first page                        */
        vpn_t              data_vpn_base;       /*! data vseg first page                        */
    vpn_t              heap_vpn_base;       /*! heap zone first page                        */

        intptr_t           entry_point;         /*! main thread entry point                     */
}
vmm_t;

/*********************************************************************************************
 * This function makes a partial initialisation of the VMM attached to an user process.
 * It intializes the STACK and MMAP allocators, the VSL lock, and the instrumentation
 * counters, but it does not register any vseg in the VSL:
 * - The "kernel" vsegs are registered, by the hal_vmm_kernel_update() function.
 * - The "args" & "envs" vsegs are registered by the process_make_exec() function.
 * - The "code" and "data" vsegs are registered by the elf_load_process() function.
 * - The "stack" vsegs are registered by process_make_exec() / thread_user_create().
 * - The "file", "anon", "remote" vsegs are dynamically registered by the mmap() syscall.
 *********************************************************************************************
 * @ process   : pointer on process descriptor
 * @ return 0 if success / return -1 if failure.
 ********************************************************************************************/
error_t vmm_user_init( struct process_s * process );

/*********************************************************************************************
 * This function is called by the process_make_exec() function to re-initialises the VMM
 * attached to an user process:
 * - It removes from the VMM of the process identified by the <process> argument all user
 *   vsegs, by calling the vmm_remove_vseg() function: the vsegs are removed from the VSL,
 *   the corresponding GPT entries are removed from the GPT, and the physical pages are 
 *   released to the relevant kmem when they are not shared.
 * - The VSL and the GPT are not modified for the kernel vsegs.
 * - Finally, it calls the vmm_user_init() function to re-initialize the STAK and MMAP
 *   allocators, and the lock protecting GPT.
 *********************************************************************************************
 * @ process   : pointer on process descriptor.
 ********************************************************************************************/
void vmm_user_reset( struct process_s * process );

/*********************************************************************************************
 * This function is called by the process_make_fork() function. It partially copies
 * the content of a remote parent process VMM to the local child process VMM:
 * - The KERNEL vsegs required by the architecture must have been previously
 *   created in the child VMM, using the hal_vmm_kernel_update() function.
 * - The DATA, ANON, REMOTE vsegs registered in the parent VSL are registered in the
 *   child VSL. All valid PTEs in parent GPT are copied to the child GPT.
 *   The WRITABLE  and COW flags are not modified, as it will be done later for those
 *   shared pages by the vmm_set_cow() function.
 * - The CODE vsegs registered in the parent VSL are registered in the child VSL, but the
 *   GPT entries are not copied in the child GPT, and will be dynamically updated from
 *   the .elf file when a page fault is reported.
 * - The FILE vsegs registered in the parent VSL are registered in the child VSL, and all
 *   valid GPT entries in parent GPT are copied to the child GPT. The COW flag is not set.
 * - No STACK vseg is copied from  parent VMM to child VMM: the child stack vseg is copied
 *   later from the cluster containing the user thread requesting the fork().
 *********************************************************************************************
 * @ child_process     : local pointer on local child process descriptor.
 * @ parent_process_xp : extended pointer on remote parent process descriptor.
 * @ return 0 if success / return -1 if failure.
 ********************************************************************************************/
error_t vmm_fork_copy( struct process_s * child_process,
                       xptr_t             parent_process_xp );

/*********************************************************************************************
 * This function is called by the process_make_fork() function to update the COW attribute
 * in the parent parent process vsegs. It set the COW flag, and reset the WRITABLE flag of
 * all GPT entries of the DATA, MMAP, and REMOTE vsegs of the <process> argument.
 * It must be called by a thread running in the reference cluster, that contains the complete
 * VSL and GPT (use the rpc_vmm_set_cow_client() when the calling thread client is remote). 
 * It updates all copies of the process in all clusters, to maintain coherence in GPT copies,
 * using the list of copies stored in the owner process, and using remote_write accesses to
 * update the remote GPTs. It atomically increment the pending_fork counter, in all involved
 * physical page descriptors. It cannot fail, as only mapped entries in GPTs are updated.
 *********************************************************************************************
 * @ process   : local pointer on local reference process descriptor.
 ********************************************************************************************/
void vmm_set_cow( struct process_s * process );

