Re: [PATCH 02/13] list-objects-filter-map: extend oidmap to collect omitted objects

[Jeff Hostetler <git@jeffhostetler.com>]

On 10/25/2017 3:10 AM, Junio C Hamano wrote:
> Jeff Hostetler <git@jeffhostetler.com> writes:
> 
>> From: Jeff Hostetler <jeffhost@microsoft.com>
>>
>> Create helper class to extend oidmap to collect a list of
>> omitted or missing objects during traversal.
> 
> The reason why oidmap itself cannot be used is because the code
> wants to record not just the object name but something else about
> the object.  And attributes that the code may care about we can see
> in this patch are the object type and the path it found.

I recently simplified the code in this version to not completely
sub-class oidmap, but to just use it along with a custom
_insert method that takes care of allocating the _entry
data.  I should update the commit message to reflect that.

> 
> Is the plan to extend this set of attributes over time as different
> "omitter"s are added?  Why was "path" chosen as a member of the
> initial set and how it will be useful (also, what path would we
> record for tags and commits)?

I envisioned this to let rev-list print the pathname of omitted
objects -- like "rev-list --objects" does for regular blobs.
I would leave the pathname NULL for tags and commits.

The pathname helps with debugging and testing, but also is
used by the sparse filter to avoid some expensive duplicate
is-excluded lookups.

Currently the 3 filters I have defined all use the same extra
data.  I suppose a future filter could want additional fields,
so maybe it would be better to refactor my "map-entry" to be
per-filter specific.

> 
> These "future plans" needs revealed upfront, instead of (or in
> addition to) "will be used in a later commit".  As it is hard to
> judge if "filter map" is an appropriate name for this thing without
> knowing _how_ it is envisioned to be used.  "filter map" sounds more
> like a map function that is consulted when we decide if we want to
> drop the object, but from the looks of the code, it is used more to
> record what was done to these objects.

Sorry, I meant a later commit in this patch series.  It is used by
commits 4, 5, 6, and 10 to actually do the filtering and collect a
list of omitted or missing objects.

> 
> Is it really a "map" (i.e. whose primary focus is to find out what
> an object name is "mapped to" when we get an object name---e.g. we
> notice an otherwise connected object is missing, and consult this
> "map" to learn what the type/path is because we want to do X)?  Or
> is it more like a "set of known-to-be-missing object" (i.e. whose
> primary point is to serve as a set of object names and what a name
> maps to is primarily for debugging)?  These are easier to answer if
> we know how it will be used.

I think of a "set" as a member? or not-member? class.
I think of a "map" as a member? or not-member? class but where each
member also has a value.  Sometimes map lookups just want to know
membership and sometimes the lookup wants the value.

Granted, having the key and value data stuffed into the same entry
(from hashmap's point of view, rather than a key having a pointer
to a value) does kind of blur the line, but I was thinking about
a map here.  (And I was building on oidmap which builds on hashmap,
so it seemed appropriate.)

> 
>> This will be used in a later commit by the list-object filtering
>> code.
>>
>> Signed-off-by: Jeff Hostetler <jeffhost@microsoft.com>
>> ---
>> diff --git a/list-objects-filter-map.c b/list-objects-filter-map.c
>> new file mode 100644
>> index 0000000..7e496b3
>> --- /dev/null
>> +++ b/list-objects-filter-map.c
>> @@ -0,0 +1,63 @@
>> +#include "cache.h"
>> +#include "list-objects-filter-map.h"
>> +
>> +int list_objects_filter_map_insert(struct oidmap *map,
>> +				   const struct object_id *oid,
>> +				   const char *pathname, enum object_type type)
>> +{
>> +	size_t len, size;
>> +	struct list_objects_filter_map_entry *e;
>> +
>> +	if (oidmap_get(map, oid))
>> +		return 1;
> 
> It is OK for the existing entry to record a path that is totally
> different from what the caller has.  It is hard to judge without
> knowing what pathname the callers are expected to call this function
> with, but I am guessing that it is similar to the path shown in the
> output from "rev-list --objects"---and if that is the case, it is
> correct that the same object may be reached at different paths
> depending on what tree the traversal begins at, so pathname recorded
> in the map is merely "there is one tree somewhere that has this
> object at this path".

Right, the first observed pathname is as good as any.

> 
> For that matter, the caller may have a completely different type
> from the object we saw earlier; not checking and flagging it as a
> possible error makes me feel somewhat uneasy, but there probably is
> little you can do at this layer of the code if you noticed such a
> discrepancy so it may be OK to punt.

I could assert() that the types match, but right there's not much
we can do about it at this layer.

