/* * vmm.c - virtual memory manager related operations interface. * * Authors Ghassan Almaless (2008,2009,2010,2011, 2012) * Mohamed Lamine Karaoui (2015) * Alain Greiner (2016,2017,2018) * * 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 */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include ////////////////////////////////////////////////////////////////////////////////// // Extern global variables ////////////////////////////////////////////////////////////////////////////////// extern process_t process_zero; // allocated in cluster.c /////////////////////////////////////// error_t vmm_init( process_t * process ) { error_t error; vseg_t * vseg_kentry; vseg_t * vseg_args; vseg_t * vseg_envs; intptr_t base; intptr_t size; #if DEBUG_VMM_INIT thread_t * this = CURRENT_THREAD; uint32_t cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_INIT ) printk("\n[%s] thread[%x,%x] enter for process %x / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, process->pid , cycle ); #endif // get pointer on VMM vmm_t * vmm = &process->vmm; // initialize local list of vsegs vmm->vsegs_nr = 0; xlist_root_init( XPTR( local_cxy , &vmm->vsegs_root ) ); remote_rwlock_init( XPTR( local_cxy , &vmm->vsegs_lock ) , LOCK_VMM_VSL ); assert( ((CONFIG_VMM_KENTRY_SIZE + CONFIG_VMM_ARGS_SIZE + CONFIG_VMM_ENVS_SIZE) <= CONFIG_VMM_ELF_BASE) , "UTILS zone too small\n" ); assert( (CONFIG_THREADS_MAX_PER_CLUSTER <= 32) , "no more than 32 threads per cluster for a single process\n"); assert( ((CONFIG_VMM_STACK_SIZE * CONFIG_THREADS_MAX_PER_CLUSTER) <= (CONFIG_VMM_VSPACE_SIZE - CONFIG_VMM_STACK_BASE)) , "STACK zone too small\n"); // register kentry vseg in VSL base = CONFIG_VMM_KENTRY_BASE << CONFIG_PPM_PAGE_SHIFT; size = CONFIG_VMM_KENTRY_SIZE << CONFIG_PPM_PAGE_SHIFT; vseg_kentry = vmm_create_vseg( process, VSEG_TYPE_CODE, base, size, 0, // file_offset unused 0, // file_size unused XPTR_NULL, // mapper_xp unused local_cxy ); if( vseg_kentry == NULL ) { printk("\n[ERROR] in %s : cannot register kentry vseg\n", __FUNCTION__ ); return -1; } vmm->kent_vpn_base = base; // register args vseg in VSL base = (CONFIG_VMM_KENTRY_BASE + CONFIG_VMM_KENTRY_SIZE ) << CONFIG_PPM_PAGE_SHIFT; size = CONFIG_VMM_ARGS_SIZE << CONFIG_PPM_PAGE_SHIFT; vseg_args = vmm_create_vseg( process, VSEG_TYPE_DATA, base, size, 0, // file_offset unused 0, // file_size unused XPTR_NULL, // mapper_xp unused local_cxy ); if( vseg_args == NULL ) { printk("\n[ERROR] in %s : cannot register args vseg\n", __FUNCTION__ ); return -1; } vmm->args_vpn_base = base; // register the envs vseg in VSL base = (CONFIG_VMM_KENTRY_BASE + CONFIG_VMM_KENTRY_SIZE + CONFIG_VMM_ARGS_SIZE ) << CONFIG_PPM_PAGE_SHIFT; size = CONFIG_VMM_ENVS_SIZE << CONFIG_PPM_PAGE_SHIFT; vseg_envs = vmm_create_vseg( process, VSEG_TYPE_DATA, base, size, 0, // file_offset unused 0, // file_size unused XPTR_NULL, // mapper_xp unused local_cxy ); if( vseg_envs == NULL ) { printk("\n[ERROR] in %s : cannot register envs vseg\n", __FUNCTION__ ); return -1; } vmm->envs_vpn_base = base; // create GPT (empty) error = hal_gpt_create( &vmm->gpt ); if( error ) printk("\n[ERROR] in %s : cannot create GPT\n", __FUNCTION__ ); // initialize GPT lock remote_rwlock_init( XPTR( local_cxy , &vmm->gpt_lock ) , LOCK_VMM_GPT ); // architecture specic GPT initialisation // (For TSAR, identity map the kentry_vseg) error = hal_vmm_init( vmm ); if( error ) printk("\n[ERROR] in %s : cannot initialize GPT\n", __FUNCTION__ ); // initialize STACK allocator vmm->stack_mgr.bitmap = 0; vmm->stack_mgr.vpn_base = CONFIG_VMM_STACK_BASE; busylock_init( &vmm->stack_mgr.lock , LOCK_VMM_STACK ); // initialize MMAP allocator vmm->mmap_mgr.vpn_base = CONFIG_VMM_HEAP_BASE; vmm->mmap_mgr.vpn_size = CONFIG_VMM_STACK_BASE - CONFIG_VMM_HEAP_BASE; vmm->mmap_mgr.first_free_vpn = CONFIG_VMM_HEAP_BASE; busylock_init( &vmm->mmap_mgr.lock , LOCK_VMM_MMAP ); uint32_t i; for( i = 0 ; i < 32 ; i++ ) list_root_init( &vmm->mmap_mgr.zombi_list[i] ); // initialize instrumentation counters vmm->pgfault_nr = 0; hal_fence(); #if DEBUG_VMM_INIT cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_INIT ) printk("\n[%s] thread[%x,%x] exit / process %x / entry_point %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid, process->vmm.entry_point, cycle ); #endif return 0; } // end vmm_init() ////////////////////////////////////// void vmm_display( process_t * process, bool_t mapping ) { vmm_t * vmm = &process->vmm; gpt_t * gpt = &vmm->gpt; printk("\n***** VSL and GPT(%x) for process %x in cluster %x\n\n", process->vmm.gpt.ptr , process->pid , local_cxy ); // get lock protecting the VSL and the GPT remote_rwlock_rd_acquire( XPTR( local_cxy , &vmm->vsegs_lock ) ); remote_rwlock_rd_acquire( XPTR( local_cxy , &vmm->gpt_lock ) ); // scan the list of vsegs xptr_t root_xp = XPTR( local_cxy , &vmm->vsegs_root ); xptr_t iter_xp; xptr_t vseg_xp; vseg_t * vseg; XLIST_FOREACH( root_xp , iter_xp ) { vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist ); vseg = GET_PTR( vseg_xp ); printk(" - %s : base = %X / size = %X / npages = %d\n", vseg_type_str( vseg->type ) , vseg->min , vseg->max - vseg->min , vseg->vpn_size ); if( mapping ) { vpn_t vpn; ppn_t ppn; uint32_t attr; vpn_t base = vseg->vpn_base; vpn_t size = vseg->vpn_size; for( vpn = base ; vpn < (base+size) ; vpn++ ) { hal_gpt_get_pte( XPTR( local_cxy , gpt ) , vpn , &attr , &ppn ); if( attr & GPT_MAPPED ) { printk(" . vpn = %X / attr = %X / ppn = %X\n", vpn , attr , ppn ); } } } } // release the locks remote_rwlock_rd_release( XPTR( local_cxy , &vmm->vsegs_lock ) ); remote_rwlock_rd_release( XPTR( local_cxy , &vmm->gpt_lock ) ); } // vmm_display() /////////////////////////////////// void vmm_vseg_attach( vmm_t * vmm, vseg_t * vseg ) { // build extended pointer on rwlock protecting VSL xptr_t lock_xp = XPTR( local_cxy , &vmm->vsegs_lock ); // get rwlock in write mode remote_rwlock_wr_acquire( lock_xp ); // update vseg descriptor vseg->vmm = vmm; // add vseg in vmm list xlist_add_last( XPTR( local_cxy , &vmm->vsegs_root ), XPTR( local_cxy , &vseg->xlist ) ); // release rwlock in write mode remote_rwlock_wr_release( lock_xp ); } /////////////////////////////////// void vmm_vseg_detach( vmm_t * vmm, vseg_t * vseg ) { // build extended pointer on rwlock protecting VSL xptr_t lock_xp = XPTR( local_cxy , &vmm->vsegs_lock ); // get rwlock in write mode remote_rwlock_wr_acquire( lock_xp ); // update vseg descriptor vseg->vmm = NULL; // remove vseg from vmm list xlist_unlink( XPTR( local_cxy , &vseg->xlist ) ); // release rwlock in write mode remote_rwlock_wr_release( lock_xp ); } //////////////////////////////////////////////// void vmm_global_update_pte( process_t * process, vpn_t vpn, uint32_t attr, ppn_t ppn ) { xlist_entry_t * process_root_ptr; xptr_t process_root_xp; xptr_t process_iter_xp; xptr_t remote_process_xp; cxy_t remote_process_cxy; process_t * remote_process_ptr; xptr_t remote_gpt_xp; pid_t pid; cxy_t owner_cxy; lpid_t owner_lpid; #if DEBUG_VMM_UPDATE_PTE uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_UPDATE_PTE < cycle ) printk("\n[%s] thread[%x,%x] enter