kern/kernel/mm/page.c

/* See the end of this file for copyright and license terms. */

#include <arch/page.h>

#include <gay/clist.h>
#include <gay/config.h>
#include <gay/kprintf.h>
#include <gay/mm.h>
#include <gay/types.h>
#include <gay/util.h>

#include <limits.h>
#include <string.h>

#ifndef __HAVE_HUGEPAGES
#error "Systems without huge pages are currently unsupported because i'm a dumb bitch"
#endif

#if DMAP_OFFSET % HUGEPAGE_SIZE != 0
#error "DMAP_OFFSET must be an integral multiple of HUGEPAGE_SIZE"
#endif

/* this should be impossible because arch/page.h must also define PAGE_SHIFT
 * and HUGEPAGE_SHIFT, meaning the two are definitively powers of 2 */
#if HUGEPAGE_SIZE % PAGE_SIZE != 0
#error "HUGEPAGE_SIZE must be an integral multiple of PAGE_SIZE"
#endif

#if PAGE_SIZE % LONG_BIT != 0
#error "PAGE_SIZE must be an integral multiple of LONG_BIT"
#endif

#if CFG_DEBUG_PAGE_ALLOCS
#define page_debug(msg, ...) kprintf("[page] " msg, ##__VA_ARGS__)
#else
#define page_debug(msg, ...)
#endif

/**
 * We have cache levels for areas ranging from a single page up to a huge page
 * on a logarithmic scale.  Every level covers double the pages per entry than
 * the one below it, starting at one page per entry.  The effective result is
 * that a single entry in the cache on level L covers `(1 << L)` pages.
 */
#define CACHE_LEVELS (HUGEPAGE_SHIFT - PAGE_SHIFT + 1)

struct cache_pool {
	struct clist freelist;
	unsigned long *bitmap;
	usize free_entries;
	usize bitmap_len;
};
static struct cache_pool caches[CACHE_LEVELS];

#define LONG_MASK ( ~(usize)(sizeof(long) - 1) )

uintptr_t phys_start;
uintptr_t phys_end;

/**
 * @brief Split a block and return the lower half.
 * The block is assumed to already have been removed from its freelist.
 * The high half (i.e. the block that is *not* returned) is inserted into the
 * freelist one level below `level`.
 */
static void *split_buddy(void *ptr, int level);
/**
 * @brief Attempt to coalesce a block with its buddy.
 * If coalition is possible, the buddy is removed from its freelist at
 * `level` and the union block is inserted at `level + 1`.
 *
 * @param ptr Pointer to the block
 * @param level Cache level, must be less than `CACHE_LEVELS - 1` (because you
 *	can't join blocks at the highest cache level)
 * @return The joined block, or `nil` if coalition was not possible
 */
static void *try_join_buddy(void *ptr, int level);

static usize get_bit_number(void *ptr, int level);

static void set_bits(unsigned long *bitfield, usize first, usize count);
static void clr_bits(unsigned long *bitfield, usize first, usize count);

static bool get_bit(const unsigned long *bitfield, usize bit_number);
static void set_bit(unsigned long *bitfield, usize bit_number);

static int sanity_check(void)
{
	if (phys_end != HUGEPAGE_ALIGN(phys_end) || phys_start != HUGEPAGE_ALIGN(phys_start)) {
		kprintf("Unaligned memory, this should never be possible\n");
		return 1;
	}

	if ((phys_end - phys_start) < (32 * 1024 * 1024)) {
		kprintf("Less than 32 MB of usable RAM, this wouldn't go well\n");
		return 1;
	}

	if (phys_start > phys_end) {
		kprintf("Hey, this is funny. pages_init() was called with parameters "
			"such that phys_start > phys_end (%p > %p), which "
			"should absolutely never be possible. I can't really continue "
			"like this, so have a nice day.\n", (void *)phys_start, (void *)phys_end);
		return 1;
	}

	return 0;
}

static void init_freelist(void)
{
	for (int i = 0; i < CACHE_LEVELS; i++) {
		clist_init(&caches[i].freelist);
		caches[i].free_entries = 0;
	}

	struct cache_pool *pool = &caches[CACHE_LEVELS - 1];
	const usize step = 1 << (CACHE_LEVELS + PAGE_SHIFT);
	for (void *pos = kheap_start; pos < kheap_end; pos += step) {
		struct clist *entry = pos;
		clist_add(&pool->freelist, entry);
		pool->free_entries += 1;
	}
}

int pages_init(void)
{
	usize phys_size = phys_end - phys_start;

	if (sanity_check() != 0)
		return 1;

