/* * lpage_fault - handle a fault on a specific lpage. If the page is * not resident, get a physical page from coremap and swap it in. * * You do not yet need to distinguish a readonly fault from a write * fault. When we implement sharing, there will be a difference. * * Synchronization: Lock the lpage while checking if it's in memory. * If it's not, unlock the page while allocting space and loading the * page in. This only works because lpages are not currently sharable. * The page should be locked again as soon as it is loaded, but be * careful of interactions with other locks while modifying the coremap. * * After it has been loaded, the page must be pinned so that it is not * evicted while changes are made to the TLB. It can be unpinned as soon * as the TLB is updated. */ int lpage_fault(struct lpage *lp, struct addrspace *as, int faulttype, vaddr_t va) { paddr_t pa = lp->lp_paddr & PAGE_FRAME; off_t swap = lp->lp_swapaddr; int writable = 0; //lock the page lpage_lock_and_pin(lp); //If the page is not in RAM, load into RAM if(pa == INVALID_PADDR) { //unlock the page if its not lpage_unlock(lp); //allocate a page and pin it pa = coremap_allocuser(lp); if(pa == INVALID_PADDR) { coremap_unpin(lp->lp_paddr & PAGE_FRAME); return ENOMEM; } //assert the page is pinned and lock KASSERT(coremap_pageispinned(pa)); lock_acquire(global_paging_lock); //fetch from disk and put in RAM swap_pagein(pa, swap); //release locks lpage_lock(lp); lock_release(global_paging_lock); //make sure nobody else paged in the page KASSERT((lp->lp_paddr & PAGE_FRAME) == INVALID_PADDR); //set the pages new phyiscal address lp->lp_paddr = pa; } if(faulttype == VM_FAULT_WRITE || faulttype == VM_FAULT_READONLY) { LP_SET(lp, LPF_DIRTY); writable = 1; } //put a mapping into the TLB /*if(coremap_pageispinned(lp->lp_paddr) == 0) { DEBUG(DB_VM, "Page is unpinned!"); }*/ mmu_map(as, va, pa, writable); lpage_unlock(lp); return 0; }
/* * lpage_evict: Evict an lpage from physical memory. * * Synchronization: lock the lpage while evicting it. We come here * from the coremap and should * have pinned the physical page. This is why we must not hold lpage * locks while entering the coremap code. */ void lpage_evict(struct lpage *lp) { KASSERT(lp != NULL); lpage_lock(lp); KASSERT(lp->lp_paddr != INVALID_PADDR); KASSERT(lp->lp_swapaddr != INVALID_SWAPADDR); /* if the page is dirty, swap_pageout */ if (LP_ISDIRTY(lp)) { lpage_unlock(lp); // release lock before doing I/O KASSERT(lock_do_i_hold(global_paging_lock)); KASSERT(coremap_pageispinned(lp->lp_paddr)); swap_pageout((lp->lp_paddr & PAGE_FRAME), lp->lp_swapaddr); lpage_lock(lp); KASSERT((lp->lp_paddr & PAGE_FRAME) != INVALID_PADDR); /* update stats */ spinlock_acquire(&stats_spinlock); ct_write_evictions++; DEBUG (DB_VM, "lpage_evict: evicting Dirty page 0x%x\n", (lp->lp_paddr & PAGE_FRAME)); spinlock_release(&stats_spinlock); } else { /* if page is clean, just update stats */ spinlock_acquire(&stats_spinlock); ct_discard_evictions++; DEBUG (DB_VM, "lpage_evict: evicting Clean page 0x%x\n", (lp->lp_paddr & PAGE_FRAME)); spinlock_release(&stats_spinlock); } /* modify PTE to indicate that the page is no longer in memory. */ lp->lp_paddr = INVALID_PADDR; lpage_unlock(lp); }
