git/csum-file.c

242 lines
5.5 KiB
C
Raw Normal View History

/*
* csum-file.c
*
* Copyright (C) 2005 Linus Torvalds
*
* Simple file write infrastructure for writing SHA1-summed
* files. Useful when you write a file that you want to be
* able to verify hasn't been messed with afterwards.
*/
#include "git-compat-util.h"
#include "progress.h"
#include "csum-file.h"
#include "hash.h"
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
static void verify_buffer_or_die(struct hashfile *f,
const void *buf,
unsigned int count)
{
ssize_t ret = read_in_full(f->check_fd, f->check_buffer, count);
if (ret < 0)
die_errno("%s: sha1 file read error", f->name);
if (ret != count)
die("%s: sha1 file truncated", f->name);
if (memcmp(buf, f->check_buffer, count))
die("sha1 file '%s' validation error", f->name);
}
static void flush(struct hashfile *f, const void *buf, unsigned int count)
{
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
if (0 <= f->check_fd && count)
verify_buffer_or_die(f, buf, count);
if (write_in_full(f->fd, buf, count) < 0) {
if (errno == ENOSPC)
die("sha1 file '%s' write error. Out of diskspace", f->name);
die_errno("sha1 file '%s' write error", f->name);
}
f->total += count;
display_throughput(f->tp, f->total);
}
void hashflush(struct hashfile *f)
{
unsigned offset = f->offset;
if (offset) {
hashfile: allow skipping the hash function The hashfile API is useful for generating files that include a trailing hash of the file's contents up to that point. Using such a hash is helpful for verifying the file for corruption-at-rest, such as a faulty drive causing flipped bits. Git's index file includes this trailing hash, so it uses a 'struct hashfile' to handle the I/O to the file. This was very convenient to allow using the hashfile methods during these operations. However, hashing the file contents during write comes at a performance penalty. It's slower to hash the bytes on their way to the disk than without that step. This problem is made worse by the replacement of hardware-accelerated SHA1 computations with the software-based sha1dc computation. This write cost is significant, and the checksum capability is likely not worth that cost for such a short-lived file. The index is rewritten frequently and the only time the checksum is checked is during 'git fsck'. Thus, it would be helpful to allow a user to opt-out of the hash computation. We first need to allow Git to opt-out of the hash computation in the hashfile API. The buffered writes of the API are still helpful, so it makes sense to make the change here. Introduce a new 'skip_hash' option to 'struct hashfile'. When set, the update_fn and final_fn members of the_hash_algo are skipped. When finalizing the hashfile, the trailing hash is replaced with the null hash. This use of a trailing null hash would be desireable in either case, since we do not want to special case a file format to have a different length depending on whether it was hashed or not. When the final bytes of a file are all zero, we can infer that it was written without hashing, and thus that verification is not available as a check for file consistency. This also means that we could easily toggle hashing for any file format we desire. A version of this patch has existed in the microsoft/git fork since 2017 [1] (the linked commit was rebased in 2018, but the original dates back to January 2017). Here, the change to make the index use this fast path is delayed until a later change. [1] https://github.com/microsoft/git/commit/21fed2d91410f45d85279467f21d717a2db45201 Co-authored-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2023-01-06 16:31:53 +00:00
if (!f->skip_hash)
the_hash_algo->update_fn(&f->ctx, f->buffer, offset);
flush(f, f->buffer, offset);
f->offset = 0;
}
}
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
static void free_hashfile(struct hashfile *f)
{
free(f->buffer);
free(f->check_buffer);
free(f);
}
int finalize_hashfile(struct hashfile *f, unsigned char *result,
enum fsync_component component, unsigned int flags)
{
int fd;
hashflush(f);
