wg-backend-django/acer-env/lib/python3.10/site-packages/django/template/loaders/cached.py
2022-11-30 15:58:16 +07:00

101 lines
3.6 KiB
Python

"""
Wrapper class that takes a list of template loaders as an argument and attempts
to load templates from them in order, caching the result.
"""
import hashlib
from django.template import TemplateDoesNotExist
from django.template.backends.django import copy_exception
from .base import Loader as BaseLoader
class Loader(BaseLoader):
def __init__(self, engine, loaders):
self.get_template_cache = {}
self.loaders = engine.get_template_loaders(loaders)
super().__init__(engine)
def get_dirs(self):
for loader in self.loaders:
if hasattr(loader, "get_dirs"):
yield from loader.get_dirs()
def get_contents(self, origin):
return origin.loader.get_contents(origin)
def get_template(self, template_name, skip=None):
"""
Perform the caching that gives this loader its name. Often many of the
templates attempted will be missing, so memory use is of concern here.
To keep it in check, caching behavior is a little complicated when a
template is not found. See ticket #26306 for more details.
With template debugging disabled, cache the TemplateDoesNotExist class
for every missing template and raise a new instance of it after
fetching it from the cache.
With template debugging enabled, a unique TemplateDoesNotExist object
is cached for each missing template to preserve debug data. When
raising an exception, Python sets __traceback__, __context__, and
__cause__ attributes on it. Those attributes can contain references to
all sorts of objects up the call chain and caching them creates a
memory leak. Thus, unraised copies of the exceptions are cached and
copies of those copies are raised after they're fetched from the cache.
"""
key = self.cache_key(template_name, skip)
cached = self.get_template_cache.get(key)
if cached:
if isinstance(cached, type) and issubclass(cached, TemplateDoesNotExist):
raise cached(template_name)
elif isinstance(cached, TemplateDoesNotExist):
raise copy_exception(cached)
return cached
try:
template = super().get_template(template_name, skip)
except TemplateDoesNotExist as e:
self.get_template_cache[key] = (
copy_exception(e) if self.engine.debug else TemplateDoesNotExist
)
raise
else:
self.get_template_cache[key] = template
return template
def get_template_sources(self, template_name):
for loader in self.loaders:
yield from loader.get_template_sources(template_name)
def cache_key(self, template_name, skip=None):
"""
Generate a cache key for the template name and skip.
If skip is provided, only origins that match template_name are included
in the cache key. This ensures each template is only parsed and cached
once if contained in different extend chains like:
x -> a -> a
y -> a -> a
z -> a -> a
"""
skip_prefix = ""
if skip:
matching = [
origin.name for origin in skip if origin.template_name == template_name
]
if matching:
skip_prefix = self.generate_hash(matching)
return "-".join(s for s in (str(template_name), skip_prefix) if s)
def generate_hash(self, values):
return hashlib.sha1("|".join(values).encode()).hexdigest()
def reset(self):
"Empty the template cache."
self.get_template_cache.clear()