1 files changed, 336 insertions, 0 deletions

diff --git a/src/lib/file-cache.c b/src/lib/file-cache.c
new file mode 100644
index 0000000..008021e
--- /dev/null
+++ b/src/lib/file-cache.c
@@ -0,0 +1,336 @@
+/* Copyright (c) 2004-2018 Dovecot authors, see the included COPYING file */
+
+#include "lib.h"
+#include "buffer.h"
+#include "mmap-util.h"
+#include "file-cache.h"
+
+#include <sys/stat.h>
+
+struct file_cache {
+	int fd;
+	char *path;
+	buffer_t *page_bitmask;
+
+	void *mmap_base;
+	size_t mmap_length;
+	size_t read_highwater;
+};
+
+struct file_cache *file_cache_new(int fd)
+{
+	return file_cache_new_path(fd, "");
+}
+
+struct file_cache *file_cache_new_path(int fd, const char *path)
+{
+	struct file_cache *cache;
+
+	cache = i_new(struct file_cache, 1);
+	cache->fd = fd;
+	cache->path = i_strdup(path);
+	cache->page_bitmask = buffer_create_dynamic(default_pool, 128);
+	return cache;
+}
+
+void file_cache_free(struct file_cache **_cache)
+{
+	struct file_cache *cache = *_cache;
+
+	*_cache = NULL;
+
+	if (cache->mmap_base != NULL) {
+		if (munmap_anon(cache->mmap_base, cache->mmap_length) < 0)
+			i_error("munmap_anon(%s) failed: %m", cache->path);
+	}
+	buffer_free(&cache->page_bitmask);
+	i_free(cache->path);
+	i_free(cache);
+}
+
+void file_cache_set_fd(struct file_cache *cache, int fd)
+{
+	cache->fd = fd;
+	file_cache_invalidate(cache, 0, cache->mmap_length);
+}
+
+int file_cache_set_size(struct file_cache *cache, uoff_t size)
+{
+	size_t page_size = mmap_get_page_size();
+	uoff_t diff;
+	void *new_base;
+
+	i_assert(page_size > 0);
+
+	diff = size % page_size;
+	if (diff != 0)
+		size += page_size - diff;
+
+	i_assert((size % page_size) == 0);
+	if (size <= cache->mmap_length)
+		return 0;
+
+	if (size > SIZE_MAX) {
+		i_error("file_cache_set_size(%s, %"PRIuUOFF_T"): size too large",
+			cache->path, size);
+		return -1;
+	}
+
+	/* grow mmaping */
+	if (cache->mmap_base == NULL) {
+		cache->mmap_base = mmap_anon(size);
+		if (cache->mmap_base == MAP_FAILED) {
+			i_error("mmap_anon(%s, %"PRIuUOFF_T") failed: %m",
+				cache->path, size);
+			cache->mmap_base = NULL;
+			cache->mmap_length = 0;
+			return -1;
+		}
+	} else {
+		new_base = mremap_anon(cache->mmap_base, cache->mmap_length,
+				       size, MREMAP_MAYMOVE);
+		if (new_base == MAP_FAILED) {
+			i_error("mremap_anon(%s, %"PRIuUOFF_T") failed: %m",
+				cache->path, size);
+			return -1;
+		}
+
+		cache->mmap_base = new_base;
+	}
+	cache->mmap_length = size;
+	return 0;
+}
+
+ssize_t file_cache_read(struct file_cache *cache, uoff_t offset, size_t size)
+{
+	size_t page_size = mmap_get_page_size();
+	size_t poffset, psize, dest_offset, dest_size;
+	unsigned char *bits, *dest;
+	ssize_t ret;
+
+	i_assert(page_size > 0);
+
+	if (size > SSIZE_T_MAX) {
+		/* make sure our calculations won't overflow. most likely
+		   we'll be reading less data, but allow it anyway so caller
+		   doesn't have to deal with any extra checks. */
+		size = SSIZE_T_MAX;
+	}
+	if (offset >= UOFF_T_MAX - size)
+		size = UOFF_T_MAX - offset;
+
+	if (offset + size > cache->mmap_length &&
+	    offset + size - cache->mmap_length > 1024*1024) {
+		/* growing more than a megabyte, make sure that the
+		   file is large enough so we don't allocate memory
+		   more than needed */
+		struct stat st;
+
+                if (fstat(cache->fd, &st) < 0) {
+                        if (errno != ESTALE)
+                                i_error("fstat(%s) failed: %m", cache->path);
+			return -1;
+		}
+
+		if (offset + size > (uoff_t)st.st_size) {
+			if (offset >= (uoff_t)st.st_size)
+				return 0;
+			size = (uoff_t)st.st_size - offset;
+		}
+	}
+
+	if (file_cache_set_size(cache, offset + size) < 0)
+		return -1;
+
+	poffset = offset / page_size;
+	psize = (offset + size + page_size-1) / page_size - poffset;
+	i_assert(psize > 0);
+
+	bits = buffer_get_space_unsafe(cache->page_bitmask, 0,
+				       (poffset + psize + CHAR_BIT - 1) /
+				       CHAR_BIT);
+
+	dest_offset = poffset * page_size;
+	dest = PTR_OFFSET(cache->mmap_base, dest_offset);
+	dest_size = page_size;
+
+	while (psize > 0) {
+		if ((bits[poffset / CHAR_BIT] & (1 << (poffset % CHAR_BIT))) != 0) {
+			/* page is already in cache */
+			dest_offset += page_size;
+			if (dest_offset <= cache->read_highwater) {
+				psize--; poffset++;
+				dest += page_size;
+				continue;
+			}
+
+			/* this is the last partially cached block.
