/*
* __wt_block_read_off_blind --
* Read the block at an offset, return the size and checksum, debugging
* only.
*/
int
__wt_block_read_off_blind(WT_SESSION_IMPL *session,
WT_BLOCK *block, wt_off_t offset, uint32_t *sizep, uint32_t *checksump)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
*sizep = 0;
*checksump = 0;
/*
* Make sure the buffer is large enough for the header and read the
* the first allocation-size block.
*/
WT_RET(__wt_scr_alloc(session, block->allocsize, &tmp));
WT_ERR(__wt_read(
session, block->fh, offset, (size_t)block->allocsize, tmp->mem));
blk = WT_BLOCK_HEADER_REF(tmp->mem);
__wt_block_header_byteswap(blk);
*sizep = blk->disk_size;
*checksump = blk->checksum;
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_block_read_off_blind --
* Read the block at an offset, try to figure out what it looks like,
* debugging only.
*/
int
__wt_block_read_off_blind(
WT_SESSION_IMPL *session, WT_BLOCK *block, WT_ITEM *buf, off_t offset)
{
WT_BLOCK_HEADER *blk;
uint32_t cksum, size;
/*
* Make sure the buffer is large enough for the header and read the
* the first allocation-size block.
*/
WT_RET(__wt_buf_init(session, buf, block->allocsize));
WT_RET(__wt_read(
session, block->fh, offset, (size_t)block->allocsize, buf->mem));
blk = WT_BLOCK_HEADER_REF(buf->mem);
/*
* Copy out the size and checksum (we're about to re-use the buffer),
* and if the size isn't insane, read the rest of the block.
*/
size = blk->disk_size;
cksum = blk->cksum;
if (__wt_block_offset_invalid(block, offset, size))
WT_RET_MSG(session, EINVAL,
"block at offset %" PRIuMAX " cannot be a valid block, no "
"read attempted",
(uintmax_t)offset);
return (__wt_block_read_off(session, block, buf, offset, size, cksum));
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, wt_off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk;
size_t bufsize;
uint32_t page_cksum;
WT_RET(__wt_verbose(session, WT_VERB_READ,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum));
WT_STAT_FAST_CONN_INCR(session, block_read);
WT_STAT_FAST_CONN_INCRV(session, block_byte_read, size);
/*
* Grow the buffer as necessary and read the block. Buffers should be
* aligned for reading, but there are lots of buffers (for example, file
* cursors have two buffers each, key and value), and it's difficult to
* be sure we've found all of them. If the buffer isn't aligned, it's
* an easy fix: set the flag and guarantee we reallocate it. (Most of
* the time on reads, the buffer memory has not yet been allocated, so
* we're not adding any additional processing time.)
*/
if (F_ISSET(buf, WT_ITEM_ALIGNED))
bufsize = size;
else {
F_SET(buf, WT_ITEM_ALIGNED);
bufsize = WT_MAX(size, buf->memsize + 10);
}
WT_RET(__wt_buf_init(session, buf, bufsize));
WT_RET(__wt_read(session, block->fh, offset, size, buf->mem));
buf->size = size;
blk = WT_BLOCK_HEADER_REF(buf->mem);
page_cksum = blk->cksum;
if (page_cksum == cksum) {
blk->cksum = 0;
page_cksum = __wt_cksum(buf->mem,
F_ISSET(blk, WT_BLOCK_DATA_CKSUM) ?
size : WT_BLOCK_COMPRESS_SKIP);
if (page_cksum == cksum)
return (0);
}
if (!F_ISSET(session, WT_SESSION_SALVAGE_CORRUPT_OK))
__wt_errx(session,
"read checksum error [%" PRIu32 "B @ %" PRIuMAX ", %"
PRIu32 " != %" PRIu32 "]",
size, (uintmax_t)offset, cksum, page_cksum);
/* Panic if a checksum fails during an ordinary read. */
return (block->verify ||
F_ISSET(session, WT_SESSION_SALVAGE_CORRUPT_OK) ?
WT_ERROR : __wt_illegal_value(session, block->name));
}
/*
* copy --
* Copy the created page to the end of the salvage file.
*/
void
copy(u_int gen, u_int recno)
{
FILE *ifp, *ofp;
WT_PAGE_HEADER *dsk;
WT_BLOCK_HEADER *blk;
char buf[PSIZE];
CHECK((ifp = fopen(LOAD, "r")) != NULL);
/*
* If the salvage file doesn't exist, then we're creating it:
* copy the first sector (the file description).
* Otherwise, we are appending to an existing file.
*/
if (file_exists(SLVG))
CHECK((ofp = fopen(SLVG, "a")) != NULL);
else {
CHECK((ofp = fopen(SLVG, "w")) != NULL);
CHECK(fread(buf, 1, PSIZE, ifp) == PSIZE);
CHECK(fwrite(buf, 1, PSIZE, ofp) == PSIZE);
}
/*
* If there's data, copy/update the first formatted page.
*/
if (gen != 0) {
CHECK(fseek(ifp, (long)PSIZE, SEEK_SET) == 0);
CHECK(fread(buf, 1, PSIZE, ifp) == PSIZE);
dsk = (void *)buf;
if (page_type != WT_PAGE_ROW_LEAF)
dsk->recno = recno;
dsk->write_gen = gen;
blk = WT_BLOCK_HEADER_REF(buf);
blk->checksum = 0;
blk->checksum = __wt_checksum(dsk, PSIZE);
CHECK(fwrite(buf, 1, PSIZE, ofp) == PSIZE);
}
CHECK(fclose(ifp) == 0);
CHECK(fclose(ofp) == 0);
}
/*
* __wt_block_salvage_next --
* Return the address for the next potential block from the file.