for process %x / vpn %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid , vpn , cycle ); #endif // check cluster is reference assert( (GET_CXY( process->ref_xp ) == local_cxy) , "not called in reference cluster\n"); // get extended pointer on root of process copies xlist in owner cluster pid = process->pid; owner_cxy = CXY_FROM_PID( pid ); owner_lpid = LPID_FROM_PID( pid ); process_root_ptr = &LOCAL_CLUSTER->pmgr.copies_root[owner_lpid]; process_root_xp = XPTR( owner_cxy , process_root_ptr ); // loop on destination process copies XLIST_FOREACH( process_root_xp , process_iter_xp ) { // get cluster and local pointer on remote process remote_process_xp = XLIST_ELEMENT( process_iter_xp , process_t , copies_list ); remote_process_ptr = GET_PTR( remote_process_xp ); remote_process_cxy = GET_CXY( remote_process_xp ); #if (DEBUG_VMM_UPDATE_PTE & 0x1) if( DEBUG_VMM_UPDATE_PTE < cycle ) printk("\n[%s] threadr[%x,%x] handling vpn %x for process %x in cluster %x\n", __FUNCTION__, this->process->pid, this->trdid, vpn, process->pid, remote_process_cxy ); #endif // get extended pointer on remote gpt remote_gpt_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.gpt ); // update remote GPT hal_gpt_update_pte( remote_gpt_xp, vpn, attr, ppn ); } #if DEBUG_VMM_UPDATE_PTE cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_UPDATE_PTE < cycle ) printk("\n[%s] thread[%x,%x] exit for process %x / vpn %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid , vpn , cycle ); #endif } // end vmm_global_update_pte() /////////////////////////////////////// void vmm_set_cow( process_t * process ) { vmm_t * vmm; xlist_entry_t * process_root_ptr; xptr_t process_root_xp; xptr_t process_iter_xp; xptr_t remote_process_xp; cxy_t remote_process_cxy; process_t * remote_process_ptr; xptr_t remote_gpt_xp; xptr_t vseg_root_xp; xptr_t vseg_iter_xp; xptr_t vseg_xp; vseg_t * vseg; pid_t pid; cxy_t owner_cxy; lpid_t owner_lpid; #if DEBUG_VMM_SET_COW uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_SET_COW < cycle ) printk("\n[%s] thread[%x,%x] enter for process %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid , cycle ); #endif // check cluster is reference assert( (GET_CXY( process->ref_xp ) == local_cxy) , "local cluster is not process reference cluster\n"); // get pointer on reference VMM vmm = &process->vmm; // get extended pointer on root of process copies xlist in owner cluster pid = process->pid; owner_cxy = CXY_FROM_PID( pid ); owner_lpid = LPID_FROM_PID( pid ); process_root_ptr = &LOCAL_CLUSTER->pmgr.copies_root[owner_lpid]; process_root_xp = XPTR( owner_cxy , process_root_ptr ); // get extended pointer on root of vsegs xlist from reference VMM vseg_root_xp = XPTR( local_cxy , &vmm->vsegs_root ); // loop on destination process copies XLIST_FOREACH( process_root_xp , process_iter_xp ) { // get cluster and local pointer on remote process remote_process_xp = XLIST_ELEMENT( process_iter_xp , process_t , copies_list ); remote_process_ptr = GET_PTR( remote_process_xp ); remote_process_cxy = GET_CXY( remote_process_xp ); #if (DEBUG_VMM_SET_COW & 1) if( DEBUG_VMM_SET_COW < cycle ) printk("\n[%s] thread[%x,%x] handling process %x in cluster %x\n", __FUNCTION__, this->process->pid, this->trdid, process->pid , remote_process_cxy ); #endif // get extended pointer on remote gpt remote_gpt_xp = XPTR( remote_process_cxy , &remote_process_ptr->vmm.gpt ); // loop on vsegs in (local) reference process VSL XLIST_FOREACH( vseg_root_xp , vseg_iter_xp ) { // get pointer on vseg vseg_xp = XLIST_ELEMENT( vseg_iter_xp , vseg_t , xlist ); vseg = GET_PTR( vseg_xp ); assert( (GET_CXY( vseg_xp ) == local_cxy) , "all vsegs in reference VSL must be local\n" ); // get vseg type, base and size uint32_t type = vseg->type; vpn_t vpn_base = vseg->vpn_base; vpn_t vpn_size = vseg->vpn_size; #if (DEBUG_VMM_SET_COW & 1) if( DEBUG_VMM_SET_COW < cycle ) printk("\n[%s] thread[%x,%x] handling vseg %s / vpn_base = %x / vpn_size = %x\n", __FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), vpn_base, vpn_size ); #endif // only DATA, ANON and REMOTE vsegs if( (type == VSEG_TYPE_DATA) || (type == VSEG_TYPE_ANON) || (type == VSEG_TYPE_REMOTE) ) { vpn_t vpn; uint32_t attr; ppn_t ppn; xptr_t page_xp; cxy_t page_cxy; page_t * page_ptr; xptr_t forks_xp; xptr_t lock_xp; // update flags in remote GPT hal_gpt_set_cow( remote_gpt_xp, vpn_base, vpn_size ); // atomically increment pending forks counter in physical pages, // for all vseg pages that are mapped in reference cluster if( remote_process_cxy == local_cxy ) { // scan all pages in vseg for( vpn = vpn_base ; vpn < (vpn_base + vpn_size) ; vpn++ ) { // get page attributes and PPN from reference GPT hal_gpt_get_pte( remote_gpt_xp , vpn , &attr , &ppn ); // atomically update pending forks counter if page is mapped if( attr & GPT_MAPPED ) { // get pointers and cluster on page descriptor page_xp = ppm_ppn2page( ppn ); page_cxy = GET_CXY( page_xp ); page_ptr = GET_PTR( page_xp ); // get extended pointers on "forks" and "lock" forks_xp = XPTR( page_cxy , &page_ptr->forks ); lock_xp = XPTR( page_cxy , &page_ptr->lock ); // take lock protecting "forks" counter remote_busylock_acquire( lock_xp ); // increment "forks" hal_remote_atomic_add( forks_xp , 1 ); // release lock protecting "forks" counter remote_busylock_release( lock_xp ); } } // end loop on vpn } // end if local } // end if vseg type } // end loop on vsegs } // end loop on process copies #if DEBUG_VMM_SET_COW cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_SET_COW < cycle ) printk("\n[%s] thread[%x,%x] exit for process %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid , cycle ); #endif } // end vmm_set-cow() ///////////////////////////////////////////////// error_t vmm_fork_copy( process_t * child_process, xptr_t parent_process_xp ) { error_t error; cxy_t parent_cxy; process_t * parent_process; vmm_t * parent_vmm; xptr_t parent_lock_xp; vmm_t * child_vmm; xptr_t iter_xp; xptr_t parent_vseg_xp; vseg_t * parent_vseg; vseg_t * child_vseg; uint32_t type; bool_t cow; vpn_t vpn; vpn_t vpn_base; vpn_t vpn_size; xptr_t page_xp; // extended pointer on page descriptor page_t * page_ptr; cxy_t page_cxy; xptr_t forks_xp; // extended pointer on forks counter in page descriptor xptr_t lock_xp; // extended pointer on lock protecting the forks counter xptr_t parent_root_xp; bool_t mapped; ppn_t ppn; #if DEBUG_VMM_FORK_COPY uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_FORK_COPY < cycle ) printk("\n[%s] thread %x enter / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, cycle ); #endif // get parent process cluster and local pointer parent_cxy = GET_CXY( parent_process_xp ); parent_process = GET_PTR( parent_process_xp ); // get local pointers on parent and child VMM parent_vmm = &parent_process->vmm; child_vmm = &child_process->vmm; // get extended pointer on lock protecting the parent VSL parent_lock_xp = XPTR( parent_cxy , &parent_vmm->vsegs_lock ); // initialize the lock protecting the child VSL remote_rwlock_init( XPTR( local_cxy , &child_vmm->vsegs_lock ), LOCK_VMM_STACK ); // initialize