Git Source Code Mirror - This is a publish-only repository and all pull requests are ignored. Please follow Documentation/SubmittingPatches procedure for any of your improvements. https://git-scm.com/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
git/read-cache.c

3801 lines
103 KiB

/*
* GIT - The information manager from hell
*
* Copyright (C) Linus Torvalds, 2005
*/
#include "cache.h"
#include "config.h"
#include "diff.h"
#include "diffcore.h"
#include "tempfile.h"
#include "lockfile.h"
#include "cache-tree.h"
#include "refs.h"
#include "dir.h"
#include "object-store.h"
#include "tree.h"
#include "commit.h"
#include "blob.h"
#include "resolve-undo.h"
#include "run-command.h"
#include "strbuf.h"
#include "varint.h"
#include "split-index.h"
#include "utf8.h"
#include "fsmonitor.h"
#include "thread-utils.h"
#include "progress.h"
sparse-index: convert from full to sparse If we have a full index, then we can convert it to a sparse index by replacing directories outside of the sparse cone with sparse directory entries. The convert_to_sparse() method does this, when the situation is appropriate. For now, we avoid converting the index to a sparse index if: 1. the index is split. 2. the index is already sparse. 3. sparse-checkout is disabled. 4. sparse-checkout does not use cone mode. Finally, we currently limit the conversion to when the GIT_TEST_SPARSE_INDEX environment variable is enabled. A mode using Git config will be added in a later change. The trickiest thing about this conversion is that we might not be able to mark a directory as a sparse directory just because it is outside the sparse cone. There might be unmerged files within that directory, so we need to look for those. Also, if there is some strange reason why a file is not marked with CE_SKIP_WORKTREE, then we should give up on converting that directory. There is still hope that some of its subdirectories might be able to convert to sparse, so we keep looking deeper. The conversion process is assisted by the cache-tree extension. This is calculated from the full index if it does not already exist. We then abandon the cache-tree as it no longer applies to the newly-sparse index. Thus, this cache-tree will be recalculated in every sparse-full-sparse round-trip until we integrate the cache-tree extension with the sparse index. Some Git commands use the index after writing it. For example, 'git add' will update the index, then write it to disk, then read its entries to report information. To keep the in-memory index in a full state after writing, we re-expand it to a full one after the write. This is wasteful for commands that only write the index and do not read from it again, but that is only the case until we make those commands "sparse aware." We can compare the behavior of the sparse-index in t1092-sparse-checkout-compability.sh by using GIT_TEST_SPARSE_INDEX=1 when operating on the 'sparse-index' repo. We can also compare the two sparse repos directly, such as comparing their indexes (when expanded to full in the case of the 'sparse-index' repo). We also verify that the index is actually populated with sparse directory entries. The 'checkout and reset (mixed)' test is marked for failure when comparing a sparse repo to a full repo, but we can compare the two sparse-checkout cases directly to ensure that we are not changing the behavior when using a sparse index. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2 years ago
#include "sparse-index.h"
read-cache: use hashfile instead of git_hash_ctx The do_write_index() method in read-cache.c has its own hashing logic and buffering mechanism. Specifically, the ce_write() method was introduced by 4990aadc (Speed up index file writing by chunking it nicely, 2005-04-20) and similar mechanisms were introduced a few months later in c38138cd (git-pack-objects: write the pack files with a SHA1 csum, 2005-06-26). Based on the timing, in the early days of the Git codebase, I figured that these roughly equivalent code paths were never unified only because it got lost in the shuffle. The hashfile API has since been used extensively in other file formats, such as pack-indexes, multi-pack-indexes, and commit-graphs. Therefore, it seems prudent to unify the index writing code to use the same mechanism. I discovered this disparity while trying to create a new index format that uses the chunk-format API. That API uses a hashfile as its base, so it is incompatible with the custom code in read-cache.c. This rewrite is rather straightforward. It replaces all writes to the temporary file with writes to the hashfile struct. This takes care of many of the direct interactions with the_hash_algo. There are still some git_hash_ctx uses remaining: the extension headers are hashed for use in the End of Index Entries (EOIE) extension. This use of the git_hash_ctx is left as-is. There are multiple reasons to not use a hashfile here, including the fact that the data is not actually writing to a file, just a hash computation. These hashes do not block our adoption of the chunk-format API in a future change to the index, so leave it as-is. The internals of the algorithms are mostly identical. Previously, the hashfile API used a smaller 8KB buffer instead of the 128KB buffer from read-cache.c. The previous change already unified these sizes. There is one subtle point: we do not pass the CSUM_FSYNC to the finalize_hashfile() method, which differs from most consumers of the hashfile API. The extra fsync() call indicated by this flag causes a significant peformance degradation that is noticeable for quick commands that write the index, such as "git add". Other consumers can absorb this cost with their more complicated data structure organization, and further writing structures such as pack-files and commit-graphs is rarely in the critical path for common user interactions. Some static methods become orphaned in this diff, so I marked them as MAYBE_UNUSED. The diff is much harder to read if they are deleted during this change. Instead, they will be deleted in the following change. In addition to the test suite passing, I computed indexes using the previous binaries and the binaries compiled after this change, and found the index data to be exactly equal. Finally, I did extensive performance testing of "git update-index --force-write" on repos of various sizes, including one with over 2 million paths at HEAD. These tests demonstrated less than 1% difference in behavior. As expected, the performance should be considered unchanged. The previous changes to increase the hashfile buffer size from 8K to 128K ensured this change would not create a peformance regression. Signed-off-by: Derrick Stolee <dstolee@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
1 year ago
#include "csum-file.h"
#include "promisor-remote.h"
#include "hook.h"
/* Mask for the name length in ce_flags in the on-disk index */
#define CE_NAMEMASK (0x0fff)
/* Index extensions.
*
* The first letter should be 'A'..'Z' for extensions that are not
* necessary for a correct operation (i.e. optimization data).
* When new extensions are added that _needs_ to be understood in
* order to correctly interpret the index file, pick character that
* is outside the range, to cause the reader to abort.