*/
int
__wt_block_salvage_next(WT_SESSION_IMPL *session,
WT_BLOCK *block, uint8_t *addr, size_t *addr_sizep, bool *eofp)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_FH *fh;
wt_off_t max, offset;
uint32_t allocsize, checksum, size;
uint8_t *endp;
*eofp = 0;
fh = block->fh;
allocsize = block->allocsize;
WT_ERR(__wt_scr_alloc(session, allocsize, &tmp));
/* Read through the file, looking for pages. */
for (max = block->size;;) {
offset = block->slvg_off;
if (offset >= max) { /* Check eof. */
*eofp = 1;
goto done;
}
/*
* Read the start of a possible page (an allocation-size block),
* and get a page length from it. Move to the next allocation
* sized boundary, we'll never consider this one again.
*/
WT_ERR(__wt_read(
session, fh, offset, (size_t)allocsize, tmp->mem));
blk = WT_BLOCK_HEADER_REF(tmp->mem);
__wt_block_header_byteswap(blk);
size = blk->disk_size;
checksum = blk->checksum;
/*
* Check the block size: if it's not insane, read the block.
* Reading the block validates any checksum; if reading the
* block succeeds, return its address as a possible page,
* otherwise, move past it.
*/
if (!__wt_block_offset_invalid(block, offset, size) &&
__wt_block_read_off(
session, block, tmp, offset, size, checksum) == 0)
break;
/* Free the allocation-size block. */
__wt_verbose(session, WT_VERB_SALVAGE,
"skipping %" PRIu32 "B at file offset %" PRIuMAX,
allocsize, (uintmax_t)offset);
WT_ERR(__wt_block_off_free(
session, block, offset, (wt_off_t)allocsize));
block->slvg_off += allocsize;
}
/* Re-create the address cookie that should reference this block. */
endp = addr;
WT_ERR(__wt_block_addr_to_buffer(block, &endp, offset, size, checksum));
*addr_sizep = WT_PTRDIFF(endp, addr);
done:
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_block_salvage_next --
* Return the next block from the file.
*/
int
__wt_block_salvage_next(
WT_SESSION_IMPL *session, WT_BLOCK *block, WT_ITEM *buf,
uint8_t *addr, uint32_t *addr_sizep, uint64_t *write_genp, int *eofp)
{
WT_BLOCK_HEADER *blk;
WT_FH *fh;
off_t max, offset;
uint32_t allocsize, cksum, size;
uint8_t *endp;
*eofp = 0;
offset = block->slvg_off;
fh = block->fh;
allocsize = block->allocsize;
WT_RET(__wt_buf_initsize(session, buf, allocsize));
/* Read through the file, looking for pages with valid checksums. */
for (max = fh->file_size;;) {
if (offset >= max) { /* Check eof. */
*eofp = 1;
return (0);
}
/*
* Read the start of a possible page (an allocation-size block),
* and get a page length from it.
*/
WT_RET(__wt_read(session, fh, offset, allocsize, buf->mem));
blk = WT_BLOCK_HEADER_REF(buf->mem);
/*
* The page can't be more than the min/max page size, or past
* the end of the file.
*/
size = blk->disk_size;
cksum = blk->cksum;
if (size == 0 ||
size % allocsize != 0 ||
size > WT_BTREE_PAGE_SIZE_MAX ||
offset + (off_t)size > max)
goto skip;
/*
* The page size isn't insane, read the entire page: reading the
* page validates the checksum and then decompresses the page as
* needed. If reading the page fails, it's probably corruption,
* we ignore this block.
*/
if (__wt_block_read_off(
session, block, buf, offset, size, cksum)) {
skip: WT_VERBOSE_RET(session, salvage,
"skipping %" PRIu32 "B at file offset %" PRIuMAX,
allocsize, (uintmax_t)offset);
/*
* Free the block and make sure we don't return it more
* than once.
*/
WT_RET(__wt_block_off_free(
session, block, offset, (off_t)allocsize));
block->slvg_off = offset += allocsize;
continue;
}
/*
* Valid block, return to our caller.
*
* The buffer may have grown: make sure we read from the full
* page image.
*/
blk = WT_BLOCK_HEADER_REF(buf->mem);
break;
}
/*
* Track the largest write-generation we've seen in the file so future
* writes, done after salvage completes, are preferred to these blocks.
*/
*write_genp = blk->write_gen;
if (block->live.write_gen < blk->write_gen)
block->live.write_gen = blk->write_gen;
/* Re-create the address cookie that should reference this block. */
endp = addr;
WT_RET(__wt_block_addr_to_buffer(block, &endp, offset, size, cksum));
*addr_sizep = WT_PTRDIFF32(endp, addr);
/* We're successfully returning the page, move past it. */
block->slvg_off = offset + size;
return (0);
}
/*
* __wt_block_write_off --
* Write a buffer into a block, returning the block's addr/size and
* checksum.
*/
int
__wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, off_t *offsetp, uint32_t *sizep, uint32_t *cksump,
int data_cksum, int locked)
{
WT_BLOCK_HEADER *blk;
WT_DECL_RET;
WT_FH *fh;
off_t offset;
uint32_t align_size;
blk = WT_BLOCK_HEADER_REF(buf->mem);
fh = block->fh;
/* Buffers should be aligned for writing. */
if (!F_ISSET(buf, WT_ITEM_ALIGNED)) {
WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
WT_RET_MSG(session, EINVAL,
"direct I/O check: write buffer incorrectly allocated");
}
/*
* Align the size to an allocation unit.