/*********************************************************************************************
 * This function modifies the vseg identified by <process> and <base> arguments in all
 * clusters containing a VSL copy, as defined by <new_base> and <new_size> arguments.
 * The new vseg, defined by the <new_base> and <new_size> arguments must be included
 * in the existing vseg. The target VSL size and base fields are modified in the VSL.
 * This is done in all clusters containing a VMM copy to maintain VMM coherence. 
 * It is called by the sys_munmap() and dev_fbf_resize_window() functions.
 * It can be called by a thread running in any cluster, as it uses the vmm_resize_vseg() in
 * the local cluster, and parallel RPC_VMM_RESIZE_VSEG for remote clusters.
 * It cannot fail, as only vseg registered  in VSL copies are updated.
 *********************************************************************************************
 * @ process   : local pointer on process descriptor.
 * @ base      : current vseg base address in user space.
 * @ new_base  : new vseg base.
 * @ new_size  : new vseg size.
 ********************************************************************************************/
void vmm_global_resize_vseg( struct process_s * process,
                             intptr_t           base,
                             intptr_t           new_base,
                             intptr_t           new_size );

/*********************************************************************************************
 * This function removes the vseg identified by the <process> and <base> arguments from
 * the VSL and remove all associated PTE entries from the GPT.
 * This is done in all clusters containing a VMM copy to maintain VMM coherence. 
 * It is called by the sys_munmap() and dev_fbf_resize_window() functions.
 * It can be called by a thread running in any cluster, as it uses the vmm_remove_vseg() in
 * the local cluster, and parallel RPC_VMM_REMOVE_VSEG for remote clusters.
 * It cannot fail, as only vseg registered  in VSL copies are deleted.
 *********************************************************************************************
 * @ pid      : local pointer on process identifier.
 * @ base     : vseg base address in user space.
 ********************************************************************************************/
void vmm_global_delete_vseg( struct process_s * process,
                             intptr_t           base );

/*********************************************************************************************
 * This function modifies one GPT entry identified by the <process> and <vpn> arguments
 * in all clusters containing a process copy. It maintains coherence in GPT copies,
 * using remote_write accesses.
 * It cannot fail, as only mapped PTE2 in GPT copies are updated.
 *********************************************************************************************
 * @ process   : local pointer on local process descriptor.
 * @ vpn       : PTE index.
 * @ attr      : PTE / attributes.
 * @ ppn       : PTE / physical page index.
 ********************************************************************************************/
void vmm_global_update_pte( struct process_s * process,
                            vpn_t              vpn,
                            uint32_t           attr,
                            ppn_t              ppn );

/*********************************************************************************************
 * This function deletes, in the local cluster, all vsegs registered in the VSL
 * of the process identified by the <process> argument. For each vseg:
 * - it unmaps all vseg PTEs from the GPT (release the physical pages when required). 
 * - it removes the vseg from the local VSL.
 * - it releases the memory allocated to the local vseg descriptors.
 * - it releases the memory allocated to the GPT itself.
 *********************************************************************************************
 * @ process   : pointer on process descriptor.
 ********************************************************************************************/
void vmm_destroy( struct process_s * process );

/*********************************************************************************************
 * This function scans the list of vsegs registered in the VMM of a given process descriptor
 * to check if a given virtual region (defined by a base and size) overlap an existing vseg.
 *********************************************************************************************
 * @ process  : pointer on process descriptor.
 * @ base     : region virtual base address.
 * @ size     : region size (bytes).
 * @ returns NULL if no conflict / return conflicting vseg pointer if conflict.
 ********************************************************************************************/
vseg_t * vmm_check_conflict( struct process_s * process,
                             vpn_t              base,
                             vpn_t              size );

/*********************************************************************************************
 * This function allocates memory for a vseg descriptor, initialises it, and register it
 * in the VSL of the local process descriptor.
 * - For the FILE, ANON, & REMOTE types, it does not use the <base> argument, but uses 
 *   the specific VMM MMAP allocator.
 * - For the STACK type, it does not use the <base> and <size> arguments,  but uses the
 *   the specific VMM STACK allocator.
 * It checks collision with pre-existing vsegs.
 * To comply with the "on-demand" paging policy, this function does NOT modify the GPT,
 * and does not allocate physical memory for vseg data.
 * It should be called by a local thread (could be a RPC thread if the client thread is not
 * running in the reference cluster).
 *********************************************************************************************
 * @ process     : pointer on local processor descriptor.
 * @ type        : vseg type.
 * @ base        : vseg base address (or user thread ltid for an user stack vseg).
 * @ size        : vseg size (bytes).
 * @ file_offset : offset in file for CODE, DATA, FILE types. 
 * @ file_size   : can be smaller than "size" for DATA type.
 * @ mapper_xp   : extended pointer on mapper for CODE, DATA, FILE types.
 * @ cxy         : physical mapping cluster (for non distributed vsegs).
 * @ returns pointer on vseg if success / returns NULL if no memory, or conflict.
 ********************************************************************************************/
vseg_t * vmm_create_vseg( struct process_s * process,
                          vseg_type_t        type,
                          intptr_t           base,
                              uint32_t           size,
                          uint32_t           file_offset,
                          uint32_t           file_size,
                          xptr_t             mapper_xp,
                          cxy_t              cxy );

