git/object.c

641 lines
15 KiB
C
Raw Normal View History

#include "git-compat-util.h"
#include "gettext.h"
#include "hex.h"
#include "object.h"
#include "replace-object.h"
#include "object-file.h"
#include "object-store.h"
#include "blob.h"
#include "statinfo.h"
#include "tree.h"
#include "commit.h"
#include "tag.h"
#include "alloc.h"
#include "packfile.h"
#include "commit-graph.h"
#include "loose.h"
unsigned int get_max_object_index(void)
{
return the_repository->parsed_objects->obj_hash_size;
}
struct object *get_indexed_object(unsigned int idx)
{
return the_repository->parsed_objects->obj_hash[idx];
}
static const char *object_type_strings[] = {
NULL, /* OBJ_NONE = 0 */
"commit", /* OBJ_COMMIT = 1 */
"tree", /* OBJ_TREE = 2 */
"blob", /* OBJ_BLOB = 3 */
"tag", /* OBJ_TAG = 4 */
};
const char *type_name(unsigned int type)
{
if (type >= ARRAY_SIZE(object_type_strings))
return NULL;
return object_type_strings[type];
}
int type_from_string_gently(const char *str, ssize_t len, int gentle)
{
int i;
if (len < 0)
len = strlen(str);
for (i = 1; i < ARRAY_SIZE(object_type_strings); i++)
if (!xstrncmpz(object_type_strings[i], str, len))
return i;
if (gentle)
return -1;
die(_("invalid object type \"%s\""), str);
}
/*
* Return a numerical hash value between 0 and n-1 for the object with
* the specified sha1. n must be a power of 2. Please note that the
* return value is *not* consistent across computer architectures.
*/
static unsigned int hash_obj(const struct object_id *oid, unsigned int n)
{
return oidhash(oid) & (n - 1);
}
/*
* Insert obj into the hash table hash, which has length size (which
* must be a power of 2). On collisions, simply overflow to the next
* empty bucket.
*/
static void insert_obj_hash(struct object *obj, struct object **hash, unsigned int size)
{
unsigned int j = hash_obj(&obj->oid, size);
while (hash[j]) {
j++;
if (j >= size)
j = 0;
}
hash[j] = obj;
}
/*
* Look up the record for the given sha1 in the hash map stored in
* obj_hash. Return NULL if it was not found.
*/
struct object *lookup_object(struct repository *r, const struct object_id *oid)
{
lookup_object: prioritize recently found objects The lookup_object function is backed by a hash table of all objects we have seen in the program. We manage collisions with a linear walk over the colliding entries, checking each with hashcmp(). The main cost of lookup is in these hashcmp() calls; finding our item in the first slot is cheaper than finding it in the second slot, which is cheaper than the third, and so on. If we assume that there is some locality to the object lookups (e.g., if X and Y collide, and we have just looked up X, the next lookup is more likely to be for X than for Y), then we can improve our average lookup speed by checking X before Y. This patch does so by swapping a found item to the front of the collision chain. The p0001 perf test reveals that this does indeed exploit locality in the case of "rev-list --all --objects": Test origin this tree ------------------------------------------------------------------------- 0001.1: rev-list --all 0.40(0.38+0.02) 0.40(0.36+0.03) +0.0% 0001.2: rev-list --all --objects 2.24(2.17+0.05) 1.86(1.79+0.05) -17.0% This is not surprising, as the full object traversal will hit the same tree entries over and over (e.g., for every commit that doesn't change "Documentation/", we will have to look up the same sha1 just to find out that we already processed it). The reason why this technique works (and does not violate any properties of the hash table) is subtle and bears some explanation. Let's imagine we get a lookup for sha1 `X`, and it hashes to bucket `i` in our table. That stretch of the table may look like: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | B | C | X | ... ----------------------------------- We start our probe at i, see that B does not match, nor does C, and finally find X. There may be multiple C's in the middle, but we know that there are no empty slots (or else we would not find X at all). We do not know the original index of B; it may be `i`, or it may be less than i (e.g., if it were `i-1`, it would collide with A and spill over into the `i` bucket). So it is acceptable for us to move it to the right of a contiguous stretch of entries (because we will find it from a linear walk starting anywhere at `i` or before), but never to the left (if we moved it to `i-1`, we would miss it when starting our walk at `i`). We do know the original index of X; it is `i`, so it is safe to place it anywhere in the contiguous stretch between `i` and where we found it (`i+2` in the this case). This patch does a pure swap; after finding X in the situation above, we would end with: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | C | B | ... ----------------------------------- We could instead bump X into the `i` slot, and then shift the whole contiguous chain down by one, resulting in: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | B | C | ... ----------------------------------- That puts our chain in true most-recently-used order. However, experiments show that it is not any faster (and in fact, is slightly slower due to the extra manipulation). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-01 20:34:50 +00:00
unsigned int i, first;
struct object *obj;
if (!r->parsed_objects->obj_hash)
return NULL;
first = i = hash_obj(oid, r->parsed_objects->obj_hash_size);
while ((obj = r->parsed_objects->obj_hash[i]) != NULL) {
if (oideq(oid, &obj->oid))
break;
i++;
if (i == r->parsed_objects->obj_hash_size)
i = 0;
}
lookup_object: prioritize recently found objects The lookup_object function is backed by a hash table of all objects we have seen in the program. We manage collisions with a linear walk over the colliding entries, checking each with hashcmp(). The main cost of lookup is in these hashcmp() calls; finding our item in the first slot is cheaper than finding it in the second slot, which is cheaper than the third, and so on. If we assume that there is some locality to the object lookups (e.g., if X and Y collide, and we have just looked up X, the next lookup is more likely to be for X than for Y), then we can improve our average lookup speed by checking X before Y. This patch does so by swapping a found item to the front of the collision chain. The p0001 perf test reveals that this does indeed exploit locality in the case of "rev-list --all --objects": Test origin this tree ------------------------------------------------------------------------- 0001.1: rev-list --all 0.40(0.38+0.02) 0.40(0.36+0.03) +0.0% 0001.2: rev-list --all --objects 2.24(2.17+0.05) 1.86(1.79+0.05) -17.0% This is not surprising, as the full object traversal will hit the same tree entries over and over (e.g., for every commit that doesn't change "Documentation/", we will have to look up the same sha1 just to find out that we already processed it). The reason why this technique works (and does not violate any properties of the hash table) is subtle and bears some explanation. Let's imagine we get a lookup for sha1 `X`, and it hashes to bucket `i` in our table. That stretch of the table may look like: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | B | C | X | ... ----------------------------------- We start our probe at i, see that B does not match, nor does C, and finally find X. There may be multiple C's in the middle, but we know that there are no empty slots (or else we would not find X at all). We do not know the original index of B; it may be `i`, or it may be less than i (e.g., if it were `i-1`, it would collide with A and spill over into the `i` bucket). So it is acceptable for us to move it to the right of a contiguous stretch of entries (because we will find it from a linear walk starting anywhere at `i` or before), but never to the left (if we moved it to `i-1`, we would miss it when starting our walk at `i`). We do know the original index of X; it is `i`, so it is safe to place it anywhere in the contiguous stretch between `i` and where we found it (`i+2` in the this case). This patch does a pure swap; after finding X in the situation above, we would end with: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | C | B | ... ----------------------------------- We could instead bump X into the `i` slot, and then shift the whole contiguous chain down by one, resulting in: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | B | C | ... ----------------------------------- That puts our chain in true most-recently-used order. However, experiments show that it is not any faster (and in fact, is slightly slower due to the extra manipulation). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-01 20:34:50 +00:00
if (obj && i != first) {
/*
* Move object to where we started to look for it so
* that we do not need to walk the hash table the next
* time we look for it.
*/
SWAP(r->parsed_objects->obj_hash[i],
r->parsed_objects->obj_hash[first]);
lookup_object: prioritize recently found objects The lookup_object function is backed by a hash table of all objects we have seen in the program. We manage collisions with a linear walk over the colliding entries, checking each with hashcmp(). The main cost of lookup is in these hashcmp() calls; finding our item in the first slot is cheaper than finding it in the second slot, which is cheaper than the third, and so on. If we assume that there is some locality to the object lookups (e.g., if X and Y collide, and we have just looked up X, the next lookup is more likely to be for X than for Y), then we can improve our average lookup speed by checking X before Y. This patch does so by swapping a found item to the front of the collision chain. The p0001 perf test reveals that this does indeed exploit locality in the case of "rev-list --all --objects": Test origin this tree ------------------------------------------------------------------------- 0001.1: rev-list --all 0.40(0.38+0.02) 0.40(0.36+0.03) +0.0% 0001.2: rev-list --all --objects 2.24(2.17+0.05) 1.86(1.79+0.05) -17.0% This is not surprising, as the full object traversal will hit the same tree entries over and over (e.g., for every commit that doesn't change "Documentation/", we will have to look up the same sha1 just to find out that we already processed it). The reason why this technique works (and does not violate any properties of the hash table) is subtle and bears some explanation. Let's imagine we get a lookup for sha1 `X`, and it hashes to bucket `i` in our table. That stretch of the table may look like: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | B | C | X | ... ----------------------------------- We start our probe at i, see that B does not match, nor does C, and finally find X. There may be multiple C's in the middle, but we know that there are no empty slots (or else we would not find X at all). We do not know the original index of B; it may be `i`, or it may be less than i (e.g., if it were `i-1`, it would collide with A and spill over into the `i` bucket). So it is acceptable for us to move it to the right of a contiguous stretch of entries (because we will find it from a linear walk starting anywhere at `i` or before), but never to the left (if we moved it to `i-1`, we would miss it when starting our walk at `i`). We do know the original index of X; it is `i`, so it is safe to place it anywhere in the contiguous stretch between `i` and where we found it (`i+2` in the this case). This patch does a pure swap; after finding X in the situation above, we would end with: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | C | B | ... ----------------------------------- We could instead bump X into the `i` slot, and then shift the whole contiguous chain down by one, resulting in: index | i-1 | i | i+1 | i+2 | ----------------------------------- entry ... | A | X | B | C | ... ----------------------------------- That puts our chain in true most-recently-used order. However, experiments show that it is not any faster (and in fact, is slightly slower due to the extra manipulation). Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-01 20:34:50 +00:00
}
return obj;
}
/*
* Increase the size of the hash map stored in obj_hash to the next
* power of 2 (but at least 32). Copy the existing values to the new
* hash map.
*/
static void grow_object_hash(struct repository *r)
{
int i;
/*
* Note that this size must always be power-of-2 to match hash_obj
* above.