*
* The buffer must be big enough for us to zero to the next allocsize
* boundary, this is one of the reasons the btree layer must find out
* from the block-manager layer the maximum size of the eventual write.
*/
align_size = (uint32_t)WT_ALIGN(buf->size, block->allocsize);
if (align_size > buf->memsize) {
WT_ASSERT(session, align_size <= buf->memsize);
WT_RET_MSG(session, EINVAL,
"buffer size check: write buffer incorrectly allocated");
}
/* Zero out any unused bytes at the end of the buffer. */
memset((uint8_t *)buf->mem + buf->size, 0, align_size - buf->size);
/*
* Set the disk size so we don't have to incrementally read blocks
* during salvage.
*/
blk->disk_size = align_size;
/*
* Update the block's checksum: if our caller specifies, checksum the
* complete data, otherwise checksum the leading WT_BLOCK_COMPRESS_SKIP
* bytes. The assumption is applications with good compression support
* turn off checksums and assume corrupted blocks won't decompress
* correctly. However, if compression failed to shrink the block, the
* block wasn't compressed, in which case our caller will tell us to
* checksum the data to detect corruption. If compression succeeded,
* we still need to checksum the first WT_BLOCK_COMPRESS_SKIP bytes
* because they're not compressed, both to give salvage a quick test
* of whether a block is useful and to give us a test so we don't lose
* the first WT_BLOCK_COMPRESS_SKIP bytes without noticing.
*/
blk->flags = 0;
if (data_cksum)
F_SET(blk, WT_BLOCK_DATA_CKSUM);
blk->cksum = 0;
blk->cksum = __wt_cksum(
buf->mem, data_cksum ? align_size : WT_BLOCK_COMPRESS_SKIP);
if (!locked)
__wt_spin_lock(session, &block->live_lock);
ret = __wt_block_alloc(session, block, &offset, (off_t)align_size);
if (!locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
#if defined(HAVE_POSIX_FALLOCATE) || defined(HAVE_FTRUNCATE)
/*
* Extend the file in chunks. We aren't holding a lock and we'd prefer
* to limit the number of threads extending the file at the same time,
* so choose the one thread that's crossing the extended boundary. We
* don't extend newly created files, and it's theoretically possible we
* might wait so long our extension of the file is passed by another
* thread writing single blocks, that's why there's a check in case the
* extended file size becomes too small: if the file size catches up,
* every thread will try to extend it.
*/
if (fh->extend_len != 0 &&
(fh->extend_size <= fh->size ||
(offset + fh->extend_len <= fh->extend_size &&
offset + fh->extend_len + align_size >= fh->extend_size))) {
fh->extend_size = offset + fh->extend_len * 2;
#if defined(HAVE_POSIX_FALLOCATE)
if ((ret =
posix_fallocate(fh->fd, offset, fh->extend_len * 2)) != 0)
WT_RET_MSG(
session, ret, "%s: posix_fallocate", fh->name);
#elif defined(HAVE_FTRUNCATE)
if ((ret = ftruncate(fh->fd, fh->extend_size)) != 0)
WT_RET_MSG(session, ret, "%s: ftruncate", fh->name);
#endif
}
#endif
if ((ret =
__wt_write(session, fh, offset, align_size, buf->mem)) != 0) {
if (!locked)
__wt_spin_lock(session, &block->live_lock);
WT_TRET(
__wt_block_off_free(session, block, offset, align_size));
if (!locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
}
#ifdef HAVE_SYNC_FILE_RANGE
/*
* Optionally schedule writes for dirty pages in the system buffer
* cache.
*/
if (block->os_cache_dirty_max != 0 &&
(block->os_cache_dirty += align_size) > block->os_cache_dirty_max) {
block->os_cache_dirty = 0;
if ((ret = sync_file_range(fh->fd,
(off64_t)0, (off64_t)0, SYNC_FILE_RANGE_WRITE)) != 0)
WT_RET_MSG(
session, ret, "%s: sync_file_range", block->name);
}
#endif
#ifdef HAVE_POSIX_FADVISE
/* Optionally discard blocks from the system buffer cache. */
if (block->os_cache_max != 0 &&
(block->os_cache += align_size) > block->os_cache_max) {
block->os_cache = 0;
if ((ret = posix_fadvise(fh->fd,
(off_t)0, (off_t)0, POSIX_FADV_DONTNEED)) != 0)
WT_RET_MSG(
session, ret, "%s: posix_fadvise", block->name);
}
#endif
WT_CSTAT_INCR(session, block_write);
WT_CSTAT_INCRV(session, block_byte_write, align_size);
WT_VERBOSE_RET(session, write,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, align_size, blk->cksum);
*offsetp = offset;
*sizep = align_size;
*cksump = blk->cksum;
return (ret);
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, wt_off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk, swap;
size_t bufsize;
uint32_t page_cksum;
__wt_verbose(session, WT_VERB_READ,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum);
WT_STAT_FAST_CONN_INCR(session, block_read);
WT_STAT_FAST_CONN_INCRV(session, block_byte_read, size);
/*
* Grow the buffer as necessary and read the block. Buffers should be
* aligned for reading, but there are lots of buffers (for example, file
* cursors have two buffers each, key and value), and it's difficult to
* be sure we've found all of them. If the buffer isn't aligned, it's
* an easy fix: set the flag and guarantee we reallocate it. (Most of
* the time on reads, the buffer memory has not yet been allocated, so
* we're not adding any additional processing time.)