the child VSL as empty xlist_root_init( XPTR( local_cxy, &child_vmm->vsegs_root ) ); child_vmm->vsegs_nr = 0; // create child GPT error = hal_gpt_create( &child_vmm->gpt ); if( error ) { printk("\n[ERROR] in %s : cannot create GPT\n", __FUNCTION__ ); return -1; } // build extended pointer on parent VSL parent_root_xp = XPTR( parent_cxy , &parent_vmm->vsegs_root ); // take the lock protecting the parent VSL in read mode remote_rwlock_rd_acquire( parent_lock_xp ); // loop on parent VSL xlist XLIST_FOREACH( parent_root_xp , iter_xp ) { // get local and extended pointers on current parent vseg parent_vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist ); parent_vseg = GET_PTR( parent_vseg_xp ); // get vseg type type = hal_remote_l32( XPTR( parent_cxy , &parent_vseg->type ) ); #if DEBUG_VMM_FORK_COPY cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_FORK_COPY < cycle ) printk("\n[%s] thread[%x,%x] found parent vseg %s / vpn_base = %x / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, vseg_type_str(type), hal_remote_l32( XPTR( parent_cxy , &parent_vseg->vpn_base ) ) , cycle ); #endif // all parent vsegs - but STACK - must be copied in child VSL if( type != VSEG_TYPE_STACK ) { // allocate memory for a new child vseg child_vseg = vseg_alloc(); if( child_vseg == NULL ) // release all allocated vsegs { vmm_destroy( child_process ); printk("\n[ERROR] in %s : cannot create vseg for child\n", __FUNCTION__ ); return -1; } // copy parent vseg to child vseg vseg_init_from_ref( child_vseg , parent_vseg_xp ); // register child vseg in child VSL vmm_vseg_attach( child_vmm , child_vseg ); #if DEBUG_VMM_FORK_COPY cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_FORK_COPY < cycle ) printk("\n[%s] thread[%x,%x] copied vseg %s / vpn_base = %x to child VSL / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, vseg_type_str(type), hal_remote_l32( XPTR( parent_cxy , &parent_vseg->vpn_base ) ) , cycle ); #endif // copy DATA, MMAP, REMOTE, FILE parent GPT entries to child GPT if( type != VSEG_TYPE_CODE ) { // activate the COW for DATA, MMAP, REMOTE vsegs only cow = ( type != VSEG_TYPE_FILE ); vpn_base = child_vseg->vpn_base; vpn_size = child_vseg->vpn_size; // scan pages in parent vseg for( vpn = vpn_base ; vpn < (vpn_base + vpn_size) ; vpn++ ) { error = hal_gpt_pte_copy( &child_vmm->gpt, XPTR( parent_cxy , &parent_vmm->gpt ), vpn, cow, &ppn, &mapped ); if( error ) { vmm_destroy( child_process ); printk("\n[ERROR] in %s : cannot copy GPT\n", __FUNCTION__ ); return -1; } // increment pending forks counter in page if mapped if( mapped ) { // get pointers and cluster on page descriptor page_xp = ppm_ppn2page( ppn ); page_cxy = GET_CXY( page_xp ); page_ptr = GET_PTR( page_xp ); // get extended pointers on "forks" and "lock" forks_xp = XPTR( page_cxy , &page_ptr->forks ); lock_xp = XPTR( page_cxy , &page_ptr->lock ); // get lock protecting "forks" counter remote_busylock_acquire( lock_xp ); // increment "forks" hal_remote_atomic_add( forks_xp , 1 ); // release lock protecting "forks" counter remote_busylock_release( lock_xp ); #if DEBUG_VMM_FORK_COPY cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_FORK_COPY < cycle ) printk("\n[%s] thread[%x,%x] copied vpn %x to child GPT / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid , vpn , cycle ); #endif } } } // end if no code & no stack } // end if no stack } // end loop on vsegs // release the parent VSL lock in read mode remote_rwlock_rd_release( parent_lock_xp ); // initialize child GPT (architecture specic) // => For TSAR, identity map the kentry_vseg error = hal_vmm_init( child_vmm ); if( error ) { printk("\n[ERROR] in %s : cannot create GPT\n", __FUNCTION__ ); return -1; } // initialize the child VMM STACK allocator child_vmm->stack_mgr.bitmap = 0; child_vmm->stack_mgr.vpn_base = CONFIG_VMM_STACK_BASE; // initialize the child VMM MMAP allocator uint32_t i; child_vmm->mmap_mgr.vpn_base = CONFIG_VMM_HEAP_BASE; child_vmm->mmap_mgr.vpn_size = CONFIG_VMM_STACK_BASE - CONFIG_VMM_HEAP_BASE; child_vmm->mmap_mgr.first_free_vpn = CONFIG_VMM_HEAP_BASE; for( i = 0 ; i < 32 ; i++ ) list_root_init( &child_vmm->mmap_mgr.zombi_list[i] ); // initialize instrumentation counters child_vmm->pgfault_nr = 0; // copy base addresses from parent VMM to child VMM child_vmm->kent_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->kent_vpn_base)); child_vmm->args_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->args_vpn_base)); child_vmm->envs_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->envs_vpn_base)); child_vmm->heap_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->heap_vpn_base)); child_vmm->code_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->code_vpn_base)); child_vmm->data_vpn_base = (vpn_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->data_vpn_base)); child_vmm->entry_point = (intptr_t)hal_remote_lpt(XPTR(parent_cxy, &parent_vmm->entry_point)); hal_fence(); #if DEBUG_VMM_FORK_COPY cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_FORK_COPY < cycle ) printk("\n[%s] thread[%x,%x] exit successfully / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid , cycle ); #endif return 0; } // vmm_fork_copy() /////////////////////////////////////// void vmm_destroy( process_t * process ) { xptr_t vseg_xp; vseg_t * vseg; #if DEBUG_VMM_DESTROY uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_DESTROY < cycle ) printk("\n[%s] thread[%x,%x] enter for process %x in cluster %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy, cycle ); #endif #if (DEBUG_VMM_DESTROY & 1 ) if( DEBUG_VMM_DESTROY < cycle ) vmm_display( process , true ); #endif // get pointer on local VMM vmm_t * vmm = &process->vmm; // get extended pointer on VSL root and VSL lock xptr_t root_xp = XPTR( local_cxy , &vmm->vsegs_root ); // remove all user vsegs registered in VSL while( !xlist_is_empty( root_xp ) ) { // get pointer on first vseg in VSL vseg_xp = XLIST_FIRST( root_xp , vseg_t , xlist ); vseg = GET_PTR( vseg_xp ); // unmap and release physical pages vmm_unmap_vseg( process , vseg ); // remove vseg from VSL vmm_vseg_detach( vmm , vseg ); // release memory allocated to vseg descriptor vseg_free( vseg ); #if( DEBUG_VMM_DESTROY & 1 ) if( DEBUG_VMM_DESTROY < cycle ) printk("\n[%s] %s vseg released / vpn_base %x / vpn_size %d\n", __FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size ); #endif } // remove all vsegs from zombi_lists in MMAP allocator uint32_t i; for( i = 0 ; i<32 ; i++ ) { while( !list_is_empty( &vmm->mmap_mgr.zombi_list[i] ) ) { vseg = LIST_FIRST( &vmm->mmap_mgr.zombi_list[i] , vseg_t , zlist ); #if( DEBUG_VMM_DESTROY & 1 ) if( DEBUG_VMM_DESTROY < cycle ) printk("\n[%s] found zombi vseg / vpn_base %x / vpn_size %d\n", __FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size ); #endif vmm_vseg_detach( vmm , vseg ); vseg_free( vseg ); #if( DEBUG_VMM_DESTROY & 1 ) if( DEBUG_VMM_DESTROY < cycle ) printk("\n[%s] zombi vseg released / vpn_base %x / vpn_size %d\n", __FUNCTION__ , vseg_type_str( vseg->type ), vseg->vpn_base, vseg->vpn_size ); #endif } } // release memory allocated to the GPT