hashfile: allow skipping the hash function The hashfile API is useful for generating files that include a trailing hash of the file's contents up to that point. Using such a hash is helpful for verifying the file for corruption-at-rest, such as a faulty drive causing flipped bits. Git's index file includes this trailing hash, so it uses a 'struct hashfile' to handle the I/O to the file. This was very convenient to allow using the hashfile methods during these operations. However, hashing the file contents during write comes at a performance penalty. It's slower to hash the bytes on their way to the disk than without that step. This problem is made worse by the replacement of hardware-accelerated SHA1 computations with the software-based sha1dc computation. This write cost is significant, and the checksum capability is likely not worth that cost for such a short-lived file. The index is rewritten frequently and the only time the checksum is checked is during 'git fsck'. Thus, it would be helpful to allow a user to opt-out of the hash computation. We first need to allow Git to opt-out of the hash computation in the hashfile API. The buffered writes of the API are still helpful, so it makes sense to make the change here. Introduce a new 'skip_hash' option to 'struct hashfile'. When set, the update_fn and final_fn members of the_hash_algo are skipped. When finalizing the hashfile, the trailing hash is replaced with the null hash. This use of a trailing null hash would be desireable in either case, since we do not want to special case a file format to have a different length depending on whether it was hashed or not. When the final bytes of a file are all zero, we can infer that it was written without hashing, and thus that verification is not available as a check for file consistency. This also means that we could easily toggle hashing for any file format we desire. A version of this patch has existed in the microsoft/git fork since 2017 [1] (the linked commit was rebased in 2018, but the original dates back to January 2017). Here, the change to make the index use this fast path is delayed until a later change. [1] https://github.com/microsoft/git/commit/21fed2d91410f45d85279467f21d717a2db45201 Co-authored-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2023-01-06 16:31:53 +00:00
if (f->skip_hash)
hashclr(f->buffer);
else
the_hash_algo->final_fn(f->buffer, &f->ctx);
if (result)
hashcpy(result, f->buffer);
if (flags & CSUM_HASH_IN_STREAM)
flush(f, f->buffer, the_hash_algo->rawsz);
if (flags & CSUM_FSYNC)
fsync_component_or_die(component, f->fd, f->name);
if (flags & CSUM_CLOSE) {
if (close(f->fd))
die_errno("%s: sha1 file error on close", f->name);
fd = 0;
} else
fd = f->fd;
if (0 <= f->check_fd) {
char discard;
int cnt = read_in_full(f->check_fd, &discard, 1);
if (cnt < 0)
die_errno("%s: error when reading the tail of sha1 file",
f->name);
if (cnt)
die("%s: sha1 file has trailing garbage", f->name);
if (close(f->check_fd))
die_errno("%s: sha1 file error on close", f->name);
}
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
free_hashfile(f);
return fd;
}
void hashwrite(struct hashfile *f, const void *buf, unsigned int count)
{
while (count) {
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
unsigned left = f->buffer_len - f->offset;
unsigned nr = count > left ? left : count;
if (f->do_crc)
f->crc32 = crc32(f->crc32, buf, nr);
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
if (nr == f->buffer_len) {
csum-file: make hashwrite() more readable The hashwrite() method takes an input buffer and updates a hashfile's hash function while writing the data to a file. To avoid overuse of flushes, the hashfile has an internal buffer and most writes will use memcpy() to transfer data from the input 'buf' to the hashfile's buffer of size 8 * 1024 bytes. Logic introduced by a8032d12 (sha1write: don't copy full sized buffers, 2008-09-02) reduces the number of memcpy() calls when the input buffer is sufficiently longer than the hashfile's buffer, causing nr to be the length of the full buffer. In these cases, the input buffer is used directly in chunks equal to the hashfile's buffer size. This method caught my attention while investigating some performance issues, but it turns out that these performance issues were noise within the variance of the experiment. However, during this investigation, I inspected hashwrite() and misunderstood it, even after looking closely and trying to make it faster. This change simply reorganizes some parts of the loop within hashwrite() to make it clear that each batch either uses memcpy() to the hashfile's buffer or writes directly from the input buffer. The previous code relied on indirection through local variables and essentially inlined the implementation of hashflush() to reduce lines of code. Helped-by: Jeff King <peff@peff.net> Helped-by: Junio C Hamano <gitster@pobox.com> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-26 12:38:11 +00:00
/*
* Flush a full batch worth of data directly
* from the input, skipping the memcpy() to
* the hashfile's buffer. In this block,
* f->offset is necessarily zero.