*/
if (F_ISSET(buf, WT_ITEM_ALIGNED))
bufsize = size;
else {
F_SET(buf, WT_ITEM_ALIGNED);
bufsize = WT_MAX(size, buf->memsize + 10);
}
WT_RET(__wt_buf_init(session, buf, bufsize));
WT_RET(__wt_read(session, block->fh, offset, size, buf->mem));
buf->size = size;
/*
* We incrementally read through the structure before doing a checksum,
* do little- to big-endian handling early on, and then select from the
* original or swapped structure as needed.
*/
blk = WT_BLOCK_HEADER_REF(buf->mem);
__wt_block_header_byteswap_copy(blk, &swap);
if (swap.cksum == cksum) {
blk->cksum = 0;
page_cksum = __wt_cksum(buf->mem,
F_ISSET(&swap, WT_BLOCK_DATA_CKSUM) ?
size : WT_BLOCK_COMPRESS_SKIP);
if (page_cksum == cksum) {
/*
* Swap the page-header as needed; this doesn't belong
* here, but it's the best place to catch all callers.
*/
__wt_page_header_byteswap(buf->mem);
return (0);
}
if (!F_ISSET(session, WT_SESSION_QUIET_CORRUPT_FILE))
__wt_errx(session,
"read checksum error for %" PRIu32 "B block at "
"offset %" PRIuMAX ": calculated block checksum "
"of %" PRIu32 " doesn't match expected checksum "
"of %" PRIu32,
size, (uintmax_t)offset, page_cksum, cksum);
} else
if (!F_ISSET(session, WT_SESSION_QUIET_CORRUPT_FILE))
__wt_errx(session,
"read checksum error for %" PRIu32 "B block at "
"offset %" PRIuMAX ": block header checksum "
"of %" PRIu32 " doesn't match expected checksum "
"of %" PRIu32,
size, (uintmax_t)offset, swap.cksum, cksum);
/* Panic if a checksum fails during an ordinary read. */
return (block->verify ||
F_ISSET(session, WT_SESSION_QUIET_CORRUPT_FILE) ?
WT_ERROR : __wt_illegal_value(session, block->name));
}
/*
* __wt_block_write_off --
* Write a buffer into a block, returning the block's offset, size and
* checksum.
*/
int
__wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, wt_off_t *offsetp, uint32_t *sizep, uint32_t *cksump,
int data_cksum, int caller_locked)
{
WT_BLOCK_HEADER *blk;
WT_DECL_RET;
WT_FH *fh;
size_t align_size;
wt_off_t offset;
int local_locked;
blk = WT_BLOCK_HEADER_REF(buf->mem);
fh = block->fh;
local_locked = 0;
/* Buffers should be aligned for writing. */
if (!F_ISSET(buf, WT_ITEM_ALIGNED)) {
WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
WT_RET_MSG(session, EINVAL,
"direct I/O check: write buffer incorrectly allocated");
}
/*
* Align the size to an allocation unit.
*
* The buffer must be big enough for us to zero to the next allocsize
* boundary, this is one of the reasons the btree layer must find out
* from the block-manager layer the maximum size of the eventual write.
*/
align_size = WT_ALIGN(buf->size, block->allocsize);
if (align_size > buf->memsize) {
WT_ASSERT(session, align_size <= buf->memsize);
WT_RET_MSG(session, EINVAL,
"buffer size check: write buffer incorrectly allocated");
}
if (align_size > UINT32_MAX) {
WT_ASSERT(session, align_size <= UINT32_MAX);
WT_RET_MSG(session, EINVAL,
"buffer size check: write buffer too large to write");
}
/* Zero out any unused bytes at the end of the buffer. */
memset((uint8_t *)buf->mem + buf->size, 0, align_size - buf->size);
/*
* Set the disk size so we don't have to incrementally read blocks
* during salvage.
*/
blk->disk_size = WT_STORE_SIZE(align_size);
/*
* Update the block's checksum: if our caller specifies, checksum the
* complete data, otherwise checksum the leading WT_BLOCK_COMPRESS_SKIP
* bytes. The assumption is applications with good compression support
* turn off checksums and assume corrupted blocks won't decompress
* correctly. However, if compression failed to shrink the block, the
* block wasn't compressed, in which case our caller will tell us to
* checksum the data to detect corruption. If compression succeeded,
* we still need to checksum the first WT_BLOCK_COMPRESS_SKIP bytes
* because they're not compressed, both to give salvage a quick test
* of whether a block is useful and to give us a test so we don't lose
* the first WT_BLOCK_COMPRESS_SKIP bytes without noticing.
*/
blk->flags = 0;
if (data_cksum)
F_SET(blk, WT_BLOCK_DATA_CKSUM);
blk->cksum = 0;
blk->cksum = __wt_cksum(
buf->mem, data_cksum ? align_size : WT_BLOCK_COMPRESS_SKIP);
if (!caller_locked) {
WT_RET(__wt_block_ext_prealloc(session, 5));
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
ret = __wt_block_alloc(session, block, &offset, (wt_off_t)align_size);
/*
* Extend the file in chunks. We want to limit the number of threads
* extending the file at the same time, so choose the one thread that's
* crossing the extended boundary. We don't extend newly created files,
* and it's theoretically possible we might wait so long our extension
* of the file is passed by another thread writing single blocks, that's
* why there's a check in case the extended file size becomes too small:
* if the file size catches up, every thread tries to extend it.
