arch/arm/pmap.c

/**
 * \file
 * \brief pmap management
 */

/*
 * Copyright (c) 2010-2015 ETH Zurich.
 * All rights reserved.
 *
 * This file is distributed under the terms in the attached LICENSE file.
 * If you do not find this file, copies can be found by writing to:
 * ETH Zurich D-INFK, Universitaetstrasse 6, CH-8092 Zurich. Attn: Systems Group.
 */

/*
 * There was some minor difficulty here with mapping the cpus native
 * page table arrangement onto Barrelfish. The problem lies with
 * resource bootstrapping. The bootstrap ram allocator allocates pages.
 *
 * After reworking retype to be range based, we can now select to create a
 * single 1kB vnode from a 4kB frame, so we currently waste 3kB when creating
 * ARM l2 vnodes before we have a connection to the memory server.
 *
 */

#include <barrelfish/barrelfish.h>
#include <barrelfish/caddr.h>
#include <barrelfish/invocations_arch.h>
#include <stdio.h>

// Location of VSpace managed by this system.
#define VSPACE_BEGIN ((lvaddr_t)(256UL << 20) * (disp_get_core_id() + 1))

// Amount of virtual address space reserved for mapping frames
// backing refill_slabs.
//#define META_DATA_RESERVED_SPACE (BASE_PAGE_SIZE * 128) // 64
#define META_DATA_RESERVED_SPACE (BASE_PAGE_SIZE * 1024)
// increased above value from 128 for pandaboard port

static inline uintptr_t
vregion_flags_to_kpi_paging_flags(vregion_flags_t flags)
{
    STATIC_ASSERT(0x1ff == VREGION_FLAGS_MASK, "");
    STATIC_ASSERT(0x0f == KPI_PAGING_FLAGS_MASK, "");
    STATIC_ASSERT(VREGION_FLAGS_READ    == KPI_PAGING_FLAGS_READ,    "");
    STATIC_ASSERT(VREGION_FLAGS_WRITE   == KPI_PAGING_FLAGS_WRITE,   "");
    STATIC_ASSERT(VREGION_FLAGS_EXECUTE == KPI_PAGING_FLAGS_EXECUTE, "");
    STATIC_ASSERT(VREGION_FLAGS_NOCACHE == KPI_PAGING_FLAGS_NOCACHE, "");
    if ((flags & VREGION_FLAGS_MPB) != 0) {
        // XXX: ignore MPB flag on ARM, otherwise the assert below fires -AB
        flags &= ~VREGION_FLAGS_MPB;
    }
    if ((flags & VREGION_FLAGS_WRITE_COMBINING) != 0) {
        // XXX mask out write-combining flag on ARM
        flags &= ~VREGION_FLAGS_WRITE_COMBINING;
    }
    if ((flags & VREGION_FLAGS_VTD_SNOOP) != 0) {
        // XXX mask out vtd-snooping flag on ARM
        flags &= ~VREGION_FLAGS_VTD_SNOOP;
    }
    if ((flags & VREGION_FLAGS_GUARD) != 0) {
        flags = 0;
    }
    assert(0 == (~KPI_PAGING_FLAGS_MASK & (uintptr_t)flags));
    return (uintptr_t)flags;
}

static void
set_mapping_capref(struct capref *mapping, struct vnode *root, uint32_t entry)
{
#ifdef GLOBAL_MCN
    mapping->cnode = root->u.vnode.mcnode[entry / L2_CNODE_SLOTS];
    mapping->slot  = entry % L2_CNODE_SLOTS;
#else
    errval_t err = slot_alloc(mapping);
    assert(err_is_ok(err));
#endif
    assert(!cnoderef_is_null(mapping->cnode));
}

// debug print preprocessor flag for this file
//#define LIBBARRELFISH_DEBUG_PMAP

/**
 * \brief check whether region A = [start_a .. end_a) overlaps
 * region B = [start_b .. end_b).
 * \return true iff A overlaps B
 */
static bool is_overlapping(uint16_t start_a, uint16_t end_a, uint16_t start_b, uint16_t end_b)
{
    return
        // B strict subset of A
        (start_a < start_b && end_a >= end_b)
        // start_a inside B
        || (start_a >= start_b && start_a < end_b)
        // end_a inside B
        || (end_a > start_b && end_a < end_b);
}

/**
 * \brief Check whether vnode `root' has entries in between [entry ..
 * entry+len).
 * \param root the vnode to look at
 * \param entry first entry of the region to check
 * \param len   length of the region to check
 * \param only_pages true == do not report previously allocated lower-level
 *                   page tables that are empty
 * \return true iff entries exist in region.
 */
#if defined(LIBBARRELFISH_DEBUG_PMAP)
#define DEBUG_HAS_VNODE
#endif
static bool has_vnode(struct vnode *root, uint32_t entry, size_t len,
               bool only_pages)
{
    assert(root != NULL);
    assert(root->is_vnode);
    struct vnode *n;

    uint32_t end_entry = entry + len;
#ifdef DEBUG_HAS_VNODE
    debug_printf("%s: checking region [%"PRIu32"--%"PRIu32"], only_pages = %d\n",
            __FUNCTION__, entry, end_entry, only_pages);
#endif

    for (n = root->u.vnode.children; n; n = n->next) {
        // region to check [entry .. end_entry)
        if (n->is_vnode && n->entry >= entry && n->entry < end_entry) {
            if (only_pages) {
                return has_vnode(n, 0, ARM_L2_TABLE_BYTES, true);
            }
#ifdef LIBBARRELFISH_DEBUG_PMAP
            debug_printf("1: found page table inside our region\n");
#endif
            return true;
        } else if (n->is_vnode) {
            // all other vnodes do not overlap with us, so go to next
            assert(n->entry < entry || n->entry >= end_entry);
            continue;
        } else {
            // not vnode
            uint32_t end = n->entry + n->u.frame.pte_count;
#ifdef DEBUG_HAS_VNODE
            debug_printf("%s: looking at region: [%"PRIu32"--%"PRIu32"]\n",
                    __FUNCTION__, n->entry, end);
#endif