*/
hashfile: allow skipping the hash function The hashfile API is useful for generating files that include a trailing hash of the file's contents up to that point. Using such a hash is helpful for verifying the file for corruption-at-rest, such as a faulty drive causing flipped bits. Git's index file includes this trailing hash, so it uses a 'struct hashfile' to handle the I/O to the file. This was very convenient to allow using the hashfile methods during these operations. However, hashing the file contents during write comes at a performance penalty. It's slower to hash the bytes on their way to the disk than without that step. This problem is made worse by the replacement of hardware-accelerated SHA1 computations with the software-based sha1dc computation. This write cost is significant, and the checksum capability is likely not worth that cost for such a short-lived file. The index is rewritten frequently and the only time the checksum is checked is during 'git fsck'. Thus, it would be helpful to allow a user to opt-out of the hash computation. We first need to allow Git to opt-out of the hash computation in the hashfile API. The buffered writes of the API are still helpful, so it makes sense to make the change here. Introduce a new 'skip_hash' option to 'struct hashfile'. When set, the update_fn and final_fn members of the_hash_algo are skipped. When finalizing the hashfile, the trailing hash is replaced with the null hash. This use of a trailing null hash would be desireable in either case, since we do not want to special case a file format to have a different length depending on whether it was hashed or not. When the final bytes of a file are all zero, we can infer that it was written without hashing, and thus that verification is not available as a check for file consistency. This also means that we could easily toggle hashing for any file format we desire. A version of this patch has existed in the microsoft/git fork since 2017 [1] (the linked commit was rebased in 2018, but the original dates back to January 2017). Here, the change to make the index use this fast path is delayed until a later change. [1] https://github.com/microsoft/git/commit/21fed2d91410f45d85279467f21d717a2db45201 Co-authored-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2023-01-06 16:31:53 +00:00
if (!f->skip_hash)
the_hash_algo->update_fn(&f->ctx, buf, nr);
csum-file: make hashwrite() more readable The hashwrite() method takes an input buffer and updates a hashfile's hash function while writing the data to a file. To avoid overuse of flushes, the hashfile has an internal buffer and most writes will use memcpy() to transfer data from the input 'buf' to the hashfile's buffer of size 8 * 1024 bytes. Logic introduced by a8032d12 (sha1write: don't copy full sized buffers, 2008-09-02) reduces the number of memcpy() calls when the input buffer is sufficiently longer than the hashfile's buffer, causing nr to be the length of the full buffer. In these cases, the input buffer is used directly in chunks equal to the hashfile's buffer size. This method caught my attention while investigating some performance issues, but it turns out that these performance issues were noise within the variance of the experiment. However, during this investigation, I inspected hashwrite() and misunderstood it, even after looking closely and trying to make it faster. This change simply reorganizes some parts of the loop within hashwrite() to make it clear that each batch either uses memcpy() to the hashfile's buffer or writes directly from the input buffer. The previous code relied on indirection through local variables and essentially inlined the implementation of hashflush() to reduce lines of code. Helped-by: Jeff King <peff@peff.net> Helped-by: Junio C Hamano <gitster@pobox.com> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-26 12:38:11 +00:00
flush(f, buf, nr);
} else {
csum-file: make hashwrite() more readable The hashwrite() method takes an input buffer and updates a hashfile's hash function while writing the data to a file. To avoid overuse of flushes, the hashfile has an internal buffer and most writes will use memcpy() to transfer data from the input 'buf' to the hashfile's buffer of size 8 * 1024 bytes. Logic introduced by a8032d12 (sha1write: don't copy full sized buffers, 2008-09-02) reduces the number of memcpy() calls when the input buffer is sufficiently longer than the hashfile's buffer, causing nr to be the length of the full buffer. In these cases, the input buffer is used directly in chunks equal to the hashfile's buffer size. This method caught my attention while investigating some performance issues, but it turns out that these performance issues were noise within the variance of the experiment. However, during this investigation, I inspected hashwrite() and misunderstood it, even after looking closely and trying to make it faster. This change simply reorganizes some parts of the loop within hashwrite() to make it clear that each batch either uses memcpy() to the hashfile's buffer or writes directly from the input buffer. The previous code relied on indirection through local variables and essentially inlined the implementation of hashflush() to reduce lines of code. Helped-by: Jeff King <peff@peff.net> Helped-by: Junio C Hamano <gitster@pobox.com> Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-03-26 12:38:11 +00:00
/*
* Copy to the hashfile's buffer, flushing only
* if it became full.