*
* File extension may require locking: some variants of the system call
* used to extend the file initialize the extended space. If a writing
* thread races with the extending thread, the extending thread might
* overwrite already written data, and that would be very, very bad.
*
* Some variants of the system call to extend the file fail at run-time
* based on the filesystem type, fall back to ftruncate in that case,
* and remember that ftruncate requires locking.
*/
if (ret == 0 &&
fh->extend_len != 0 &&
(fh->extend_size <= fh->size ||
(offset + fh->extend_len <= fh->extend_size &&
offset +
fh->extend_len + (wt_off_t)align_size >= fh->extend_size))) {
fh->extend_size = offset + fh->extend_len * 2;
if (fh->fallocate_available != WT_FALLOCATE_NOT_AVAILABLE) {
/*
* Release any locally acquired lock if it's not needed
* to extend the file, extending the file might require
* updating file metadata, which can be slow. (It may be
* a bad idea to configure for file extension on systems
* that require locking over the extend call.)
*/
if (!fh->fallocate_requires_locking && local_locked) {
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
/* Extend the file. */
if ((ret = __wt_fallocate(session,
fh, offset, fh->extend_len * 2)) == ENOTSUP) {
ret = 0;
goto extend_truncate;
}
} else {
extend_truncate: /*
* We may have a caller lock or a locally acquired lock,
* but we need a lock to call ftruncate.
*/
if (!caller_locked && local_locked == 0) {
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
/*
* The truncate might fail if there's a file mapping
* (if there's an open checkpoint on the file), that's
* OK.
*/
if ((ret = __wt_ftruncate(
session, fh, offset + fh->extend_len * 2)) == EBUSY)
ret = 0;
}
}
/* Release any locally acquired lock. */
if (local_locked) {
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
WT_RET(ret);
/* Write the block. */
if ((ret =
__wt_write(session, fh, offset, align_size, buf->mem)) != 0) {
if (!caller_locked)
__wt_spin_lock(session, &block->live_lock);
WT_TRET(__wt_block_off_free(
session, block, offset, (wt_off_t)align_size));
if (!caller_locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
}
#ifdef HAVE_SYNC_FILE_RANGE
/*
* Optionally schedule writes for dirty pages in the system buffer
* cache, but only if the current session can wait.
*/
if (block->os_cache_dirty_max != 0 &&
(block->os_cache_dirty += align_size) > block->os_cache_dirty_max &&
__wt_session_can_wait(session)) {
block->os_cache_dirty = 0;
WT_RET(__wt_fsync_async(session, fh));
}
#endif
#ifdef HAVE_POSIX_FADVISE
/* Optionally discard blocks from the system buffer cache. */
if (block->os_cache_max != 0 &&
(block->os_cache += align_size) > block->os_cache_max) {
block->os_cache = 0;
if ((ret = posix_fadvise(fh->fd,
(wt_off_t)0, (wt_off_t)0, POSIX_FADV_DONTNEED)) != 0)
WT_RET_MSG(
session, ret, "%s: posix_fadvise", block->name);
}
#endif
WT_STAT_FAST_CONN_INCR(session, block_write);
WT_STAT_FAST_CONN_INCRV(session, block_byte_write, align_size);
WT_RET(__wt_verbose(session, WT_VERB_WRITE,
"off %" PRIuMAX ", size %" PRIuMAX ", cksum %" PRIu32,
(uintmax_t)offset, (uintmax_t)align_size, blk->cksum));
*offsetp = offset;
*sizep = WT_STORE_SIZE(align_size);
*cksump = blk->cksum;
return (ret);
}
/*
* __wt_block_write_off --
* Write a buffer into a block, returning the block's offset, size and
* checksum.
*/
int
__wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, wt_off_t *offsetp, uint32_t *sizep, uint32_t *cksump,
bool data_cksum, bool caller_locked)
{
WT_BLOCK_HEADER *blk;
WT_DECL_RET;
WT_FH *fh;
size_t align_size;
wt_off_t offset;
uint32_t cksum;
bool local_locked;
fh = block->fh;
/*
* Clear the block header to ensure all of it is initialized, even the
* unused fields.
*/
blk = WT_BLOCK_HEADER_REF(buf->mem);
memset(blk, 0, sizeof(*blk));
/*
* Swap the page-header as needed; this doesn't belong here, but it's
* the best place to catch all callers.
*/
__wt_page_header_byteswap(buf->mem);
/* Buffers should be aligned for writing. */
if (!F_ISSET(buf, WT_ITEM_ALIGNED)) {
WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
WT_RET_MSG(session, EINVAL,
"direct I/O check: write buffer incorrectly allocated");
}
/*
* Align the size to an allocation unit.
*
* The buffer must be big enough for us to zero to the next allocsize
* boundary, this is one of the reasons the btree layer must find out
* from the block-manager layer the maximum size of the eventual write.
*/
align_size = WT_ALIGN(buf->size, block->allocsize);
if (align_size > buf->memsize) {
WT_ASSERT(session, align_size <= buf->memsize);
WT_RET_MSG(session, EINVAL,
"buffer size check: write buffer incorrectly allocated");
}
if (align_size > UINT32_MAX) {
WT_ASSERT(session, align_size <= UINT32_MAX);
WT_RET_MSG(session, EINVAL,
"buffer size check: write buffer too large to write");
}
/* Zero out any unused bytes at the end of the buffer. */
memset((uint8_t *)buf->mem + buf->size, 0, align_size - buf->size);
/*
* Set the disk size so we don't have to incrementally read blocks
* during salvage.