            // do checks
            if (is_overlapping(entry, end_entry, n->entry, end)) {
                return true;
            }
        }
    }

    return false;
}

/**
 * \brief Starting at a given root, return the vnode with entry equal to #entry
 * \return vnode at index `entry` or NULL
 */
#ifdef LIBBARRELFISH_DEBUG_PMAP
#define DEBUG_FIND_VNODE
#endif
static struct vnode *find_vnode(struct vnode *root, uint16_t entry)
{
    assert(root != NULL);
    assert(root->is_vnode);
    struct vnode *n;

#ifdef DEBUG_FIND_VNODE
    debug_printf("%s: looking for %"PRIu16"\n", __FUNCTION__, entry);
#endif

    for(n = root->u.vnode.children; n != NULL; n = n->next) {
        if (n->is_vnode &&
            is_overlapping(entry, entry + 1, n->entry, n->entry + 1)) {
#ifdef DEBUG_FIND_VNODE
            debug_printf("%s: found ptable at [%"PRIu16"--%"PRIu16"]\n",
                    __FUNCTION__, n->entry, n->entry + 1);
#endif
            return n;
        }
        else if (n->is_vnode) {
            assert(!is_overlapping(entry, entry + 1, n->entry, n->entry + 1));
            // ignore all other vnodes;
            continue;
        }

        // not vnode
        assert(!n->is_vnode);
        uint16_t end = n->entry + n->u.frame.pte_count;
#ifdef DEBUG_FIND_VNODE
        debug_printf("%s: looking at section [%"PRIu16"--%"PRIu16"]\n", __FUNCTION__, n->entry, end);
#endif
        if (n->entry <= entry && entry < end) {
#ifdef DEBUG_FIND_VNODE
            debug_printf("%d \\in [%d, %d]\n", entry, n->entry, end);
#endif
            return n;
        }
    }
    return NULL;
}

/**
 * \brief check whether region [entry, entry+npages) is contained in a child
 * of `root`.
 */
static bool inside_region(struct vnode *root, uint32_t entry, uint32_t npages)
{
    assert(root != NULL);
    assert(root->is_vnode);

    struct vnode *n;

    for (n = root->u.vnode.children; n; n = n->next) {
        if (!n->is_vnode) {
            uint16_t end = n->entry + n->u.frame.pte_count;
            if (n->entry <= entry && entry + npages <= end) {
                return true;
            }
        }
    }

    return false;
}

/**
 * \brief remove vnode `item` from linked list of children of `root`
 */
static void remove_vnode(struct vnode *root, struct vnode *item)
{
    assert(root->is_vnode);
    struct vnode *walk = root->u.vnode.children;
    struct vnode *prev = NULL;
    while (walk) {
        if (walk == item) {
            if (prev) {
                prev->next = walk->next;
                return;
            } else {
                root->u.vnode.children = walk->next;
                return;
            }
        }
        prev = walk;
        walk = walk->next;
    }
    USER_PANIC("Should not get here");
}

/**
 * \brief (recursively) remove empty page tables in region [entry ..
 * entry+len) in vnode `root`.
 */
#ifdef LIBBARRELFISH_DEBUG_PMAP
#define DEBUG_REMOVE_EMPTY_VNODES
#endif
static void remove_empty_vnodes(struct slab_allocator *vnode_alloc, struct vnode *root,
                         uint32_t entry, size_t len)
{
    // precondition: root does not have pages in [entry, entry+len)
    assert(!has_vnode(root, entry, len, true));

    errval_t err;
    uint32_t end_entry = entry + len;
    for (struct vnode *n = root->u.vnode.children; n; n = n->next) {
        // sanity check and skip leaf entries
        if (!n->is_vnode) {
            continue;
        }
        // here we know that all vnodes we're interested in are
        // page tables
        assert(n->is_vnode);

        // Unmap vnode if it is in range [entry .. entry+len)
        if (n->entry >= entry && n->entry < end_entry) {
            err = vnode_unmap(root->u.vnode.invokable, n->mapping);
            assert(err_is_ok(err));

            if (!capcmp(n->u.vnode.cap, n->u.vnode.invokable)) {
                // delete invokable pt cap if it's a real copy
                err =cap_destroy(n->u.vnode.invokable);
                assert(err_is_ok(err));
            }

            // delete last copy of pt cap
            err = cap_destroy(n->u.vnode.cap);
            assert(err_is_ok(err));

            // remove vnode from list
            remove_vnode(root, n);
            slab_free(vnode_alloc, n);
        }
    }
}

/**
 * \brief Allocates a new VNode, adding it to the page table and our metadata
 */
static errval_t alloc_vnode(struct pmap_arm *pmap_arm, struct vnode *root,
                            enum objtype type, uint32_t entry,
                            struct vnode **retvnode)
{
    assert(root->is_vnode);
    errval_t err;

    struct vnode *newvnode = slab_alloc(&pmap_arm->slab);
    if (newvnode == NULL) {
        return LIB_ERR_SLAB_ALLOC_FAIL;
    }
    newvnode->is_vnode = true;