*/
#define CACHE_EXT(s) ( (s[0]<<24)|(s[1]<<16)|(s[2]<<8)|(s[3]) )
#define CACHE_EXT_TREE 0x54524545 /* "TREE" */
#define CACHE_EXT_RESOLVE_UNDO 0x52455543 /* "REUC" */
#define CACHE_EXT_LINK 0x6c696e6b /* "link" */
#define CACHE_EXT_UNTRACKED 0x554E5452 /* "UNTR" */
#define CACHE_EXT_FSMONITOR 0x46534D4E /* "FSMN" */
#define CACHE_EXT_ENDOFINDEXENTRIES 0x454F4945 /* "EOIE" */
#define CACHE_EXT_INDEXENTRYOFFSETTABLE 0x49454F54 /* "IEOT" */
#define CACHE_EXT_SPARSE_DIRECTORIES 0x73646972 /* "sdir" */
/* changes that can be kept in $GIT_DIR/index (basically all extensions) */
#define EXTMASK (RESOLVE_UNDO_CHANGED | CACHE_TREE_CHANGED | \
CE_ENTRY_ADDED | CE_ENTRY_REMOVED | CE_ENTRY_CHANGED | \
SPLIT_INDEX_ORDERED | UNTRACKED_CHANGED | FSMONITOR_CHANGED)
block alloc: allocate cache entries from mem_pool When reading large indexes from disk, a portion of the time is dominated in malloc() calls. This can be mitigated by allocating a large block of memory and manage it ourselves via memory pools. This change moves the cache entry allocation to be on top of memory pools. Design: The index_state struct will gain a notion of an associated memory_pool from which cache_entries will be allocated from. When reading in the index from disk, we have information on the number of entries and their size, which can guide us in deciding how large our initial memory allocation should be. When an index is discarded, the associated memory_pool will be discarded as well - so the lifetime of a cache_entry is tied to the lifetime of the index_state that it was allocated for. In the case of a Split Index, the following rules are followed. 1st, some terminology is defined: Terminology: - 'the_index': represents the logical view of the index - 'split_index': represents the "base" cache entries. Read from the split index file. 'the_index' can reference a single split_index, as well as cache_entries from the split_index. `the_index` will be discarded before the `split_index` is. This means that when we are allocating cache_entries in the presence of a split index, we need to allocate the entries from the `split_index`'s memory pool. This allows us to follow the pattern that `the_index` can reference cache_entries from the `split_index`, and that the cache_entries will not be freed while they are still being referenced. Managing transient cache_entry structs: Cache entries are usually allocated for an index, but this is not always the case. Cache entries are sometimes allocated because this is the type that the existing checkout_entry function works with. Because of this, the existing code needs to handle cache entries associated with an index / memory pool, and those that only exist transiently. Several strategies were contemplated around how to handle this: Chosen approach: An extra field was added to the cache_entry type to track whether the cache_entry was allocated from a memory pool or not. This is currently an int field, as there are no more available bits in the existing ce_flags bit field. If / when more bits are needed, this new field can be turned into a proper bit field. Alternatives: 1) Do not include any information about how the cache_entry was allocated. Calling code would be responsible for tracking whether the cache_entry needed to be freed or not. Pro: No extra memory overhead to track this state Con: Extra complexity in callers to handle this correctly. The extra complexity and burden to not regress this behavior in the future was more than we wanted. 2) cache_entry would gain knowledge about which mem_pool allocated it Pro: Could (potentially) do extra logic to know when a mem_pool no longer had references to any cache_entry Con: cache_entry would grow heavier by a pointer, instead of int We didn't see a tangible benefit to this approach 3) Do not add any extra information to a cache_entry, but when freeing a cache entry, check if the memory exists in a region managed by existing mem_pools. Pro: No extra memory overhead to track state Con: Extra computation is performed when freeing cache entries We decided tracking and iterating over known memory pool regions was less desirable than adding an extra field to track this stae. Signed-off-by: Jameson Miller <jamill@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
/*
* This is an estimate of the pathname length in the index. We use
* this for V4 index files to guess the un-deltafied size of the index
* in memory because of pathname deltafication. This is not required
* for V2/V3 index formats because their pathnames are not compressed.
* If the initial amount of memory set aside is not sufficient, the
* mem pool will allocate extra memory.
*/
#define CACHE_ENTRY_PATH_LENGTH 80
enum index_search_mode {
NO_EXPAND_SPARSE = 0,
EXPAND_SPARSE = 1
};
block alloc: allocate cache entries from mem_pool When reading large indexes from disk, a portion of the time is dominated in malloc() calls. This can be mitigated by allocating a large block of memory and manage it ourselves via memory pools. This change moves the cache entry allocation to be on top of memory pools. Design: The index_state struct will gain a notion of an associated memory_pool from which cache_entries will be allocated from. When reading in the index from disk, we have information on the number of entries and their size, which can guide us in deciding how large our initial memory allocation should be. When an index is discarded, the associated memory_pool will be discarded as well - so the lifetime of a cache_entry is tied to the lifetime of the index_state that it was allocated for. In the case of a Split Index, the following rules are followed. 1st, some terminology is defined: Terminology: - 'the_index': represents the logical view of the index - 'split_index': represents the "base" cache entries. Read from the split index file. 'the_index' can reference a single split_index, as well as cache_entries from the split_index. `the_index` will be discarded before the `split_index` is. This means that when we are allocating cache_entries in the presence of a split index, we need to allocate the entries from the `split_index`'s memory pool. This allows us to follow the pattern that `the_index` can reference cache_entries from the `split_index`, and that the cache_entries will not be freed while they are still being referenced. Managing transient cache_entry structs: Cache entries are usually allocated for an index, but this is not always the case. Cache entries are sometimes allocated because this is the type that the existing checkout_entry function works with. Because of this, the existing code needs to handle cache entries associated with an index / memory pool, and those that only exist transiently. Several strategies were contemplated around how to handle this: Chosen approach: An extra field was added to the cache_entry type to track whether the cache_entry was allocated from a memory pool or not. This is currently an int field, as there are no more available bits in the existing ce_flags bit field. If / when more bits are needed, this new field can be turned into a proper bit field. Alternatives: 1) Do not include any information about how the cache_entry was allocated. Calling code would be responsible for tracking whether the cache_entry needed to be freed or not. Pro: No extra memory overhead to track this state Con: Extra complexity in callers to handle this correctly. The extra complexity and burden to not regress this behavior in the future was more than we wanted. 2) cache_entry would gain knowledge about which mem_pool allocated it Pro: Could (potentially) do extra logic to know when a mem_pool no longer had references to any cache_entry Con: cache_entry would grow heavier by a pointer, instead of int We didn't see a tangible benefit to this approach 3) Do not add any extra information to a cache_entry, but when freeing a cache entry, check if the memory exists in a region managed by existing mem_pools. Pro: No extra memory overhead to track state Con: Extra computation is performed when freeing cache entries We decided tracking and iterating over known memory pool regions was less desirable than adding an extra field to track this stae. Signed-off-by: Jameson Miller <jamill@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
