Removal of WeakMap/WeakSet clear
Mark S. Miller
erights at google.com
Wed Nov 26 11:21:26 PST 2014
On Wed, Nov 26, 2014 at 9:33 AM, Katelyn Gadd <kg at luminance.org> wrote:
> Is there a detailed rationale for this somewhere?
It is a combination of three issues.
1. The security issue.
2. The implementation issue.
3. The ES process issue.
The implementation issue is that the use cases for WeakMaps basically
divide into the following cases:
a. Those for which we're confident that the map's lifetime outlives
its typical key.
b. Those for which we're confident that the key's lifetime outlives
the typical map it is a key in.
c. Those for which we're not confident which typically lives longer.
For #a, the transposed representation of WeakMaps is strictly better.
The non-transposed implementation would promote ephemeral keys to
later GC generations, which is very expensive. (I believe the expense
of this promotion has been radically underestimated in most prior
discussions.) This is the GC pressure that really matters.
For #b, just use a Map rather than a WeakMap.
For #c, transposed rep or not is a tie. In the interests of minimizing
implementation mechanism, we should just use a transposed rep for this
Given the transposed representation, the only implementation
techniques for clear are
x. Enumerating all memory
y. Having the WeakMap encapsulate the token used to lookup the value
in the key's hidden map, and have .clear replace this token.
#x is not reasonable.
#y is equivalent to the replacing of the Map that would be the
user-level technique for emulating clear in any case. This is exactly
what the use of WeakMaps for membranes do, when the membrane should be
revoked. If .clear could be implemented for the transposed
representation more efficiently than this, membranes would benefit
from .clear as well. I have never expected they could.
The process issue is that .clear was not in the original proposal (for
all of these reasons), and it never achieved consensus in committee.
It would have violated our process to keep it in the spec. The process
issue is not "Should it be dropped?" since it was never legitimately
added. The only issue is "Should it be added?".
> Making typical
> applications pay the cost here for a specific security scenario seems
> really bizarre to me. Clearing standard library data structures is an
> incredibly common operation. If you want to ensure that someone can't
> clear the map/set, shouldn't you be handing them an encapsulated
> version of the data structure? This seems like a corner case that
> shouldn't justify removing an important primitive.
> If you have a clear method, the security problem seems solved by
> wrapping it in an object or using a proxy to deny the ability to clear
> (you hide the actual map/set, so it can't be cleared - you expose only
> the operations you want to expose).
> If you don't have a clear method, anyone wanting to clear the data
> structure has to throw it away and allocate a new one. This has
> significant disadvantages:
> The new structure starts empty at a default size, so repopulating it
> will have to grow the buffer multiple times - this is undesirable for
> cases where you are reusing a single data structure to store state for
> a long-running application.
> The allocation adds to GC and memory pressure for a long-running
> application that needs to clear data structures frequently. Were it a
> lightweight data type this would matter less, but a typical map
> instance with data in it can occupy a considerable amount of space in
> the heap.
> Being able to clear the structure now requires that all consumers have
> support for replacing their reference(s) to the old map with the new
> one. This makes it harder to maintain encapsulation because you may
> have saved a reference to the map in a private property or within a
> closure. Now you need to add accessibility points to everything that
> might retain the map so that you can update the reference. Or, you
> have to encapsulate maps and sets just to recreate the clear operation
> that should have been there to begin with.
> In either case, encapsulation or shielding the container behind a
> proxy is necessary. I insist that the common case is the one that
> shouldn't have to encapsulate, because optimizing for that case will
> benefit the vast majority of web applications that use it and the
> penalty to security-sensitive cases is small.
> es-discuss mailing list
> es-discuss at mozilla.org
More information about the es-discuss