    // The VNode capability
    err = slot_alloc(&newvnode->u.vnode.cap);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_SLOT_ALLOC);
    }

    err = vnode_create(newvnode->u.vnode.cap, type);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_VNODE_CREATE);
    }

    // XXX: do we need to put master copy in other cspace?
    newvnode->u.vnode.invokable = newvnode->u.vnode.cap;

    // The VNode meta data
    newvnode->entry            = entry;
    newvnode->next             = root->u.vnode.children;
    root->u.vnode.children     = newvnode;
    newvnode->u.vnode.children = NULL;

    err = slot_alloc(&newvnode->mapping);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_SLOT_ALLOC);
    }

    err = vnode_map(root->u.vnode.invokable, newvnode->u.vnode.cap,
            entry, KPI_PAGING_FLAGS_READ | KPI_PAGING_FLAGS_WRITE, 0, 1,
            newvnode->mapping);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_PMAP_MAP);
    }

#ifdef GLOBAL_MCN
    /* allocate mapping cnodes */
    for (int i = 0; i < MCN_COUNT; i++) {
        err = cnode_create_l2(&newvnode->u.vnode.mcn[i], &newvnode->u.vnode.mcnode[i]);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_ALLOC_CNODE);
        }
    }
#endif

    if (retvnode) {
        *retvnode = newvnode;
    }
    return SYS_ERR_OK;
}

/**
 * \brief Returns the vnode for the pagetable mapping a given vspace address
 */
#ifdef LIBBARRELFISH_DEBUG_PMAP
#define DEBUG_GET_PTABLE
#endif
static errval_t get_ptable(struct pmap_arm  *pmap,
                           genvaddr_t        vaddr,
                           struct vnode    **ptable)
{
    // NB Strictly there are 12 bits in the ARM L1, but allocations unit
    // of L2 is 1 page of L2 entries (4 tables) so we use 10 bits for the L1
    // idx here
    uintptr_t idx = ARM_L1_OFFSET(vaddr);
    if ((*ptable = find_vnode(&pmap->root, idx)) == NULL)
    {
        // L1 table entries point to L2 tables so allocate an L2
        // table for this L1 entry.

        struct vnode *tmp = NULL; // Tmp variable for passing to alloc_vnode

        errval_t err = alloc_vnode(pmap, &pmap->root, ObjType_VNode_ARM_l2,
                                   idx, &tmp);
        if (err_is_fail(err)) {
            DEBUG_ERR(err, "alloc_vnode");
            return err;
        }
        assert(tmp != NULL);
        *ptable = tmp; // Set argument to received value

        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_ALLOC_VNODE);
        }
    }
    assert(ptable);
    struct vnode *pt = *ptable;
    if (!pt->is_vnode) {
        debug_printf("found section @%d, trying to get ptable for %d\n",
                pt->entry, idx);
    }
    assert(pt->is_vnode);
#ifdef DEBUG_GET_PTABLE
    debug_printf("have ptable: %p\n", pt);
#endif

    return SYS_ERR_OK;
}

static struct vnode *find_ptable(struct pmap_arm  *pmap,
                                 genvaddr_t vaddr)
{
    // NB Strictly there are 12 bits in the ARM L1, but allocations unit
    // of L2 is 1 page of L2 entries (4 tables) so
    uintptr_t idx = ARM_L1_OFFSET(vaddr);
    return find_vnode(&pmap->root, idx);
}

static errval_t do_single_map(struct pmap_arm *pmap, genvaddr_t vaddr, genvaddr_t vend,
                              struct capref frame, size_t offset, size_t pte_count,
                              vregion_flags_t flags)
{
    errval_t err = SYS_ERR_OK;
    // Get the page table
    struct vnode *ptable;
    uintptr_t entry;
    bool is_large = false;

    struct frame_identity fi;
    err = frame_identify(frame, &fi);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_PMAP_FRAME_IDENTIFY);
    }

    if (flags & VREGION_FLAGS_LARGE &&
        (vaddr & LARGE_PAGE_MASK) == 0 &&
        fi.bytes >= LARGE_PAGE_SIZE &&
        (fi.base & LARGE_PAGE_MASK) == 0) {
        //section mapping (1MB)
        //mapped in the L1 table at root
        //
        ptable = &pmap->root;
        entry = ARM_L1_OFFSET(vaddr);
        is_large = true;
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("do_single_map: large path: entry=%zu\n", entry);
#endif
    } else {
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("%s: 4k path: mapping %"PRIxGENVADDR", %zu entries\n", __FUNCTION__, vaddr, pte_count);
        debug_printf("4k path: L1 entry: %zu\n", ARM_L1_OFFSET(vaddr));
#endif
        //4k mapping
        // XXX: reassess the following note -SG
        // NOTE: strictly speaking a l2 entry only has 8 bits, while a l1 entry
        // has 12 bits, but due to the way Barrelfish allocates l1 and l2 tables,
        // we use 10 bits for the entry here and in the map syscall
        err = get_ptable(pmap, vaddr, &ptable);
        if (err_is_fail(err)) {
            DEBUG_ERR(err, "get_ptable() in do_single_map");
            return err_push(err, LIB_ERR_PMAP_GET_PTABLE);
        }
        entry = ARM_L2_OFFSET(vaddr);
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("%s: 4k path: L2 entry=%zu\n", __FUNCTION__, entry);
        debug_printf("%s: ptable->is_vnode = %d\n",
                __FUNCTION__, ptable->is_vnode);
#endif
    }

    // convert flags
    flags &= ~(VREGION_FLAGS_LARGE | VREGION_FLAGS_HUGE);
    uintptr_t pmap_flags = vregion_flags_to_kpi_paging_flags(flags);