static inline struct cache_entry *mem_pool__ce_alloc(struct mem_pool *mem_pool, size_t len)
{
struct cache_entry *ce;
ce = mem_pool_alloc(mem_pool, cache_entry_size(len));
ce->mem_pool_allocated = 1;
return ce;
}
static inline struct cache_entry *mem_pool__ce_calloc(struct mem_pool *mem_pool, size_t len)
{
struct cache_entry * ce;
ce = mem_pool_calloc(mem_pool, 1, cache_entry_size(len));
ce->mem_pool_allocated = 1;
return ce;
}
static struct mem_pool *find_mem_pool(struct index_state *istate)
{
struct mem_pool **pool_ptr;
if (istate->split_index && istate->split_index->base)
pool_ptr = &istate->split_index->base->ce_mem_pool;
else
pool_ptr = &istate->ce_mem_pool;
mem-pool: use more standard initialization and finalization A typical memory type, such as strbuf, hashmap, or string_list can be stored on the stack or embedded within another structure. mem_pool cannot be, because of how mem_pool_init() and mem_pool_discard() are written. mem_pool_init() does essentially the following (simplified for purposes of explanation here): void mem_pool_init(struct mem_pool **pool...) { *pool = xcalloc(1, sizeof(*pool)); It seems weird to require that mem_pools can only be accessed through a pointer. It also seems slightly dangerous: unlike strbuf_release() or strbuf_reset() or string_list_clear(), all of which put the data structure into a state where it can be re-used after the call, mem_pool_discard(pool) will leave pool pointing at free'd memory. read-cache (and split-index) are the only current users of mem_pools, and they haven't fallen into a use-after-free mistake here, but it seems likely to be problematic for future users especially since several of the current callers of mem_pool_init() will only call it when the mem_pool* is not already allocated (i.e. is NULL). This type of mechanism also prevents finding synchronization points where one can free existing memory and then resume more operations. It would be natural at such points to run something like mem_pool_discard(pool...); and, if necessary, mem_pool_init(&pool...); and then carry on continuing to use the pool. However, this fails badly if several objects had a copy of the value of pool from before these commands; in such a case, those objects won't get the updated value of pool that mem_pool_init() overwrites pool with and they'll all instead be reading and writing from free'd memory. Modify mem_pool_init()/mem_pool_discard() to behave more like strbuf_init()/strbuf_release() or string_list_init()/string_list_clear() In particular: (1) make mem_pool_init() just take a mem_pool* and have it only worry about allocating struct mp_blocks, not the struct mem_pool itself, (2) make mem_pool_discard() free the memory that the pool was responsible for, but leave it in a state where it can be used to allocate more memory afterward (without the need to call mem_pool_init() again). Signed-off-by: Elijah Newren <newren@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2 years ago
if (!*pool_ptr) {
*pool_ptr = xmalloc(sizeof(**pool_ptr));
mem_pool_init(*pool_ptr, 0);
}
block alloc: allocate cache entries from mem_pool When reading large indexes from disk, a portion of the time is dominated in malloc() calls. This can be mitigated by allocating a large block of memory and manage it ourselves via memory pools. This change moves the cache entry allocation to be on top of memory pools. Design: The index_state struct will gain a notion of an associated memory_pool from which cache_entries will be allocated from. When reading in the index from disk, we have information on the number of entries and their size, which can guide us in deciding how large our initial memory allocation should be. When an index is discarded, the associated memory_pool will be discarded as well - so the lifetime of a cache_entry is tied to the lifetime of the index_state that it was allocated for. In the case of a Split Index, the following rules are followed. 1st, some terminology is defined: Terminology: - 'the_index': represents the logical view of the index - 'split_index': represents the "base" cache entries. Read from the split index file. 'the_index' can reference a single split_index, as well as cache_entries from the split_index. `the_index` will be discarded before the `split_index` is. This means that when we are allocating cache_entries in the presence of a split index, we need to allocate the entries from the `split_index`'s memory pool. This allows us to follow the pattern that `the_index` can reference cache_entries from the `split_index`, and that the cache_entries will not be freed while they are still being referenced. Managing transient cache_entry structs: Cache entries are usually allocated for an index, but this is not always the case. Cache entries are sometimes allocated because this is the type that the existing checkout_entry function works with. Because of this, the existing code needs to handle cache entries associated with an index / memory pool, and those that only exist transiently. Several strategies were contemplated around how to handle this: Chosen approach: An extra field was added to the cache_entry type to track whether the cache_entry was allocated from a memory pool or not. This is currently an int field, as there are no more available bits in the existing ce_flags bit field. If / when more bits are needed, this new field can be turned into a proper bit field. Alternatives: 1) Do not include any information about how the cache_entry was allocated. Calling code would be responsible for tracking whether the cache_entry needed to be freed or not. Pro: No extra memory overhead to track this state Con: Extra complexity in callers to handle this correctly. The extra complexity and burden to not regress this behavior in the future was more than we wanted. 2) cache_entry would gain knowledge about which mem_pool allocated it Pro: Could (potentially) do extra logic to know when a mem_pool no longer had references to any cache_entry Con: cache_entry would grow heavier by a pointer, instead of int We didn't see a tangible benefit to this approach 3) Do not add any extra information to a cache_entry, but when freeing a cache entry, check if the memory exists in a region managed by existing mem_pools. Pro: No extra memory overhead to track state Con: Extra computation is performed when freeing cache entries We decided tracking and iterating over known memory pool regions was less desirable than adding an extra field to track this stae. Signed-off-by: Jameson Miller <jamill@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
return *pool_ptr;
}
static const char *alternate_index_output;
static void set_index_entry(struct index_state *istate, int nr, struct cache_entry *ce)
{
if (S_ISSPARSEDIR(ce->ce_mode))
istate->sparse_index = INDEX_COLLAPSED;
istate->cache[nr] = ce;
add_name_hash(istate, ce);
}
Create pathname-based hash-table lookup into index This creates a hash index of every single file added to the index. Right now that hash index isn't actually used for much: I implemented a "cache_name_exists()" function that uses it to efficiently look up a filename in the index without having to do the O(logn) binary search, but quite frankly, that's not why this patch is interesting. No, the whole and only reason to create the hash of the filenames in the index is that by modifying the hash function, you can fairly easily do things like making it always hash equivalent names into the same bucket. That, in turn, means that suddenly questions like "does this name exist in the index under an _equivalent_ name?" becomes much much cheaper. Guiding principles behind this patch: - it shouldn't be too costly. In fact, my primary goal here was to actually speed up "git commit" with a fully populated kernel tree, by being faster at checking whether a file already existed in the index. I did succeed, but only barely: Best before: [torvalds@woody linux]$ time git commit > /dev/null real 0m0.255s user 0m0.168s sys 0m0.088s Best after: [torvalds@woody linux]$ time ~/git/git commit > /dev/null real 0m0.233s user 0m0.144s sys 0m0.088s so some