*/
int new_hash_size = r->parsed_objects->obj_hash_size < 32 ? 32 : 2 * r->parsed_objects->obj_hash_size;
struct object **new_hash;
CALLOC_ARRAY(new_hash, new_hash_size);
for (i = 0; i < r->parsed_objects->obj_hash_size; i++) {
struct object *obj = r->parsed_objects->obj_hash[i];
if (!obj)
continue;
insert_obj_hash(obj, new_hash, new_hash_size);
}
free(r->parsed_objects->obj_hash);
r->parsed_objects->obj_hash = new_hash;
r->parsed_objects->obj_hash_size = new_hash_size;
}
void *create_object(struct repository *r, const struct object_id *oid, void *o)
{
struct object *obj = o;
obj->parsed = 0;
obj->flags = 0;
oidcpy(&obj->oid, oid);
if (r->parsed_objects->obj_hash_size - 1 <= r->parsed_objects->nr_objs * 2)
grow_object_hash(r);
insert_obj_hash(obj, r->parsed_objects->obj_hash,
r->parsed_objects->obj_hash_size);
r->parsed_objects->nr_objs++;
return obj;
}
void *object_as_type(struct object *obj, enum object_type type, int quiet)
add object_as_type helper for casting objects When we call lookup_commit, lookup_tree, etc, the logic goes something like: 1. Look for an existing object struct. If we don't have one, allocate and return a new one. 2. Double check that any object we have is the expected type (and complain and return NULL otherwise). 3. Convert an object with type OBJ_NONE (from a prior call to lookup_unknown_object) to the expected type. We can encapsulate steps 2 and 3 in a helper function which checks whether we have the expected object type, converts OBJ_NONE as appropriate, and returns the object. Not only does this shorten the code, but it also provides one central location for converting OBJ_NONE objects into objects of other types. Future patches will use that to enforce type-specific invariants. Since this is a refactoring, we would want it to behave exactly as the current code. It takes a little reasoning to see that this is the case: - for lookup_{commit,tree,etc} functions, we are just pulling steps 2 and 3 into a function that does the same thing. - for the call in peel_object, we currently only do step 3 (but we want to consolidate it with the others, as mentioned above). However, step 2 is a noop here, as the surrounding conditional makes sure we have OBJ_NONE (which we want to keep to avoid an extraneous call to sha1_object_info). - for the call in lookup_commit_reference_gently, we are currently doing step 2 but not step 3. However, step 3 is a noop here. The object we got will have just come from deref_tag, which must have figured out the type for each object in order to know when to stop peeling. Therefore the type will never be OBJ_NONE. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2014-07-13 06:42:03 +00:00
{
if (obj->type == type)
return obj;
else if (obj->type == OBJ_NONE) {
if (type == OBJ_COMMIT)
init_commit_node((struct commit *) obj);
object_as_type: initialize commit-graph-related fields of 'struct commit' When the commit graph and generation numbers were introduced in commits 177722b344 (commit: integrate commit graph with commit parsing, 2018-04-10) and 83073cc994 (commit: add generation number to struct commit, 2018-04-25), they tried to make sure that the corresponding 'graph_pos' and 'generation' fields of 'struct commit' are initialized conservatively, as if the commit were not included in the commit-graph file. Alas, initializing those fields only in alloc_commit_node() missed the case when an object that happens to be a commit is first looked up via lookup_unknown_object(), and is then later converted to a 'struct commit' via the object_as_type() helper function (either calling it directly, or as part of a subsequent lookup_commit() call). Consequently, both of those fields incorrectly remain set to zero, which means e.g. that the commit is present in and is the first entry of the commit-graph file. This will result in wrong timestamp, parent and root tree hashes, if such a 'struct commit' instance is later filled from the commit-graph. Extract the initialization of 'struct commit's fields from alloc_commit_node() into a helper function, and call it from object_as_type() as well, to make sure that it properly initializes the two commit-graph-related fields, too. With this helper function it is hopefully less likely that any new fields added to 'struct commit' in the future would remain uninitialized. With this change alloc_commit_index() won't have any remaining callers outside of 'alloc.c', so mark it as static. Signed-off-by: SZEDER Gábor <szeder.dev@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2019-01-27 13:08:32 +00:00
else
obj->type = type;
add object_as_type helper for casting objects When we call lookup_commit, lookup_tree, etc, the logic goes something like: 1. Look for an existing object struct. If we don't have one, allocate and return a new one. 2. Double check that any object we have is the expected type (and complain and return NULL otherwise). 3. Convert an object with type OBJ_NONE (from a prior call to lookup_unknown_object) to the expected type. We can encapsulate steps 2 and 3 in a helper function which checks whether we have the expected object type, converts OBJ_NONE as appropriate, and returns the object. Not only does this shorten the code, but it also provides one central location for converting OBJ_NONE objects into objects of other types. Future patches will use that to enforce type-specific invariants. Since this is a refactoring, we would want it to behave exactly as the current code. It takes a little reasoning to see that this is the case: - for lookup_{commit,tree,etc} functions, we are just pulling steps 2 and 3 into a function that does the same thing. - for the call in peel_object, we currently only do step 3 (but we want to consolidate it with the others, as mentioned above). However, step 2 is a noop here, as the surrounding conditional makes sure we have OBJ_NONE (which we want to keep to avoid an extraneous call to sha1_object_info). - for the call in lookup_commit_reference_gently, we are currently doing step 2 but not step 3. However, step 3 is a noop here. The object we got will have just come from deref_tag, which must have figured out the type for each object in order to know when to stop peeling. Therefore the type will never be OBJ_NONE. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2014-07-13 06:42:03 +00:00
return obj;
}
else {
if (!quiet)
error(_("object %s is a %s, not a %s"),
oid_to_hex(&obj->oid),
type_name(obj->type), type_name(type));
add object_as_type helper for casting objects When we call lookup_commit, lookup_tree, etc, the logic goes something like: 1. Look for an existing object struct. If we don't have one, allocate and return a new one. 2. Double check that any object we have is the expected type (and complain and return NULL otherwise). 3. Convert an object with type OBJ_NONE (from a prior call to lookup_unknown_object) to the expected type. We can encapsulate steps 2 and 3 in a helper function which checks whether we have the expected object type, converts OBJ_NONE as appropriate, and returns the object. Not only does this shorten the code, but it also provides one central location for converting OBJ_NONE objects into objects of other types. Future patches will use that to enforce type-specific invariants. Since this is a refactoring, we would want it to behave exactly as the current code. It takes a little reasoning to see that this is the case: - for lookup_{commit,tree,etc} functions, we are just pulling steps 2 and 3 into a function that does the same thing. - for the call in peel_object, we currently only do step 3 (but we want to consolidate it with the others, as mentioned above). However, step 2 is a noop here, as the surrounding conditional makes sure we have OBJ_NONE (which we want to keep to avoid an extraneous call to sha1_object_info). - for the call in lookup_commit_reference_gently, we are currently doing step 2 but not step 3. However, step 3 is a noop here. The object we got will have just come from deref_tag, which must have figured out the type for each object in order to know when to stop peeling. Therefore the type will never be OBJ_NONE. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2014-07-13 06:42:03 +00:00
return NULL;
}
}
struct object *lookup_unknown_object(struct repository *r, const struct object_id *oid)
{
struct object *obj = lookup_object(r, oid);
if (!obj)
obj = create_object(r, oid, alloc_object_node(r));
return obj;
}
struct object *lookup_object_by_type(struct repository *r,
const struct object_id *oid,
enum object_type type)
{
switch (type) {
case OBJ_COMMIT:
return (struct object *)lookup_commit(r, oid);
case OBJ_TREE:
return (struct object *)lookup_tree(r, oid);
case OBJ_TAG:
return (struct object *)lookup_tag(r, oid);
case OBJ_BLOB:
return (struct object *)lookup_blob(r, oid);
default:
BUG("unknown object type %d", type);
}
}
struct object *parse_object_buffer(struct repository *r, const struct object_id *oid, enum object_type type, unsigned long size, void *buffer, int *eaten_p)
{
struct object *obj;
*eaten_p = 0;
obj = NULL;
if (type == OBJ_BLOB) {
struct blob *blob = lookup_blob(r, oid);
if (blob) {
blob: drop unused parts of parse_blob_buffer() Our parse_blob_buffer() takes a ptr/len combo, just like parse_tree_buffer(), etc, and returns success or failure. But it doesn't actually do anything with them; we just set the "parsed" flag in the object and return success, without even looking at the contents. There could be some value to keeping these unused parameters: - it's consistent with the parse functions for other object types. But we already lost that consistency in 837d395a5c (Replace parse_blob() with an explanatory comment, 2010-01-18). - As the comment from 837d395a5c explains, callers are supposed to make sure they have the object content available. So in theory asking for these parameters could serve as a signal. But there are only two callers, and one of them always passes NULL (after doing a streaming check of the object hash). This shows that there aren't likely to be a lot of callers (since everyone either uses the type-generic parse functions, or handles blobs individually), and that they need to take special care anyway (because we usually want to avoid loading whole blobs in memory if we can avoid it). So let's just drop these unused parameters, and likewise the useless return value. While we're touching the header file, let's move the declaration of parse_blob_buffer() right below that explanatory comment, where it's more likely to be seen by people looking for the function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-12-13 11:11:57 +00:00
parse_blob_buffer(blob);
obj = &blob->object;
}
} else if (type == OBJ_TREE) {
struct tree *tree = lookup_tree(r, oid);
if (tree) {
obj = &tree->object;
if (!tree->buffer)
tree->object.parsed = 0;
if (!tree->object.parsed) {
if (parse_tree_buffer(tree, buffer, size))
return NULL;
*eaten_p = 1;
}
}
} else if (type == OBJ_COMMIT) {
struct commit *commit = lookup_commit(r, oid);
if (commit) {
if (parse_commit_buffer(r, commit, buffer, size, 1))
return NULL;
parse_object_buffer(): respect save_commit_buffer If the global variable "save_commit_buffer" is set to 0, then parse_commit() will throw away the commit object data after parsing it, rather than sticking it into a commit slab. This goes all the way back to 60ab26de99 ([PATCH] Avoid wasting memory in git-rev-list, 2005-09-15). But there's another code path which may similarly stash the buffer: parse_object_buffer(). This is where we end up if we parse a commit via parse_object(), and it's used directly in a few other code paths like git-fsck. The original goal of 60ab26de99 was avoiding extra memory usage for rev-list. And there it's not all that important to catch parse_object(). We use that function only for looking at the tips of the traversal, and the majority of the commits are parsed by following parent links, where we use parse_commit() directly. So we were wasting some memory, but only a small portion. It's much easier to see the effect with fsck. Since we now turn off save_commit_buffer by default there, we _should_ be able to drop the freeing of the commit buffer in fsck_obj(). But if we do so (taking the first hunk of this patch without the rest), then the peak heap of "git fsck" in a clone of git.git goes from 136MB to 194MB. Teaching parse_object_buffer() to respect save_commit_buffer brings that down to 134.5MB (it's hard to tell from massif's output, but I suspect the savings comes from avoiding the overhead of the mostly-empty commit slab). Other programs should see a small improvement. Both "rev-list --all" and "fsck --connectivity-only" improve by a few hundred kilobytes, as they'd avoid loading the tip objects of their traversals. Most importantly, no code should be hurt by doing this. Any program that turns off save_commit_buffer is already making the assumption that any commit it sees may need to have its object data loaded on demand, as it doesn't know which ones were parsed by parse_commit() versus parse_object(). Not to mention that anything parsed by the commit graph may be in the same boat, even if save_commit_buffer was not disabled. This should be the only spot that needs to be fixed. Grepping for set_commit_buffer() shows that this and parse_commit() are the only relevant calls. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-22 10:15:38 +00:00
if (save_commit_buffer &&
!get_cached_commit_buffer(r, commit, NULL)) {
set_commit_buffer(r, commit, buffer, size);
*eaten_p = 1;
}
obj = &commit->object;
}
} else if (type == OBJ_TAG) {
struct tag *tag = lookup_tag(r, oid);
if (tag) {
if (parse_tag_buffer(r, tag, buffer, size))
return NULL;
obj = &tag->object;
}
} else {
warning(_("object %s has unknown type id %d"), oid_to_hex(oid), type);
obj = NULL;
}
return obj;
}