    // check if there is an overlapping mapping
    if (has_vnode(ptable, entry, pte_count, false)) {
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("has_vnode, only_pages=false  returned true\n");
#endif
        if (has_vnode(ptable, entry, pte_count, true)) {
            printf("page already exists in 0x%"
                    PRIxGENVADDR"--0x%"PRIxGENVADDR"\n", vaddr, vend);
            return LIB_ERR_PMAP_EXISTING_MAPPING;
        } else {
#ifdef LIBBARRELFISH_DEBUG_PMAP
            debug_printf("has_vnode, only_pages=true  returned false, cleaning up empty ptables\n");
#endif
            // clean out empty page tables. We do this here because we benefit
            // from having the page tables in place when doing lots of small
            // mappings
            // XXX: TODO: fix this + mapping of L2 to work on single 1k
            // chunks
            remove_empty_vnodes(&pmap->slab, ptable, entry, pte_count);
        }
    }

    // Create user level datastructure for the mapping
    struct vnode *page = slab_alloc(&pmap->slab);
    assert(page);
    page->is_vnode = false;
    page->entry = entry;
    page->next  = ptable->u.vnode.children;
    ptable->u.vnode.children = page;
    page->u.frame.cap = frame;
    page->u.frame.flags = flags;
    page->u.frame.pte_count = pte_count;

    set_mapping_capref(&page->mapping, ptable, entry);

    // Map entry into the page table
    err = vnode_map(ptable->u.vnode.invokable, frame, entry,
                    pmap_flags, offset, pte_count,
                    page->mapping);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_VNODE_MAP);
    }
    return SYS_ERR_OK;
}

static errval_t do_map(struct pmap_arm *pmap, genvaddr_t vaddr,
                       struct capref frame, size_t offset, size_t size,
                       vregion_flags_t flags, size_t *retoff, size_t *retsize)
{
    errval_t err;
    size_t page_size;
    size_t offset_level;

    // get base address and size of frame
    struct frame_identity fi;
    err = frame_identify(frame, &fi);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_PMAP_DO_MAP);
    }

    // determine mapping specific parts
    if (flags & VREGION_FLAGS_LARGE &&
        (vaddr & LARGE_PAGE_MASK) == 0 &&
        fi.bytes >= LARGE_PAGE_SIZE &&
        (fi.base & LARGE_PAGE_MASK) == 0) {
        //section mapping (1MB)
        page_size = LARGE_PAGE_SIZE;
        offset_level = ARM_L1_OFFSET(vaddr);
#ifdef LIBBARRELFISH_DEBUG_PMAP
        printf("do_map: large path\n");
        printf("page_size: %zx, size: %zx\n", page_size, size);
#endif
    } else {
        //normal 4k mapping
        page_size = BASE_PAGE_SIZE;
        offset_level = ARM_L2_OFFSET(vaddr);
    }

    size = ROUND_UP(size, page_size);
    size_t pte_count = DIVIDE_ROUND_UP(size, page_size);
    if (flags & VREGION_FLAGS_LARGE) {
#ifdef LIBBARRELFISH_DEBUG_PMAP
        printf("#pages: 0x%zu\n", pte_count);
#endif
    }
    genvaddr_t vend = vaddr + size;

    if (fi.bytes < size) {
        return LIB_ERR_PMAP_FRAME_SIZE;
    }

#ifdef LIBBARRELFISH_DEBUG_PMAP
        printf("do_map: mapping %zu pages (size=%zx), from %zu.%zu\n",
                pte_count, page_size, ARM_L1_OFFSET(vaddr), ARM_L2_OFFSET(vaddr));
        printf("page_size: %zx, size: %zx\n", page_size, size);
#endif

    //should be trivially true for section mappings
    if ((ARM_L1_OFFSET(vaddr) == ARM_L1_OFFSET(vend)) ||
        flags & VREGION_FLAGS_LARGE) {
        // fast path
        err = do_single_map(pmap, vaddr, vend, frame, offset, pte_count, flags);
        if (err_is_fail(err)) {
            DEBUG_ERR(err, "[do_map] in fast path");
            return err_push(err, LIB_ERR_PMAP_DO_MAP);
        }
    } else { // multiple leaf page tables
        // first leaf
        uint32_t c = ARM_L2_MAX_ENTRIES - offset_level;
        genvaddr_t temp_end = vaddr + c * page_size;
        err = do_single_map(pmap, vaddr, temp_end, frame, offset, c, flags);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_DO_MAP);
        }

        // map full leaves
        while (ARM_L1_OFFSET(temp_end) < ARM_L1_OFFSET(vend)) { // update vars
            vaddr = temp_end;
            temp_end = vaddr + ARM_L2_MAX_ENTRIES * page_size;
            offset += c * page_size;
            c = ARM_L2_MAX_ENTRIES;

            // do mapping
            err = do_single_map(pmap, vaddr, temp_end, frame, offset, ARM_L2_MAX_ENTRIES, flags);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_DO_MAP);
            }
        }

        // map remaining part
        offset += c * page_size;
        c = ARM_L2_OFFSET(vend) - ARM_L2_OFFSET(temp_end);
        if (c) {