things are actually faster (~8%). Caveat: that's really the best case. Other things are invariably going to be slightly slower, since we populate that index cache, and quite frankly, few things really use it to look things up. That said, the cost is really quite small. The worst case is probably doing a "git ls-files", which will do very little except puopulate the index, and never actually looks anything up in it, just lists it. Before: [torvalds@woody linux]$ time git ls-files > /dev/null real 0m0.016s user 0m0.016s sys 0m0.000s After: [torvalds@woody linux]$ time ~/git/git ls-files > /dev/null real 0m0.021s user 0m0.012s sys 0m0.008s and while the thing has really gotten relatively much slower, we're still talking about something almost unmeasurable (eg 5ms). And that really should be pretty much the worst case. So we lose 5ms on one "benchmark", but win 22ms on another. Pick your poison - this patch has the advantage that it will _likely_ speed up the cases that are complex and expensive more than it slows down the cases that are already so fast that nobody cares. But if you look at relative speedups/slowdowns, it doesn't look so good. - It should be simple and clean The code may be a bit subtle (the reasons I do hash removal the way I do etc), but it re-uses the existing hash.c files, so it really is fairly small and straightforward apart from a few odd details. Now, this patch on its own doesn't really do much, but I think it's worth looking at, if only because if done correctly, the name hashing really can make an improvement to the whole issue of "do we have a filename that looks like this in the index already". And at least it gets real testing by being used even by default (ie there is a real use-case for it even without any insane filesystems). NOTE NOTE NOTE! The current hash is a joke. I'm ashamed of it, I'm just not ashamed of it enough to really care. I took all the numbers out of my nether regions - I'm sure it's good enough that it works in practice, but the whole point was that you can make a really much fancier hash that hashes characters not directly, but by their upper-case value or something like that, and thus you get a case-insensitive hash, while still keeping the name and the index itself totally case sensitive. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
15 years ago
static void replace_index_entry(struct index_state *istate, int nr, struct cache_entry *ce)
{
struct cache_entry *old = istate->cache[nr];
replace_index_entry_in_base(istate, old, ce);
name-hash.c: fix endless loop with core.ignorecase=true With core.ignorecase=true, name-hash.c builds a case insensitive index of all tracked directories. Currently, the existing cache entry structures are added multiple times to the same hashtable (with different name lengths and hash codes). However, there's only one dir_next pointer, which gets completely messed up in case of hash collisions. In the worst case, this causes an endless loop if ce == ce->dir_next (see t7062). Use a separate hashtable and separate structures for the directory index so that each directory entry has its own next pointer. Use reference counting to track which directory entry contains files. There are only slight changes to the name-hash.c API: - new free_name_hash() used by read_cache.c::discard_index() - remove_name_hash() takes an additional index_state parameter - index_name_exists() for a directory (trailing '/') may return a cache entry that has been removed (CE_UNHASHED). This is not a problem as the return value is only used to check if the directory exists (dir.c) or to normalize casing of directory names (read-cache.c). Getting rid of cache_entry.dir_next reduces memory consumption, especially with core.ignorecase=false (which doesn't use that member at all). With core.ignorecase=true, building the directory index is slightly faster as we add / check the parent directory first (instead of going through all directory levels for each file in the index). E.g. with WebKit (~200k files, ~7k dirs), time spent in lazy_init_name_hash is reduced from 176ms to 130ms. Signed-off-by: Karsten Blees <blees@dcon.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
10 years ago
remove_name_hash(istate, old);
block alloc: add lifecycle APIs for cache_entry structs It has been observed that the time spent loading an index with a large number of entries is partly dominated by malloc() calls. This change is in preparation for using memory pools to reduce the number of malloc() calls made to allocate cahce entries when loading an index. Add an API to allocate and discard cache entries, abstracting the details of managing the memory backing the cache entries. This commit does actually change how memory is managed - this will be done in a later commit in the series. This change makes the distinction between cache entries that are associated with an index and cache entries that are not associated with an index. A main use of cache entries is with an index, and we can optimize the memory management around this. We still have other cases where a cache entry is not persisted with an index, and so we need to handle the "transient" use case as well. To keep the congnitive overhead of managing the cache entries, there will only be a single discard function. This means there must be enough information kept with the cache entry so that we know how to discard them. A summary of the main functions in the API is: make_cache_entry: create cache entry for use in an index. Uses specified parameters to populate cache_entry fields. make_empty_cache_entry: Create an empty cache entry for use in an index. Returns cache entry with empty fields. make_transient_cache_entry: create cache entry that is not used in an index. Uses specified parameters to populate cache_entry fields. make_empty_transient_cache_entry: create cache entry that is not used in an index. Returns cache entry with empty fields. discard_cache_entry: A single function that knows how to discard a cache entry regardless of how it was allocated. Signed-off-by: Jameson Miller <jamill@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
discard_cache_entry(old);
ce->ce_flags &= ~CE_HASHED;
Fix name re-hashing semantics We handled the case of removing and re-inserting cache entries badly, which is something that merging commonly needs to do (removing the different stages, and then re-inserting one of them as the merged state). We even had a rather ugly special case for this failure case, where replace_index_entry() basically turned itself into a no-op if the new and the old entries were the same, exactly because the hash routines didn't handle it on their own. So what this patch does is to not just have the UNHASHED bit, but a HASHED bit too, and when you insert an entry into the name hash, that involves: - clear the UNHASHED bit, because now it's valid again for lookup (which is really all that UNHASHED meant) - if we're being lazy, we're done here (but we still want to clear the UNHASHED bit regardless of lazy mode, since we can become unlazy later, and so we need the UNHASHED bit to always be set correctly, even if we never actually insert the entry into the hash list) - if it was already hashed, we just leave it on the list - otherwise mark it HASHED and insert it into the list this all means that unhashing and rehashing a name all just works automatically. Obviously, you cannot change the name of an entry (that would be a serious bug), but nothing can validly do that anyway (you'd have to allocate a new struct cache_entry anyway since the name length could change), so that's not a new limitation. The code actually gets simpler in many ways, although the lazy hashing does mean that there are a few odd cases (ie something can be marked unhashed even though it was never on the hash in the first place, and isn't actually marked hashed!). Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
15 years ago
set_index_entry(istate, nr, ce);
ce->ce_flags |= CE_UPDATE_IN_BASE;
mark_fsmonitor_invalid(istate, ce);
istate->cache_changed |= CE_ENTRY_CHANGED;