+			   use the caching only if we don't want to
+			   read past read_highwater */
+			if (offset + size <= cache->read_highwater) {
+				i_assert(psize == 1);
+				break;
+			}
+
+			/* mark the block noncached again and
+			   read it */
+			bits[poffset / CHAR_BIT] &=
+				~(1 << (poffset % CHAR_BIT));
+			dest_offset -= page_size;
+		}
+
+		ret = pread(cache->fd, dest, dest_size, dest_offset);
+		if (ret <= 0) {
+			if (ret < 0)
+				return -1;
+
+			/* EOF. mark the last block as cached even if it
+			   isn't completely. read_highwater tells us how far
+			   we've actually made. */
+			if (dest_offset == cache->read_highwater) {
+				i_assert(poffset ==
+					 cache->read_highwater / page_size);
+				bits[poffset / CHAR_BIT] |=
+					1 << (poffset % CHAR_BIT);
+			}
+			return dest_offset <= offset ? 0 :
+				dest_offset - offset < size ?
+				dest_offset - offset : size;
+		}
+
+		dest += ret;
+		dest_offset += ret;
+
+		if (cache->read_highwater < dest_offset) {
+			unsigned int high_poffset =
+				cache->read_highwater / page_size;
+
+			/* read_highwater needs to be updated. if we didn't
+			   just read that block, we can't trust anymore that
+			   we have it cached */
+			bits[high_poffset / CHAR_BIT] &=
+				~(1 << (high_poffset % CHAR_BIT));
+			cache->read_highwater = dest_offset;
+		}
+
+		if ((size_t)ret != dest_size) {
+			/* partial read - probably EOF but make sure. */
+			dest_size -= ret;
+			continue;
+		}
+
+		bits[poffset / CHAR_BIT] |= 1 << (poffset % CHAR_BIT);
+		dest_size = page_size;
+		psize--; poffset++;
+	}
+
+	return size;
+}
+
+const void *file_cache_get_map(struct file_cache *cache, size_t *size_r)
+{
+	*size_r = cache->read_highwater;
+	return cache->mmap_base;
+}
+
+void file_cache_write(struct file_cache *cache, const void *data, size_t size,
+		      uoff_t offset)
+{
+	size_t page_size = mmap_get_page_size();
+	unsigned char *bits;
+	unsigned int first_page, last_page;
+
+	i_assert(page_size > 0);
+	i_assert(UOFF_T_MAX - offset > size);
+
+	if (file_cache_set_size(cache, offset + size) < 0) {
+		/* couldn't grow mapping. just make sure the written memory
+		   area is invalidated then. */
+		file_cache_invalidate(cache, offset, size);
+		return;
+	}
+
+	memcpy(PTR_OFFSET(cache->mmap_base, offset), data, size);
+
+	if (cache->read_highwater < offset + size) {
+		unsigned int page = cache->read_highwater / page_size;
+
+		bits = buffer_get_space_unsafe(cache->page_bitmask,
+					       page / CHAR_BIT, 1);
+		*bits &= ~(1 << (page % CHAR_BIT));
+		cache->read_highwater = offset + size;
+	}
+
+	/* mark fully written pages cached */
+	if (size >= page_size) {
+		first_page = offset / page_size;
+		last_page = (offset + size) / page_size;
+		if ((offset % page_size) != 0)
+			first_page++;
+
+		bits = buffer_get_space_unsafe(cache->page_bitmask, 0,
+					       last_page / CHAR_BIT + 1);
+		for (; first_page < last_page; first_page++) {
+			bits[first_page / CHAR_BIT] |=
+				1 << (first_page % CHAR_BIT);
+		}
+	}
+}
+
+void file_cache_invalidate(struct file_cache *cache, uoff_t offset, uoff_t size)
+{
+	size_t page_size = mmap_get_page_size();
+	unsigned char *bits, mask;
+	unsigned int i;
+
+	if (offset >= cache->read_highwater || size == 0)
+		return;
+
+	i_assert(page_size > 0);
+
+	if (size > cache->read_highwater - offset) {
+		/* ignore anything after read highwater */
+		size = cache->read_highwater - offset;
+	}
+	if (size >= cache->read_highwater) {
+		/* we're invalidating everything up to read highwater.
+		   drop the highwater position. */
+		cache->read_highwater = offset & ~(page_size-1);
+	}
+
+	size = (offset + size + page_size-1) / page_size;
+	offset /= page_size;
+	i_assert(size > offset);
+	size -= offset;
+
+	if (size != 1) {
+		/* tell operating system that we don't need the memory anymore
+		   and it may free it. don't bother to do it for single pages,
+		   there's a good chance that they get re-read back
+		   immediately. */
+		(void)madvise(PTR_OFFSET(cache->mmap_base, offset * page_size),
+			      size * page_size, MADV_DONTNEED);
+	}
+
+	bits = buffer_get_space_unsafe(cache->page_bitmask, offset / CHAR_BIT,
+				       1 + (size + CHAR_BIT - 1) / CHAR_BIT);
+
+	/* set the first byte */
+	for (i = offset % CHAR_BIT, mask = 0; i < CHAR_BIT && size > 0; i++) {
+		mask |= 1 << i;
+		size--;
+	}
+	*bits++ &= ~mask;
+
+	/* set the middle bytes */
+	memset(bits, 0, size / CHAR_BIT);
+	bits += size / CHAR_BIT;
+	size %= CHAR_BIT;
+
+	/* set the last byte */
+	if (size > 0) {
+		for (i = 0, mask = 0; i < size; i++)
+			mask |= 1 << i;
+		*bits &= ~mask;
+	}
+}