itself hal_gpt_destroy( &vmm->gpt ); #if DEBUG_VMM_DESTROY cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_DESTROY < cycle ) printk("\n[%s] thread[%x,%x] exit for process %x in cluster %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid, local_cxy , cycle ); #endif } // end vmm_destroy() ///////////////////////////////////////////////// vseg_t * vmm_check_conflict( process_t * process, vpn_t vpn_base, vpn_t vpn_size ) { vmm_t * vmm = &process->vmm; // scan the VSL vseg_t * vseg; xptr_t iter_xp; xptr_t vseg_xp; xptr_t root_xp = XPTR( local_cxy , &vmm->vsegs_root ); XLIST_FOREACH( root_xp , iter_xp ) { vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist ); vseg = GET_PTR( vseg_xp ); if( ((vpn_base + vpn_size) > vseg->vpn_base) && (vpn_base < (vseg->vpn_base + vseg->vpn_size)) ) return vseg; } return NULL; } // end vmm_check_conflict() //////////////////////////////////////////////////////////////////////////////////////////// // This static function is called by the vmm_create_vseg() function, and implements // the VMM stack_vseg specific allocator. //////////////////////////////////////////////////////////////////////////////////////////// // @ vmm : pointer on VMM. // @ vpn_base : (return value) first allocated page // @ vpn_size : (return value) number of allocated pages //////////////////////////////////////////////////////////////////////////////////////////// static error_t vmm_stack_alloc( vmm_t * vmm, vpn_t * vpn_base, vpn_t * vpn_size ) { // get stack allocator pointer stack_mgr_t * mgr = &vmm->stack_mgr; // get lock on stack allocator busylock_acquire( &mgr->lock ); // get first free slot index in bitmap int32_t index = bitmap_ffc( &mgr->bitmap , 4 ); if( (index < 0) || (index > 31) ) { busylock_release( &mgr->lock ); return 0xFFFFFFFF; } // update bitmap bitmap_set( &mgr->bitmap , index ); // release lock on stack allocator busylock_release( &mgr->lock ); // returns vpn_base, vpn_size (one page non allocated) *vpn_base = mgr->vpn_base + index * CONFIG_VMM_STACK_SIZE + 1; *vpn_size = CONFIG_VMM_STACK_SIZE - 1; return 0; } // end vmm_stack_alloc() //////////////////////////////////////////////////////////////////////////////////////////// // This static function is called by the vmm_create_vseg() function, and implements // the VMM MMAP specific allocator. //////////////////////////////////////////////////////////////////////////////////////////// // @ vmm : [in] pointer on VMM. // @ npages : [in] requested number of pages. // @ vpn_base : [out] first allocated page. // @ vpn_size : [out] actual number of allocated pages. //////////////////////////////////////////////////////////////////////////////////////////// static error_t vmm_mmap_alloc( vmm_t * vmm, vpn_t npages, vpn_t * vpn_base, vpn_t * vpn_size ) { uint32_t index; vseg_t * vseg; vpn_t base; vpn_t size; vpn_t free; // mmap vseg size must be power of 2 // compute actual size and index in zombi_list array size = POW2_ROUNDUP( npages ); index = bits_log2( size ); // get mmap allocator pointer mmap_mgr_t * mgr = &vmm->mmap_mgr; // get lock on mmap allocator busylock_acquire( &mgr->lock ); // get vseg from zombi_list or from mmap zone if( list_is_empty( &mgr->zombi_list[index] ) ) // from mmap zone { // check overflow free = mgr->first_free_vpn; if( (free + size) > mgr->vpn_size ) return ENOMEM; // update STACK allocator mgr->first_free_vpn += size; // compute base base = free; } else // from zombi_list { // get pointer on zombi vseg from zombi_list vseg = LIST_FIRST( &mgr->zombi_list[index] , vseg_t , zlist ); // remove vseg from free-list list_unlink( &vseg->zlist ); // compute base base = vseg->vpn_base; } // release lock on mmap allocator busylock_release( &mgr->lock ); // returns vpn_base, vpn_size *vpn_base = base; *vpn_size = size; return 0; } // end vmm_mmap_alloc() //////////////////////////////////////////////// vseg_t * vmm_create_vseg( process_t * 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 ) { vseg_t * vseg; // created vseg pointer vpn_t vpn_base; // first page index vpn_t vpn_size; // number of pages covered by vseg error_t error; #if DEBUG_VMM_CREATE_VSEG thread_t * this = CURRENT_THREAD; uint32_t cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_CREATE_VSEG < cycle ) printk("\n[%s] thread[%x,%x] enter / %s / cxy %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), cxy, cycle ); #endif // get pointer on VMM vmm_t * vmm = &process->vmm; // compute base, size, vpn_base, vpn_size, depending on vseg type // we use the VMM specific allocators for "stack", "file", "anon", & "remote" vsegs if( type == VSEG_TYPE_STACK ) { // get vpn_base and vpn_size from STACK allocator error = vmm_stack_alloc( vmm , &vpn_base , &vpn_size ); if( error ) { printk("\n[ERROR] in %s : no space for stack vseg / process %x in cluster %x\n", __FUNCTION__ , process->pid , local_cxy ); return NULL; } // compute vseg base and size from vpn_base and vpn_size base = vpn_base << CONFIG_PPM_PAGE_SHIFT; size = vpn_size << CONFIG_PPM_PAGE_SHIFT; } else if( type == VSEG_TYPE_FILE ) { // compute page index (in mapper) for first byte vpn_t vpn_min = file_offset >> CONFIG_PPM_PAGE_SHIFT; // compute page index (in mapper) for last byte vpn_t vpn_max = (file_offset + size - 1) >> CONFIG_PPM_PAGE_SHIFT; // compute offset in first page uint32_t offset = file_offset & CONFIG_PPM_PAGE_MASK; // compute number of pages required in virtual space vpn_t npages = vpn_max - vpn_min + 1; // get vpn_base and vpn_size from MMAP allocator error = vmm_mmap_alloc( vmm , npages , &vpn_base , &vpn_size ); if( error ) { printk("\n[ERROR] in %s : no vspace for mmap vseg / process %x in cluster %x\n", __FUNCTION__ , process->pid , local_cxy ); return NULL; } // set the vseg base (not always aligned for FILE) base = (vpn_base << CONFIG_PPM_PAGE_SHIFT) + offset; } else if( (type == VSEG_TYPE_ANON) || (type == VSEG_TYPE_REMOTE) ) { // compute number of required pages in virtual space vpn_t npages = size >> CONFIG_PPM_PAGE_SHIFT; if( size & CONFIG_PPM_PAGE_MASK) npages++; // get vpn_base and vpn_size from MMAP allocator error = vmm_mmap_alloc( vmm , npages , &vpn_base , &vpn_size ); if( error ) { printk("\n[ERROR] in %s : no vspace for mmap vseg / process %x in cluster %x\n", __FUNCTION__ , process->pid , local_cxy ); return NULL; } // set vseg base (always aligned for ANON or REMOTE) base = vpn_base << CONFIG_PPM_PAGE_SHIFT; } else // VSEG_TYPE_DATA or VSEG_TYPE_CODE { uint32_t vpn_min = base >> CONFIG_PPM_PAGE_SHIFT; uint32_t vpn_max = (base + size - 1) >> CONFIG_PPM_PAGE_SHIFT; vpn_base = vpn_min; vpn_size = vpn_max - vpn_min + 1; } // check collisions vseg = vmm_check_conflict( process , vpn_base , vpn_size ); if( vseg != NULL ) { printk("\n[ERROR] in %s for process %x : new vseg [vpn_base = %x / vpn_size = %x]\n" " overlap existing vseg [vpn_base = %x / vpn_size = %x]\n", __FUNCTION__ , process->pid, vpn_base, vpn_size, vseg->vpn_base, vseg->vpn_size ); return NULL; } // allocate physical memory for vseg descriptor vseg = vseg_alloc(); if( vseg == NULL ) { printk("\n[ERROR] in %s for process %x : cannot