*/
memcpy(f->buffer + f->offset, buf, nr);
f->offset += nr;
left -= nr;
if (!left)
hashflush(f);
}
count -= nr;
buf = (char *) buf + nr;
}
}
struct hashfile *hashfd_check(const char *name)
{
int sink, check;
struct hashfile *f;
sink = xopen("/dev/null", O_WRONLY);
check = xopen(name, O_RDONLY);
f = hashfd(sink, name);
f->check_fd = check;
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
f->check_buffer = xmalloc(f->buffer_len);
return f;
}
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
static struct hashfile *hashfd_internal(int fd, const char *name,
struct progress *tp,
size_t buffer_len)
{
struct hashfile *f = xmalloc(sizeof(*f));
f->fd = fd;
f->check_fd = -1;
f->offset = 0;
f->total = 0;
f->tp = tp;
f->name = name;
compute a CRC32 for each object as stored in a pack The most important optimization for performance when repacking is the ability to reuse data from a previous pack as is and bypass any delta or even SHA1 computation by simply copying the raw data from one pack to another directly. The problem with this is that any data corruption within a copied object would go unnoticed and the new (repacked) pack would be self-consistent with its own checksum despite containing a corrupted object. This is a real issue that already happened at least once in the past. In some attempt to prevent this, we validate the copied data by inflating it and making sure no error is signaled by zlib. But this is still not perfect as a significant portion of a pack content is made of object headers and references to delta base objects which are not deflated and therefore not validated when repacking actually making the pack data reuse still not as safe as it could be. Of course a full SHA1 validation could be performed, but that implies full data inflating and delta replaying which is extremely costly, which cost the data reuse optimization was designed to avoid in the first place. So the best solution to this is simply to store a CRC32 of the raw pack data for each object in the pack index. This way any object in a pack can be validated before being copied as is in another pack, including header and any other non deflated data. Why CRC32 instead of a faster checksum like Adler32? Quoting Wikipedia: Jonathan Stone discovered in 2001 that Adler-32 has a weakness for very short messages. He wrote "Briefly, the problem is that, for very short packets, Adler32 is guaranteed to give poor coverage of the available bits. Don't take my word for it, ask Mark Adler. :-)" The problem is that sum A does not wrap for short messages. The maximum value of A for a 128-byte message is 32640, which is below the value 65521 used by the modulo operation. An extended explanation can be found in RFC 3309, which mandates the use of CRC32 instead of Adler-32 for SCTP, the Stream Control Transmission Protocol. In the context of a GIT pack, we have lots of small objects, especially deltas, which are likely to be quite small and in a size range for which Adler32 is dimed not to be sufficient. Another advantage of CRC32 is the possibility for recovery from certain types of small corruptions like single bit errors which are the most probable type of corruptions. OK what this patch does is to compute the CRC32 of each object written to a pack within pack-objects. It is not written to the index yet and it is obviously not validated when reusing pack data yet either. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2007-04-09 05:06:31 +00:00
f->do_crc = 0;
hashfile: allow skipping the hash function The hashfile API is useful for generating files that include a trailing hash of the file's contents up to that point. Using such a hash is helpful for verifying the file for corruption-at-rest, such as a faulty drive causing flipped bits. Git's index file includes this trailing hash, so it uses a 'struct hashfile' to handle the I/O to the file. This was very convenient to allow using the hashfile methods during these operations. However, hashing the file contents during write comes at a performance penalty. It's slower to hash the bytes on their way to the disk than without that step. This problem is made worse by the replacement of hardware-accelerated SHA1 computations with the software-based sha1dc computation. This write cost is significant, and the checksum capability is