struct object *parse_object_or_die(const struct object_id *oid,
const char *name)
{
struct object *o = parse_object(the_repository, oid);
if (o)
return o;
die(_("unable to parse object: %s"), name ? name : oid_to_hex(oid));
}
parse_object(): allow skipping hash check The parse_object() function checks the object hash of any object it parses. This is a nice feature, as it means we may catch bit corruption during normal use, rather than waiting for specific fsck operations. But it also can be slow. It's particularly noticeable for blobs, where except for the hash check, we could return without loading the object contents at all. Now one may wonder what is the point of calling parse_object() on a blob in the first place then, but usually it's not intentional: we were fed an oid from somewhere, don't know the type, and want an object struct. For commits and trees, the parsing is usually helpful; we're about to look at the contents anyway. But this is less true for blobs, where we may be collecting them as part of a reachability traversal, etc, and don't actually care what's in them. And blobs, of course, tend to be larger. We don't want to just throw out the hash-checks for blobs, though. We do depend on them in some circumstances (e.g., rev-list --verify-objects uses parse_object() to check them). It's only the callers that know how they're going to use the result. And so we can help them by providing a special flag to skip the hash check. We could just apply this to blobs, as they're going to be the main source of performance improvement. But if a caller doesn't care about checking the hash, we might as well skip it for other object types, too. Even though we can't avoid reading the object contents, we can still skip the actual hash computation. If this seems like it is making Git a little bit less safe against corruption, it may be. But it's part of a series of tradeoffs we're already making. For instance, "rev-list --objects" does not open the contents of blobs it prints. And when a commit graph is present, we skip opening most commits entirely. The important thing will be to use this flag in cases where it's safe to skip the check. For instance, when serving a pack for a fetch, we know the client will fully index the objects and do a connectivity check itself. There's little to be gained from the server side re-hashing a blob itself. And indeed, most of the time we don't! The revision machinery won't open up a blob reached by traversal, but only one requested directly with a "want" line. So applied properly, this new feature shouldn't make anything less safe in practice. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-06 23:01:34 +00:00
struct object *parse_object_with_flags(struct repository *r,
const struct object_id *oid,
enum parse_object_flags flags)
{
parse_object(): allow skipping hash check The parse_object() function checks the object hash of any object it parses. This is a nice feature, as it means we may catch bit corruption during normal use, rather than waiting for specific fsck operations. But it also can be slow. It's particularly noticeable for blobs, where except for the hash check, we could return without loading the object contents at all. Now one may wonder what is the point of calling parse_object() on a blob in the first place then, but usually it's not intentional: we were fed an oid from somewhere, don't know the type, and want an object struct. For commits and trees, the parsing is usually helpful; we're about to look at the contents anyway. But this is less true for blobs, where we may be collecting them as part of a reachability traversal, etc, and don't actually care what's in them. And blobs, of course, tend to be larger. We don't want to just throw out the hash-checks for blobs, though. We do depend on them in some circumstances (e.g., rev-list --verify-objects uses parse_object() to check them). It's only the callers that know how they're going to use the result. And so we can help them by providing a special flag to skip the hash check. We could just apply this to blobs, as they're going to be the main source of performance improvement. But if a caller doesn't care about checking the hash, we might as well skip it for other object types, too. Even though we can't avoid reading the object contents, we can still skip the actual hash computation. If this seems like it is making Git a little bit less safe against corruption, it may be. But it's part of a series of tradeoffs we're already making. For instance, "rev-list --objects" does not open the contents of blobs it prints. And when a commit graph is present, we skip opening most commits entirely. The important thing will be to use this flag in cases where it's safe to skip the check. For instance, when serving a pack for a fetch, we know the client will fully index the objects and do a connectivity check itself. There's little to be gained from the server side re-hashing a blob itself. And indeed, most of the time we don't! The revision machinery won't open up a blob reached by traversal, but only one requested directly with a "want" line. So applied properly, this new feature shouldn't make anything less safe in practice. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-06 23:01:34 +00:00
int skip_hash = !!(flags & PARSE_OBJECT_SKIP_HASH_CHECK);
upload-pack: free tree buffers after parsing When a client sends us a "want" or "have" line, we call parse_object() to get an object struct. If the object is a tree, then the parsed state means that tree->buffer points to the uncompressed contents of the tree. But we don't really care about it. We only really need to parse commits and tags; for trees and blobs, the important output is just a "struct object" with the correct type. But much worse, we do not ever free that tree buffer. It's not leaked in the traditional sense, in that we still have a pointer to it from the global object hash. But if the client requests many trees, we'll hold all of their contents in memory at the same time. Nobody really noticed because it's rare for clients to directly request a tree. It might happen for a lightweight tag pointing straight at a tree, or it might happen for a "tree:depth" partial clone filling in missing trees. But it's also possible for a malicious client to request a lot of trees, causing upload-pack's memory to balloon. For example, without this patch, requesting every tree in git.git like: pktline() { local msg="$*" printf "%04x%s\n" $((1+4+${#msg})) "$msg" } want_trees() { pktline command=fetch printf 0001 git cat-file --batch-all-objects --batch-check='%(objectname) %(objecttype)' | while read oid type; do test "$type" = "tree" || continue pktline want $oid done pktline done printf 0000 } want_trees | GIT_PROTOCOL=version=2 valgrind --tool=massif ./git upload-pack . >/dev/null shows a peak heap usage of ~3.7GB. Which is just about the sum of the sizes of all of the uncompressed trees. For linux.git, it's closer to 17GB. So the obvious thing to do is to call free_tree_buffer() after we realize that we've parsed a tree. We know that upload-pack won't need it later. But let's push the logic into parse_object_with_flags(), telling it to discard the tree buffer immediately. There are two reasons for this. One, all of the relevant call-sites already call the with_options variant to pass the SKIP_HASH flag. So it actually ends up as less code than manually free-ing in each spot. And two, it enables an extra optimization that I'll discuss below. I've touched all of the sites that currently use SKIP_HASH in upload-pack. That drops the peak heap of the upload-pack invocation above from 3.7GB to ~24MB. I've also modified the caller in get_reference(); a partial clone benefits from its use in pack-objects for the reasons given in 0bc2557951 (upload-pack: skip parse-object re-hashing of "want" objects, 2022-09-06), where we were measuring blob requests. But note that the results of get_reference() are used for traversing, as well; so we really would _eventually_ use the tree contents. That makes this at first glance a space/time tradeoff: we won't hold all of the trees in memory at once, but we'll have to reload them each when it comes time to traverse. And here's where our extra optimization comes in. If the caller is not going to immediately look at the tree contents, and it doesn't care about checking the hash, then parse_object() can simply skip loading the tree entirely, just like we do for blobs! And now it's not a space/time tradeoff in get_reference() anymore. It's just a lazy-load: we're delaying reading the tree contents until it's time to actually traverse them one by one. And of course for upload-pack, this optimization means we never load the trees at all, saving lots of CPU time. Timing the "every tree from git.git" request above shows upload-pack dropping from 32 seconds of CPU to 19 (the remainder is mostly due to pack-objects actually sending the pack; timing just the upload-pack portion shows we go from 13s to ~0.28s). These are all highly gamed numbers, of course. For real-world partial-clone requests we're saving only a small bit of time in practice. But it does help harden upload-pack against malicious denial-of-service attacks. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-02-28 22:39:07 +00:00
int discard_tree = !!(flags & PARSE_OBJECT_DISCARD_TREE);
unsigned long size;
enum object_type type;
int eaten;
const struct object_id *repl = lookup_replace_object(r, oid);
parse_object: try internal cache before reading object db When parse_object is called, we do the following: 1. read the object data into a buffer via read_sha1_file 2. call parse_object_buffer, which then: a. calls the appropriate lookup_{commit,tree,blob,tag} to either create a new "struct object", or to find an existing one. We know the appropriate type from the lookup in step 1. b. calls the appropriate parse_{commit,tree,blob,tag} to parse the buffer for the new (or existing) object In step 2b, all of the called functions are no-ops for object "X" if "X->object.parsed" is set. I.e., when we have already parsed an object, we end up going to a lot of work just to find out at a low level that there is nothing left for us to do (and we throw away the data from read_sha1_file unread). We can optimize this by moving the check for "do we have an in-memory object" from 2a before the expensive call to read_sha1_file in step 1. This might seem circular, since step 2a uses the type information determined in step 1 to call the appropriate lookup function. However, we can notice that all of the lookup_* functions are backed by lookup_object. In other words, all of the objects are kept in a master hash table, and we don't actually need the type to do the "do we have it" part of the lookup, only to do the "and create it if it doesn't exist" part. This can save time whenever we call parse_object on the same sha1 twice in a single program. Some code paths already perform this optimization manually, with either: if (!obj->parsed) obj = parse_object(obj->sha1); if you already have a "struct object", or: struct object *obj = lookup_unknown_object(sha1); if (!obj || !obj->parsed) obj = parse_object(sha1); if you don't. This patch moves the optimization into parse_object itself. Most git operations won't notice any impact. Either they don't parse a lot of duplicate sha1s, or the calling code takes special care not to re-parse objects. I timed two code paths that do benefit (there may be more, but these two were immediately obvious and easy to time). The first is fast-export, which calls parse_object on each object it outputs, like this: object = parse_object(sha1); if (!object) die(...); if (object->flags & SHOWN) return; which means that just to realize we have already shown an object, we will read the whole object from disk! With this patch, my best-of-five time for "fast-export --all" on git.git dropped from 26.3s to 21.3s. The second case is upload-pack, which will call parse_object for each advertised ref (because it needs to peel tags to show "^{}" entries). This doesn't matter for most repositories, because they don't have a lot of refs pointing to the same objects. However, if you have a big alternates repository with a shared object db for a number of child repositories, then the alternates repository will have duplicated refs representing each of its children. For example, GitHub's alternates repository for git.git has ~120,000 refs, of which only ~3200 are unique. The time for upload-pack to print its list of advertised refs dropped from 3.4s to 0.76s. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2012-01-05 21:00:01 +00:00
void *buffer;
struct object *obj;
obj = lookup_object(r, oid);