allocate memory for vseg\n", __FUNCTION__ , process->pid ); return NULL; } // initialize vseg descriptor vseg_init( vseg, type, base, size, vpn_base, vpn_size, file_offset, file_size, mapper_xp, cxy ); // attach vseg to VSL vmm_vseg_attach( vmm , vseg ); #if DEBUG_VMM_CREATE_VSEG cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_CREATE_VSEG < cycle ) printk("\n[%s] thread[%x,%x] exit / %s / cxy %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, vseg_type_str(type), cxy, cycle ); #endif return vseg; } // vmm_create_vseg() ///////////////////////////////////// void vmm_remove_vseg( vseg_t * vseg ) { // get pointers on calling process and VMM thread_t * this = CURRENT_THREAD; vmm_t * vmm = &this->process->vmm; uint32_t type = vseg->type; // detach vseg from VSL vmm_vseg_detach( vmm , vseg ); // release the stack slot to VMM stack allocator if STACK type if( type == VSEG_TYPE_STACK ) { // get pointer on stack allocator stack_mgr_t * mgr = &vmm->stack_mgr; // compute slot index uint32_t index = ((vseg->vpn_base - mgr->vpn_base - 1) / CONFIG_VMM_STACK_SIZE); // update stacks_bitmap busylock_acquire( &mgr->lock ); bitmap_clear( &mgr->bitmap , index ); busylock_release( &mgr->lock ); } // release the vseg to VMM mmap allocator if MMAP type if( (type == VSEG_TYPE_ANON) || (type == VSEG_TYPE_FILE) || (type == VSEG_TYPE_REMOTE) ) { // get pointer on mmap allocator mmap_mgr_t * mgr = &vmm->mmap_mgr; // compute zombi_list index uint32_t index = bits_log2( vseg->vpn_size ); // update zombi_list busylock_acquire( &mgr->lock ); list_add_first( &mgr->zombi_list[index] , &vseg->zlist ); busylock_release( &mgr->lock ); } // release physical memory allocated for vseg descriptor if no MMAP type if( (type != VSEG_TYPE_ANON) && (type != VSEG_TYPE_FILE) && (type != VSEG_TYPE_REMOTE) ) { vseg_free( vseg ); } } // end vmm_remove_vseg() ///////////////////////////////////////// void vmm_unmap_vseg( process_t * process, vseg_t * vseg ) { vpn_t vpn; // VPN of current PTE vpn_t vpn_min; // VPN of first PTE vpn_t vpn_max; // VPN of last PTE (excluded) ppn_t ppn; // current PTE ppn value uint32_t attr; // current PTE attributes kmem_req_t req; // request to release memory xptr_t page_xp; // extended pointer on page descriptor cxy_t page_cxy; // page descriptor cluster page_t * page_ptr; // page descriptor pointer xptr_t forks_xp; // extended pointer on pending forks counter xptr_t lock_xp; // extended pointer on lock protecting forks counter uint32_t forks; // actual number of pendinf forks #if DEBUG_VMM_UNMAP_VSEG uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_UNMAP_VSEG < cycle ) printk("\n[%s] thread[%x,%x] enter / process %x / vseg %s / base %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid, vseg_type_str( vseg->type ), vseg->vpn_base, cycle ); #endif // get pointer on local GPT gpt_t * gpt = &process->vmm.gpt; // loop on pages in vseg vpn_min = vseg->vpn_base; vpn_max = vpn_min + vseg->vpn_size; for( vpn = vpn_min ; vpn < vpn_max ; vpn++ ) { // get GPT entry hal_gpt_get_pte( XPTR( local_cxy , gpt ) , vpn , &attr , &ppn ); if( attr & GPT_MAPPED ) // entry is mapped { #if( DEBUG_VMM_UNMAP_VSEG & 1 ) if( DEBUG_VMM_UNMAP_VSEG < cycle ) printk("- vpn %x / ppn %x\n" , vpn , ppn ); #endif // check small page assert( (attr & GPT_SMALL) , "an user vseg must use small pages" ); // unmap GPT entry in local GPT hal_gpt_reset_pte( gpt , vpn ); // handle pending forks counter if // 1) not identity mapped // 2) reference cluster if( ((vseg->flags & VSEG_IDENT) == 0) && (GET_CXY( process->ref_xp ) == local_cxy) ) { // get extended pointer on physical page descriptor page_xp = ppm_ppn2page( ppn ); page_cxy = GET_CXY( page_xp ); page_ptr = GET_PTR( page_xp ); // get extended pointers on forks and lock fields forks_xp = XPTR( page_cxy , &page_ptr->forks ); lock_xp = XPTR( page_cxy , &page_ptr->lock ); // get pending forks counter forks = hal_remote_l32( forks_xp ); if( forks ) // decrement pending forks counter { hal_remote_atomic_add( forks_xp , -1 ); } else // release physical page to relevant cluster { if( page_cxy == local_cxy ) // local cluster { req.type = KMEM_PAGE; req.ptr = page_ptr; kmem_free( &req ); } else // remote cluster { rpc_pmem_release_pages_client( page_cxy , page_ptr ); } } } } } #if DEBUG_VMM_UNMAP_VSEG cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_UNMAP_VSEG < cycle ) printk("\n[%s] thread[%x,%x] exit / process %x / vseg %s / base %x / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, process->pid, vseg_type_str( vseg->type ), vseg->vpn_base, cycle ); #endif } // end vmm_unmap_vseg() ////////////////////////////////////////////////////////////////////////////////////////// // This low-level static function is called by the vmm_get_vseg(), vmm_get_pte(), // and vmm_resize_vseg() functions. It scan the local VSL to find the unique vseg // containing a given virtual address. ////////////////////////////////////////////////////////////////////////////////////////// // @ vmm : pointer on the process VMM. // @ vaddr : virtual address. // @ return vseg pointer if success / return NULL if not found. ////////////////////////////////////////////////////////////////////////////////////////// static vseg_t * vmm_vseg_from_vaddr( vmm_t * vmm, intptr_t vaddr ) { xptr_t iter_xp; xptr_t vseg_xp; vseg_t * vseg; // get extended pointers on VSL lock and root xptr_t lock_xp = XPTR( local_cxy , &vmm->vsegs_lock ); xptr_t root_xp = XPTR( local_cxy , &vmm->vsegs_root ); // get lock protecting the VSL remote_rwlock_rd_acquire( lock_xp ); // scan the list of vsegs in VSL XLIST_FOREACH( root_xp , iter_xp ) { vseg_xp = XLIST_ELEMENT( iter_xp , vseg_t , xlist ); vseg = GET_PTR( vseg_xp ); if( (vaddr >= vseg->min) && (vaddr < vseg->max) ) { // return success remote_rwlock_rd_release( lock_xp ); return vseg; } } // return failure remote_rwlock_rd_release( lock_xp ); return NULL; } // end vmm_vseg_from_vaddr() ///////////////////////////////////////////// error_t vmm_resize_vseg( process_t * process, intptr_t base, intptr_t size ) { error_t error; vseg_t * new; vpn_t vpn_min; vpn_t vpn_max; // get pointer on process VMM vmm_t * vmm = &process->vmm; intptr_t addr_min = base; intptr_t addr_max = base + size; // get pointer on vseg vseg_t * vseg = vmm_vseg_from_vaddr( vmm , base ); if( vseg == NULL) return EINVAL; // get extended pointer on VSL lock xptr_t lock_xp = XPTR( local_cxy , &vmm->vsegs_lock ); // get lock protecting VSL remote_rwlock_wr_acquire( lock_xp ); if( (vseg->min > addr_min) || (vseg->max < addr_max) ) // region not included in vseg { error = EINVAL; } else if( (vseg->min == addr_min) && (vseg->max == addr_max) ) // vseg must be removed { vmm_remove_vseg( vseg ); error = 0; } else if( vseg->min == addr_min ) // vseg must be resized { // update vseg base address vseg->min = addr_max; // update vpn_base and vpn_size vpn_min = vseg->min >> CONFIG_PPM_PAGE_SHIFT; vpn_max = (vseg->max - 1) >> CONFIG_PPM_PAGE_SHIFT; vseg->vpn_base = vpn_min; vseg->vpn_size = vpn_max - vpn_min + 1; error = 0; } else if( vseg->max == addr_max ) // vseg must be resized { // update