/*********************************************************************************************
 * This function removes from the VMM of a local process descriptor, identified by
 * the <process> argument, the vseg identified by the <vseg> argument.  
 * It is called by the vmm_user_reset(), vmm_global_delete_vseg(), vmm_destroy() functions.
 * It must be called by a local thread, running in the cluster containing the modified VMM.
 * Use the RPC_VMM_REMOVE_VSEG if required.
 * It makes a kernel panic if the process is not registered in the local cluster.
 * For all vseg types, the vseg is detached from local VSL, and all associated PTEs are
 * unmapped from local GPT. Other actions depend on the vseg type:
 * Regarding the vseg descriptor release:
 *   . for ANON and REMOTE, the vseg is not released, but registered in local zombi_list.
 *   . for STACK the vseg is released to the local stack allocator.
 *   . for all other types, the vseg descriptor is released to the local kmem.
 * Regarding the physical pages release:
 *   . for KERNEL and FILE, the pages are not released to kmem.
 *   . for CODE and STACK, the pages are released to local kmem.
 *   . for DATA, ANON and REMOTE, the pages are released to relevant kmem only when
 *     the local cluster is the reference cluster.
 * The VSL lock protecting the VSL must be taken by the caller.
 *********************************************************************************************
 * @ process  : local pointer on process descriptor.
 * @ vseg     : local pointer on target vseg. 
 ********************************************************************************************/
void vmm_remove_vseg( struct process_s * process,
                      struct vseg_s    * vseg );

/*********************************************************************************************
 * This function resize a local vseg identified by the <process> and <vseg> arguments.
 * Both the "size" and "base" fields are modified in the process VSL. When the new vseg
 * contains less pages than the target vseg, the relevant pages are removed from the GPT. 
 * It is called by the vmm_global_resize() and dev_fbf_resize_window() functions.
 * It must be called by a local thread, running in the cluster containing the modified VSL.
 * Use the RPC_VMM_RESIZE_VSEG if required.
 * The VSL lock protecting the VSL must be taken by the caller.
 *********************************************************************************************
 * @ process   : local pointer on process descriptor
 * @ vseg      : local pointer on target vseg
 * @ new_base  : vseg base address
 * @ new_size  : vseg size (bytes)
 ********************************************************************************************/
void vmm_resize_vseg( struct process_s * process,
                      struct vseg_s    * vseg,
                      intptr_t           new_base,
                      intptr_t           new_size );

/*********************************************************************************************
 * This function checks that a given virtual address <vaddr> in a given <process> is
 * contained in a registered vseg. It can be called by any thread running in any cluster.
 * - if the vseg is registered in the local process VSL, it returns the local vseg pointer.
 * - if the vseg is missing in local VSL, it access directly the reference VSL.
 * - if the vseg is found in reference VSL, it updates the local VSL and returns this pointer.
 * It returns an error when the vseg is missing in the reference VMM, or when there is 
 * not enough memory for a new vseg descriptor in the calling thread cluster.
 * For both the local and the reference VSL, it takes the VSL lock before scanning the VSL.
 *********************************************************************************************
 * @ process   : [in] pointer on process descriptor.
 * @ vaddr     : [in] virtual address.
 * @ vseg      : [out] local pointer on local vseg.
 * @ returns 0 if success / returns -1 if user error 
 ********************************************************************************************/
error_t vmm_get_vseg( struct process_s  * process,
                      intptr_t            vaddr,
                      vseg_t           ** vseg );            