likely not worth that cost for such a short-lived file. The index is rewritten frequently and the only time the checksum is checked is during 'git fsck'. Thus, it would be helpful to allow a user to opt-out of the hash computation. We first need to allow Git to opt-out of the hash computation in the hashfile API. The buffered writes of the API are still helpful, so it makes sense to make the change here. Introduce a new 'skip_hash' option to 'struct hashfile'. When set, the update_fn and final_fn members of the_hash_algo are skipped. When finalizing the hashfile, the trailing hash is replaced with the null hash. This use of a trailing null hash would be desireable in either case, since we do not want to special case a file format to have a different length depending on whether it was hashed or not. When the final bytes of a file are all zero, we can infer that it was written without hashing, and thus that verification is not available as a check for file consistency. This also means that we could easily toggle hashing for any file format we desire. A version of this patch has existed in the microsoft/git fork since 2017 [1] (the linked commit was rebased in 2018, but the original dates back to January 2017). Here, the change to make the index use this fast path is delayed until a later change. [1] https://github.com/microsoft/git/commit/21fed2d91410f45d85279467f21d717a2db45201 Co-authored-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Kevin Willford <kewillf@microsoft.com> Signed-off-by: Derrick Stolee <derrickstolee@github.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2023-01-06 16:31:53 +00:00
f->skip_hash = 0;
the_hash_algo->init_fn(&f->ctx);
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
f->buffer_len = buffer_len;
f->buffer = xmalloc(buffer_len);
f->check_buffer = NULL;
return f;
}
csum-file.h: increase hashfile buffer size The hashfile API uses a hard-coded buffer size of 8KB and has ever since it was introduced in c38138c (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). It performs a similar function to the hashing buffers in read-cache.c, but that code was updated from 8KB to 128KB in f279894 (read-cache: make the index write buffer size 128K, 2021-02-18). The justification there was that do_write_index() improves from 1.02s to 0.72s. Since our end goal is to have the index writing code use the hashfile API, we need to unify this buffer size to avoid a performance regression. There is a buffer, 'check_buffer', that is used to verify the check_fd file descriptor. When this buffer increases to 128K to fit the data being flushed, it causes the stack to overflow the limits placed in the test suite. To avoid issues with stack size, move both 'buffer' and 'check_buffer' to be heap pointers within 'struct hashfile'. The 'check_buffer' member is left as NULL unless check_fd is set in hashfd_check(). Both buffers are cleared as part of finalize_hashfile() which also frees the full structure. Since these buffers are now on the heap, we can adjust their size based on the needs of the consumer. In particular, callers to hashfd_throughput() are expecting to report progress indicators as the buffer flushes. These callers would prefer the smaller 8k buffer to avoid large delays between updates, especially for users with slower networks. When the progress indicator is not used, the larger buffer is preferrable. By adding a new trace2 region in the chunk-format API, we can see that the writing portion of 'git multi-pack-index write' lowers from ~1.49s to ~1.47s on a Linux machine. These effects may be more pronounced or diminished on other filesystems. The end-to-end timing is too noisy to have a definitive change either way. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-05-18 18:32:46 +00:00
struct hashfile *hashfd(int fd, const char *name)
{
/*
* Since we are not going to use a progress meter to
* measure the rate of data passing through this hashfile,
* use a larger buffer size to reduce fsync() calls.
*/
return hashfd_internal(fd, name, NULL, 128 * 1024);
}
struct hashfile *hashfd_throughput(int fd, const char *name, struct progress *tp)
{
/*
* Since we are expecting to report progress of the
* write into this hashfile, use a smaller buffer
* size so the progress indicators arrive at a more
* frequent rate.