Create pathname-based hash-table lookup into index This creates a hash index of every single file added to the index. Right now that hash index isn't actually used for much: I implemented a "cache_name_exists()" function that uses it to efficiently look up a filename in the index without having to do the O(logn) binary search, but quite frankly, that's not why this patch is interesting. No, the whole and only reason to create the hash of the filenames in the index is that by modifying the hash function, you can fairly easily do things like making it always hash equivalent names into the same bucket. That, in turn, means that suddenly questions like "does this name exist in the index under an _equivalent_ name?" becomes much much cheaper. Guiding principles behind this patch: - it shouldn't be too costly. In fact, my primary goal here was to actually speed up "git commit" with a fully populated kernel tree, by being faster at checking whether a file already existed in the index. I did succeed, but only barely: Best before: [torvalds@woody linux]$ time git commit > /dev/null real 0m0.255s user 0m0.168s sys 0m0.088s Best after: [torvalds@woody linux]$ time ~/git/git commit > /dev/null real 0m0.233s user 0m0.144s sys 0m0.088s so some things are actually faster (~8%). Caveat: that's really the best case. Other things are invariably going to be slightly slower, since we populate that index cache, and quite frankly, few things really use it to look things up. That said, the cost is really quite small. The worst case is probably doing a "git ls-files", which will do very little except puopulate the index, and never actually looks anything up in it, just lists it. Before: [torvalds@woody linux]$ time git ls-files > /dev/null real 0m0.016s user 0m0.016s sys 0m0.000s After: [torvalds@woody linux]$ time ~/git/git ls-files > /dev/null real 0m0.021s user 0m0.012s sys 0m0.008s and while the thing has really gotten relatively much slower, we're still talking about something almost unmeasurable (eg 5ms). And that really should be pretty much the worst case. So we lose 5ms on one "benchmark", but win 22ms on another. Pick your poison - this patch has the advantage that it will _likely_ speed up the cases that are complex and expensive more than it slows down the cases that are already so fast that nobody cares. But if you look at relative speedups/slowdowns, it doesn't look so good. - It should be simple and clean The code may be a bit subtle (the reasons I do hash removal the way I do etc), but it re-uses the existing hash.c files, so it really is fairly small and straightforward apart from a few odd details. Now, this patch on its own doesn't really do much, but I think it's worth looking at, if only because if done correctly, the name hashing really can make an improvement to the whole issue of "do we have a filename that looks like this in the index already". And at least it gets real testing by being used even by default (ie there is a real use-case for it even without any insane filesystems). NOTE NOTE NOTE! The current hash is a joke. I'm ashamed of it, I'm just not ashamed of it enough to really care. I took all the numbers out of my nether regions - I'm sure it's good enough that it works in practice, but the whole point was that you can make a really much fancier hash that hashes characters not directly, but by their upper-case value or something like that, and thus you get a case-insensitive hash, while still keeping the name and the index itself totally case sensitive. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Junio C Hamano <gitster@pobox.com>
15 years ago
}
void rename_index_entry_at(struct index_state *istate, int nr, const char *new_name)
{
struct cache_entry *old_entry = istate->cache[nr], *new_entry, *refreshed;
int namelen = strlen(new_name);
block alloc: add lifecycle APIs for cache_entry structs It has been observed that the time spent loading an index with a large number of entries is partly dominated by malloc() calls. This change is in preparation for using memory pools to reduce the number of malloc() calls made to allocate cahce entries when loading an index. Add an API to allocate and discard cache entries, abstracting the details of managing the memory backing the cache entries. This commit does actually change how memory is managed - this will be done in a later commit in the series. This change makes the distinction between cache entries that are associated with an index and cache entries that are not associated with an index. A main use of cache entries is with an index, and we can optimize the memory management around this. We still have other cases where a cache entry is not persisted with an index, and so we need to handle the "transient" use case as well. To keep the congnitive overhead of managing the cache entries, there will only be a single discard function. This means there must be enough information kept with the cache entry so that we know how to discard them. A summary of the main functions in the API is: make_cache_entry: create cache entry for use in an index. Uses specified parameters to populate cache_entry fields. make_empty_cache_entry: Create an empty cache entry for use in an index. Returns cache entry with empty fields. make_transient_cache_entry: create cache entry that is not used in an index. Uses specified parameters to populate cache_entry fields. make_empty_transient_cache_entry: create cache entry that is not used in an index. Returns cache entry with empty fields. discard_cache_entry: A single function that knows how to discard a cache entry regardless of how it was allocated. Signed-off-by: Jameson Miller <jamill@microsoft.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
new_entry = make_empty_cache_entry(istate, namelen);
copy_cache_entry(new_entry, old_entry);
new_entry->ce_flags &= ~CE_HASHED;
new_entry->ce_namelen = namelen;
new_entry->index = 0;
memcpy(new_entry->name, new_name, namelen + 1);
cache_tree_invalidate_path(istate, old_entry->name);
untracked_cache_remove_from_index(istate, old_entry->name);
remove_index_entry_at(istate, nr);
/*
* Refresh the new index entry. Using 'refresh_cache_entry' ensures
* we only update stat info if the entry is otherwise up-to-date (i.e.,
* the contents/mode haven't changed). This ensures that we reflect the
* 'ctime' of the rename in the index without (incorrectly) updating
* the cached stat info to reflect unstaged changes on disk.
*/
refreshed = refresh_cache_entry(istate, new_entry, CE_MATCH_REFRESH);
if (refreshed && refreshed != new_entry) {
add_index_entry(istate, refreshed, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE);
discard_cache_entry(new_entry);
} else
add_index_entry(istate, new_entry, ADD_CACHE_OK_TO_ADD|ADD_CACHE_OK_TO_REPLACE);
}
void fill_stat_data(struct stat_data *sd, struct stat *st)
{
sd->sd_ctime.sec = (unsigned int)st->st_ctime;
sd->sd_mtime.sec = (unsigned int)st->st_mtime;
sd->sd_ctime.nsec = ST_CTIME_NSEC(*st);
sd->sd_mtime.nsec = ST_MTIME_NSEC(*st);
sd->sd_dev = st->st_dev;
sd->sd_ino = st->st_ino;
sd->sd_uid = st->st_uid;
sd->sd_gid = st->st_gid;
sd->sd_size = st->st_size;
}
int match_stat_data(const struct stat_data *sd, struct stat *st)
{
int changed = 0;
if (sd->sd_mtime.sec != (unsigned int)st->st_mtime)
changed |= MTIME_CHANGED;
if (trust_ctime && check_stat &&
sd->sd_ctime.sec != (unsigned int)st->st_ctime)
changed |= CTIME_CHANGED;
#ifdef USE_NSEC
if (check_stat && sd->sd_mtime.nsec != ST_MTIME_NSEC(*st))
changed |= MTIME_CHANGED;
if (trust_ctime && check_stat &&
sd->sd_ctime.nsec != ST_CTIME_NSEC(*st))
changed |= CTIME_CHANGED;
#endif
if (check_stat) {
if (sd->sd_uid != (unsigned int) st->st_uid ||
sd->sd_gid != (unsigned int) st->st_gid)
changed |= OWNER_CHANGED;
if (sd->sd_ino != (unsigned int) st->st_ino)
changed |= INODE_CHANGED;
}
#ifdef USE_STDEV
/*
* st_dev breaks on network filesystems where different
* clients will have different views of what "device"
* the filesystem is on
*/
if (check_stat && sd->sd_dev != (unsigned int) st->st_dev)
changed |= INODE_CHANGED;
#endif
if (sd->sd_size != (unsigned int) st->st_size)
changed |= DATA_CHANGED;
return changed;
}
/*
* This only updates the "non-critical" parts of the directory
* cache, ie the parts that aren't tracked by GIT, and only used
* to validate the cache.