parse_object: try internal cache before reading object db When parse_object is called, we do the following: 1. read the object data into a buffer via read_sha1_file 2. call parse_object_buffer, which then: a. calls the appropriate lookup_{commit,tree,blob,tag} to either create a new "struct object", or to find an existing one. We know the appropriate type from the lookup in step 1. b. calls the appropriate parse_{commit,tree,blob,tag} to parse the buffer for the new (or existing) object In step 2b, all of the called functions are no-ops for object "X" if "X->object.parsed" is set. I.e., when we have already parsed an object, we end up going to a lot of work just to find out at a low level that there is nothing left for us to do (and we throw away the data from read_sha1_file unread). We can optimize this by moving the check for "do we have an in-memory object" from 2a before the expensive call to read_sha1_file in step 1. This might seem circular, since step 2a uses the type information determined in step 1 to call the appropriate lookup function. However, we can notice that all of the lookup_* functions are backed by lookup_object. In other words, all of the objects are kept in a master hash table, and we don't actually need the type to do the "do we have it" part of the lookup, only to do the "and create it if it doesn't exist" part. This can save time whenever we call parse_object on the same sha1 twice in a single program. Some code paths already perform this optimization manually, with either: if (!obj->parsed) obj = parse_object(obj->sha1); if you already have a "struct object", or: struct object *obj = lookup_unknown_object(sha1); if (!obj || !obj->parsed) obj = parse_object(sha1); if you don't. This patch moves the optimization into parse_object itself. Most git operations won't notice any impact. Either they don't parse a lot of duplicate sha1s, or the calling code takes special care not to re-parse objects. I timed two code paths that do benefit (there may be more, but these two were immediately obvious and easy to time). The first is fast-export, which calls parse_object on each object it outputs, like this: object = parse_object(sha1); if (!object) die(...); if (object->flags & SHOWN) return; which means that just to realize we have already shown an object, we will read the whole object from disk! With this patch, my best-of-five time for "fast-export --all" on git.git dropped from 26.3s to 21.3s. The second case is upload-pack, which will call parse_object for each advertised ref (because it needs to peel tags to show "^{}" entries). This doesn't matter for most repositories, because they don't have a lot of refs pointing to the same objects. However, if you have a big alternates repository with a shared object db for a number of child repositories, then the alternates repository will have duplicated refs representing each of its children. For example, GitHub's alternates repository for git.git has ~120,000 refs, of which only ~3200 are unique. The time for upload-pack to print its list of advertised refs dropped from 3.4s to 0.76s. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2012-01-05 21:00:01 +00:00
if (obj && obj->parsed)
return obj;
if (skip_hash) {
struct commit *commit = lookup_commit_in_graph(r, repl);
if (commit)
return &commit->object;
}
if ((!obj || obj->type == OBJ_BLOB) &&
parse_object(): check on-disk type of suspected blob In parse_object(), we try to handle blobs by streaming rather than loading them entirely into memory. The most common case here will be that we haven't seen the object yet and check oid_object_info(), which tells us we have a blob. But we trigger this code on one other case: when we have an in-memory object struct with type OBJ_BLOB (and without its "parsed" flag set, since otherwise we'd return early from the function). This indicates that some other part of the code suspected we have a blob (e.g., it was mentioned by a tree or tag) but we haven't yet looked at the on-disk copy. In this case before hitting the streaming path, we check if we have the object on-disk at all. This is mostly pointless extra work, as the streaming path would complain if it couldn't open the object (albeit with the message "hash mismatch", which is a little misleading). But it's also insufficient to catch all problems. The streaming code will only tell us "yes, the on-disk object matches the oid". But it doesn't actually confirm that what we found was indeed a blob, and neither does repo_has_object_file(). One way to improve this would be to teach stream_object_signature() to check the type (either by returning it to us to check, or taking an "expected" type). But there's an even simpler fix here: if we suspect the object is a blob, just call oid_object_info() to confirm that we have it on-disk, and that it really is a blob. This is slightly less efficient than teaching stream_object_signature() to do it (since it has to open the object already). But this case very rarely comes up. In practice, we usually don't have any clue what the type is, in which case we already call oid_object_info(). This "suspected" case happens only when some other code created an object struct but didn't actually parse the blob, which is actually tricky to trigger at all (see the discussion of the test below). I reworked the conditional a bit so that instead of: if ((suspected_blob && oid_object_info() == OBJ_BLOB) (no_clue && oid_object_info() == OBJ_BLOB) we have the simpler: if ((suspected_blob || no_clue) && oid_object_info() == OBJ_BLOB) This is shorter, but also reflects what we really want say, which is "have we ruled out this being a blob; if not, check it on-disk". In either case, if oid_object_info() fails to tell us it's a blob, we'll skip the streaming code path and call repo_read_object_file(), just as before. And if we really do have a mismatch with the existing object struct, we'll eventually call lookup_commit(), etc, via parse_object_buffer(), which will complain that it doesn't match our existing obj->type. So this fixes one of the lingering expect_failure cases from 0616617c7e (t: introduce tests for unexpected object types, 2019-04-09). That test works by peeling a tag that claims to point to a blob (triggering us to create the struct), but really points to something else, which we later discover when we call parse_object() as part of the actual traversal). Prior to this commit, we'd quietly check the sha1 and mark the blob as "parsed". Now we correctly complain about the mismatch. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Taylor Blau <me@ttaylorr.com>
2022-11-17 22:41:16 +00:00
oid_object_info(r, oid, NULL) == OBJ_BLOB) {
parse_object(): allow skipping hash check The parse_object() function checks the object hash of any object it parses. This is a nice feature, as it means we may catch bit corruption during normal use, rather than waiting for specific fsck operations. But it also can be slow. It's particularly noticeable for blobs, where except for the hash check, we could return without loading the object contents at all. Now one may wonder what is the point of calling parse_object() on a blob in the first place then, but usually it's not intentional: we were fed an oid from somewhere, don't know the type, and want an object struct. For commits and trees, the parsing is usually helpful; we're about to look at the contents anyway. But this is less true for blobs, where we may be collecting them as part of a reachability traversal, etc, and don't actually care what's in them. And blobs, of course, tend to be larger. We don't want to just throw out the hash-checks for blobs, though. We do depend on them in some circumstances (e.g., rev-list --verify-objects uses parse_object() to check them). It's only the callers that know how they're going to use the result. And so we can help them by providing a special flag to skip the hash check. We could just apply this to blobs, as they're going to be the main source of performance improvement. But if a caller doesn't care about checking the hash, we might as well skip it for other object types, too. Even though we can't avoid reading the object contents, we can still skip the actual hash computation. If this seems like it is making Git a little bit less safe against corruption, it may be. But it's part of a series of tradeoffs we're already making. For instance, "rev-list --objects" does not open the contents of blobs it prints. And when a commit graph is present, we skip opening most commits entirely. The important thing will be to use this flag in cases where it's safe to skip the check. For instance, when serving a pack for a fetch, we know the client will fully index the objects and do a connectivity check itself. There's little to be gained from the server side re-hashing a blob itself. And indeed, most of the time we don't! The revision machinery won't open up a blob reached by traversal, but only one requested directly with a "want" line. So applied properly, this new feature shouldn't make anything less safe in practice. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-06 23:01:34 +00:00
if (!skip_hash && stream_object_signature(r, repl) < 0) {
error(_("hash mismatch %s"), oid_to_hex(oid));
return NULL;
}
blob: drop unused parts of parse_blob_buffer() Our parse_blob_buffer() takes a ptr/len combo, just like parse_tree_buffer(), etc, and returns success or failure. But it doesn't actually do anything with them; we just set the "parsed" flag in the object and return success, without even looking at the contents. There could be some value to keeping these unused parameters: - it's consistent with the parse functions for other object types. But we already lost that consistency in 837d395a5c (Replace parse_blob() with an explanatory comment, 2010-01-18). - As the comment from 837d395a5c explains, callers are supposed to make sure they have the object content available. So in theory asking for these parameters could serve as a signal. But there are only two callers, and one of them always passes NULL (after doing a streaming check of the object hash). This shows that there aren't likely to be a lot of callers (since everyone either uses the type-generic parse functions, or handles blobs individually), and that they need to take special care anyway (because we usually want to avoid loading whole blobs in memory if we can avoid it). So let's just drop these unused parameters, and likewise the useless return value. While we're touching the header file, let's move the declaration of parse_blob_buffer() right below that explanatory comment, where it's more likely to be seen by people looking for the function. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-12-13 11:11:57 +00:00
parse_blob_buffer(lookup_blob(r, oid));
return lookup_object(r, oid);
}
upload-pack: free tree buffers after parsing When a client sends us a "want" or "have" line, we call parse_object() to get an object struct. If the object is a tree, then the parsed state means that tree->buffer points to the uncompressed contents of the tree. But we don't really care about it. We only really need to parse commits and tags; for trees and blobs, the important output is just a "struct object" with the correct type. But much worse, we do not ever free that tree buffer. It's not leaked in the traditional sense, in that we still have a pointer to it from the global object hash. But if the client requests many trees, we'll hold all of their contents in memory at the same time. Nobody really noticed because it's rare for clients to directly request a tree. It might happen for a lightweight tag pointing straight at a tree, or it might happen for a "tree:depth" partial clone filling in missing trees. But it's also possible for a malicious client to request a lot of trees, causing upload-pack's memory to balloon. For example, without this patch, requesting every tree in git.git like: pktline() { local msg="$*" printf "%04x%s\n" $((1+4+${#msg})) "$msg" } want_trees() { pktline command=fetch printf 0001 git cat-file --batch-all-objects --batch-check='%(objectname) %(objecttype)' | while read oid type; do test "$type" = "tree" || continue pktline want $oid done pktline done printf 0000 } want_trees | GIT_PROTOCOL=version=2 valgrind --tool=massif ./git upload-pack . >/dev/null shows a peak heap usage of ~3.7GB. Which is just about the sum of the sizes of all of the uncompressed trees. For linux.git, it's closer to 17GB. So the obvious thing to do is to call free_tree_buffer() after we realize that we've parsed a tree. We know that upload-pack won't need it later. But let's push the logic into parse_object_with_flags(), telling it to discard the tree buffer immediately. There are two reasons for this. One, all of the relevant call-sites already call the with_options variant to pass the SKIP_HASH flag. So it actually ends up as less code than manually free-ing in each spot. And two, it enables an extra optimization that I'll discuss below. I've touched all of the sites that currently use SKIP_HASH in upload-pack. That drops the peak heap of the upload-pack invocation above from 3.7GB to ~24MB. I've also modified the caller in get_reference(); a partial clone benefits from its use in pack-objects for the reasons given in 0bc2557951 (upload-pack: skip parse-object re-hashing of "want" objects, 2022-09-06), where we were measuring blob requests. But note that the results of get_reference() are used for traversing, as well; so we really would _eventually_ use the tree contents. That makes this at first glance a space/time tradeoff: we won't hold all of the trees in memory at once, but we'll have to reload them each when it comes time to traverse. And here's where our extra optimization comes in. If the caller is not going to immediately look at the tree contents, and it doesn't care about checking the hash, then parse_object() can simply skip loading the tree entirely, just like we do for blobs! And now it's not a space/time tradeoff in get_reference() anymore. It's just a lazy-load: we're delaying reading the tree contents until it's time to actually traverse them one by one. And of course for upload-pack, this optimization means we never load the trees at all, saving lots of CPU time. Timing the "every tree from git.git" request above shows upload-pack dropping from 32 seconds of CPU to 19 (the remainder is mostly due to pack-objects actually sending the pack; timing just the upload-pack portion shows we go from 13s to ~0.28s). These are all highly gamed numbers, of course. For real-world partial-clone requests we're saving only a small bit of time in practice. But it does help harden upload-pack against malicious denial-of-service attacks. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-02-28 22:39:07 +00:00
/*
* If the caller does not care about the tree buffer and does not
* care about checking the hash, we can simply verify that we
* have the on-disk object with the correct type.