*/
blk->disk_size = WT_STORE_SIZE(align_size);
/*
* Update the block's checksum: if our caller specifies, checksum the
* complete data, otherwise checksum the leading WT_BLOCK_COMPRESS_SKIP
* bytes. The assumption is applications with good compression support
* turn off checksums and assume corrupted blocks won't decompress
* correctly. However, if compression failed to shrink the block, the
* block wasn't compressed, in which case our caller will tell us to
* checksum the data to detect corruption. If compression succeeded,
* we still need to checksum the first WT_BLOCK_COMPRESS_SKIP bytes
* because they're not compressed, both to give salvage a quick test
* of whether a block is useful and to give us a test so we don't lose
* the first WT_BLOCK_COMPRESS_SKIP bytes without noticing.
*
* Checksum a little-endian version of the header, and write everything
* in little-endian format. The checksum is (potentially) returned in a
* big-endian format, swap it into place in a separate step.
*/
blk->flags = 0;
if (data_cksum)
F_SET(blk, WT_BLOCK_DATA_CKSUM);
blk->cksum = 0;
__wt_block_header_byteswap(blk);
blk->cksum = cksum = __wt_cksum(
buf->mem, data_cksum ? align_size : WT_BLOCK_COMPRESS_SKIP);
#ifdef WORDS_BIGENDIAN
blk->cksum = __wt_bswap32(blk->cksum);
#endif
/* Pre-allocate some number of extension structures. */
WT_RET(__wt_block_ext_prealloc(session, 5));
/*
* Acquire a lock, if we don't already hold one.
* Allocate space for the write, and optionally extend the file (note
* the block-extend function may release the lock).
* Release any locally acquired lock.
*/
local_locked = false;
if (!caller_locked) {
__wt_spin_lock(session, &block->live_lock);
local_locked = true;
}
ret = __wt_block_alloc(session, block, &offset, (wt_off_t)align_size);
if (ret == 0)
ret = __wt_block_extend(
session, block, fh, offset, align_size, &local_locked);
if (local_locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
/* Write the block. */
if ((ret =
__wt_write(session, fh, offset, align_size, buf->mem)) != 0) {
if (!caller_locked)
__wt_spin_lock(session, &block->live_lock);
WT_TRET(__wt_block_off_free(
session, block, offset, (wt_off_t)align_size));
if (!caller_locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
}
#ifdef HAVE_SYNC_FILE_RANGE
/*
* Optionally schedule writes for dirty pages in the system buffer
* cache, but only if the current session can wait.
*/
if (block->os_cache_dirty_max != 0 &&
(block->os_cache_dirty += align_size) > block->os_cache_dirty_max &&
__wt_session_can_wait(session)) {
block->os_cache_dirty = 0;
WT_RET(__wt_fsync_async(session, fh));
}
#endif
#ifdef HAVE_POSIX_FADVISE
/* Optionally discard blocks from the system buffer cache. */
if (block->os_cache_max != 0 &&
(block->os_cache += align_size) > block->os_cache_max) {
block->os_cache = 0;
if ((ret = posix_fadvise(fh->fd,
(wt_off_t)0, (wt_off_t)0, POSIX_FADV_DONTNEED)) != 0)
WT_RET_MSG(
session, ret, "%s: posix_fadvise", block->name);
}
#endif
WT_STAT_FAST_CONN_INCR(session, block_write);
WT_STAT_FAST_CONN_INCRV(session, block_byte_write, align_size);
WT_RET(__wt_verbose(session, WT_VERB_WRITE,
"off %" PRIuMAX ", size %" PRIuMAX ", cksum %" PRIu32,
(uintmax_t)offset, (uintmax_t)align_size, cksum));
*offsetp = offset;
*sizep = WT_STORE_SIZE(align_size);
*cksump = cksum;
return (0);
}
/*
* __wt_block_salvage_next --
* Return the address for the next potential block from the file.
*/
int
__wt_block_salvage_next(WT_SESSION_IMPL *session,
WT_BLOCK *block, uint8_t *addr, uint32_t *addr_sizep, int *eofp)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_FH *fh;
off_t max, offset;
uint32_t allocsize, cksum, size;
uint8_t *endp;
*eofp = 0;
fh = block->fh;
allocsize = block->allocsize;
WT_ERR(__wt_scr_alloc(session, allocsize, &tmp));
/* Read through the file, looking for pages. */
for (max = fh->size;;) {
offset = block->slvg_off;
if (offset >= max) { /* Check eof. */
*eofp = 1;
goto done;
}
/*
* Read the start of a possible page (an allocation-size block),
* and get a page length from it. Move to the next allocation
* sized boundary, we'll never consider this one again.
*/
WT_ERR(__wt_read(session, fh, offset, allocsize, tmp->mem));
blk = WT_BLOCK_HEADER_REF(tmp->mem);
block->slvg_off += allocsize;
/*
* The page can't be more than the min/max page size, or past
* the end of the file.
*/
size = blk->disk_size;
cksum = blk->cksum;
if (size == 0 ||
size % allocsize != 0 ||
size > WT_BTREE_PAGE_SIZE_MAX ||
offset + (off_t)size > max)
goto skip;
/*
* The block size isn't insane, read the entire block. Reading
* the block validates the checksum; if reading the block fails,
* ignore it. If reading the block succeeds, return its address
* as a possible page.