*/
void fill_stat_cache_info(struct index_state *istate, struct cache_entry *ce, struct stat *st)
{
fill_stat_data(&ce->ce_stat_data, st);
"Assume unchanged" git This adds "assume unchanged" logic, started by this message in the list discussion recently: <Pine.LNX.4.64.0601311807470.7301@g5.osdl.org> This is a workaround for filesystems that do not have lstat() that is quick enough for the index mechanism to take advantage of. On the paths marked as "assumed to be unchanged", the user needs to explicitly use update-index to register the object name to be in the next commit. You can use two new options to update-index to set and reset the CE_VALID bit: git-update-index --assume-unchanged path... git-update-index --no-assume-unchanged path... These forms manipulate only the CE_VALID bit; it does not change the object name recorded in the index file. Nor they add a new entry to the index. When the configuration variable "core.ignorestat = true" is set, the index entries are marked with CE_VALID bit automatically after: - update-index to explicitly register the current object name to the index file. - when update-index --refresh finds the path to be up-to-date. - when tools like read-tree -u and apply --index update the working tree file and register the current object name to the index file. The flag is dropped upon read-tree that does not check out the index entry. This happens regardless of the core.ignorestat settings. Index entries marked with CE_VALID bit are assumed to be unchanged most of the time. However, there are cases that CE_VALID bit is ignored for the sake of safety and usability: - while "git-read-tree -m" or git-apply need to make sure that the paths involved in the merge do not have local modifications. This sacrifices performance for safety. - when git-checkout-index -f -q -u -a tries to see if it needs to checkout the paths. Otherwise you can never check anything out ;-). - when git-update-index --really-refresh (a new flag) tries to see if the index entry is up to date. You can start with everything marked as CE_VALID and run this once to drop CE_VALID bit for paths that are modified. Most notably, "update-index --refresh" honours CE_VALID and does not actively stat, so after you modified a file in the working tree, update-index --refresh would not notice until you tell the index about it with "git-update-index path" or "git-update-index --no-assume-unchanged path". This version is not expected to be perfect. I think diff between index and/or tree and working files may need some adjustment, and there probably needs other cases we should automatically unmark paths that are marked to be CE_VALID. But the basics seem to work, and ready to be tested by people who asked for this feature. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
if (assume_unchanged)
ce->ce_flags |= CE_VALID;
if (S_ISREG(st->st_mode)) {
ce_mark_uptodate(ce);
mark_fsmonitor_valid(): mark the index as changed if needed Without this bug fix, t7519's four "status doesn't detect unreported modifications" test cases would fail occasionally (and, oddly enough, *a lot* more frequently on Windows). The reason is that these test cases intentionally use the side effect of `git status` to re-write the index if any updates were detected: they first clean the worktree, run `git status` to update the index as well as show the output to the casual reader, then make the worktree dirty again and expect no changes to reported if running with a mocked fsmonitor hook. The problem with this strategy was that the index was written during said `git status` on the clean worktree for the *wrong* reason: not because the index was marked as changed (it wasn't), but because the recorded mtimes were racy with the index' own mtime. As the mtime granularity on Windows is 100 nanoseconds (see e.g. https://docs.microsoft.com/en-us/windows/desktop/SysInfo/file-times), the mtimes of the files are often enough *not* racy with the index', so that that `git status` call currently does not always update the index (including the fsmonitor extension), causing the test case to fail. The obvious fix: if we change *any* index entry's `CE_FSMONITOR_VALID` flag, we should also mark the index as changed. That will cause the index to be written upon `git status`, *including* an updated fsmonitor extension. Side note: Even though the reader might think that the t7519 issue should be *much* more prevalent on Linux, given that the ext4 filesystem (that seems to be used by every Linux distribution) stores mtimes in nanosecond precision. However, ext4 uses `current_kernel_time()` (see https://unix.stackexchange.com/questions/11599#comment762968_11599; it is *amazingly* hard to find any proper source of information about such ext4 questions) whose accuracy seems to depend on many factors but is safely worse than the 100-nanosecond granularity of NTFS (again, it is *horribly* hard to find anything remotely authoritative about this question). So it seems that the racy index condition that hid the bug fixed by this patch simply is a lot more likely on Linux than on Windows. But not impossible ;-) Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> Signed-off-by: Junio C Hamano <gitster@pobox.com>
3 years ago
mark_fsmonitor_valid(istate, ce);
}
}
static int ce_compare_data(struct index_state *istate,
const struct cache_entry *ce,
struct stat *st)
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
{
int match = -1;
int fd = git_open_cloexec(ce->name, O_RDONLY);
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
if (fd >= 0) {
struct object_id oid;
if (!index_fd(istate, &oid, fd, st, OBJ_BLOB, ce->name, 0))
match = !oideq(&oid, &ce->oid);
/* index_fd() closed the file descriptor already */
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
}
return match;
}
static int ce_compare_link(const struct cache_entry *ce, size_t expected_size)
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
{
int match = -1;
void *buffer;
unsigned long size;
enum object_type type;
struct strbuf sb = STRBUF_INIT;
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
if (strbuf_readlink(&sb, ce->name, expected_size))
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
return -1;
buffer = read_object_file(&ce->oid, &type, &size);
if (buffer) {
if (size == sb.len)
match = memcmp(buffer, sb.buf, size);
free(buffer);
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
}
strbuf_release(&sb);
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
return match;
}
static int ce_compare_gitlink(const struct cache_entry *ce)
{
struct object_id oid;
/*
* We don't actually require that the .git directory
* under GITLINK directory be a valid git directory. It
* might even be missing (in case nobody populated that
* sub-project).
*
* If so, we consider it always to match.
*/
if (resolve_gitlink_ref(ce->name, "HEAD", &oid) < 0)
return 0;
return !oideq(&oid, &ce->oid);
}
static int ce_modified_check_fs(struct index_state *istate,
const struct cache_entry *ce,
struct stat *st)
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
{
switch (st->st_mode & S_IFMT) {
case S_IFREG:
if (ce_compare_data(istate, ce, st))
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
return DATA_CHANGED;
break;
case S_IFLNK:
if (ce_compare_link(ce, xsize_t(st->st_size)))
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
return DATA_CHANGED;
break;
case S_IFDIR:
if (S_ISGITLINK(ce->ce_mode))
return ce_compare_gitlink(ce) ? DATA_CHANGED : 0;
consistently use "fallthrough" comments in switches Gcc 7 adds -Wimplicit-fallthrough, which can warn when a switch case falls through to the next case. The general idea is that the compiler can't tell if this was intentional or not, so you should annotate any intentional fall-throughs as such, leaving it to complain about any unannotated ones. There's a GNU __attribute__ which can be used for annotation, but of course we'd have to #ifdef it away on non-gcc compilers. Gcc will also recognize specially-formatted comments, which matches our current practice. Let's extend that practice to all of the unannotated sites (which I did look over and verify that they were behaving as intended). Ideally in each case we'd actually give some reasons in the comment about why we're falling through, or what we're falling through to. And gcc does support that with -Wimplicit-fallthrough=2, which relaxes the comment pattern matching to anything that contains "fallthrough" (or a variety of spelling variants). However, this isn't the default for -Wimplicit-fallthrough, nor for -Wextra. In the name of simplicity, it's probably better for us to support the default level, which requires "fallthrough" to be the only thing in the comment (modulo some window dressing like "else" and some punctuation; see the gcc manual for the complete set of patterns). This patch suppresses all warnings due to -Wimplicit-fallthrough. We might eventually want to add that to the DEVELOPER Makefile knob, but we should probably wait until gcc 7 is more widely adopted (since earlier versions will complain about the unknown warning type). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