*/
return hashfd_internal(fd, name, tp, 8 * 1024);
}
void hashfile_checkpoint(struct hashfile *f, struct hashfile_checkpoint *checkpoint)
{
hashflush(f);
checkpoint->offset = f->total;
the_hash_algo->clone_fn(&checkpoint->ctx, &f->ctx);
}
int hashfile_truncate(struct hashfile *f, struct hashfile_checkpoint *checkpoint)
{
off_t offset = checkpoint->offset;
if (ftruncate(f->fd, offset) ||
lseek(f->fd, offset, SEEK_SET) != offset)
return -1;
f->total = offset;
the_hash_algo->clone_fn(&f->ctx, &checkpoint->ctx);
f->offset = 0; /* hashflush() was called in checkpoint */
return 0;
}
void crc32_begin(struct hashfile *f)
compute a CRC32 for each object as stored in a pack The most important optimization for performance when repacking is the ability to reuse data from a previous pack as is and bypass any delta or even SHA1 computation by simply copying the raw data from one pack to another directly. The problem with this is that any data corruption within a copied object would go unnoticed and the new (repacked) pack would be self-consistent with its own checksum despite containing a corrupted object. This is a real issue that already happened at least once in the past. In some attempt to prevent this, we validate the copied data by inflating it and making sure no error is signaled by zlib. But this is still not perfect as a significant portion of a pack content is made of object headers and references to delta base objects which are not deflated and therefore not validated when repacking actually making the pack data reuse still not as safe as it could be. Of course a full SHA1 validation could be performed, but that implies full data inflating and delta replaying which is extremely costly, which cost the data reuse optimization was designed to avoid in the first place. So the best solution to this is simply to store a CRC32 of the raw pack data for each object in the pack index. This way any object in a pack can be validated before being copied as is in another pack, including header and any other non deflated data. Why CRC32 instead of a faster checksum like Adler32? Quoting Wikipedia: Jonathan Stone discovered in 2001 that Adler-32 has a weakness for very short messages. He wrote "Briefly, the problem is that, for very short packets, Adler32 is guaranteed to give poor coverage of the available bits. Don't take my word for it, ask Mark Adler. :-)" The problem is that sum A does not wrap for short messages. The maximum value of A for a 128-byte message is 32640, which is below the value 65521 used by the modulo operation. An extended explanation can be found in RFC 3309, which mandates the use of CRC32 instead of Adler-32 for SCTP, the Stream Control Transmission Protocol. In the context of a GIT pack, we have lots of small objects, especially deltas, which are likely to be quite small and in a size range for which Adler32 is dimed not to be sufficient. Another advantage of CRC32 is the possibility for recovery from certain types of small corruptions like single bit errors which are the most probable type of corruptions. OK what this patch does is to compute the CRC32 of each object written to a pack within pack-objects. It is not written to the index yet and it is obviously not validated when reusing pack data yet either. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2007-04-09 05:06:31 +00:00
{
f->crc32 = crc32(0, NULL, 0);
compute a CRC32 for each object as stored in a pack The most important optimization for performance when repacking is the ability to reuse data from a previous pack as is and bypass any delta or even SHA1 computation by simply copying the raw data from one pack to another directly. The problem with this is that any data corruption within a copied object would go unnoticed and the new (repacked) pack would be self-consistent with its own checksum despite containing a corrupted object. This is a real issue that already happened at least once in the past. In some attempt to prevent this, we validate the copied data by inflating it and making sure no error is signaled by zlib. But this is still not perfect as a significant portion of a pack content is made of object headers and references to delta base objects which are not deflated and therefore not validated when repacking actually making the pack data reuse still not as safe as it could be. Of course a full SHA1 validation could be performed, but that implies full data inflating and delta replaying which is extremely costly, which cost the data reuse optimization was designed to avoid in the first place. So the best solution to this is simply to store a CRC32 of the raw pack data for each object in the pack index. This