/* * lpage_lock_and_pin * * Lock the lpage and also pin the underlying physical page (if any) * in the coremap. This requires a silly retry dance, because we need * to pin first but also need the physical address from the lpage to * do that. If the physical address changes while we were pinning the * page, retry. * * Note that you can't in general hold another lpage lock when calling * this, because it acquires the coremap spinlock, and then perhaps * waits to pin the physical page. The eviction path holds the coremap * spinlock and holds a page pinned while locking the lpage; so if * someone's trying to swap the other page out you can deadlock. * * However, if you've got the other lpage locked *and* its physical * page pinned, that can't happen, so it's safe to lock and pin * multiple pages. */ void lpage_lock_and_pin(struct lpage *lp) { paddr_t pa, pinned; pinned = INVALID_PADDR; lpage_lock(lp); while (1) { pa = lp->lp_paddr & PAGE_FRAME; /* * If the lpage matches what we have (including on the * first pass with INVALID_PADDR) we're done. */ if (pa == pinned) { break; } /* * Otherwise we need to unpin, which means unlock the * lpage too. */ lpage_unlock(lp); if (pinned != INVALID_PADDR) { coremap_unpin(pinned); } /* * If what we just got out of the lpage is *now* * invalid, because the page was paged out on us, * we're done. The page can't be paged in again behind * our back, so assert it hasn't after regrabbing the * lpage lock. */ if (pa == INVALID_PADDR) { lpage_lock(lp); KASSERT((lp->lp_paddr & PAGE_FRAME) == INVALID_PADDR); break; } /* Pin what we got and try again. */ coremap_pin(pa); pinned = pa; lpage_lock(lp); } }
/* * lpage_evict: Evict an lpage from physical memory. * * Synchronization: lock the lpage while accessing it. We come here * from the coremap and should have the global paging lock and should * have pinned the physical page (see coremap.c:do_evict()). * This is why we must not hold lpage locks while entering the coremap code. * * Similar to lpage_fault, the lpage lock should not be held while performing * the page out (if one is needed). */ void lpage_evict(struct lpage *lp) { paddr_t pa; off_t swapaddr; KASSERT(lock_do_i_hold(global_paging_lock)); KASSERT(lp != NULL); lpage_lock(lp); swapaddr = lp->lp_swapaddr; pa = lp->lp_paddr & PAGE_FRAME; KASSERT(pa != INVALID_PADDR); if (LP_ISDIRTY(lp)) { lpage_unlock(lp); LP_CLEAR(lp, LPF_DIRTY); swap_pageout(pa, swapaddr); lpage_lock(lp); } lp->lp_paddr = INVALID_PADDR; lpage_unlock(lp); }
/* * lpage_evict: Evict an lpage from physical memory. * * Synchronization: lock the lpage while accessing it. We come here * from the coremap and should have the global paging lock and should * have pinned the physical page (see coremap.c:do_evict()). * This is why we must not hold lpage locks while entering the coremap code. * * Similar to lpage_fault, the lpage lock should not be held while performing * the page out (if one is needed). */ void lpage_evict(struct lpage *lp) { paddr_t physical_address; off_t swap_address; // Lock the lpage while accessing it. KASSERT(lp != NULL); lpage_lock(lp); // Obtain the physical & swap address' physical_address = lp->lp_paddr & PAGE_FRAME; swap_address = lp->lp_swapaddr; // If the page is stored in RAM memory... if (physical_address != INVALID_PADDR) { DEBUG(DB_VM, "lpage_evict: Moving page from paddr 0x%x to swapaddr 0x%llx\n", physical_address, swap_address); // If page is dirty.. if (LP_ISDIRTY(lp)) { // Move page into swapspace. lpage_unlock(lp); swap_pageout(physical_address, swap_address); LP_CLEAR(lp, LPF_DIRTY); lpage_lock(lp); } // Remove page from physical memory. lp->lp_paddr = INVALID_PADDR; lpage_unlock(lp); } else { lpage_unlock(lp); } }