            // do mapping
            err = do_single_map(pmap, temp_end, vend, frame, offset, c, flags);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_DO_MAP);
            }
        }
    }
    if (retoff) {
        *retoff = offset;
    }
    if (retsize) {
        *retsize = size;
    }
    //has_vnode_debug = false;
    return SYS_ERR_OK;
#if 0
    errval_t err;
    uintptr_t pmap_flags = vregion_flags_to_kpi_paging_flags(flags);

    for (size_t i = offset; i < offset + size; i += BASE_PAGE_SIZE) {

        vaddr += BASE_PAGE_SIZE;
    }

    if (retoff) {
        *retoff = offset;
    }
    if (retsize) {
        *retsize = size;
    }
    return SYS_ERR_OK;
#endif
}

static size_t
max_slabs_required(size_t bytes)
{
    // Perform a slab allocation for every page (do_map -> slab_alloc)
    size_t pages     = DIVIDE_ROUND_UP(bytes, BASE_PAGE_SIZE);
    // Perform a slab allocation for every L2 (get_ptable -> find_vnode)
    size_t l2entries = DIVIDE_ROUND_UP(pages, ARM_L2_MAX_ENTRIES);

    return 2 * l2entries;
}
static size_t max_slabs_required_large(size_t bytes)
{
    // always need only 1 slab, as we can represent any size section mapping
    // in a single struct vnode.
    return 1;
}

/**
 * \brief Refill slabs used for metadata
 *
 * \param pmap     The pmap to refill in
 * \param request  The number of slabs the allocator must have
 * when the function returns
 *
 * When the current pmap is initialized,
 * it reserves some virtual address space for metadata.
 * This reserved address space is used here
 *
 * Can only be called for the current pmap
 * Will recursively call into itself till it has enough slabs
 */
#include <stdio.h>
static errval_t refill_slabs(struct pmap_arm *pmap, size_t request)
{
    errval_t err;

    /* Keep looping till we have #request slabs */
    while (slab_freecount(&pmap->slab) < request) {
        // Amount of bytes required for #request
        size_t bytes = SLAB_STATIC_SIZE(request - slab_freecount(&pmap->slab),
                                        sizeof(struct vnode));

        /* Get a frame of that size */
        struct capref cap;
        err = frame_alloc(&cap, bytes, &bytes);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_FRAME_ALLOC);
        }

        /* If we do not have enough slabs to map the frame in, recurse */
        size_t required_slabs_for_frame = max_slabs_required(bytes);
        if (slab_freecount(&pmap->slab) < required_slabs_for_frame) {
            // If we recurse, we require more slabs than to map a single page
            assert(required_slabs_for_frame > 4);

            err = refill_slabs(pmap, required_slabs_for_frame);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_SLAB_REFILL);
            }
        }

        /* Perform mapping */
        genvaddr_t genvaddr = pmap->vregion_offset;
        pmap->vregion_offset += (genvaddr_t)bytes;

        // if this assert fires, increase META_DATA_RESERVED_SPACE
        assert(pmap->vregion_offset < (vregion_get_base_addr(&pmap->vregion) +
               vregion_get_size(&pmap->vregion)));

        err = do_map(pmap, genvaddr, cap, 0, bytes,
                     VREGION_FLAGS_READ_WRITE, NULL, NULL);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_DO_MAP);
        }

        /* Grow the slab */
        lvaddr_t buf = vspace_genvaddr_to_lvaddr(genvaddr);
        slab_grow(&pmap->slab, (void*)buf, bytes);
    }

    return SYS_ERR_OK;
}

/**
 * \brief Create page mappings
 *
 * \param pmap     The pmap object
 * \param vaddr    The virtual address to create the mapping for
 * \param frame    The frame cap to map in
 * \param offset   Offset into the frame cap
 * \param size     Size of the mapping
 * \param flags    Flags for the mapping
 * \param retoff   If non-NULL, filled in with adjusted offset of mapped region
 * \param retsize  If non-NULL, filled in with adjusted size of mapped region
 */
static errval_t
map(struct pmap     *pmap,
    genvaddr_t       vaddr,
    struct capref    frame,
    size_t           offset,
    size_t           size,
    vregion_flags_t  flags,
    size_t          *retoff,
    size_t          *retsize)
{
    struct pmap_arm *pmap_arm = (struct pmap_arm *)pmap;

    errval_t err;
    size_t base;
    size_t page_size;
    size_t slabs_required;

    struct frame_identity fi;
    err = frame_identify(frame, &fi);
    if (err_is_fail(err)) {
        return err_push(err, LIB_ERR_PMAP_FRAME_IDENTIFY);
    }

    // adjust the mapping to be on page boundaries
    if (flags & VREGION_FLAGS_LARGE &&
        (vaddr & LARGE_PAGE_MASK) == 0 &&
        fi.bytes >= LARGE_PAGE_SIZE &&
        (fi.base & LARGE_PAGE_MASK) == 0) {
        //section mapping (1MB)
        base = LARGE_PAGE_OFFSET(offset);
        page_size = LARGE_PAGE_SIZE;
        slabs_required = max_slabs_required_large(size);
#ifdef LIBBARRELFISH_DEBUG_PMAP
        printf("map: large path, page_size: %i, base: %i, slabs: %i, size: %i,"
                "frame size: %zu\n", page_size, base, slabs_required, size, fi.bytes);
#endif
    } else {
        //4k mapping
        base = BASE_PAGE_OFFSET(offset);
        page_size = BASE_PAGE_SIZE;
        slabs_required = max_slabs_required(size);
    }
    size   += base;
    size    = ROUND_UP(size, page_size);
    offset -= base;

    const size_t slabs_reserve = 3; // == max_slabs_required(1)
    uint64_t  slabs_free       = slab_freecount(&pmap_arm->slab);

    slabs_required += slabs_reserve;

    if (slabs_required > slabs_free) {
        if (get_current_pmap() == pmap) {
            err = refill_slabs(pmap_arm, slabs_required);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_SLAB_REFILL);
            }
        }
        else {
            size_t bytes = SLAB_STATIC_SIZE(slabs_required - slabs_free,
                                            sizeof(struct vnode));
            void *buf = malloc(bytes);
            if (!buf) {
                return LIB_ERR_MALLOC_FAIL;
            }
            slab_grow(&pmap_arm->slab, buf, bytes);
        }
    }