vseg max address vseg->max = addr_min; // update vpn_base and vpn_size vpn_min = vseg->min >> CONFIG_PPM_PAGE_SHIFT; vpn_max = (vseg->max - 1) >> CONFIG_PPM_PAGE_SHIFT; vseg->vpn_base = vpn_min; vseg->vpn_size = vpn_max - vpn_min + 1; error = 0; } else // vseg cut in three regions { // resize existing vseg vseg->max = addr_min; // update vpn_base and vpn_size vpn_min = vseg->min >> CONFIG_PPM_PAGE_SHIFT; vpn_max = (vseg->max - 1) >> CONFIG_PPM_PAGE_SHIFT; vseg->vpn_base = vpn_min; vseg->vpn_size = vpn_max - vpn_min + 1; // create new vseg new = vmm_create_vseg( process, vseg->type, addr_min, (vseg->max - addr_max), vseg->file_offset, vseg->file_size, vseg->mapper_xp, vseg->cxy ); if( new == NULL ) error = EINVAL; else error = 0; } // release VMM lock remote_rwlock_wr_release( lock_xp ); return error; } // vmm_resize_vseg() /////////////////////////////////////////// error_t vmm_get_vseg( process_t * process, intptr_t vaddr, vseg_t ** found_vseg ) { xptr_t vseg_xp; vseg_t * vseg; vmm_t * vmm; error_t error; // get pointer on local VMM vmm = &process->vmm; // try to get vseg from local VMM vseg = vmm_vseg_from_vaddr( vmm , vaddr ); if( vseg == NULL ) // vseg not found in local cluster => try to get it from ref { // get extended pointer on reference process xptr_t ref_xp = process->ref_xp; // get cluster and local pointer on reference process cxy_t ref_cxy = GET_CXY( ref_xp ); process_t * ref_ptr = GET_PTR( ref_xp ); if( local_cxy == ref_cxy ) return -1; // local cluster is the reference // get extended pointer on reference vseg rpc_vmm_get_vseg_client( ref_cxy , ref_ptr , vaddr , &vseg_xp , &error ); if( error ) return -1; // vseg not found => illegal user vaddr // allocate a vseg in local cluster vseg = vseg_alloc(); if( vseg == NULL ) return -1; // cannot allocate a local vseg // initialise local vseg from reference vseg_init_from_ref( vseg , vseg_xp ); // register local vseg in local VMM vmm_vseg_attach( vmm , vseg ); } // success *found_vseg = vseg; return 0; } // end vmm_get_vseg() ////////////////////////////////////////////////////////////////////////////////////// // This static function compute the target cluster to allocate a physical page // for a given in a given , allocates the page (with an RPC if required) // and returns an extended pointer on the allocated page descriptor. // It can be called by a thread running in any cluster. // The vseg cannot have the FILE type. ////////////////////////////////////////////////////////////////////////////////////// static xptr_t vmm_page_allocate( vseg_t * vseg, vpn_t vpn ) { #if DEBUG_VMM_ALLOCATE_PAGE uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_ALLOCATE_PAGE < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] enter for vpn %x / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, vpn, cycle ); #endif // compute target cluster page_t * page_ptr; cxy_t page_cxy; kmem_req_t req; uint32_t index; uint32_t type = vseg->type; uint32_t flags = vseg->flags; uint32_t x_size = LOCAL_CLUSTER->x_size; uint32_t y_size = LOCAL_CLUSTER->y_size; // check vseg type assert( ( type != VSEG_TYPE_FILE ) , "illegal vseg type\n" ); if( flags & VSEG_DISTRIB ) // distributed => cxy depends on vpn LSB { index = vpn & ((x_size * y_size) - 1); page_cxy = HAL_CXY_FROM_XY( (index / y_size) , (index % y_size) ); // If the cluster selected from VPN's LSBs is empty, we select one randomly if ( cluster_is_active( page_cxy ) == false ) { page_cxy = cluster_random_select(); } } else // other cases => cxy specified in vseg { page_cxy = vseg->cxy; } // allocate a physical page from target cluster if( page_cxy == local_cxy ) // target cluster is the local cluster { req.type = KMEM_PAGE; req.size = 0; req.flags = AF_NONE; page_ptr = (page_t *)kmem_alloc( &req ); } else // target cluster is not the local cluster { rpc_pmem_get_pages_client( page_cxy , 0 , &page_ptr ); } #if DEBUG_VMM_ALLOCATE_PAGE cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_ALLOCATE_PAGE < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] exit for vpn %x / ppn %x / cycle %d\n", __FUNCTION__ , this->process->pid, this->trdid, vpn, ppm_page2ppn( XPTR( page_cxy , page_ptr ) , cycle ); #endif if( page_ptr == NULL ) return XPTR_NULL; else return XPTR( page_cxy , page_ptr ); } // end vmm_page_allocate() //////////////////////////////////////// error_t vmm_get_one_ppn( vseg_t * vseg, vpn_t vpn, ppn_t * ppn ) { error_t error; xptr_t page_xp; // extended pointer on physical page descriptor page_t * page_ptr; // local pointer on physical page descriptor uint32_t index; // missing page index in vseg mapper uint32_t type; // vseg type; type = vseg->type; index = vpn - vseg->vpn_base; #if DEBUG_VMM_GET_ONE_PPN uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_GET_ONE_PPN < cycle ) printk("\n[%s] thread[%x,%x] enter for vpn %x / type %s / index %d / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, vpn, vseg_type_str(type), index, cycle ); #endif // FILE type : get the physical page from the file mapper if( type == VSEG_TYPE_FILE ) { // get extended pointer on mapper xptr_t mapper_xp = vseg->mapper_xp; assert( (mapper_xp != XPTR_NULL), "mapper not defined for a FILE vseg\n" ); // get mapper cluster and local pointer cxy_t mapper_cxy = GET_CXY( mapper_xp ); mapper_t * mapper_ptr = GET_PTR( mapper_xp ); // get page descriptor from mapper if( mapper_cxy == local_cxy ) // mapper is local { page_ptr = mapper_get_page( mapper_ptr , index ); } else // mapper is remote { rpc_mapper_get_page_client( mapper_cxy , mapper_ptr , index , &page_ptr ); } if ( page_ptr == NULL ) return EINVAL; page_xp = XPTR( mapper_cxy , page_ptr ); } // Other types : allocate a physical page from target cluster, // as defined by vseg type and vpn value else { // allocate one physical page page_xp = vmm_page_allocate( vseg , vpn ); if( page_xp == XPTR_NULL ) return ENOMEM; // initialise missing page from .elf file mapper for DATA and CODE types // the vseg->mapper_xp field is an extended pointer on the .elf file mapper if( (type == VSEG_TYPE_CODE) || (type == VSEG_TYPE_DATA) ) { // get extended pointer on mapper xptr_t mapper_xp = vseg->mapper_xp; assert( (mapper_xp != XPTR_NULL), "mapper not defined for a CODE or DATA vseg\n" ); // get mapper cluster and local pointer cxy_t mapper_cxy = GET_CXY( mapper_xp ); mapper_t * mapper_ptr = GET_PTR( mapper_xp ); // compute missing page offset in vseg uint32_t offset = index << CONFIG_PPM_PAGE_SHIFT; // compute missing page offset in .elf file uint32_t elf_offset = vseg->file_offset + offset; #if (DEBUG_VMM_GET_ONE_PPN & 0x1) if( DEBUG_VMM_GET_ONE_PPN < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] for vpn = %x / elf_offset = %x\n", __FUNCTION__, this->process->pid, this->trdid, vpn, elf_offset ); #endif // compute extended pointer on page base xptr_t base_xp = ppm_page2base( page_xp ); // file_size (in .elf mapper) can be smaller than vseg_size (BSS) uint32_t file_size = vseg->file_size; if( file_size < offset ) // missing page fully in BSS { #if (DEBUG_VMM_GET_ONE_PPN & 0x1) if( DEBUG_VMM_GET_ONE_PPN < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] for vpn %x / fully in BSS\n", __FUNCTION__, this->process->pid, this->trdid, vpn ); #endif if( GET_CXY( page_xp ) == local_cxy ) { memset( GET_PTR( base_xp ) , 0 , CONFIG_PPM_PAGE_SIZE ); } else { hal_remote_memset( base_xp , 0 , CONFIG_PPM_PAGE_SIZE ); } } else if( file_size >= (offset + CONFIG_PPM_PAGE_SIZE) ) // fully in mapper { #if (DEBUG_VMM_GET_ONE_PPN & 0x1) if( DEBUG_VMM_GET_ONE_PPN < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] for vpn %x / fully in mapper\n", __FUNCTION__, this->process->pid, this->trdid, vpn ); #endif if( mapper_cxy == local_cxy ) { error = mapper_move_kernel( mapper_ptr, true, // to_buffer elf_offset, base_xp, CONFIG_PPM_PAGE_SIZE ); } else { rpc_mapper_move_buffer_client( mapper_cxy, mapper_ptr, true, // to buffer false, // kernel buffer elf_offset, base_xp, CONFIG_PPM_PAGE_SIZE, &error ); } if( error ) return EINVAL; } else // both in mapper and in BSS : // - (file_size - offset) bytes from mapper // - (page_size + offset - file_size) bytes from BSS { #if (DEBUG_VMM_GET_ONE_PPN & 0x1) if( DEBUG_VMM_GET_ONE_PPN < (uint32_t)hal_get_cycles() ) printk("\n[%s] thread[%x,%x] for vpn %x / both mapper & BSS\n", " %d bytes from mapper / %d bytes from BSS\n", __FUNCTION__, this->process->pid, this->trdid, vpn, file_size - offset , offset + CONFIG_PPM_PAGE_SIZE - file_size ); #endif // initialize mapper part if( mapper_cxy == local_cxy ) { error = mapper_move_kernel( mapper_ptr, true, // to buffer elf_offset, base_xp, file_size - offset ); } else { rpc_mapper_move_buffer_client( mapper_cxy, mapper_ptr, true, // to buffer false, // kernel buffer elf_offset, base_xp, file_size - offset, &error ); } if( error ) return EINVAL; // initialize BSS part if( GET_CXY( page_xp ) == local_cxy ) { memset( GET_PTR( base_xp ) + file_size - offset , 0 , offset + CONFIG_PPM_PAGE_SIZE - file_size ); } else { hal_remote_memset( base_xp + file_size - offset , 0 , offset + CONFIG_PPM_PAGE_SIZE - file_size ); } } } // end initialisation for CODE or DATA types } // return ppn *ppn = ppm_page2ppn( page_xp ); #if DEBUG_VMM_GET_ONE_PPN cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_GET_ONE_PPN < cycle ) printk("\n[%s] thread[%x,%x] exit for vpn %x / ppn %x / cycle\n", __FUNCTION__ , this->process->pid, this->trdid , vpn , *ppn, cycle ); #endif return 0; } // end vmm_get_one_ppn() /////////////////////////////////////////////////// error_t vmm_handle_page_fault( process_t * process, vpn_t vpn ) { vseg_t * vseg; // vseg containing vpn uint32_t new_attr; // new PTE_ATTR value ppn_t new_ppn; // new PTE_PPN value uint32_t ref_attr; // PTE_ATTR value in reference GPT ppn_t ref_ppn; // PTE_PPN value in reference GPT cxy_t ref_cxy; // reference cluster for missing vpn process_t * ref_ptr; // reference process for missing vpn xptr_t local_gpt_xp; // extended pointer on local GPT xptr_t local_lock_xp; // extended pointer on local GPT lock xptr_t ref_gpt_xp; // extended pointer on reference GPT xptr_t ref_lock_xp; // extended pointer on reference GPT lock error_t error; // value returned by called functions // get local vseg (access to reference VSL can be required) error = vmm_get_vseg( process, (intptr_t)vpn<pid ); return EXCP_USER_ERROR; } #if DEBUG_VMM_HANDLE_PAGE_FAULT uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_HANDLE_PAGE_FAULT < cycle ) printk("\n[%s] threadr[%x,%x] enter for vpn %x / %s / cycle %d\n", __FUNCTION__, this->process->pid, this->trdid, vpn, vseg_type_str(vseg->type), cycle ); #endif //////////////// private vseg => access only the local GPT if( (vseg->type == VSEG_TYPE_STACK) || (vseg->type == VSEG_TYPE_CODE) ) { // build extended pointer on local GPT and local GPT lock local_gpt_xp = XPTR( local_cxy , &process->vmm.gpt ); local_lock_xp = XPTR( local_cxy , &process->vmm.gpt_lock ); // take local GPT lock in write mode remote_rwlock_wr_acquire( local_lock_xp ); // check VPN still unmapped in local GPT // do nothing if VPN has been mapped by a a concurrent page_fault hal_gpt_get_pte( local_gpt_xp, vpn, &new_attr, &new_ppn ); if( (new_attr & GPT_MAPPED) == 0 ) // VPN still unmapped { // allocate and initialise a physical page depending on the vseg type error = vmm_get_one_ppn( vseg , vpn , &new_ppn ); if( error ) { printk("\n[ERROR] in %s : no memory / process = %x / vpn = %x\n", __FUNCTION__ , process->pid , vpn ); // release local GPT lock in write mode remote_rwlock_wr_release( local_lock_xp ); return EXCP_KERNEL_PANIC; } // define new_attr from vseg flags new_attr = GPT_MAPPED | GPT_SMALL; if( vseg->flags & VSEG_USER ) new_attr |= GPT_USER; if( vseg->flags & VSEG_WRITE ) new_attr |= GPT_WRITABLE; if( vseg->flags & VSEG_EXEC ) new_attr |= GPT_EXECUTABLE; if( vseg->flags & VSEG_CACHE ) new_attr |= GPT_CACHABLE; // set PTE (PPN & attribute) to local GPT error = hal_gpt_set_pte( local_gpt_xp, vpn, new_attr, new_ppn ); if ( error ) { printk("\n[ERROR] in %s : cannot update local GPT / process %x / vpn = %x\n", __FUNCTION__ , process->pid , vpn ); // release local GPT lock in write mode remote_rwlock_wr_release( local_lock_xp ); return EXCP_KERNEL_PANIC; } } // release local GPT lock in write mode remote_rwlock_wr_release( local_lock_xp ); #if DEBUG_VMM_HANDLE_PAGE_FAULT cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_HANDLE_PAGE_FAULT < cycle ) printk("\n[%s] private page fault handled / vpn %x / ppn %x / attr %x / cycle %d\n", __FUNCTION__, vpn, new_ppn, new_attr, cycle ); #endif return EXCP_NON_FATAL; } // end local GPT access //////////// public vseg => access reference GPT else { // get reference process cluster and local pointer ref_cxy = GET_CXY( process->ref_xp ); ref_ptr = GET_PTR( process->ref_xp ); // build extended pointer on reference GPT and reference GPT lock ref_gpt_xp = XPTR( ref_cxy , &ref_ptr->vmm.gpt ); ref_lock_xp = XPTR( ref_cxy , &ref_ptr->vmm.gpt_lock ); // build extended pointer on local GPT and local GPT lock local_gpt_xp = XPTR( local_cxy , &process->vmm.gpt ); local_lock_xp = XPTR( local_cxy , &process->vmm.gpt_lock ); // take reference GPT lock in read mode remote_rwlock_rd_acquire( ref_lock_xp ); // get directly PPN & attributes from reference GPT // this can avoids a costly RPC for a false page fault hal_gpt_get_pte( ref_gpt_xp, vpn, &ref_attr, &ref_ppn ); // release reference GPT lock in read mode remote_rwlock_rd_release( ref_lock_xp ); if( ref_attr & GPT_MAPPED ) // false page fault => update local GPT { // take local GPT lock in write mode remote_rwlock_wr_acquire( local_lock_xp ); // check VPN still unmapped in local GPT hal_gpt_get_pte( local_gpt_xp, vpn, &new_attr, &new_ppn ); if( (new_attr & GPT_MAPPED) == 0 ) // VPN still unmapped { // update local GPT from reference GPT error = hal_gpt_set_pte( local_gpt_xp, vpn, ref_attr, ref_ppn ); if( error ) { printk("\n[ERROR] in %s : cannot update local GPT / process %x / vpn %x\n", __FUNCTION__ , process->pid , vpn ); // release local GPT lock in write mode remote_rwlock_wr_release( local_lock_xp ); return EXCP_KERNEL_PANIC; } } else // VPN has been mapped by a a concurrent page_fault { // keep PTE from