	/*
	 * map entire physical memory into the direct contiguous area
	 */
	for (uintptr_t physptr = phys_start; physptr < phys_end; physptr += HUGEPAGE_SIZE) {
		const enum mm_page_flags pflags = MM_PAGE_HUGE | MM_PAGE_RW | MM_PAGE_GLOBAL;
		map_page(physptr, (void *)(physptr + DMAP_OFFSET), pflags);
	}
	vm_flush();

	/*
	 * calculate the size of each bitmap, as well as their combined size
	 */
	usize cache_bytes = 0;
	for (int i = 0; i < CACHE_LEVELS; i++) {
		usize bits = phys_size >> (PAGE_SHIFT + i);
		/* round up to the next full long */
		if (bits % LONG_BIT) {
			bits &= LONG_MASK;
			bits += LONG_BIT;
		}
		cache_bytes += bits / 8;
		caches[i].bitmap_len = bits / LONG_BIT;
	}
	/* smol buffer in case we overshoot for whatever reason */
	cache_bytes += sizeof(long);

	page_debug("Page frame overhead = %zu bytes\n", cache_bytes);

	/*
	 * zero out all bitmaps
	 */
	uintptr_t cache_start_phys = phys_end - cache_bytes;
	unsigned long *cache_start = __v(cache_start_phys);
	memset(cache_start, 0, cache_bytes);

	/*
	 * populate the caches array and preallocate pages that can't be handed
	 * out (i.e. the cache bitmaps)
	 */
	unsigned long *cache_pos = cache_start;
	for (int i = 0; i < CACHE_LEVELS; i++) {
		usize total_bits = caches[i].bitmap_len * LONG_BIT;
		usize wasted_bits = total_bits - (cache_bytes >> (PAGE_SHIFT + i));
		if (wasted_bits == 0)
			wasted_bits = 1;
		set_bits(cache_pos, total_bits - wasted_bits, wasted_bits);

		caches[i].bitmap = cache_pos;
		cache_pos += caches[i].bitmap_len;
	}

	/* kheap_start and kheap_end are globals */
	kheap_start = __v(phys_start);
	kheap_end = cache_start;

	init_freelist();

	return 0;
}

static int get_level(usize count)
{
	int level;
	for (level = 0; level < CACHE_LEVELS; level++) {
		if ((1 << level) >= count)
			break;
	}
	return level;
}

void *get_pages(usize count, enum mm_flags flags)
{
	int level = get_level(count);
	if (level == CACHE_LEVELS)
		return nil;

	struct clist *entry;
	int entry_level;
	for (entry_level = level; entry_level < CACHE_LEVELS; entry_level++) {
		if (caches[entry_level].free_entries > 0) {
			entry = caches[entry_level].freelist.next;
			break;
		}
	}
	if (entry_level == CACHE_LEVELS)
		return nil;

	clist_del(entry);
	caches[entry_level].free_entries--;

	usize bit_number = get_bit_number(entry, level);
	while (entry_level > level) {
		entry = split_buddy(entry, entry_level);
		set_bit(caches[entry_level].bitmap, bit_number);
		entry_level--;
		bit_number <<= 1;
	}

	do {
		usize bit_count = 1 << (level - entry_level);
		set_bits(caches[entry_level].bitmap, bit_number, bit_count);
	} while (entry_level-- != 0);

	return (void *)entry;
}

void free_pages(void *ptr, usize count)
{
	int level = get_level(count);
	if (level == CACHE_LEVELS)
		return;

	usize bit_number = get_bit_number(ptr, level);
	usize bit_count = 1;
	for (int i = level; i >= 0; i--) {
		clr_bits(caches[i].bitmap, bit_number, bit_count);
		bit_number <<= 1;
		bit_count <<= 1;
	}

	while (ptr != nil && level < CACHE_LEVELS - 1) {
		ptr = try_join_buddy(ptr, level);
		level++;
	}
}

static inline usize get_bit_number(void *ptr, int level)
{
	return ((uintptr_t)ptr - (uintptr_t)kheap_start) >> (PAGE_SHIFT + level);
}