*/
if (skip_hash && discard_tree &&
(!obj || obj->type == OBJ_TREE) &&
oid_object_info(r, oid, NULL) == OBJ_TREE) {
return &lookup_tree(r, oid)->object;
}
buffer = repo_read_object_file(r, oid, &type, &size);
if (buffer) {
parse_object(): allow skipping hash check The parse_object() function checks the object hash of any object it parses. This is a nice feature, as it means we may catch bit corruption during normal use, rather than waiting for specific fsck operations. But it also can be slow. It's particularly noticeable for blobs, where except for the hash check, we could return without loading the object contents at all. Now one may wonder what is the point of calling parse_object() on a blob in the first place then, but usually it's not intentional: we were fed an oid from somewhere, don't know the type, and want an object struct. For commits and trees, the parsing is usually helpful; we're about to look at the contents anyway. But this is less true for blobs, where we may be collecting them as part of a reachability traversal, etc, and don't actually care what's in them. And blobs, of course, tend to be larger. We don't want to just throw out the hash-checks for blobs, though. We do depend on them in some circumstances (e.g., rev-list --verify-objects uses parse_object() to check them). It's only the callers that know how they're going to use the result. And so we can help them by providing a special flag to skip the hash check. We could just apply this to blobs, as they're going to be the main source of performance improvement. But if a caller doesn't care about checking the hash, we might as well skip it for other object types, too. Even though we can't avoid reading the object contents, we can still skip the actual hash computation. If this seems like it is making Git a little bit less safe against corruption, it may be. But it's part of a series of tradeoffs we're already making. For instance, "rev-list --objects" does not open the contents of blobs it prints. And when a commit graph is present, we skip opening most commits entirely. The important thing will be to use this flag in cases where it's safe to skip the check. For instance, when serving a pack for a fetch, we know the client will fully index the objects and do a connectivity check itself. There's little to be gained from the server side re-hashing a blob itself. And indeed, most of the time we don't! The revision machinery won't open up a blob reached by traversal, but only one requested directly with a "want" line. So applied properly, this new feature shouldn't make anything less safe in practice. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-06 23:01:34 +00:00
if (!skip_hash &&
check_object_signature(r, repl, buffer, size, type) < 0) {
free(buffer);
error(_("hash mismatch %s"), oid_to_hex(repl));
return NULL;
}
obj = parse_object_buffer(r, oid, type, size,
buffer, &eaten);
if (!eaten)
free(buffer);
upload-pack: free tree buffers after parsing When a client sends us a "want" or "have" line, we call parse_object() to get an object struct. If the object is a tree, then the parsed state means that tree->buffer points to the uncompressed contents of the tree. But we don't really care about it. We only really need to parse commits and tags; for trees and blobs, the important output is just a "struct object" with the correct type. But much worse, we do not ever free that tree buffer. It's not leaked in the traditional sense, in that we still have a pointer to it from the global object hash. But if the client requests many trees, we'll hold all of their contents in memory at the same time. Nobody really noticed because it's rare for clients to directly request a tree. It might happen for a lightweight tag pointing straight at a tree, or it might happen for a "tree:depth" partial clone filling in missing trees. But it's also possible for a malicious client to request a lot of trees, causing upload-pack's memory to balloon. For example, without this patch, requesting every tree in git.git like: pktline() { local msg="$*" printf "%04x%s\n" $((1+4+${#msg})) "$msg" } want_trees() { pktline command=fetch printf 0001 git cat-file --batch-all-objects --batch-check='%(objectname) %(objecttype)' | while read oid type; do test "$type" = "tree" || continue pktline want $oid done pktline done printf 0000 } want_trees | GIT_PROTOCOL=version=2 valgrind --tool=massif ./git upload-pack . >/dev/null shows a peak heap usage of ~3.7GB. Which is just about the sum of the sizes of all of the uncompressed trees. For linux.git, it's closer to 17GB. So the obvious thing to do is to call free_tree_buffer() after we realize that we've parsed a tree. We know that upload-pack won't need it later. But let's push the logic into parse_object_with_flags(), telling it to discard the tree buffer immediately. There are two reasons for this. One, all of the relevant call-sites already call the with_options variant to pass the SKIP_HASH flag. So it actually ends up as less code than manually free-ing in each spot. And two, it enables an extra optimization that I'll discuss below. I've touched all of the sites that currently use SKIP_HASH in upload-pack. That drops the peak heap of the upload-pack invocation above from 3.7GB to ~24MB. I've also modified the caller in get_reference(); a partial clone benefits from its use in pack-objects for the reasons given in 0bc2557951 (upload-pack: skip parse-object re-hashing of "want" objects, 2022-09-06), where we were measuring blob requests. But note that the results of get_reference() are used for traversing, as well; so we really would _eventually_ use the tree contents. That makes this at first glance a space/time tradeoff: we won't hold all of the trees in memory at once, but we'll have to reload them each when it comes time to traverse. And here's where our extra optimization comes in. If the caller is not going to immediately look at the tree contents, and it doesn't care about checking the hash, then parse_object() can simply skip loading the tree entirely, just like we do for blobs! And now it's not a space/time tradeoff in get_reference() anymore. It's just a lazy-load: we're delaying reading the tree contents until it's time to actually traverse them one by one. And of course for upload-pack, this optimization means we never load the trees at all, saving lots of CPU time. Timing the "every tree from git.git" request above shows upload-pack dropping from 32 seconds of CPU to 19 (the remainder is mostly due to pack-objects actually sending the pack; timing just the upload-pack portion shows we go from 13s to ~0.28s). These are all highly gamed numbers, of course. For real-world partial-clone requests we're saving only a small bit of time in practice. But it does help harden upload-pack against malicious denial-of-service attacks. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2024-02-28 22:39:07 +00:00
if (discard_tree && type == OBJ_TREE)
free_tree_buffer((struct tree *)obj);
return obj;
}
return NULL;
}
parse_object(): allow skipping hash check The parse_object() function checks the object hash of any object it parses. This is a nice feature, as it means we may catch bit corruption during normal use, rather than waiting for specific fsck operations. But it also can be slow. It's particularly noticeable for blobs, where except for the hash check, we could return without loading the object contents at all. Now one may wonder what is the point of calling parse_object() on a blob in the first place then, but usually it's not intentional: we were fed an oid from somewhere, don't know the type, and want an object struct. For commits and trees, the parsing is usually helpful; we're about to look at the contents anyway. But this is less true for blobs, where we may be collecting them as part of a reachability traversal, etc, and don't actually care what's in them. And blobs, of course, tend to be larger. We don't want to just throw out the hash-checks for blobs, though. We do depend on them in some circumstances (e.g., rev-list --verify-objects uses parse_object() to check them). It's only the callers that know how they're going to use the result. And so we can help them by providing a special flag to skip the hash check. We could just apply this to blobs, as they're going to be the main source of performance improvement. But if a caller doesn't care about checking the hash, we might as well skip it for other object types, too. Even though we can't avoid reading the object contents, we can still skip the actual hash computation. If this seems like it is making Git a little bit less safe against corruption, it may be. But it's part of a series of tradeoffs we're already making. For instance, "rev-list --objects" does not open the contents of blobs it prints. And when a commit graph is present, we skip opening most commits entirely. The important thing will be to use this flag in cases where it's safe to skip the check. For instance, when serving a pack for a fetch, we know the client will fully index the objects and do a connectivity check itself. There's little to be gained from the server side re-hashing a blob itself. And indeed, most of the time we don't! The revision machinery won't open up a blob reached by traversal, but only one requested directly with a "want" line. So applied properly, this new feature shouldn't make anything less safe in practice. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2022-09-06 23:01:34 +00:00
struct object *parse_object(struct repository *r, const struct object_id *oid)
{
return parse_object_with_flags(r, oid, 0);
}
struct object_list *object_list_insert(struct object *item,
struct object_list **list_p)
{
struct object_list *new_list = xmalloc(sizeof(struct object_list));
new_list->item = item;
new_list->next = *list_p;
*list_p = new_list;
return new_list;
}
int object_list_contains(struct object_list *list, struct object *obj)
{
while (list) {
if (list->item == obj)
return 1;
list = list->next;
}
return 0;
}
Add "named object array" concept We've had this notion of a "object_list" for a long time, which eventually grew a "name" member because some users (notably git-rev-list) wanted to name each object as it is generated. That object_list is great for some things, but it isn't all that wonderful for others, and the "name" member is generally not used by everybody. This patch splits the users of the object_list array up into two: the traditional list users, who want the list-like format, and who don't actually use or want the name. And another class of users that really used the list as an extensible array, and generally wanted to name the objects. The patch is fairly straightforward, but it's also biggish. Most of it really just cleans things up: switching the revision parsing and listing over to the array makes things like the builtin-diff usage much simpler (we now see exactly how many members the array has, and we don't get the objects reversed from the order they were on the command line). One of the main reasons for doing this at all is that the malloc overhead of the simple object list was actually pretty high, and the array is just a lot denser. So this patch brings down memory usage by git-rev-list by just under 3% (on top of all the other memory use optimizations) on the mozilla archive. It does add more lines than it removes, and more importantly, it adds a whole new infrastructure for maintaining lists of objects, but on the other hand, the new dynamic array code is pretty obvious. The change to builtin-diff-tree.c shows a fairly good example of why an array interface is sometimes more natural, and just much simpler for everybody. Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-20 00:42:35 +00:00
void object_list_free(struct object_list **list)
{
while (*list) {
struct object_list *p = *list;
*list = p->next;
free(p);
}
}
object_array_entry: fix memory handling of the name field Previously, the memory management of the object_array_entry::name field was inconsistent and undocumented. object_array_entries are ultimately created by a single function, add_object_array_with_mode(), which has an argument "const char *name". This function used to simply set the name field to reference the string pointed to by the name parameter, and nobody on the object_array side ever freed the memory. Thus, it assumed that the memory for the name field would be managed by the caller, and that the lifetime of that string would be at least as long as the lifetime of the object_array_entry. But callers were inconsistent: * Some passed pointers to constant strings or argv entries, which was OK. * Some passed pointers to newly-allocated memory, but didn't arrange for the memory ever to be freed. * Some passed the return value of sha1_to_hex(), which is a pointer to a statically-allocated buffer that can be overwritten at any time. * Some passed pointers to refnames that they received from a for_each_ref()-type iteration, but the lifetimes of such refnames is not guaranteed by the refs API. Bring consistency to this mess by changing object_array to make its own copy for the object_array_entry::name field and free this memory when an object_array_entry is deleted from the array. Many callers were passing the empty string as the name parameter, so as a performance optimization, treat the empty string specially. Instead of making a copy, store a pointer to a statically-allocated empty string to object_array_entry::name. When deleting such an entry, skip the free(). Change the callers that were already passing copies to add_object_array_with_mode() to either skip the copy, or (if the memory needed to be allocated anyway) freeing the memory itself. A part of this commit effectively reverts 70d26c6e76 read_revisions_from_stdin: make copies for handle_revision_arg because the copying introduced by that commit (which is still necessary) is now done at a deeper level. Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-25 09:08:14 +00:00
/*
* A zero-length string to which object_array_entry::name can be
* initialized without requiring a malloc/free.