/* * lpage_fault - handle a fault on a specific lpage. If the page is * not resident, get a physical page from coremap and swap it in. * * You do not yet need to distinguish a readonly fault from a write * fault. When we implement sharing, there will be a difference. * * Synchronization: Lock the lpage while checking if it's in memory. * If it's not, unlock the page while allocting space and loading the * page in. This only works because lpages are not currently sharable. * The page should be locked again as soon as it is loaded, but be * careful of interactions with other locks while modifying the coremap. * * After it has been loaded, the page must be pinned so that it is not * evicted while changes are made to the TLB. It can be unpinned as soon * as the TLB is updated. */ int lpage_fault(struct lpage *lp, struct addrspace *as, int faulttype, vaddr_t va) { paddr_t pa, swa; /* Pin the physical page and lock the lpage. */ lpage_lock_and_pin(lp); // Get the physical address pa = lp->lp_paddr & PAGE_FRAME; // If the page is not in memeory, get it from swap if (pa == INVALID_PADDR) { swa = lp->lp_swapaddr; lpage_unlock(lp); // Have a page frame allocated pa = coremap_allocuser(lp); if (pa == INVALID_PADDR) { coremap_unpin(lp->lp_paddr & PAGE_FRAME); lpage_destroy(lp); return ENOMEM; } KASSERT(coremap_pageispinned(pa)); lock_acquire(global_paging_lock); // Add page contents from swap to physical memory swap_pagein(pa, swa); lpage_lock(lp); lock_release(global_paging_lock); /* Assert nobody else did the pagein. */ KASSERT((lp->lp_paddr & PAGE_FRAME) == INVALID_PADDR); lp->lp_paddr = pa; } //Update TLB switch (faulttype){ case VM_FAULT_READONLY: mmu_map(as, va, pa, 0); break; case VM_FAULT_READ: case VM_FAULT_WRITE: // Set it to dirty LP_SET(lp, LPF_DIRTY); mmu_map(as, va, pa, 1); } // Already unpinned in mmu_map lpage_unlock(lp); return 0; }
static int lpage_materialize(struct lpage **lpret, paddr_t *paret) { struct lpage *lp; paddr_t pa; off_t swa; lp = lpage_create(); if (lp == NULL) { return ENOMEM; } swa = swap_alloc(); if (swa == INVALID_SWAPADDR) { lpage_destroy(lp); return ENOSPC; } lp->lp_swapaddr = swa; pa = coremap_allocuser(lp); if (pa == INVALID_PADDR) { /* lpage_destroy will clean up the swap */ lpage_destroy(lp); return ENOSPC; } lpage_lock(lp); lp->lp_paddr = pa | LPF_DIRTY; KASSERT(coremap_pageispinned(pa)); *lpret = lp; *paret = pa; return 0; }
/* * lpage_fault - handle a fault on a specific lpage. If the page is * not resident, get a physical page from coremap and swap it in. * * You do not yet need to distinguish a readonly fault from a write * fault. When we implement sharing, there will be a difference. * * Synchronization: Lock the lpage while checking if it's in memory. * If it's not, unlock the page while allocating space and loading the * page in. This only works because lpages are not currently sharable. * The page should be locked again as soon as it is loaded, but be * careful of interactions with other locks while modifying the coremap. * * After it has been loaded, the page must be pinned so that it is not * evicted while changes are made to the TLB. It can be unpinned as soon * as the TLB is updated. */ int lpage_fault(struct lpage *lp, struct addrspace *as, int faulttype, vaddr_t va) { KASSERT(lp != NULL); // kernel pages never get paged out, thus never fault lock_acquire(global_paging_lock); if ((lp->lp_paddr & PAGE_FRAME) != INVALID_PADDR) { lpage_lock_and_pin(lp); } else { lpage_lock(lp); } lock_release(global_paging_lock); KASSERT(lp->lp_swapaddr != INVALID_SWAPADDR); paddr_t pa = lp->lp_paddr; int writable; // 0 if page is read-only, 1 if page is writable /* case 1 - minor fault: the frame is still in memory */ if ((pa & PAGE_FRAME) != INVALID_PADDR) { /* make sure it's a minor fault */ KASSERT(pa != INVALID_PADDR); /* Setting the TLB entry's dirty bit */ writable = (faulttype != VM_FAULT_READ); /* update stats */ spinlock_acquire(&stats_spinlock); ct_minfaults++; DEBUG(DB_VM, "\nlpage_fault: minor faults = %d.", ct_minfaults); spinlock_release(&stats_spinlock); } else { /* case 2 - major fault: the frame was swapped out to disk */ /* make sure it is a major fault */ KASSERT(pa == INVALID_PADDR); /* allocate a new frame */ lpage_unlock(lp); // must not hold lpage locks before entering coremap pa = coremap_allocuser(lp); // do evict if needed, also pin coremap if ((pa & PAGE_FRAME)== INVALID_PADDR) { DEBUG(DB_VM, "lpage_fault: ENOMEM: va=0x%x\n", va); return ENOMEM; } KASSERT(coremap_pageispinned(pa)); /* retrieving the content from disk */ lock_acquire(global_paging_lock); // because swap_pagein needs it swap_pagein((pa & PAGE_FRAME), lp->lp_swapaddr); // coremap is already pinned above lpage_lock(lp); lock_release(global_paging_lock); /* assert that nobody else did the pagein */ KASSERT((lp->lp_paddr & PAGE_FRAME) == INVALID_PADDR); /* now update PTE with new PFN */ lp->lp_paddr = pa ; // page is clean /* Setting the TLB entry's dirty bit */ writable = 0; // this way we can detect the first write to a page /* update stats */ spinlock_acquire(&stats_spinlock); ct_majfaults++; DEBUG(DB_VM, "\nlpage_fault: MAJOR faults = %d", ct_majfaults); spinlock_release(&stats_spinlock); } /* check preconditions before update TLB/PTE */ KASSERT(coremap_pageispinned(lp->lp_paddr)); KASSERT(spinlock_do_i_hold(&lp->lp_spinlock)); /* PTE entry is dirty if the instruction is a write */ if (writable) { LP_SET(lp, LPF_DIRTY); } /* Put the new TLB entry into the TLB */ KASSERT(coremap_pageispinned(lp->lp_paddr)); // done in both cases of above IF clause mmu_map(as, va, lp->lp_paddr, writable); // update TLB and unpin coremap lpage_unlock(lp); return 0; }
/* * lpage_copy: create a new lpage and copy data from another lpage. * * The synchronization for this is kind of unpleasant. We do it like * this: * * 1. Create newlp. * 2. Materialize a page for newlp, so it's locked and pinned. * 3. Lock and pin oldlp. * 4. Extract the physical address and swap address. * 5. If oldlp wasn't present, * 5a. Unlock oldlp. * 5b. Page in. * 5c. This pins the page in the coremap. * 5d. Leave the page pinned and relock oldlp. * 5e. Assert nobody else paged the page in. * 6. Copy. * 7. Unlock the lpages first, so we can enter the coremap. * 8. Unpin the physical pages. * */ int lpage_copy(struct lpage *oldlp, struct lpage **lpret) { struct lpage *newlp; paddr_t newpa, oldpa; off_t swa; int result; result = lpage_materialize(&newlp, &newpa); if (result) { return result; } KASSERT(coremap_pageispinned(newpa)); /* Pin the physical page and lock the lpage. */ lpage_lock_and_pin(oldlp); oldpa = oldlp->lp_paddr & PAGE_FRAME; /* * If there is no physical page, we allocate one, which pins * it, and then (re)lock the lpage. Since we are single- * threaded (if we weren't, we'd hold the address space lock * to exclude sibling threads) nobody else should have paged * the page in behind our back. */ if (oldpa == INVALID_PADDR) { /* * XXX this is mostly copied from lpage_fault */ swa = oldlp->lp_swapaddr; lpage_unlock(oldlp); oldpa = coremap_allocuser(oldlp); if (oldpa == INVALID_PADDR) { coremap_unpin(newlp->lp_paddr & PAGE_FRAME); lpage_destroy(newlp); return ENOMEM; } KASSERT(coremap_pageispinned(oldpa)); lock_acquire(global_paging_lock); swap_pagein(oldpa, swa); lpage_lock(oldlp); lock_release(global_paging_lock); /* Assert nobody else did the pagein. */ KASSERT((oldlp->lp_paddr & PAGE_FRAME) == INVALID_PADDR); oldlp->lp_paddr = oldpa; } KASSERT(coremap_pageispinned(oldpa)); coremap_copy_page(oldpa, newpa); KASSERT(LP_ISDIRTY(newlp)); lpage_unlock(oldlp); lpage_unlock(newlp); coremap_unpin(newpa); coremap_unpin(oldpa); *lpret = newlp; return 0; }