/**
 * @brief Set a range of bits in a bitfield.
 *
 * @param bitfield Pointer to the beginning of the bitfield
 * @param first Number of the first bit to set, counting from 0
 * @param count Amount of bits to set
 */
static void set_bits(unsigned long *bitfield, usize first, usize count)
{
	bitfield += first / LONG_BIT;
	unsigned int bit = first % LONG_BIT;

	if (bit != 0) {
		unsigned long mask = (1lu << bit) - 1;
		*bitfield++ |= ~mask;
		count -= bit;
	}

	while (count >= LONG_BIT) {
		*bitfield++ = ULONG_MAX;
		count -= LONG_BIT;
	}

	if (count != 0) {
		unsigned long mask = (1lu << count) - 1;
		*bitfield |= mask;
	}
}

/**
 * @brief Clear a range of bits in a bitfield.
 *
 * The algorithm is similar to `set_bits()`, it just does the inverse.
 *
 * @param bitfield Pointer to the beginning of the bitfield
 * @param first Number of the first bit to clear, counting from 0
 * @param count Amount of bits to clear
 */
static void clr_bits(unsigned long *bitfield, usize first, usize count)
{
	bitfield += first / LONG_BIT;
	unsigned int bit = first % LONG_BIT;

	if (bit != 0) {
		unsigned long mask = (1lu << bit) - 1;
		*bitfield++ &= mask;
		count -= bit;
	}

	while (count >= LONG_BIT) {
		*bitfield++ = 0;
		count -= LONG_BIT;
	}

	if (count != 0) {
		unsigned long mask = (1lu << bit) - 1;
		*bitfield &= ~mask;
	}
}

static inline bool get_bit(const unsigned long *bitfield, usize bit_number)
{
	unsigned long longword = bitfield[bit_number / LONG_BIT];
	unsigned long mask = 1lu << (bit_number % LONG_BIT);
	return (longword & mask) != 0;
}

static inline void set_bit(unsigned long *bitfield, usize bit_number)
{
	unsigned long mask = 1lu << (bit_number % LONG_BIT);
	bitfield[bit_number / LONG_BIT] |= mask;
}

static inline void *split_buddy(void *ptr, int level)
{
#	if CFG_DEBUG_PAGE_ALLOCS
		if ((uintptr_t)ptr % (1 << (PAGE_SHIFT + level))) {
			kprintf("split_buddy(%p, %d): unaligned ptr!\n", ptr, level);
			return nil;
		}
		if (level < 1 || level >= CACHE_LEVELS) {
			kprintf("split_buddy(%p, %d): invalid level!\n", ptr, level);
			return nil;
		}
#	endif

	struct clist *high_buddy = ptr + (1 << (PAGE_SHIFT + level - 1));
	clist_add(&caches[level - 1].freelist, high_buddy);
	caches[level - 1].free_entries++;

	return ptr;
}

static void *try_join_buddy(void *ptr, int level)
{
	const usize entry_size = 1 << (PAGE_SHIFT + level);

#	if CFG_DEBUG_PAGE_ALLOCS
		if ((uintptr_t)ptr % entry_size) {
			kprintf("try_join_buddy(%p, %d): unaligned ptr!\n", ptr, level);
			return nil;
		}
		/* level must be < CACHE_LEVELS - 1 because you
		 * can't join blocks on the topmost level */
		if (level >= CACHE_LEVELS - 1) {
			kprintf("try_join_buddy(%p, %d): level >= CACHE_LEVELS - 1!\n", ptr, level);
			return nil;
		}
#	endif

	/* test if the buddy block is allocated and return nil if it is */
	uintptr_t buddy = (uintptr_t)ptr ^ entry_size;
	usize buddy_bitnum = get_bit_number((void *)buddy, level);
	if (get_bit(caches[level].bitmap, buddy_bitnum))
		return nil;

	/* if it is not, remove it from the freelist */
	clist_del((struct clist *)buddy);
	caches[level].free_entries--;

	/* add the coalesced block to the freelist one level above */
	struct clist *even = (struct clist *)((uintptr_t)ptr & ~entry_size);
	clist_add(&caches[level + 1].freelist, even);
	caches[level + 1].free_entries++;
	return even;
}

/*
 * This file is part of GayBSD.
 * Copyright (c) 2021 fef <owo@fef.moe>.
 *
 * GayBSD is nonviolent software: you may only use, redistribute, and/or
 * modify it under the terms of the Cooperative Nonviolent Public License
 * (CNPL) as found in the LICENSE file in the source code root directory
 * or at <https://git.pixie.town/thufie/npl-builder>; either version 7
 * of the license, or (at your option) any later version.
 *
 * GayBSD comes with ABSOLUTELY NO WARRANTY, to the extent
 * permitted by applicable law.  See the CNPL for details.
 */