*/
if (__wt_block_read_off(
session, block, tmp, offset, size, cksum) == 0)
break;
skip: WT_VERBOSE_ERR(session, salvage,
"skipping %" PRIu32 "B at file offset %" PRIuMAX,
allocsize, (uintmax_t)offset);
/* Free the allocation-size block. */
WT_ERR(__wt_block_off_free(
session, block, offset, (off_t)allocsize));
}
/* Re-create the address cookie that should reference this block. */
endp = addr;
WT_ERR(__wt_block_addr_to_buffer(block, &endp, offset, size, cksum));
*addr_sizep = WT_PTRDIFF32(endp, addr);
done:
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session,
WT_BLOCK *block, WT_ITEM *buf, off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk;
uint32_t alloc_size, page_cksum;
WT_VERBOSE_RET(session, read,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum);
#ifdef HAVE_DIAGNOSTIC
/*
* In diagnostic mode, verify the block we're about to read isn't on
* either the available or discard lists.
*
* Don't check during salvage, it's possible we're reading an already
* freed overflow page.
*/
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
WT_RET(
__wt_block_misplaced(session, block, "read", offset, size));
#endif
/*
* Grow the buffer as necessary and read the block. Buffers should be
* aligned for reading, but there are lots of buffers (for example, file
* cursors have two buffers each, key and value), and it's difficult to
* be sure we've found all of them. If the buffer isn't aligned, it's
* an easy fix: set the flag and guarantee we reallocate it. (Most of
* the time on reads, the buffer memory has not yet been allocated, so
* we're not adding any additional processing time.)
*/
if (F_ISSET(buf, WT_ITEM_ALIGNED))
alloc_size = size;
else {
F_SET(buf, WT_ITEM_ALIGNED);
alloc_size = (uint32_t)WT_MAX(size, buf->memsize + 10);
}
WT_RET(__wt_buf_init(session, buf, alloc_size));
WT_RET(__wt_read(session, block->fh, offset, size, buf->mem));
buf->size = size;
blk = WT_BLOCK_HEADER_REF(buf->mem);
blk->cksum = 0;
page_cksum = __wt_cksum(buf->mem,
F_ISSET(blk, WT_BLOCK_DATA_CKSUM) ? size : WT_BLOCK_COMPRESS_SKIP);
if (cksum != page_cksum) {
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
__wt_errx(session,
"read checksum error [%"
PRIu32 "B @ %" PRIuMAX ", %"
PRIu32 " != %" PRIu32 "]",
size, (uintmax_t)offset, cksum, page_cksum);
return (WT_ERROR);
}
WT_CSTAT_INCR(session, block_read);
WT_CSTAT_INCRV(session, block_byte_read, size);
return (0);
}
/*将buffer的数据写入到block对应的文件中，并计算checksum和size*/
int __wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block, WT_ITEM *buf, wt_off_t *offsetp,
uint32_t *sizep, uint32_t *cksump, int data_cksum, int caller_locked)
{
WT_BLOCK_HEADER *blk;
WT_DECL_RET;
WT_FH *fh;
size_t align_size;
wt_off_t offset;
int local_locked;
blk = WT_BLOCK_HEADER_REF(buf->mem);
fh = block->fh;
local_locked = 0;
/*buf不是对齐模式，不能进行写,因为这个是和磁盘相关的写入，必须是对齐的*/
if(!F_ISSET(buf, WT_ITEM_ALIGNED)){
WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
WT_RET_MSG(session, EINVAL, "direct I/O check: write buffer incorrectly allocated");
}
/*计算buf->size按block对齐,对齐后有可能会比现有的buf->memsize大，如果大的话，不能进行写，有可能会缓冲区溢出*/
align_size = WT_ALIGN(buf->size, block->allocsize);
if (align_size > buf->memsize) {
WT_ASSERT(session, align_size <= buf->memsize);
WT_RET_MSG(session, EINVAL, "buffer size check: write buffer incorrectly allocated");
}
/*超过4G*/
if (align_size > UINT32_MAX) {
WT_ASSERT(session, align_size <= UINT32_MAX);
WT_RET_MSG(session, EINVAL, "buffer size check: write buffer too large to write");
}
/*将对其后pading的buffer位置进行清0*/
memset((uint8_t*)buf->mem + buf->size, 0, align_size - buf->size);
/*设置block header,计算存储的数据长度*/
blk->disk_size = WT_STORE_SIZE(align_size);
blk->flags = 0;
if(data_cksum)
F_SET(blk, WT_BLOCK_DATA_CKSUM);
/*计算buf的cksum*/
blk->cksum = __wt_cksum(buf->mem, data_cksum ? align_size : WT_BLOCK_COMPRESS_SKIP);
if (!caller_locked) {
WT_RET(__wt_block_ext_prealloc(session, 5));
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
ret = __wt_block_alloc(session, block, &offset, (wt_off_t)align_size);
/*判断文件是否需要进行扩大,如果不扩大就有可能存不下写入的block数据*/
if(ret == 0 && fh->extend_len != 0 && (fh->extend_size <= fh->size ||
(offset + fh->extend_len <= fh->extend_size && offset + fh->extend_len + (wt_off_t)align_size >= fh->extend_size))){
/*调整extend_size为原来的offset + extend_len的两倍*/
fh->extend_size = offset + fh->extend_len * 2;
if (fh->fallocate_available != WT_FALLOCATE_NOT_AVAILABLE) {
/*释放block->live_lock的自旋锁，因为重设文件大小会时间比较长，需要先释放自旋锁，防止CPU空转*/
if (!fh->fallocate_requires_locking && local_locked) {
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
/*扩大文件的占用空间*/
if ((ret = __wt_fallocate(session,fh, offset, fh->extend_len * 2)) == ENOTSUP) {
ret = 0;
goto extend_truncate;
}
}
else{
extend_truncate:
if (!caller_locked && local_locked == 0) {
