Not sure it works with readonly X. Can you test this against a read-only csr matrix (Xcsr.data.setflags(write=False)) ?

Can confirm it works with read-only CSR matrices:

from scipy import sparse
from sklearn import linear_model
import numpy as np

clf = linear_model.ElasticNet(alpha=0.1)
X = sparse.random(100,10)
X.data.setflags(write=False)
y = np.random.random(100)
clf.fit(X, y)

By default ElasticNet does a copy of the input. Also, my first comment was misleading, it will also convert if it's not in csc format. If I change your snippet to

from scipy import sparse
from sklearn import linear_model
import numpy as np

clf = linear_model.ElasticNet(alpha=0.1, copy_X=False)
X = sparse.random(100,10, format="csc")
X.data.setflags(write=False)
y = np.random.random(100)
clf.fit(X, y)

Then I get ValueError: buffer source array is read-only

Ah gotcha. In that case, since sparse_enet_coordinate_descent doesn't actually modify X_data, I went ahead and implemented the fix suggested in #10624, which is the ReadonlyArrayWrapper added in #20903. I think this should work, but I'm still quite new to the Cython side of things, so please let me know if there are problems with this approach -- thanks :)

Looks good. I guess we can do the same for enet_coordinate_descent because currently we have the same error for the dense case:

from scipy import sparse
from sklearn import linear_model
import numpy as np

clf = linear_model.ElasticNet(alpha=0.1, copy_X=False)
X = np.asfortranarray(np.random.uniform(size=(100, 10)))
X.setflags(write=False)
y = np.random.random(100)
clf.fit(X, y)

and probably also enet_coordinate_descent_multi_task. It would be good to also add a test for these read-only cases because it doesn't seem tested currently

Can indices or indptr be read only?

I think testing is a must here. If we pass actual read only data with ReadOnlyArrayWrapper, python crashes if the Cython function writes to the memoryview:

%%cython
from cython cimport floating

def f(floating[:] a):
    a[0] = 1

from sklearn.utils._readonly_array_wrapper import ReadonlyArrayWrapper
from sklearn.utils._testing import create_memmap_backed_data
import numpy as np

X = np.asarray([1, 2.3])
X_mapped = create_memmap_backed_data(X)

# crashes
f(ReadonlyArrayWrapper(X_mapped))

Added a test for this behavior and confirmed the algorithm doesn't modify indices or indptr so extended the wrapper to them as well.