Calling update with memmapped data can create invalid data in memmap views

[braingram]

If an update causes a memmapped block to be moved, any views taken of the array that uses the block will point to invalid data.

Running the following example code (that adds a large amount of text to the tree to cause the memmapped blocks to be moved):

import asdf
import numpy


fn = 'test.asdf'
a = numpy.zeros(10, dtype='uint8')
b = numpy.ones(10, dtype='uint8')

af = asdf.AsdfFile({'a': a, 'b': b})
af.write_to(fn)

with asdf.open(fn, mode='rw') as af:
    va = af['a'][:3]
    print(f"A: {af['a'][:]}")
    print(f"b: {af['b'][:]}")
    print(f"va: {va}")
    af['c'] = 'a' * 10000
    af.update()
    print(f"A: {af['a'][:]}")
    print(f"b: {af['b'][:]}")
    print(f"va: {va}")

Produces the following output which shows that the view of 'a', here 'va', is incorrect following the update.

A: [0 0 0 0 0 0 0 0 0 0]
b: [1 1 1 1 1 1 1 1 1 1]
va: [0 0 0]
A: [0 0 0 0 0 0 0 0 0 0]
b: [1 1 1 1 1 1 1 1 1 1]
va: [97 97 97]

[braingram]

Removal of the source array results in a segfault when the view is accessed.

import asdf
import numpy


fn = 'test.asdf'
a = numpy.zeros(10, dtype='uint8')
b = numpy.ones(10, dtype='uint8')

af = asdf.AsdfFile({'a': a, 'b': b})
af.write_to(fn)

with asdf.open(fn, mode='rw') as af:
    va = af['a'][:3]
    print(f"A: {af['a'][:]}")
    print(f"b: {af['b'][:]}")
    print(f"va: {va}")
    af['c'] = 'a' * 10000
    del af.tree['a']
    del af.tree['b']
    af.update()
    print(f"va: {va}")

results in

A: [0 0 0 0 0 0 0 0 0 0]
b: [1 1 1 1 1 1 1 1 1 1]
va: [0 0 0]
Segmentation fault: 11