5 years ago
/* else fallthrough */
Racy GIT This fixes the longstanding "Racy GIT" problem, which was pretty much there from the beginning of time, but was first demonstrated by Pasky in this message on October 24, 2005: http://marc.theaimsgroup.com/?l=git&m=113014629716878 If you run the following sequence of commands: echo frotz >infocom git update-index --add infocom echo xyzzy >infocom so that the second update to file "infocom" does not change st_mtime, what is recorded as the stat information for the cache entry "infocom" exactly matches what is on the filesystem (owner, group, inum, mtime, ctime, mode, length). After this sequence, we incorrectly think "infocom" file still has string "frotz" in it, and get really confused. E.g. git-diff-files would say there is no change, git-update-index --refresh would not even look at the filesystem to correct the situation. Some ways of working around this issue were already suggested by Linus in the same thread on the same day, including waiting until the next second before returning from update-index if a cache entry written out has the current timestamp, but that means we can make at most one commit per second, and given that the e-mail patch workflow used by Linus needs to process at least 5 commits per second, it is not an acceptable solution. Linus notes that git-apply is primarily used to update the index while processing e-mailed patches, which is true, and git-apply's up-to-date check is fooled by the same problem but luckily in the other direction, so it is not really a big issue, but still it is disturbing. The function ce_match_stat() is called to bypass the comparison against filesystem data when the stat data recorded in the cache entry matches what stat() returns from the filesystem. This patch tackles the problem by changing it to actually go to the filesystem data for cache entries that have the same mtime as the index file itself. This works as long as the index file and working tree files are on the filesystems that share the same monotonic clock. Files on network mounted filesystems sometimes get skewed timestamps compared to "date" output, but as long as working tree files' timestamps are skewed the same way as the index file's, this approach still works. The only problematic files are the ones that have the same timestamp as the index file's, because two file updates that sandwitch the index file update must happen within the same second to trigger the problem. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
default:
return TYPE_CHANGED;
}
return 0;
}
static int ce_match_stat_basic(const struct cache_entry *ce, struct stat *st)
{
unsigned int changed = 0;
if (ce->ce_flags & CE_REMOVE)
return MODE_CHANGED | DATA_CHANGED | TYPE_CHANGED;
switch (ce->ce_mode & S_IFMT) {
case S_IFREG:
changed |= !S_ISREG(st->st_mode) ? TYPE_CHANGED : 0;
/* We consider only the owner x bit to be relevant for
* "mode changes"
*/
if (trust_executable_bit &&
(0100 & (ce->ce_mode ^ st->st_mode)))
changed |= MODE_CHANGED;
break;
case S_IFLNK:
if (!S_ISLNK(st->st_mode) &&
(has_symlinks || !S_ISREG(st->st_mode)))
changed |= TYPE_CHANGED;
break;
case S_IFGITLINK:
/* We ignore most of the st_xxx fields for gitlinks */
if (!S_ISDIR(st->st_mode))
changed |= TYPE_CHANGED;
else if (ce_compare_gitlink(ce))
changed |= DATA_CHANGED;
return changed;
default:
BUG("unsupported ce_mode: %o", ce->ce_mode);
}
changed |= match_stat_data(&ce->ce_stat_data, st);
/* Racily smudged entry? */
if (!ce->ce_stat_data.sd_size) {
if (!is_empty_blob_sha1(ce->oid.hash))
changed |= DATA_CHANGED;
}
return changed;
}
static int is_racy_stat(const struct index_state *istate,
const struct stat_data *sd)
{
return (istate->timestamp.sec &&
#ifdef USE_NSEC
/* nanosecond timestamped files can also be racy! */
(istate->timestamp.sec < sd->sd_mtime.sec ||
(istate->timestamp.sec == sd->sd_mtime.sec &&
istate->timestamp.nsec <= sd->sd_mtime.nsec))
#else
istate->timestamp.sec <= sd->sd_mtime.sec
#endif
);
}
split-index: smudge and add racily clean cache entries to split index Ever since the split index feature was introduced [1], refreshing a split index is prone to a variant of the classic racy git problem. Consider the following sequence of commands updating the split index when the shared index contains a racily clean cache entry, i.e. an entry whose cached stat data matches with the corresponding file in the worktree and the cached mtime matches that of the index: echo "cached content" >file git update-index --split-index --add file echo "dirty worktree" >file # size stays the same! # ... wait ... git update-index --add other-file Normally, when a non-split index is updated, then do_write_index() (the function responsible for writing all kinds of indexes, "regular", split, and shared) recognizes racily clean cache entries, and writes them with smudged stat data, i.e. with file size set to 0. When subsequent git commands read the index, they will notice that the smudged stat data doesn't match with the file in the worktree, and then go on to check the file's content and notice its dirtiness. In the above example, however, in the second 'git update-index' prepare_to_write_split_index() decides which cache entries stored only in the shared index should be replaced in the new split index. Alas, this function never looks out for racily clean cache entries, and since the file's stat data in the worktree hasn't changed since the shared index was written, it won't be replaced in the new split index. Consequently, do_write_index() doesn't even get this racily clean cache entry, and can't smudge its stat data. Subsequent git commands will then see that the index has more recent mtime than the file and that the (not smudged) cached stat data still matches with the file in the worktree, and, ultimately, will erroneously consider the file clean. Modify prepare_to_write_split_index() to recognize racily clean cache entries, and mark them to be added to the split index. Note that there are two places where it should check raciness: first those cache entries that are only stored in the shared index, and then those that have been copied by unpack_trees() from the shared index while it constructed a new index. This way do_write_index() will get these racily clean cache entries as well, and will then write them with smudged stat data to the new split index. This change makes all tests in 't1701-racy-split-index.sh' pass, so flip the two 'test_expect_failure' tests to success. Also add the '#' (as in nr. of trial) to those tests' description that were omitted when the tests expected failure. Note that after this change if the index is split when it contains a racily clean cache entry, then a smudged cache entry will be written both to the new shared and to the new split indexes. This doesn't affect regular git commands: as far as they are concerned this is just an entry in the split index replacing an outdated entry in the shared index. It did affect a few tests in 't1700-split-index.sh', though, because they actually check which entries are stored in the split index; a previous patch in this series has already made the necessary adjustments in 't1700'. And racily clean cache entries and index splitting are rare enough to not worry about the resulting duplicated smudged cache entries, and the additional complexity required to prevent them is not worth it. Several tests failed occasionally when the test suite was run with 'GIT_TEST_SPLIT_INDEX=yes'. Here are those that I managed to trace back to this racy split index problem, starting with those failing more frequently, with a link to a failing Travis CI build job for each. The highlighted line [2] shows when the racy file was written, which is not always in the failing test but in a preceeding setup test. t3903-stash.sh: https://travis-ci.org/git/git/jobs/385542084#L5858 t4024-diff-optimize-common.sh: https://travis-ci.org/git/git/jobs/386531969#L3174 t4015-diff-whitespace.sh: https://travis-ci.org/git/git/jobs/360797600#L8215 t2200-add-update.sh: https://travis-ci.org/git/git/jobs/382543426#L3051 t0090-cache-tree.sh: https://travis-ci.org/git/git/jobs/416583010#L3679 There might be others, e.g. perhaps 't1000-read-tree-m-3way.sh' and others using 'lib-read-tree-m-3way.sh', but I couldn't confirm yet. [1] In the branch leading to the merge commit v2.1.0-rc0~45 (Merge branch 'nd/split-index', 2014-07-16). [2] Note that those highlighted lines are in the 'after failure' fold, and your browser might unhelpfully fold it up before you could take a good look. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
4 years ago