> 
>> +	len = ((pathname && *pathname) ? strlen(pathname) : 0);
>> +	size = (offsetof(struct list_objects_filter_map_entry, pathname) + len + 1);
>> +	e = xcalloc(1, size);
>> +
>> +	oidcpy(&e->entry.oid, oid);
>> +	e->type = type;
>> +	if (pathname && *pathname)
>> +		strcpy(e->pathname, pathname);
>> +
>> +	oidmap_put(map, e);
>> +	return 0;
>> +}
> 
> The return value from the function needs to be documented in the
> header to help callers.  It is not apparent why "we did already have
> one" and "we now newly added" is interesting to the callers, for
> example.  An obvious alternative implementation of this function
> would return the pointer to an entry that records the object id
> (i.e. either the one that was already there, or the one we created
> because we saw this object for the first time), so that the caller
> can do something interesting to it---again, because the reason why
> we want this "filter map" is not explained at this stage, it is hard
> to tell what that "sometehing interesting" would be.

good point.  thanks.

> 
>> +static int my_cmp(const void *a, const void *b)
>> +{
>> +	const struct oidmap_entry *ea, *eb;
>> +
>> +	ea = *(const struct oidmap_entry **)a;
>> +	eb = *(const struct oidmap_entry **)b;
>> +
>> +	return oidcmp(&ea->oid, &eb->oid);
>> +}
>> +
>> +void list_objects_filter_map_foreach(struct oidmap *map,
>> +				     list_objects_filter_map_foreach_cb cb,
> 
> Name a typedef of a function as something_fn, not something_cb;
> something_cb is often the type of a struct to be fed to the callback
> function.  And call such a parameter of type something_fn just fn.
> 

ok.


>> +				     void *cb_data)
>> +{
>> +	struct hashmap_iter iter;
>> +	struct list_objects_filter_map_entry **array;
>> +	struct list_objects_filter_map_entry *e;
>> +	int k, nr;
>> +
>> +	nr = hashmap_get_size(&map->map);
>> +	if (!nr)
>> +		return;
>> +
>> +	array = xcalloc(nr, sizeof(*e));
>> +
>> +	k = 0;
>> +	hashmap_iter_init(&map->map, &iter);
>> +	while ((e = hashmap_iter_next(&iter)))
>> +		array[k++] = e;
>> +
>> +	QSORT(array, nr, my_cmp);
> 
> It is entirely unclear why foreach() must return the object names in
> order.
> 
>> +	for (k = 0; k < nr; k++)
>> +		cb(k, nr, array[k], cb_data);
> 
> Also it is not clear if you wanted to expose the type of the
> entry to the callback function.
> 
> An obvious alternative
> 
> 	fn(&array[k].entry.oid, cb_data);
> 
> would allow you to keep the type of map-entry private to the map,
> and also the callback does not need to know about k or nr.
> 
> I guess you are giving k and nr to allow the callers to do a
> progress bar?  If that is the case, that's fine by me.  I still do
> not understand why we uncondtionally sort, though.

The thought was that we would sort the OIDs so that things
like rev-list could print the omitted/missing objects in OID
order.  Not critical that we do it here, but I thought it would
help callers.

I included the {k, nr} so that the callback could dump header/trailer
information when reporting the results or pre-allocate an array.
I'll look at refactoring this -- I never quite liked how it turned
out anyway -- especially with the oidmap simplifications.

>> +
>> +	free(array);
>> +}
>> diff --git a/list-objects-filter-map.h b/list-objects-filter-map.h
>> new file mode 100644
>> index 0000000..794fc81
>> --- /dev/null
>> +++ b/list-objects-filter-map.h
>> @@ -0,0 +1,26 @@
>> +#ifndef LIST_OBJECTS_FILTER_MAP_H
>> +#define LIST_OBJECTS_FILTER_MAP_H
>> +
>> +#include "oidmap.h"
>> +
>> +struct list_objects_filter_map_entry {
>> +	struct oidmap_entry entry; /* must be first */
>> +	enum object_type type;
>> +	char pathname[FLEX_ARRAY];
>> +};
>> +
>> +extern int list_objects_filter_map_insert(
>> +	struct oidmap *map,
>> +	const struct object_id *oid,
>> +	const char *pathname, enum object_type type);
>> +
>> +typedef void (*list_objects_filter_map_foreach_cb)(
>> +	int i, int i_limit,
>> +	struct list_objects_filter_map_entry *e, void *cb_data);
>> +
>> +extern void list_objects_filter_map_foreach(
>> +	struct oidmap *map,
>> +	list_objects_filter_map_foreach_cb cb,
>> +	void *cb_data);
>> +
>> +#endif /* LIST_OBJECTS_FILTER_MAP_H */