way any object in a pack can be validated before being copied as is in another pack, including header and any other non deflated data. Why CRC32 instead of a faster checksum like Adler32? Quoting Wikipedia: Jonathan Stone discovered in 2001 that Adler-32 has a weakness for very short messages. He wrote "Briefly, the problem is that, for very short packets, Adler32 is guaranteed to give poor coverage of the available bits. Don't take my word for it, ask Mark Adler. :-)" The problem is that sum A does not wrap for short messages. The maximum value of A for a 128-byte message is 32640, which is below the value 65521 used by the modulo operation. An extended explanation can be found in RFC 3309, which mandates the use of CRC32 instead of Adler-32 for SCTP, the Stream Control Transmission Protocol. In the context of a GIT pack, we have lots of small objects, especially deltas, which are likely to be quite small and in a size range for which Adler32 is dimed not to be sufficient. Another advantage of CRC32 is the possibility for recovery from certain types of small corruptions like single bit errors which are the most probable type of corruptions. OK what this patch does is to compute the CRC32 of each object written to a pack within pack-objects. It is not written to the index yet and it is obviously not validated when reusing pack data yet either. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2007-04-09 05:06:31 +00:00
f->do_crc = 1;
}
uint32_t crc32_end(struct hashfile *f)
compute a CRC32 for each object as stored in a pack The most important optimization for performance when repacking is the ability to reuse data from a previous pack as is and bypass any delta or even SHA1 computation by simply copying the raw data from one pack to another directly. The problem with this is that any data corruption within a copied object would go unnoticed and the new (repacked) pack would be self-consistent with its own checksum despite containing a corrupted object. This is a real issue that already happened at least once in the past. In some attempt to prevent this, we validate the copied data by inflating it and making sure no error is signaled by zlib. But this is still not perfect as a significant portion of a pack content is made of object headers and references to delta base objects which are not deflated and therefore not validated when repacking actually making the pack data reuse still not as safe as it could be. Of course a full SHA1 validation could be performed, but that implies full data inflating and delta replaying which is extremely costly, which cost the data reuse optimization was designed to avoid in the first place. So the best solution to this is simply to store a CRC32 of the raw pack data for each object in the pack index. This way any object in a pack can be validated before being copied as is in another pack, including header and any other non deflated data. Why CRC32 instead of a faster checksum like Adler32? Quoting Wikipedia: Jonathan Stone discovered in 2001 that Adler-32 has a weakness for very short messages. He wrote "Briefly, the problem is that, for very short packets, Adler32 is guaranteed to give poor coverage of the available bits. Don't take my word for it, ask Mark Adler. :-)" The problem is that sum A does not wrap for short messages. The maximum value of A for a 128-byte message is 32640, which is below the value 65521 used by the modulo operation. An extended explanation can be found in RFC 3309, which mandates the use of CRC32 instead of Adler-32 for SCTP, the Stream Control Transmission Protocol. In the context of a GIT pack, we have lots of small objects, especially deltas, which are likely to be quite small and in a size range for which Adler32 is dimed not to be sufficient. Another advantage of CRC32 is the possibility for recovery from certain types of small corruptions like single bit errors which are the most probable type of corruptions. OK what this patch does is to compute the CRC32 of each object written to a pack within pack-objects. It is not written to the index yet and it is obviously not validated when reusing pack data yet either. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2007-04-09 05:06:31 +00:00
{
f->do_crc = 0;
return f->crc32;
}
int hashfile_checksum_valid(const unsigned char *data, size_t total_len)
{
unsigned char got[GIT_MAX_RAWSZ];
git_hash_ctx ctx;
size_t data_len = total_len - the_hash_algo->rawsz;
if (total_len < the_hash_algo->rawsz)
return 0; /* say "too short"? */
the_hash_algo->init_fn(&ctx);
the_hash_algo->update_fn(&ctx, data, data_len);
the_hash_algo->final_fn(got, &ctx);
return hasheq(got, data + data_len);
}