*/
static char object_array_slopbuf[1];
void object_array_init(struct object_array *array)
{
struct object_array blank = OBJECT_ARRAY_INIT;
memcpy(array, &blank, sizeof(*array));
}
void add_object_array_with_path(struct object *obj, const char *name,
struct object_array *array,
unsigned mode, const char *path)
Add "named object array" concept We've had this notion of a "object_list" for a long time, which eventually grew a "name" member because some users (notably git-rev-list) wanted to name each object as it is generated. That object_list is great for some things, but it isn't all that wonderful for others, and the "name" member is generally not used by everybody. This patch splits the users of the object_list array up into two: the traditional list users, who want the list-like format, and who don't actually use or want the name. And another class of users that really used the list as an extensible array, and generally wanted to name the objects. The patch is fairly straightforward, but it's also biggish. Most of it really just cleans things up: switching the revision parsing and listing over to the array makes things like the builtin-diff usage much simpler (we now see exactly how many members the array has, and we don't get the objects reversed from the order they were on the command line). One of the main reasons for doing this at all is that the malloc overhead of the simple object list was actually pretty high, and the array is just a lot denser. So this patch brings down memory usage by git-rev-list by just under 3% (on top of all the other memory use optimizations) on the mozilla archive. It does add more lines than it removes, and more importantly, it adds a whole new infrastructure for maintaining lists of objects, but on the other hand, the new dynamic array code is pretty obvious. The change to builtin-diff-tree.c shows a fairly good example of why an array interface is sometimes more natural, and just much simpler for everybody. Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-20 00:42:35 +00:00
{
unsigned nr = array->nr;
unsigned alloc = array->alloc;
struct object_array_entry *objects = array->objects;
object_array_entry: fix memory handling of the name field Previously, the memory management of the object_array_entry::name field was inconsistent and undocumented. object_array_entries are ultimately created by a single function, add_object_array_with_mode(), which has an argument "const char *name". This function used to simply set the name field to reference the string pointed to by the name parameter, and nobody on the object_array side ever freed the memory. Thus, it assumed that the memory for the name field would be managed by the caller, and that the lifetime of that string would be at least as long as the lifetime of the object_array_entry. But callers were inconsistent: * Some passed pointers to constant strings or argv entries, which was OK. * Some passed pointers to newly-allocated memory, but didn't arrange for the memory ever to be freed. * Some passed the return value of sha1_to_hex(), which is a pointer to a statically-allocated buffer that can be overwritten at any time. * Some passed pointers to refnames that they received from a for_each_ref()-type iteration, but the lifetimes of such refnames is not guaranteed by the refs API. Bring consistency to this mess by changing object_array to make its own copy for the object_array_entry::name field and free this memory when an object_array_entry is deleted from the array. Many callers were passing the empty string as the name parameter, so as a performance optimization, treat the empty string specially. Instead of making a copy, store a pointer to a statically-allocated empty string to object_array_entry::name. When deleting such an entry, skip the free(). Change the callers that were already passing copies to add_object_array_with_mode() to either skip the copy, or (if the memory needed to be allocated anyway) freeing the memory itself. A part of this commit effectively reverts 70d26c6e76 read_revisions_from_stdin: make copies for handle_revision_arg because the copying introduced by that commit (which is still necessary) is now done at a deeper level. Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-25 09:08:14 +00:00
struct object_array_entry *entry;
Add "named object array" concept We've had this notion of a "object_list" for a long time, which eventually grew a "name" member because some users (notably git-rev-list) wanted to name each object as it is generated. That object_list is great for some things, but it isn't all that wonderful for others, and the "name" member is generally not used by everybody. This patch splits the users of the object_list array up into two: the traditional list users, who want the list-like format, and who don't actually use or want the name. And another class of users that really used the list as an extensible array, and generally wanted to name the objects. The patch is fairly straightforward, but it's also biggish. Most of it really just cleans things up: switching the revision parsing and listing over to the array makes things like the builtin-diff usage much simpler (we now see exactly how many members the array has, and we don't get the objects reversed from the order they were on the command line). One of the main reasons for doing this at all is that the malloc overhead of the simple object list was actually pretty high, and the array is just a lot denser. So this patch brings down memory usage by git-rev-list by just under 3% (on top of all the other memory use optimizations) on the mozilla archive. It does add more lines than it removes, and more importantly, it adds a whole new infrastructure for maintaining lists of objects, but on the other hand, the new dynamic array code is pretty obvious. The change to builtin-diff-tree.c shows a fairly good example of why an array interface is sometimes more natural, and just much simpler for everybody. Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-20 00:42:35 +00:00
if (nr >= alloc) {
alloc = (alloc + 32) * 2;
REALLOC_ARRAY(objects, alloc);
Add "named object array" concept We've had this notion of a "object_list" for a long time, which eventually grew a "name" member because some users (notably git-rev-list) wanted to name each object as it is generated. That object_list is great for some things, but it isn't all that wonderful for others, and the "name" member is generally not used by everybody. This patch splits the users of the object_list array up into two: the traditional list users, who want the list-like format, and who don't actually use or want the name. And another class of users that really used the list as an extensible array, and generally wanted to name the objects. The patch is fairly straightforward, but it's also biggish. Most of it really just cleans things up: switching the revision parsing and listing over to the array makes things like the builtin-diff usage much simpler (we now see exactly how many members the array has, and we don't get the objects reversed from the order they were on the command line). One of the main reasons for doing this at all is that the malloc overhead of the simple object list was actually pretty high, and the array is just a lot denser. So this patch brings down memory usage by git-rev-list by just under 3% (on top of all the other memory use optimizations) on the mozilla archive. It does add more lines than it removes, and more importantly, it adds a whole new infrastructure for maintaining lists of objects, but on the other hand, the new dynamic array code is pretty obvious. The change to builtin-diff-tree.c shows a fairly good example of why an array interface is sometimes more natural, and just much simpler for everybody. Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-20 00:42:35 +00:00
array->alloc = alloc;
array->objects = objects;
}
object_array_entry: fix memory handling of the name field Previously, the memory management of the object_array_entry::name field was inconsistent and undocumented. object_array_entries are ultimately created by a single function, add_object_array_with_mode(), which has an argument "const char *name". This function used to simply set the name field to reference the string pointed to by the name parameter, and nobody on the object_array side ever freed the memory. Thus, it assumed that the memory for the name field would be managed by the caller, and that the lifetime of that string would be at least as long as the lifetime of the object_array_entry. But callers were inconsistent: * Some passed pointers to constant strings or argv entries, which was OK. * Some passed pointers to newly-allocated memory, but didn't arrange for the memory ever to be freed. * Some passed the return value of sha1_to_hex(), which is a pointer to a statically-allocated buffer that can be overwritten at any time. * Some passed pointers to refnames that they received from a for_each_ref()-type iteration, but the lifetimes of such refnames is not guaranteed by the refs API. Bring consistency to this mess by changing object_array to make its own copy for the object_array_entry::name field and free this memory when an object_array_entry is deleted from the array. Many callers were passing the empty string as the name parameter, so as a performance optimization, treat the empty string specially. Instead of making a copy, store a pointer to a statically-allocated empty string to object_array_entry::name. When deleting such an entry, skip the free(). Change the callers that were already passing copies to add_object_array_with_mode() to either skip the copy, or (if the memory needed to be allocated anyway) freeing the memory itself. A part of this commit effectively reverts 70d26c6e76 read_revisions_from_stdin: make copies for handle_revision_arg because the copying introduced by that commit (which is still necessary) is now done at a deeper level. Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-25 09:08:14 +00:00
entry = &objects[nr];
entry->item = obj;
if (!name)
entry->name = NULL;
else if (!*name)
/* Use our own empty string instead of allocating one: */
entry->name = object_array_slopbuf;
else
entry->name = xstrdup(name);
entry->mode = mode;
if (path)
entry->path = xstrdup(path);
else
entry->path = NULL;
Add "named object array" concept We've had this notion of a "object_list" for a long time, which eventually grew a "name" member because some users (notably git-rev-list) wanted to name each object as it is generated. That object_list is great for some things, but it isn't all that wonderful for others, and the "name" member is generally not used by everybody. This patch splits the users of the object_list array up into two: the traditional list users, who want the list-like format, and who don't actually use or want the name. And another class of users that really used the list as an extensible array, and generally wanted to name the objects. The patch is fairly straightforward, but it's also biggish. Most of it really just cleans things up: switching the revision parsing and listing over to the array makes things like the builtin-diff usage much simpler (we now see exactly how many members the array has, and we don't get the objects reversed from the order they were on the command line). One of the main reasons for doing this at all is that the malloc overhead of the simple object list was actually pretty high, and the array is just a lot denser. So this patch brings down memory usage by git-rev-list by just under 3% (on top of all the other memory use optimizations) on the mozilla archive. It does add more lines than it removes, and more importantly, it adds a whole new infrastructure for maintaining lists of objects, but on the other hand, the new dynamic array code is pretty obvious. The change to builtin-diff-tree.c shows a fairly good example of why an array interface is sometimes more natural, and just much simpler for everybody. Signed-off-by: Linus Torvalds <torvalds@osdl.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
2006-06-20 00:42:35 +00:00
array->nr = ++nr;
}
void add_object_array(struct object *obj, const char *name, struct object_array *array)
{
add_object_array_with_path(obj, name, array, S_IFINVALID, NULL);
}
/*
* Free all memory associated with an entry; the result is
* in an unspecified state and should not be examined.
*/
static void object_array_release_entry(struct object_array_entry *ent)
{
if (ent->name != object_array_slopbuf)
free(ent->name);
free(ent->path);
}
object_array: add and use `object_array_pop()` In a couple of places, we pop objects off an object array `foo` by decreasing `foo.nr`. We access `foo.nr` in many places, but most if not all other times we do so read-only, e.g., as we iterate over the array. But when we change `foo.nr` behind the array's back, it feels a bit nasty and looks like it might leak memory. Leaks happen if the popped element has an allocated `name` or `path`. At the moment, that is not the case. Still, 1) the object array might gain more fields that want to be freed, 2) a code path where we pop might start using names or paths, 3) one of these code paths might be copied to somewhere where we do, and 4) using a dedicated function for popping is conceptually cleaner. Introduce and use `object_array_pop()` instead. Release memory in the new function. Document that popping an object leaves the associated elements in limbo. The converted places were identified by grepping for "\.nr\>" and looking for "--". Make the new function return NULL on an empty array. This is consistent with `pop_commit()` and allows the following: while ((o = object_array_pop(&foo)) != NULL) { // do something } But as noted above, we don't need to go out of our way to avoid reading `foo.nr`. This is probably more readable: while (foo.nr) { ... o = object_array_pop(&foo); // do something } The name of `object_array_pop()` does not quite align with `add_object_array()`. That is unfortunate. On the other hand, it matches `object_array_clear()`. Arguably it's `add_...` that is the odd one out, since it reads like it's used to "add" an "object array". For that reason, side with `object_array_clear()`. Signed-off-by: Martin Ågren <martin.agren@gmail.com> Reviewed-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2017-09-22 23:34:53 +00:00
struct object *object_array_pop(struct object_array *array)
{
struct object *ret;
if (!array->nr)
return NULL;
ret = array->objects[array->nr - 1].item;
object_array_release_entry(&array->objects[array->nr - 1]);
array->nr--;
return ret;
}
void object_array_filter(struct object_array *array,
object_array_each_func_t want, void *cb_data)
{
unsigned nr = array->nr, src, dst;
struct object_array_entry *objects = array->objects;
for (src = dst = 0; src < nr; src++) {
if (want(&objects[src], cb_data)) {
if (src != dst)
objects[dst] = objects[src];
dst++;
object_array_entry: fix memory handling of the name field Previously, the memory management of the object_array_entry::name field was inconsistent and undocumented. object_array_entries are ultimately created by a single function, add_object_array_with_mode(), which has an argument "const char *name". This function used to simply set the name field to reference the string pointed to by the name parameter, and nobody on the object_array side ever freed the memory. Thus, it assumed that the memory for the name field would be managed by the caller, and that the lifetime of that string would be at least as long as the lifetime of the object_array_entry. But callers were inconsistent: * Some passed pointers to constant strings or argv entries, which was OK. * Some passed pointers to newly-allocated memory, but didn't arrange for the memory ever to be freed. * Some passed the return value of sha1_to_hex(), which is a pointer to a statically-allocated buffer that can be overwritten at any time. * Some passed pointers to refnames that they received from a for_each_ref()-type iteration, but the lifetimes of such refnames is not guaranteed by the refs API. Bring consistency to this mess by changing object_array to make its own copy for the object_array_entry::name field and free this memory when an object_array_entry is deleted from the array. Many callers were passing the empty string as the name parameter, so as a performance optimization, treat the empty string specially. Instead of making a copy, store a pointer to a statically-allocated empty string to object_array_entry::name. When deleting such an entry, skip the free(). Change the callers that were already passing copies to add_object_array_with_mode() to either skip the copy, or (if the memory needed to be allocated anyway) freeing the memory itself. A part of this commit effectively reverts 70d26c6e76 read_revisions_from_stdin: make copies for handle_revision_arg because the copying introduced by that commit (which is still necessary) is now done at a deeper level. Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-25 09:08:14 +00:00
} else {
object_array_release_entry(&objects[src]);
}
}
array->nr = dst;
}
void object_array_clear(struct object_array *array)
{
int i;
for (i = 0; i < array->nr; i++)
object_array_release_entry(&array->objects[i]);
FREE_AND_NULL(array->objects);
array->nr = array->alloc = 0;
}
/*
bundle: lost objects when removing duplicate pendings `git rev-list` will list one commit for the following command: $ git rev-list 'main^!' <tip-commit-of-main-branch> But providing the same rev-list args to `git bundle`, fail to create a bundle file. $ git bundle create - 'main^!' # v2 git bundle -<OID> <one-line-message> fatal: Refusing to create empty bundle. This is because when removing duplicate objects in function `object_array_remove_duplicates()`, one unique pending object which has the same name is deleted by mistake. The revision arg 'main^!' in the above example is parsed by `handle_revision_arg()`, and at lease two different objects will be appended to `revs.pending`, one points to the parent commit of the "main" branch, and the other points to the tip commit of the "main" branch. These two objects have the same name "main". Only one object is left with the name "main" after calling the function `object_array_remove_duplicates()`. And what's worse, when adding boundary commits into pending list, we use one-line commit message as names, and the arbitory names may surprise git-bundle. Only comparing objects themselves (".item") is also not good enough, because user may want to create a bundle with two identical objects but with different reference names, such as: "HEAD" and "refs/heads/main". Add new function `contains_object()` which compare both the address and the name of the object. Signed-off-by: Jiang Xin <zhiyou.jx@alibaba-inc.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-12 02:27:02 +00:00
* Return true if array already contains an entry.