    return do_map(pmap_arm, vaddr, frame, offset, size, flags,
                  retoff, retsize);
}

static errval_t do_single_unmap(struct pmap_arm *pmap, genvaddr_t vaddr,
                                size_t pte_count)
{
#ifdef LIBBARRELFISH_DEBUG_PMAP
    debug_printf("%s: vaddr=0x%"PRIxGENVADDR", pte_count=%zu\n",
             __FUNCTION__, vaddr, pte_count);
#endif
    errval_t err;
    struct vnode *pt = find_ptable(pmap, vaddr);
    // pt->is_vnode == non-large mapping
    if (pt && pt->is_vnode) {
        // analog to do_single_map we use 10 bits for tracking pages in user space -SG
        struct vnode *page = find_vnode(pt, ARM_L2_OFFSET(vaddr));
        if (page && page->u.frame.pte_count == pte_count) {
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("page unmap: pt entry: %zu, entry = %zu, pte_count = %hu\n",
                pt->entry, page->entry, page->u.frame.pte_count);
#endif
            err = vnode_unmap(pt->u.vnode.cap, page->mapping);
            if (err_is_fail(err)) {
                DEBUG_ERR(err, "vnode_unmap");
                return err_push(err, LIB_ERR_VNODE_UNMAP);
            }

            // cleanup mapping cap
            err = cap_delete(page->mapping);
            if (err_is_fail(err)) {
                DEBUG_ERR(err, "cap_delete");
                return err_push(err, LIB_ERR_CAP_DELETE);
            }
#ifndef GLOBAL_MCN
            err = slot_free(page->mapping);
            if (err_is_fail(err)) {
                debug_printf("remove_empty_vnodes: slot_free (mapping): %s\n",
                        err_getstring(err));
            }
#endif

            remove_vnode(pt, page);
            slab_free(&pmap->slab, page);
        }
        else {
            return LIB_ERR_PMAP_FIND_VNODE;
        }
    } else if (pt) {
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("section unmap: entry = %zu\n", pt->entry);
#endif
        err = vnode_unmap(pmap->root.u.vnode.cap, pt->mapping);
        if (err_is_fail(err)) {
            DEBUG_ERR(err, "vnode_unmap");
            return err_push(err, LIB_ERR_VNODE_UNMAP);
        }

        // cleanup mapping cap
        err = cap_delete(pt->mapping);
        if (err_is_fail(err)) {
            DEBUG_ERR(err, "cap_delete");
            return err_push(err, LIB_ERR_CAP_DELETE);
        }
        // we need to free slots when unmapping page tables, as we don't have
        // mapping cnodes for the root vnode in armv7
        err = slot_free(pt->mapping);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_SLOT_FREE);
        }

        remove_vnode(&pmap->root, pt);
        slab_free(&pmap->slab, pt);
    } else {
        return LIB_ERR_PMAP_FIND_VNODE;
    }

    return SYS_ERR_OK;
}

/**
 * \brief Remove page mappings
 *
 * \param pmap     The pmap object
 * \param vaddr    The start of the virtual addres to remove
 * \param size     The size of virtual address to remove
 * \param retsize  If non-NULL, filled in with the actual size removed
 */
static errval_t
unmap(struct pmap *pmap,
      genvaddr_t   vaddr,
      size_t       size,
      size_t      *retsize)
{
    errval_t err, ret = SYS_ERR_OK;
    struct pmap_arm *pmap_arm = (struct pmap_arm*)pmap;
    size = ROUND_UP(size, BASE_PAGE_SIZE);
    size_t pte_count = size / BASE_PAGE_SIZE;
    genvaddr_t vend = vaddr + size;

    if (ARM_L1_OFFSET(vaddr) == ARM_L1_OFFSET(vend-1)) {
        // fast path
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("%s: fast path vaddr=0x%"PRIxGENVADDR", pte_count=%zu\n",
                __FUNCTION__, vaddr, pte_count);
#endif
        err = do_single_unmap(pmap_arm, vaddr, pte_count);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_UNMAP);
        }
    } else { // slow path
        // unmap first leaf
        uint32_t c = ARM_L2_MAX_ENTRIES - ARM_L2_OFFSET(vaddr);
#ifdef LIBBARRELFISH_DEBUG_PMAP
        debug_printf("%s: slow path 1st leaf vaddr=0x%"PRIxGENVADDR", pte_count=%zu\n",
                __FUNCTION__, vaddr, c);
#endif
        err = do_single_unmap(pmap_arm, vaddr, c);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_UNMAP);
        }

        // unmap full leaves
        vaddr += c * BASE_PAGE_SIZE;
        while (ARM_L1_OFFSET(vaddr) < ARM_L1_OFFSET(vend)) {
            c = ARM_L2_MAX_ENTRIES;
#ifdef LIBBARRELFISH_DEBUG_PMAP
            debug_printf("%s: slow path full leaf vaddr=0x%"PRIxGENVADDR", pte_count=%zu\n",
                    __FUNCTION__, vaddr, c);
#endif
            err = do_single_unmap(pmap_arm, vaddr, c);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_UNMAP);
            }
            vaddr += c * BASE_PAGE_SIZE;
        }