int is_racy_timestamp(const struct index_state *istate,
const struct cache_entry *ce)
{
return (!S_ISGITLINK(ce->ce_mode) &&
is_racy_stat(istate, &ce->ce_stat_data));
}
int match_stat_data_racy(const struct index_state *istate,
const struct stat_data *sd, struct stat *st)
{
if (is_racy_stat(istate, sd))
return MTIME_CHANGED;
return match_stat_data(sd, st);
}
int ie_match_stat(struct index_state *istate,
const struct cache_entry *ce, struct stat *st,
unsigned int options)
{
"Assume unchanged" git This adds "assume unchanged" logic, started by this message in the list discussion recently: <Pine.LNX.4.64.0601311807470.7301@g5.osdl.org> This is a workaround for filesystems that do not have lstat() that is quick enough for the index mechanism to take advantage of. On the paths marked as "assumed to be unchanged", the user needs to explicitly use update-index to register the object name to be in the next commit. You can use two new options to update-index to set and reset the CE_VALID bit: git-update-index --assume-unchanged path... git-update-index --no-assume-unchanged path... These forms manipulate only the CE_VALID bit; it does not change the object name recorded in the index file. Nor they add a new entry to the index. When the configuration variable "core.ignorestat = true" is set, the index entries are marked with CE_VALID bit automatically after: - update-index to explicitly register the current object name to the index file. - when update-index --refresh finds the path to be up-to-date. - when tools like read-tree -u and apply --index update the working tree file and register the current object name to the index file. The flag is dropped upon read-tree that does not check out the index entry. This happens regardless of the core.ignorestat settings. Index entries marked with CE_VALID bit are assumed to be unchanged most of the time. However, there are cases that CE_VALID bit is ignored for the sake of safety and usability: - while "git-read-tree -m" or git-apply need to make sure that the paths involved in the merge do not have local modifications. This sacrifices performance for safety. - when git-checkout-index -f -q -u -a tries to see if it needs to checkout the paths. Otherwise you can never check anything out ;-). - when git-update-index --really-refresh (a new flag) tries to see if the index entry is up to date. You can start with everything marked as CE_VALID and run this once to drop CE_VALID bit for paths that are modified. Most notably, "update-index --refresh" honours CE_VALID and does not actively stat, so after you modified a file in the working tree, update-index --refresh would not notice until you tell the index about it with "git-update-index path" or "git-update-index --no-assume-unchanged path". This version is not expected to be perfect. I think diff between index and/or tree and working files may need some adjustment, and there probably needs other cases we should automatically unmark paths that are marked to be CE_VALID. But the basics seem to work, and ready to be tested by people who asked for this feature. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
unsigned int changed;
int ignore_valid = options & CE_MATCH_IGNORE_VALID;
int ignore_skip_worktree = options & CE_MATCH_IGNORE_SKIP_WORKTREE;
int assume_racy_is_modified = options & CE_MATCH_RACY_IS_DIRTY;
int ignore_fsmonitor = options & CE_MATCH_IGNORE_FSMONITOR;
"Assume unchanged" git This adds "assume unchanged" logic, started by this message in the list discussion recently: <Pine.LNX.4.64.0601311807470.7301@g5.osdl.org> This is a workaround for filesystems that do not have lstat() that is quick enough for the index mechanism to take advantage of. On the paths marked as "assumed to be unchanged", the user needs to explicitly use update-index to register the object name to be in the next commit. You can use two new options to update-index to set and reset the CE_VALID bit: git-update-index --assume-unchanged path... git-update-index --no-assume-unchanged path... These forms manipulate only the CE_VALID bit; it does not change the object name recorded in the index file. Nor they add a new entry to the index. When the configuration variable "core.ignorestat = true" is set, the index entries are marked with CE_VALID bit automatically after: - update-index to explicitly register the current object name to the index file. - when update-index --refresh finds the path to be up-to-date. - when tools like read-tree -u and apply --index update the working tree file and register the current object name to the index file. The flag is dropped upon read-tree that does not check out the index entry. This happens regardless of the core.ignorestat settings. Index entries marked with CE_VALID bit are assumed to be unchanged most of the time. However, there are cases that CE_VALID bit is ignored for the sake of safety and usability: - while "git-read-tree -m" or git-apply need to make sure that the paths involved in the merge do not have local modifications. This sacrifices performance for safety. - when git-checkout-index -f -q -u -a tries to see if it needs to checkout the paths. Otherwise you can never check anything out ;-). - when git-update-index --really-refresh (a new flag) tries to see if the index entry is up to date. You can start with everything marked as CE_VALID and run this once to drop CE_VALID bit for paths that are modified. Most notably, "update-index --refresh" honours CE_VALID and does not actively stat, so after you modified a file in the working tree, update-index --refresh would not notice until you tell the index about it with "git-update-index path" or "git-update-index --no-assume-unchanged path". This version is not expected to be perfect. I think diff between index and/or tree and working files may need some adjustment, and there probably needs other cases we should automatically unmark paths that are marked to be CE_VALID. But the basics seem to work, and ready to be tested by people who asked for this feature. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
if (!ignore_fsmonitor)
refresh_fsmonitor(istate);
"Assume unchanged" git This adds "assume unchanged" logic, started by this message in the list discussion recently: <Pine.LNX.4.64.0601311807470.7301@g5.osdl.org> This is a workaround for filesystems that do not have lstat() that is quick enough for the index mechanism to take advantage of. On the paths marked as "assumed to be unchanged", the user needs to explicitly use update-index to register the object name to be in the next commit. You can use two new options to update-index to set and reset the CE_VALID bit: git-update-index --assume-unchanged path... git-update-index --no-assume-unchanged path... These forms manipulate only the CE_VALID bit; it does not change the object name recorded in the index file. Nor they add a new entry to the index. When the configuration variable "core.ignorestat = true" is set, the index entries are marked with CE_VALID bit automatically after: - update-index to explicitly register the current object name to the index file. - when update-index --refresh finds the path to be up-to-date. - when tools like read-tree -u and apply --index update the working tree file and register the current object name to the index file. The flag is dropped upon read-tree that does not check out the index entry. This happens regardless of the core.ignorestat settings. Index entries marked with CE_VALID bit are assumed to be unchanged most of the time. However, there are cases that CE_VALID bit is ignored for the sake of safety and usability: - while "git-read-tree -m" or git-apply need to make sure that the paths involved in the merge do not have local modifications. This sacrifices performance for safety. - when git-checkout-index -f -q -u -a tries to see if it needs to checkout the paths. Otherwise you can never check anything out ;-). - when git-update-index --really-refresh (a new flag) tries to see if the index entry is up to date. You can start with everything marked as CE_VALID and run this once to drop CE_VALID bit for paths that are modified. Most notably, "update-index --refresh" honours CE_VALID and does not actively stat, so after you modified a file in the working tree, update-index --refresh would not notice until you tell the index about it with "git-update-index path" or "git-update-index --no-assume-unchanged path". This version is not expected to be perfect. I think diff between index and/or tree and working files may need some adjustment, and there probably needs other cases we should automatically unmark paths that are marked to be CE_VALID. But the basics seem to work, and ready to be tested by people who asked for this feature. Signed-off-by: Junio C Hamano <junkio@cox.net>
17 years ago
/*
* If it's marked as always valid in the index, it's
* valid whatever the checked-out copy says.