*/
bundle: lost objects when removing duplicate pendings `git rev-list` will list one commit for the following command: $ git rev-list 'main^!' <tip-commit-of-main-branch> But providing the same rev-list args to `git bundle`, fail to create a bundle file. $ git bundle create - 'main^!' # v2 git bundle -<OID> <one-line-message> fatal: Refusing to create empty bundle. This is because when removing duplicate objects in function `object_array_remove_duplicates()`, one unique pending object which has the same name is deleted by mistake. The revision arg 'main^!' in the above example is parsed by `handle_revision_arg()`, and at lease two different objects will be appended to `revs.pending`, one points to the parent commit of the "main" branch, and the other points to the tip commit of the "main" branch. These two objects have the same name "main". Only one object is left with the name "main" after calling the function `object_array_remove_duplicates()`. And what's worse, when adding boundary commits into pending list, we use one-line commit message as names, and the arbitory names may surprise git-bundle. Only comparing objects themselves (".item") is also not good enough, because user may want to create a bundle with two identical objects but with different reference names, such as: "HEAD" and "refs/heads/main". Add new function `contains_object()` which compare both the address and the name of the object. Signed-off-by: Jiang Xin <zhiyou.jx@alibaba-inc.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-12 02:27:02 +00:00
static int contains_object(struct object_array *array,
const struct object *item, const char *name)
{
unsigned nr = array->nr, i;
struct object_array_entry *object = array->objects;
for (i = 0; i < nr; i++, object++)
bundle: lost objects when removing duplicate pendings `git rev-list` will list one commit for the following command: $ git rev-list 'main^!' <tip-commit-of-main-branch> But providing the same rev-list args to `git bundle`, fail to create a bundle file. $ git bundle create - 'main^!' # v2 git bundle -<OID> <one-line-message> fatal: Refusing to create empty bundle. This is because when removing duplicate objects in function `object_array_remove_duplicates()`, one unique pending object which has the same name is deleted by mistake. The revision arg 'main^!' in the above example is parsed by `handle_revision_arg()`, and at lease two different objects will be appended to `revs.pending`, one points to the parent commit of the "main" branch, and the other points to the tip commit of the "main" branch. These two objects have the same name "main". Only one object is left with the name "main" after calling the function `object_array_remove_duplicates()`. And what's worse, when adding boundary commits into pending list, we use one-line commit message as names, and the arbitory names may surprise git-bundle. Only comparing objects themselves (".item") is also not good enough, because user may want to create a bundle with two identical objects but with different reference names, such as: "HEAD" and "refs/heads/main". Add new function `contains_object()` which compare both the address and the name of the object. Signed-off-by: Jiang Xin <zhiyou.jx@alibaba-inc.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-12 02:27:02 +00:00
if (item == object->item && !strcmp(object->name, name))
return 1;
return 0;
}
void object_array_remove_duplicates(struct object_array *array)
{
unsigned nr = array->nr, src;
struct object_array_entry *objects = array->objects;
array->nr = 0;
for (src = 0; src < nr; src++) {
bundle: lost objects when removing duplicate pendings `git rev-list` will list one commit for the following command: $ git rev-list 'main^!' <tip-commit-of-main-branch> But providing the same rev-list args to `git bundle`, fail to create a bundle file. $ git bundle create - 'main^!' # v2 git bundle -<OID> <one-line-message> fatal: Refusing to create empty bundle. This is because when removing duplicate objects in function `object_array_remove_duplicates()`, one unique pending object which has the same name is deleted by mistake. The revision arg 'main^!' in the above example is parsed by `handle_revision_arg()`, and at lease two different objects will be appended to `revs.pending`, one points to the parent commit of the "main" branch, and the other points to the tip commit of the "main" branch. These two objects have the same name "main". Only one object is left with the name "main" after calling the function `object_array_remove_duplicates()`. And what's worse, when adding boundary commits into pending list, we use one-line commit message as names, and the arbitory names may surprise git-bundle. Only comparing objects themselves (".item") is also not good enough, because user may want to create a bundle with two identical objects but with different reference names, such as: "HEAD" and "refs/heads/main". Add new function `contains_object()` which compare both the address and the name of the object. Signed-off-by: Jiang Xin <zhiyou.jx@alibaba-inc.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2021-01-12 02:27:02 +00:00
if (!contains_object(array, objects[src].item,
objects[src].name)) {
if (src != array->nr)
objects[array->nr] = objects[src];
array->nr++;
object_array_entry: fix memory handling of the name field Previously, the memory management of the object_array_entry::name field was inconsistent and undocumented. object_array_entries are ultimately created by a single function, add_object_array_with_mode(), which has an argument "const char *name". This function used to simply set the name field to reference the string pointed to by the name parameter, and nobody on the object_array side ever freed the memory. Thus, it assumed that the memory for the name field would be managed by the caller, and that the lifetime of that string would be at least as long as the lifetime of the object_array_entry. But callers were inconsistent: * Some passed pointers to constant strings or argv entries, which was OK. * Some passed pointers to newly-allocated memory, but didn't arrange for the memory ever to be freed. * Some passed the return value of sha1_to_hex(), which is a pointer to a statically-allocated buffer that can be overwritten at any time. * Some passed pointers to refnames that they received from a for_each_ref()-type iteration, but the lifetimes of such refnames is not guaranteed by the refs API. Bring consistency to this mess by changing object_array to make its own copy for the object_array_entry::name field and free this memory when an object_array_entry is deleted from the array. Many callers were passing the empty string as the name parameter, so as a performance optimization, treat the empty string specially. Instead of making a copy, store a pointer to a statically-allocated empty string to object_array_entry::name. When deleting such an entry, skip the free(). Change the callers that were already passing copies to add_object_array_with_mode() to either skip the copy, or (if the memory needed to be allocated anyway) freeing the memory itself. A part of this commit effectively reverts 70d26c6e76 read_revisions_from_stdin: make copies for handle_revision_arg because the copying introduced by that commit (which is still necessary) is now done at a deeper level. Signed-off-by: Michael Haggerty <mhagger@alum.mit.edu> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2013-05-25 09:08:14 +00:00
} else {
object_array_release_entry(&objects[src]);
}
}
}
void clear_object_flags(unsigned flags)
{
int i;
for (i=0; i < the_repository->parsed_objects->obj_hash_size; i++) {
struct object *obj = the_repository->parsed_objects->obj_hash[i];
if (obj)
obj->flags &= ~flags;
}
}
void repo_clear_commit_marks(struct repository *r, unsigned int flags)
{
int i;
for (i = 0; i < r->parsed_objects->obj_hash_size; i++) {
struct object *obj = r->parsed_objects->obj_hash[i];
if (obj && obj->type == OBJ_COMMIT)
obj->flags &= ~flags;
}
}
struct parsed_object_pool *parsed_object_pool_new(void)
{
struct parsed_object_pool *o = xmalloc(sizeof(*o));
memset(o, 0, sizeof(*o));
o->blob_state = allocate_alloc_state();
o->tree_state = allocate_alloc_state();
o->commit_state = allocate_alloc_state();
o->tag_state = allocate_alloc_state();
o->object_state = allocate_alloc_state();
o->is_shallow = -1;
CALLOC_ARRAY(o->shallow_stat, 1);
o->buffer_slab = allocate_commit_buffer_slab();
return o;
}
struct raw_object_store *raw_object_store_new(void)
{
struct raw_object_store *o = xmalloc(sizeof(*o));
memset(o, 0, sizeof(*o));
INIT_LIST_HEAD(&o->packed_git_mru);
hashmap_init(&o->pack_map, pack_map_entry_cmp, NULL, 0);
replace-object: make replace operations thread-safe replace-object functions are very close to being thread-safe: the only current racy section is the lazy initialization at prepare_replace_object(). The following patches will protect some object reading operations to be called threaded, but before that, replace functions must be protected. To do so, add a mutex to struct raw_object_store and acquire it before lazy initializing the replace_map. This won't cause any noticeable performance drop as the mutex will no longer be used after the replace_map is initialized. Later, when the replace functions are called in parallel, thread debuggers might point our use of the added replace_map_initialized flag as a data race. However, as this boolean variable is initialized as false and it's only updated once, there's no real harm. It's perfectly fine if the value is updated right after a thread read it in replace-map.h:lookup_replace_object() (there'll only be a performance penalty for the affected threads at that moment). We could cease the debugger warning protecting the variable reading at the said function. However, this would negatively affect performance for all threads calling it, at any time, so it's not really worthy since the warning doesn't represent a real problem. Instead, to make sure we don't get false positives (at ThreadSanitizer, at least) an entry for the respective function is added to .tsan-suppressions. Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-01-16 02:39:52 +00:00
pthread_mutex_init(&o->replace_mutex, NULL);
return o;
}
void free_object_directory(struct object_directory *odb)
{
sha1-file: use an object_directory for the main object dir Our handling of alternate object directories is needlessly different from the main object directory. As a result, many places in the code basically look like this: do_something(r->objects->objdir); for (odb = r->objects->alt_odb_list; odb; odb = odb->next) do_something(odb->path); That gets annoying when do_something() is non-trivial, and we've resorted to gross hacks like creating fake alternates (see find_short_object_filename()). Instead, let's give each raw_object_store a unified list of object_directory structs. The first will be the main store, and everything after is an alternate. Very few callers even care about the distinction, and can just loop over the whole list (and those who care can just treat the first element differently). A few observations: - we don't need r->objects->objectdir anymore, and can just mechanically convert that to r->objects->odb->path - object_directory's path field needs to become a real pointer rather than a FLEX_ARRAY, in order to fill it with expand_base_dir() - we'll call prepare_alt_odb() earlier in many functions (i.e., outside of the loop). This may result in us calling it even when our function would be satisfied looking only at the main odb. But this doesn't matter in practice. It's not a very expensive operation in the first place, and in the majority of cases it will be a noop. We call it already (and cache its results) in prepare_packed_git(), and we'll generally check packs before loose objects. So essentially every program is going to call it immediately once per program. Arguably we should just prepare_alt_odb() immediately upon setting up the repository's object directory, which would save us sprinkling calls throughout the code base (and forgetting to do so has been a source of subtle bugs in the past). But I've stopped short of that here, since there are already a lot of other moving parts in this patch. - Most call sites just get shorter. The check_and_freshen() functions are an exception, because they have entry points to handle local and nonlocal directories separately. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-11-12 14:50:39 +00:00