/*********************************************************************************************
 * This function is called by the generic exception handler in case of page-fault event,
 * detected for a given <vpn>. The <process> argument is used to access the relevant VMM.
 * It checks the missing VPN and returns an user error if it is not in a registered vseg.
 * For a legal VPN, there is actually 3 cases:
 * 1) if the missing VPN belongs to a private vseg (STACK or CODE segment types, non
 *    replicated in all clusters), it allocates a new physical page, computes the attributes,
 *    depending on vseg type, and updates directly the local GPT.
 * 2) if the missing VPN belongs to a public vseg, it can be a false page-fault, when the VPN
 *    is mapped in the reference GPT, but not in the local GPT. For this false page-fault,
 *    the local GPT is simply updated from the reference GPT.
 * 3) if the missing VPN is public, and unmapped in the ref GPT, it is a true page fault.
 *    The calling thread  allocates a new physical page, computes the attributes, depending
 *    on vseg type, and updates directly (without RPC) the local GPT and the reference GPT.
 *    Other GPT copies  will updated on demand.
 * Concurrent accesses to the GPT(s) are handled, by locking the target PTE before accessing
 * the local and/or reference GPT(s).
 *********************************************************************************************
 * @ process  : local pointer on local process.
 * @ vpn      : VPN of the missing PTE.
 * @ returns EXCP_NON_FATAL / EXCP_USER_ERROR / EXCP_KERNEL_PANIC after analysis
 ********************************************************************************************/
error_t vmm_handle_page_fault( struct process_s * process,
                               vpn_t              vpn );

/*********************************************************************************************
 * This function is called by the generic exception handler in case of WRITE violation event, 
 * detected for a given <vpn>. The <process> argument is used to access the relevant VMM.
 * It returns a kernel panic if the faulty VPN is not in a registered vseg, or is not mapped.
 * For a legal mapped vseg there is two cases:
 * 1) If the missing VPN belongs to a private vseg (STACK), it access only the local GPT.
 *    It access the forks counter in the current physical page descriptor.
 *    If there is a pending fork, it allocates a new physical page from the cluster defined 
 *    by the vseg type, copies the old physical page content to the new physical page,
 *    and decrements the pending_fork counter in old physical page descriptor.
 *    Finally, it reset the COW flag and set the WRITE flag in local GPT.
 * 2) If the missing VPN is public, it access only the reference GPT.
 *    It access the forks counter in the current physical page descriptor.
 *    If there is a pending fork, it allocates a new physical page from the cluster defined 
 *    by the vseg type, copies the old physical page content to the new physical page,
 *    and decrements the pending_fork counter in old physical page descriptor.
 *    Finally it calls the vmm_global_update_pte() function to reset the COW flag and set
 *    the WRITE flag in all the GPT copies, using a RPC if the reference cluster is remote.
 * In both cases, concurrent accesses to the GPT are handled by locking the target PTE
 * before accessing the GPT.
 *********************************************************************************************
 * @ process   : pointer on local process descriptor copy.
 * @ vpn       : VPN of the faulting PTE.
 * @ returns EXCP_NON_FATAL / EXCP_USER_ERROR / EXCP_KERNEL_PANIC after analysis
 ********************************************************************************************/
error_t vmm_handle_cow( struct process_s * process,
                        vpn_t              vpn );

/*********************************************************************************************
 * This function is called by the vmm_get_pte() function when a page is unmapped.
 * Depending on the vseg type, defined by the <vseg> argument, it returns the PPN 
 * (Physical Page Number) associated to a missing page defined by the <vpn> argument.
 * - For the FILE type, it returns directly the physical page from the file mapper.
 * - For the CODE and DATA types, it allocates a new physical page from the cluster defined
 *   by the <vseg->cxy> field, or by the <vpn> MSB bits for a distributed vseg,
 *   and initialize this page from the .elf file mapper.
 * - For all other types, it allocates a new physical page from the cluster defined
 *   by the <vseg->cxy> field, or by the <vpn> MSB bits for a distributed vseg,
 *   but the new page is not initialized.
 *********************************************************************************************
 * @ vseg   : local pointer on vseg containing the mising page.
 * @ vpn    : Virtual Page Number identifying the missing page.
 * @ ppn    : [out] returned Physical Page Number.
 * return 0 if success / return EINVAL or ENOMEM if error.
 ********************************************************************************************/
error_t vmm_get_one_ppn( vseg_t * vseg,
                         vpn_t    vpn,
                         ppn_t  * ppn );

#endif /* _VMM_H_ */