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
/*直接调整文件大小,这个比__wt_fallocate更慢*/
if ((ret = __wt_ftruncate(session, fh, offset + fh->extend_len * 2)) == EBUSY)
ret = 0;
}
}
if(local_locked){
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
WT_RET(ret);
/*进行block的数据写入*/
ret =__wt_write(session, fh, offset, align_size, buf->mem);
if (ret != 0) {
if (!caller_locked)
__wt_spin_lock(session, &block->live_lock);
/*没写成功，将ext对应的数据返回给avail list*/
WT_TRET(__wt_block_off_free(session, block, offset, (wt_off_t)align_size));
if (!caller_locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
}
#ifdef HAVE_SYNC_FILE_RANGE
/*需要进行fsync操作,脏页太多,进行一次异步刷盘*/
if (block->os_cache_dirty_max != 0 && (block->os_cache_dirty += align_size) > block->os_cache_dirty_max && __wt_session_can_wait(session)) {
block->os_cache_dirty = 0;
WT_RET(__wt_fsync_async(session, fh));
}
#endif
#ifdef HAVE_POSIX_FADVISE
/*清理fh->fd文件对应的system page cache中的数据,这个过程可能会有IO操作,相当于同步的sync调用*/
if (block->os_cache_max != 0 && (block->os_cache += align_size) > block->os_cache_max) {
block->os_cache = 0;
if ((ret = posix_fadvise(fh->fd, (wt_off_t)0, (wt_off_t)0, POSIX_FADV_DONTNEED)) != 0)
WT_RET_MSG( session, ret, "%s: posix_fadvise", block->name);
}
#endif
WT_STAT_FAST_CONN_INCR(session, block_write);
WT_STAT_FAST_CONN_INCRV(session, block_byte_write, align_size);
WT_RET(__wt_verbose(session, WT_VERB_WRITE, "off %" PRIuMAX ", size %" PRIuMAX ", cksum %" PRIu32,
(uintmax_t)offset, (uintmax_t)align_size, blk->cksum));
*offsetp = offset;
*sizep = WT_STORE_SIZE(align_size);
*cksump = blk->cksum;
return ret;
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t result_len;
uint32_t page_cksum;
WT_VERBOSE_RET(session, read,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum);
#ifdef HAVE_DIAGNOSTIC
/*
* In diagnostic mode, verify the block we're about to read isn't on
* either the available or discard lists.
*
* Don't check during salvage, it's possible we're reading an already
* freed overflow page.
*/
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
WT_RET(
__wt_block_misplaced(session, block, "read", offset, size));
#endif
/*
* If we're compressing the file blocks, place the initial read into a
* scratch buffer, we're going to have to re-allocate more memory for
* decompression. Else check the caller's buffer size and grow it as
* necessary, there will only be one buffer.
*/
if (block->compressor == NULL) {
F_SET(buf, WT_ITEM_ALIGNED);
WT_RET(__wt_buf_init(session, buf, size));
buf->size = size;
dsk = buf->mem;
} else {
WT_RET(__wt_scr_alloc(session, size, &tmp));
tmp->size = size;
dsk = tmp->mem;
}
/* Read. */
WT_ERR(__wt_read(session, block->fh, offset, size, dsk));
blk = WT_BLOCK_HEADER_REF(dsk);
/* Validate the checksum. */
if (block->checksum &&
cksum != WT_BLOCK_CHECKSUM_NOT_SET &&
blk->cksum != WT_BLOCK_CHECKSUM_NOT_SET) {
blk->cksum = 0;
page_cksum = __wt_cksum(dsk, size);
if (page_cksum == WT_BLOCK_CHECKSUM_NOT_SET)
++page_cksum;
if (cksum != page_cksum) {
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
__wt_errx(session,
"read checksum error [%"
PRIu32 "B @ %" PRIuMAX ", %"
PRIu32 " != %" PRIu32 "]",
size, (uintmax_t)offset, cksum, page_cksum);
WT_ERR(WT_ERROR);
}
}
/*
* If the in-memory block size is larger than the on-disk block size,
* the block is compressed. Size the user's buffer, copy the skipped
* bytes of the original image into place, then decompress.
*
* If the in-memory block size is less than or equal to the on-disk
* block size, the block is not compressed.
*/
if (blk->disk_size < dsk->size) {
if (block->compressor == NULL)
WT_ERR(__wt_illegal_value(session, block->name));
WT_ERR(__wt_buf_init(session, buf, dsk->size));
buf->size = dsk->size;
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, tmp->mem, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(block->compressor->decompress(
block->compressor, &session->iface,
(uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->size - WT_BLOCK_COMPRESS_SKIP,
&result_len));
if (result_len != dsk->size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(__wt_illegal_value(session, block->name));
} else
if (block->compressor == NULL)
buf->size = dsk->size;
else
/*
* We guessed wrong: there was a compressor, but this
* block was not compressed, and now the page is in the
* wrong buffer and the buffer may be of the wrong size.
* This should be rare, why configure a compressor that
* doesn't work? Allocate a buffer of the right size
* (we used a scratch buffer which might be large), and
* copy the data into place.
*/
WT_ERR(
__wt_buf_set(session, buf, tmp->data, dsk->size));
WT_BSTAT_INCR(session, page_read);
WT_CSTAT_INCR(session, block_read);
err: __wt_scr_free(&tmp);
return (ret);
}