        // unmap remaining part
        c = ARM_L2_OFFSET(vend) - ARM_L2_OFFSET(vaddr);
        if (c) {
#ifdef LIBBARRELFISH_DEBUG_PMAP
            debug_printf("%s: slow path last leaf vaddr=0x%"PRIxGENVADDR", pte_count=%zu\n",
                    __FUNCTION__, vaddr, c);
#endif
            err = do_single_unmap(pmap_arm, vaddr, c);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_UNMAP);
            }
        }
    }

    if (retsize) {
        *retsize = size;
    }

    return ret;
}

/**
 * \brief Determine a suitable address for a given memory object
 *
 * \param pmap    The pmap object
 * \param memobj  The memory object to determine the address for
 * \param alignment Minimum alignment
 * \param vaddr   Pointer to return the determined address
 *
 * Relies on vspace.c code maintaining an ordered list of vregions
 */
static errval_t
determine_addr(struct pmap   *pmap,
               struct memobj *memobj,
               size_t        alignment,
               genvaddr_t    *vaddr)
{
    assert(pmap->vspace->head);

    if (alignment == 0) {
        alignment = BASE_PAGE_SIZE;
    } else {
        alignment = ROUND_UP(alignment, BASE_PAGE_SIZE);
    }
    size_t size = ROUND_UP(memobj->size, alignment);

    struct vregion *walk = pmap->vspace->head;
    while (walk->next) { // Try to insert between existing mappings
        genvaddr_t walk_base = vregion_get_base_addr(walk);
        genvaddr_t walk_size = ROUND_UP(vregion_get_size(walk), BASE_PAGE_SIZE);
        genvaddr_t walk_end  = ROUND_UP(walk_base + walk_size, alignment);
        genvaddr_t next_base = vregion_get_base_addr(walk->next);

        if (next_base > walk_end + size &&
            walk_base + walk_size > VSPACE_BEGIN) { // Ensure mappings are larger than VSPACE_BEGIN
            *vaddr = walk_end;
            return SYS_ERR_OK;
        }
        walk = walk->next;
    }

    *vaddr = ROUND_UP((vregion_get_base_addr(walk)
                       + ROUND_UP(vregion_get_size(walk), alignment)),
                       alignment);
    return SYS_ERR_OK;
}

/** \brief Retrieves an address that can currently be used for large mappings
  *
  */
static errval_t determine_addr_raw(struct pmap *pmap, size_t size,
                                   size_t alignment, genvaddr_t *retvaddr)
{
    struct pmap_arm *pmap_arm = (struct pmap_arm *)pmap;

    struct vnode *walk_pdir = pmap_arm->root.u.vnode.children;
    assert(walk_pdir != NULL); // assume there's always at least one existing entry

    if (alignment == 0) {
        alignment = BASE_PAGE_SIZE;
    } else {
        alignment = ROUND_UP(alignment, BASE_PAGE_SIZE);
    }
    size = ROUND_UP(size, alignment);

    size_t free_count = DIVIDE_ROUND_UP(size, LARGE_PAGE_SIZE);
    //debug_printf("need %zu contiguous free pdirs\n", free_count);

    // compile pdir free list
    // barrelfish treats L1 as 1024 entries
    bool f[ARM_L1_MAX_ENTRIES];
    for (int i = 0; i < ARM_L1_MAX_ENTRIES; i++) {
        f[i] = true;
    }
    f[walk_pdir->entry] = false;
    while (walk_pdir) {
        assert(walk_pdir->is_vnode);
        f[walk_pdir->entry] = false;
        walk_pdir = walk_pdir->next;
    }
    genvaddr_t first_free = 384;
    for (; first_free < 512; first_free++) {
        if (f[first_free]) {
            for (int i = 1; i < free_count; i++) {
                if (!f[first_free + i]) {
                    // advance pointer
                    first_free = first_free+i;
                    goto next;
                }
            }
            break;
        }
next:
        assert(1 == 1);// make compiler shut up about label
    }
    //printf("first free: %li\n", (uint32_t)first_free);
    if (first_free + free_count <= 512) {
        *retvaddr = first_free << 22;
        return SYS_ERR_OK;
    } else {
        return LIB_ERR_OUT_OF_VIRTUAL_ADDR;
    }
}


static errval_t do_single_modify_flags(struct pmap_arm *pmap, genvaddr_t vaddr,
                                       size_t pages, vregion_flags_t flags)
{
    errval_t err = SYS_ERR_OK;
    struct vnode *ptable = find_ptable(pmap, vaddr);
    uint16_t ptentry = ARM_L2_OFFSET(vaddr);
    if (ptable) {
        struct vnode *page = find_vnode(ptable, ptentry);
        if (page) {
            if (inside_region(ptable, ptentry, pages)) {
                // we're modifying part of a valid mapped region
                // arguments to invocation: invoke frame cap, first affected
                // page (as offset from first page in mapping), #affected
                // pages, new flags. Invocation should check compatibility of
                // new set of flags with cap permissions.
                size_t off = ptentry - page->entry;
                uintptr_t pmap_flags = vregion_flags_to_kpi_paging_flags(flags);
                // VA hinting NYI on ARM, so we always pass 0 for va_hint
                err = invoke_mapping_modify_flags(page->mapping,
                        off, pages, pmap_flags, 0);
                printf("invoke_frame_modify_flags returned error: %s (%"PRIuERRV")\n",
                        err_getstring(err), err);
                return err;
            } else {
                // overlaps some region border
                return LIB_ERR_PMAP_EXISTING_MAPPING;
            }
        }
    }
    return SYS_ERR_OK;
}