free(odb->path);
odb_clear_loose_cache(odb);
loose_object_map_clear(&odb->loose_map);
free(odb);
}
sha1-file: use an object_directory for the main object dir Our handling of alternate object directories is needlessly different from the main object directory. As a result, many places in the code basically look like this: do_something(r->objects->objdir); for (odb = r->objects->alt_odb_list; odb; odb = odb->next) do_something(odb->path); That gets annoying when do_something() is non-trivial, and we've resorted to gross hacks like creating fake alternates (see find_short_object_filename()). Instead, let's give each raw_object_store a unified list of object_directory structs. The first will be the main store, and everything after is an alternate. Very few callers even care about the distinction, and can just loop over the whole list (and those who care can just treat the first element differently). A few observations: - we don't need r->objects->objectdir anymore, and can just mechanically convert that to r->objects->odb->path - object_directory's path field needs to become a real pointer rather than a FLEX_ARRAY, in order to fill it with expand_base_dir() - we'll call prepare_alt_odb() earlier in many functions (i.e., outside of the loop). This may result in us calling it even when our function would be satisfied looking only at the main odb. But this doesn't matter in practice. It's not a very expensive operation in the first place, and in the majority of cases it will be a noop. We call it already (and cache its results) in prepare_packed_git(), and we'll generally check packs before loose objects. So essentially every program is going to call it immediately once per program. Arguably we should just prepare_alt_odb() immediately upon setting up the repository's object directory, which would save us sprinkling calls throughout the code base (and forgetting to do so has been a source of subtle bugs in the past). But I've stopped short of that here, since there are already a lot of other moving parts in this patch. - Most call sites just get shorter. The check_and_freshen() functions are an exception, because they have entry points to handle local and nonlocal directories separately. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-11-12 14:50:39 +00:00
static void free_object_directories(struct raw_object_store *o)
{
sha1-file: use an object_directory for the main object dir Our handling of alternate object directories is needlessly different from the main object directory. As a result, many places in the code basically look like this: do_something(r->objects->objdir); for (odb = r->objects->alt_odb_list; odb; odb = odb->next) do_something(odb->path); That gets annoying when do_something() is non-trivial, and we've resorted to gross hacks like creating fake alternates (see find_short_object_filename()). Instead, let's give each raw_object_store a unified list of object_directory structs. The first will be the main store, and everything after is an alternate. Very few callers even care about the distinction, and can just loop over the whole list (and those who care can just treat the first element differently). A few observations: - we don't need r->objects->objectdir anymore, and can just mechanically convert that to r->objects->odb->path - object_directory's path field needs to become a real pointer rather than a FLEX_ARRAY, in order to fill it with expand_base_dir() - we'll call prepare_alt_odb() earlier in many functions (i.e., outside of the loop). This may result in us calling it even when our function would be satisfied looking only at the main odb. But this doesn't matter in practice. It's not a very expensive operation in the first place, and in the majority of cases it will be a noop. We call it already (and cache its results) in prepare_packed_git(), and we'll generally check packs before loose objects. So essentially every program is going to call it immediately once per program. Arguably we should just prepare_alt_odb() immediately upon setting up the repository's object directory, which would save us sprinkling calls throughout the code base (and forgetting to do so has been a source of subtle bugs in the past). But I've stopped short of that here, since there are already a lot of other moving parts in this patch. - Most call sites just get shorter. The check_and_freshen() functions are an exception, because they have entry points to handle local and nonlocal directories separately. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-11-12 14:50:39 +00:00
while (o->odb) {
struct object_directory *next;
sha1-file: use an object_directory for the main object dir Our handling of alternate object directories is needlessly different from the main object directory. As a result, many places in the code basically look like this: do_something(r->objects->objdir); for (odb = r->objects->alt_odb_list; odb; odb = odb->next) do_something(odb->path); That gets annoying when do_something() is non-trivial, and we've resorted to gross hacks like creating fake alternates (see find_short_object_filename()). Instead, let's give each raw_object_store a unified list of object_directory structs. The first will be the main store, and everything after is an alternate. Very few callers even care about the distinction, and can just loop over the whole list (and those who care can just treat the first element differently). A few observations: - we don't need r->objects->objectdir anymore, and can just mechanically convert that to r->objects->odb->path - object_directory's path field needs to become a real pointer rather than a FLEX_ARRAY, in order to fill it with expand_base_dir() - we'll call prepare_alt_odb() earlier in many functions (i.e., outside of the loop). This may result in us calling it even when our function would be satisfied looking only at the main odb. But this doesn't matter in practice. It's not a very expensive operation in the first place, and in the majority of cases it will be a noop. We call it already (and cache its results) in prepare_packed_git(), and we'll generally check packs before loose objects. So essentially every program is going to call it immediately once per program. Arguably we should just prepare_alt_odb() immediately upon setting up the repository's object directory, which would save us sprinkling calls throughout the code base (and forgetting to do so has been a source of subtle bugs in the past). But I've stopped short of that here, since there are already a lot of other moving parts in this patch. - Most call sites just get shorter. The check_and_freshen() functions are an exception, because they have entry points to handle local and nonlocal directories separately. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-11-12 14:50:39 +00:00
next = o->odb->next;
free_object_directory(o->odb);
o->odb = next;
}
kh_destroy_odb_path_map(o->odb_by_path);
o->odb_by_path = NULL;
}
void raw_object_store_clear(struct raw_object_store *o)
{
FREE_AND_NULL(o->alternate_db);
oidmap_free(o->replace_map, 1);
FREE_AND_NULL(o->replace_map);
replace-object: make replace operations thread-safe replace-object functions are very close to being thread-safe: the only current racy section is the lazy initialization at prepare_replace_object(). The following patches will protect some object reading operations to be called threaded, but before that, replace functions must be protected. To do so, add a mutex to struct raw_object_store and acquire it before lazy initializing the replace_map. This won't cause any noticeable performance drop as the mutex will no longer be used after the replace_map is initialized. Later, when the replace functions are called in parallel, thread debuggers might point our use of the added replace_map_initialized flag as a data race. However, as this boolean variable is initialized as false and it's only updated once, there's no real harm. It's perfectly fine if the value is updated right after a thread read it in replace-map.h:lookup_replace_object() (there'll only be a performance penalty for the affected threads at that moment). We could cease the debugger warning protecting the variable reading at the said function. However, this would negatively affect performance for all threads calling it, at any time, so it's not really worthy since the warning doesn't represent a real problem. Instead, to make sure we don't get false positives (at ThreadSanitizer, at least) an entry for the respective function is added to .tsan-suppressions. Signed-off-by: Matheus Tavares <matheus.bernardino@usp.br> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2020-01-16 02:39:52 +00:00
pthread_mutex_destroy(&o->replace_mutex);
free_commit_graph(o->commit_graph);
o->commit_graph = NULL;
o->commit_graph_attempted = 0;
sha1-file: use an object_directory for the main object dir Our handling of alternate object directories is needlessly different from the main object directory. As a result, many places in the code basically look like this: do_something(r->objects->objdir); for (odb = r->objects->alt_odb_list; odb; odb = odb->next) do_something(odb->path); That gets annoying when do_something() is non-trivial, and we've resorted to gross hacks like creating fake alternates (see find_short_object_filename()). Instead, let's give each raw_object_store a unified list of object_directory structs. The first will be the main store, and everything after is an alternate. Very few callers even care about the distinction, and can just loop over the whole list (and those who care can just treat the first element differently). A few observations: - we don't need r->objects->objectdir anymore, and can just mechanically convert that to r->objects->odb->path - object_directory's path field needs to become a real pointer rather than a FLEX_ARRAY, in order to fill it with expand_base_dir() - we'll call prepare_alt_odb() earlier in many functions (i.e., outside of the loop). This may result in us calling it even when our function would be satisfied looking only at the main odb. But this doesn't matter in practice. It's not a very expensive operation in the first place, and in the majority of cases it will be a noop. We call it already (and cache its results) in prepare_packed_git(), and we'll generally check packs before loose objects. So essentially every program is going to call it immediately once per program. Arguably we should just prepare_alt_odb() immediately upon setting up the repository's object directory, which would save us sprinkling calls throughout the code base (and forgetting to do so has been a source of subtle bugs in the past). But I've stopped short of that here, since there are already a lot of other moving parts in this patch. - Most call sites just get shorter. The check_and_freshen() functions are an exception, because they have entry points to handle local and nonlocal directories separately. Signed-off-by: Jeff King <peff@peff.net> Signed-off-by: Junio C Hamano <gitster@pobox.com>
2018-11-12 14:50:39 +00:00
free_object_directories(o);
o->odb_tail = NULL;
o->loaded_alternates = 0;
INIT_LIST_HEAD(&o->packed_git_mru);
close_object_store(o);
o->packed_git = NULL;
hashmap_clear(&o->pack_map);
}
void parsed_object_pool_clear(struct parsed_object_pool *o)
{
/*
* As objects are allocated in slabs (see alloc.c), we do
* not need to free each object, but each slab instead.
*
* Before doing so, we need to free any additional memory
* the objects may hold.
*/
unsigned i;
for (i = 0; i < o->obj_hash_size; i++) {
struct object *obj = o->obj_hash[i];
if (!obj)
continue;
if (obj->type == OBJ_TREE)
free_tree_buffer((struct tree*)obj);
else if (obj->type == OBJ_COMMIT)
release_commit_memory(o, (struct commit*)obj);
else if (obj->type == OBJ_TAG)
release_tag_memory((struct tag*)obj);
}
FREE_AND_NULL(o->obj_hash);
o->obj_hash_size = 0;
free_commit_buffer_slab(o->buffer_slab);
o->buffer_slab = NULL;
clear_alloc_state(o->blob_state);
clear_alloc_state(o->tree_state);
clear_alloc_state(o->commit_state);
clear_alloc_state(o->tag_state);
clear_alloc_state(o->object_state);
stat_validity_clear(o->shallow_stat);
FREE_AND_NULL(o->blob_state);
FREE_AND_NULL(o->tree_state);
FREE_AND_NULL(o->commit_state);
FREE_AND_NULL(o->tag_state);
FREE_AND_NULL(o->object_state);
FREE_AND_NULL(o->shallow_stat);
}