local GPT ref_attr = new_attr; ref_ppn = new_ppn; } // release local GPT lock in write mode remote_rwlock_wr_release( local_lock_xp ); #if DEBUG_VMM_HANDLE_PAGE_FAULT cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_HANDLE_PAGE_FAULT < cycle ) printk("\n[%s] false page fault handled / vpn %x / ppn %x / attr %x / cycle %d\n", __FUNCTION__, vpn, ref_ppn, ref_attr, cycle ); #endif return EXCP_NON_FATAL; } else // true page fault => update reference GPT { // take reference GPT lock in write mode remote_rwlock_wr_acquire( ref_lock_xp ); // check VPN still unmapped in reference GPT // do nothing if VPN has been mapped by a a concurrent page_fault hal_gpt_get_pte( ref_gpt_xp, vpn, &ref_attr, &ref_ppn ); if( (ref_attr & GPT_MAPPED) == 0 ) // VPN actually unmapped { // allocate and initialise a physical page depending on the vseg type error = vmm_get_one_ppn( vseg , vpn , &new_ppn ); if( error ) { printk("\n[ERROR] in %s : no memory / process = %x / vpn = %x\n", __FUNCTION__ , process->pid , vpn ); // release reference GPT lock in write mode remote_rwlock_wr_release( ref_lock_xp ); return EXCP_KERNEL_PANIC; } // define new_attr from vseg flags new_attr = GPT_MAPPED | GPT_SMALL; if( vseg->flags & VSEG_USER ) new_attr |= GPT_USER; if( vseg->flags & VSEG_WRITE ) new_attr |= GPT_WRITABLE; if( vseg->flags & VSEG_EXEC ) new_attr |= GPT_EXECUTABLE; if( vseg->flags & VSEG_CACHE ) new_attr |= GPT_CACHABLE; // update reference GPT error = hal_gpt_set_pte( ref_gpt_xp, vpn, new_attr, new_ppn ); // update local GPT (protected by reference GPT lock) error |= hal_gpt_set_pte( local_gpt_xp, vpn, new_attr, new_ppn ); if( error ) { printk("\n[ERROR] in %s : cannot update GPT / process %x / vpn = %x\n", __FUNCTION__ , process->pid , vpn ); // release reference GPT lock in write mode remote_rwlock_wr_release( ref_lock_xp ); return EXCP_KERNEL_PANIC; } } // release reference GPT lock in write mode remote_rwlock_wr_release( ref_lock_xp ); #if DEBUG_VMM_HANDLE_PAGE_FAULT cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_HANDLE_PAGE_FAULT < cycle ) printk("\n[%s] true page fault handled / vpn %x / ppn %x / attr %x / cycle %d\n", __FUNCTION__, vpn, new_ppn, new_attr, cycle ); #endif return EXCP_NON_FATAL; } } } // end vmm_handle_page_fault() //////////////////////////////////////////// error_t vmm_handle_cow( process_t * process, vpn_t vpn ) { vseg_t * vseg; // vseg containing vpn cxy_t ref_cxy; // reference cluster for missing vpn process_t * ref_ptr; // reference process for missing vpn xptr_t gpt_xp; // extended pointer on GPT xptr_t gpt_lock_xp; // extended pointer on GPT lock uint32_t old_attr; // current PTE_ATTR value ppn_t old_ppn; // current PTE_PPN value uint32_t new_attr; // new PTE_ATTR value ppn_t new_ppn; // new PTE_PPN value error_t error; #if DEBUG_VMM_HANDLE_COW uint32_t cycle = (uint32_t)hal_get_cycles(); thread_t * this = CURRENT_THREAD; if( DEBUG_VMM_HANDLE_COW < cycle ) printk("\n[%s] thread[%x,%x] enter for vpn %x / core[%x,%d] / cycle %d\n", __FUNCTION__, process->pid, this->trdid, vpn, local_cxy, this->core->lid, cycle ); #endif // get local vseg error = vmm_get_vseg( process, (intptr_t)vpn<pid ); return EXCP_KERNEL_PANIC; } // get reference GPT cluster and local pointer ref_cxy = GET_CXY( process->ref_xp ); ref_ptr = GET_PTR( process->ref_xp ); // build relevant extended pointers on GPT and GPT lock // - access local GPT for a private vseg // - access reference GPT for a public vseg if( (vseg->type == VSEG_TYPE_STACK) || (vseg->type == VSEG_TYPE_CODE) ) { gpt_xp = XPTR( local_cxy , &process->vmm.gpt ); gpt_lock_xp = XPTR( local_cxy , &process->vmm.gpt_lock ); } else { gpt_xp = XPTR( ref_cxy , &ref_ptr->vmm.gpt ); gpt_lock_xp = XPTR( ref_cxy , &ref_ptr->vmm.gpt_lock ); } // take GPT lock in write mode remote_rwlock_wr_acquire( gpt_lock_xp ); // get current PTE from reference GPT hal_gpt_get_pte( gpt_xp, vpn, &old_attr, &old_ppn ); // the PTE must be mapped for a COW if( (old_attr & GPT_MAPPED) == 0 ) { printk("\n[PANIC] in %s : VPN %x in process %x unmapped\n", __FUNCTION__, vpn, process->pid ); // release GPT lock in write mode remote_rwlock_wr_acquire( gpt_lock_xp ); return EXCP_KERNEL_PANIC; } // get extended pointer, cluster and local pointer on physical page descriptor xptr_t page_xp = ppm_ppn2page( old_ppn ); cxy_t page_cxy = GET_CXY( page_xp ); page_t * page_ptr = GET_PTR( page_xp ); // get extended pointers on forks and lock field in page descriptor xptr_t forks_xp = XPTR( page_cxy , &page_ptr->forks ); xptr_t forks_lock_xp = XPTR( page_cxy , &page_ptr->lock ); // take lock protecting "forks" counter remote_busylock_acquire( forks_lock_xp ); // get number of pending forks from page descriptor uint32_t forks = hal_remote_l32( forks_xp ); if( forks ) // pending fork => allocate a new page, and copy old to new { // allocate a new physical page page_xp = vmm_page_allocate( vseg , vpn ); if( page_xp == XPTR_NULL ) { printk("\n[PANIC] in %s : no memory for vpn %x in process %x\n", __FUNCTION__ , vpn, process->pid ); // release GPT lock in write mode remote_rwlock_wr_acquire( gpt_lock_xp ); // release lock protecting "forks" counter remote_busylock_release( forks_lock_xp ); return EXCP_KERNEL_PANIC; } // compute allocated page PPN new_ppn = ppm_page2ppn( page_xp ); // copy old page content to new page xptr_t old_base_xp = ppm_ppn2base( old_ppn ); xptr_t new_base_xp = ppm_ppn2base( new_ppn ); memcpy( GET_PTR( new_base_xp ), GET_PTR( old_base_xp ), CONFIG_PPM_PAGE_SIZE ); // decrement pending forks counter in page descriptor hal_remote_atomic_add( forks_xp , -1 ); #if(DEBUG_VMM_HANDLE_COW & 1) if( DEBUG_VMM_HANDLE_COW < cycle ) printk("\n[%s] thread[%x,%x] : pending forks => allocate a new PPN %x\n", __FUNCTION__, process->pid, this->trdid, new_ppn ); #endif } else // no pending fork => keep the existing page { #if(DEBUG_VMM_HANDLE_COW & 1) if( DEBUG_VMM_HANDLE_COW < cycle ) printk("\n[%s] thread[%x,%x] no pending forks => keep existing PPN %x\n", __FUNCTION__, process->pid, this->trdid, new_ppn ); #endif new_ppn = old_ppn; } // release lock protecting "forks" counter remote_busylock_release( forks_lock_xp ); // build new_attr : reset COW and set WRITABLE, new_attr = (old_attr | GPT_WRITABLE) & (~GPT_COW); // update the relevan GPT // - private vseg => update local GPT // - public vseg => update all GPT copies if( (vseg->type == VSEG_TYPE_STACK) || (vseg->type == VSEG_TYPE_CODE) ) { hal_gpt_set_pte( gpt_xp, vpn, new_attr, new_ppn ); } else { if( ref_cxy == local_cxy ) // reference cluster is local { vmm_global_update_pte( process, vpn, new_attr, new_ppn ); } else // reference cluster is remote { rpc_vmm_global_update_pte_client( ref_cxy, ref_ptr, vpn, new_attr, new_ppn ); } } // release GPT lock in write mode remote_rwlock_wr_release( gpt_lock_xp ); #if DEBUG_VMM_HANDLE_COW cycle = (uint32_t)hal_get_cycles(); if( DEBUG_VMM_HANDLE_COW < cycle ) printk("\n[%s] thread[%x,%x] exit for vpn %x / core[%x,%d] / cycle %d\n", __FUNCTION__, process->pid, this->trdid, vpn, local_cxy, this->core->lid, cycle ); #endif return EXCP_NON_FATAL; } // end vmm_handle_cow()