/**
 * \brief Modify page mapping
 *
 * \param pmap     The pmap object
 * \param vaddr    The virtual address to unmap
 * \param flags    New flags for the mapping
 * \param retsize  If non-NULL, filled in with the actual size modified
 */
static errval_t
modify_flags(struct pmap     *pmap,
             genvaddr_t       vaddr,
             size_t           size,
             vregion_flags_t  flags,
             size_t          *retsize)
{
    errval_t err, ret = SYS_ERR_OK;
    struct pmap_arm *pmap_arm = (struct pmap_arm*)pmap;
    size = ROUND_UP(size, BASE_PAGE_SIZE);
    size_t pte_count = size / BASE_PAGE_SIZE;
    genvaddr_t vend = vaddr + size;

    if (ARM_L1_OFFSET(vaddr) == ARM_L1_OFFSET(vend-1)) {
        // fast path
        err = do_single_modify_flags(pmap_arm, vaddr, pte_count, flags);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_UNMAP);
        }
    }
    else { // slow path
        // modify flags in first leaf
        uint32_t c = ARM_L2_MAX_ENTRIES - ARM_L2_OFFSET(vaddr);
        err = do_single_modify_flags(pmap_arm, vaddr, c, flags);
        if (err_is_fail(err)) {
            return err_push(err, LIB_ERR_PMAP_UNMAP);
        }

        // modify flags in full leaves
        vaddr += c * BASE_PAGE_SIZE;
        while (ARM_L1_OFFSET(vaddr) < ARM_L1_OFFSET(vend)) {
            c = ARM_L2_MAX_ENTRIES;
            err = do_single_modify_flags(pmap_arm, vaddr, c, flags);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_UNMAP);
            }
            vaddr += c * BASE_PAGE_SIZE;
        }

        // modify flags in remaining part
        c = ARM_L2_OFFSET(vend) - ARM_L2_OFFSET(vaddr);
        if (c) {
            err = do_single_modify_flags(pmap_arm, vaddr, c, flags);
            if (err_is_fail(err)) {
                return err_push(err, LIB_ERR_PMAP_UNMAP);
            }
        }
    }

    if (retsize) {
        *retsize = size;
    }

    return ret;
}

/**
 * \brief Query existing page mapping
 *
 * \param pmap     The pmap object
 * \param vaddr    The virtual address to query
 * \param retvaddr Returns the base virtual address of the mapping
 * \param retsize  Returns the actual size of the mapping
 * \param retcap   Returns the cap mapped at this address
 * \param retoffset Returns the offset within the cap that is mapped
 * \param retflags Returns the flags for this mapping
 *
 * All of the ret parameters are optional.
 */
static errval_t lookup(struct pmap *pmap, genvaddr_t vaddr,
                       struct pmap_mapping_info *info)
{
    USER_PANIC("NYI");
    return 0;
}


static errval_t
serialise(struct pmap *pmap, void *buf, size_t buflen)
{
    // Unimplemented: ignored
    return SYS_ERR_OK;
}

static errval_t
deserialise(struct pmap *pmap, void *buf, size_t buflen)
{
    // Unimplemented: we start with an empty pmap, and avoid the bottom of the A/S
    return SYS_ERR_OK;
}

static struct pmap_funcs pmap_funcs = {
    .determine_addr = determine_addr,
    .determine_addr_raw = determine_addr_raw,
    .map = map,
    .unmap = unmap,
    .modify_flags = modify_flags,
    .lookup = lookup,
    .serialise = serialise,
    .deserialise = deserialise,
};

/**
 * \brief Initialize the pmap object
 */
errval_t
pmap_init(struct pmap   *pmap,
          struct vspace *vspace,
          struct capref  vnode,
          struct slot_allocator *opt_slot_alloc)
{
    struct pmap_arm* pmap_arm = (struct pmap_arm*)pmap;

    /* Generic portion */
    pmap->f = pmap_funcs;
    pmap->vspace = vspace;

    // Slab allocator for vnodes
    slab_init(&pmap_arm->slab, sizeof(struct vnode), NULL);
    slab_grow(&pmap_arm->slab,
              pmap_arm->slab_buffer,
              sizeof(pmap_arm->slab_buffer));

    pmap_arm->root.is_vnode         = true;
    pmap_arm->root.u.vnode.cap      = vnode;
    if (get_croot_addr(vnode) != CPTR_ROOTCN) {
        /* non invokable root cnode; copy */
        errval_t err = slot_alloc(&pmap_arm->root.u.vnode.invokable);
        assert(err_is_ok(err));
        err = cap_copy(pmap_arm->root.u.vnode.invokable, vnode);
        assert(err_is_ok(err));
    } else {
        pmap_arm->root.u.vnode.invokable= vnode;
    }
    pmap_arm->root.next             = NULL;
    pmap_arm->root.u.vnode.children = NULL;

    return SYS_ERR_OK;
}

errval_t pmap_current_init(bool init_domain)
{
    struct pmap_arm *pmap_arm = (struct pmap_arm*)get_current_pmap();

    // To reserve a block of virtual address space,
    // a vregion representing the address space is required.
    // We construct a superficial one here and add it to the vregion list.
    struct vregion *vregion = &pmap_arm->vregion;
    assert((void*)vregion > (void*)pmap_arm);
    assert((void*)vregion < (void*)(pmap_arm + 1));
    vregion->vspace = NULL;
    vregion->memobj = NULL;
    vregion->base   = VSPACE_BEGIN;
    vregion->offset = 0;
    vregion->size   = META_DATA_RESERVED_SPACE;
    vregion->flags  = 0;
    vregion->next = NULL;

    struct vspace *vspace = pmap_arm->p.vspace;
    assert(!vspace->head);
    vspace->head = vregion;

    pmap_arm->vregion_offset = pmap_arm->vregion.base;

    return SYS_ERR_OK;
}