On Wed, Feb 26, 2025 at 08:00:16PM +0000, Eric Biggers wrote:

On Wed, Feb 26, 2025 at 06:56:25PM +0000, Yosry Ahmed wrote:

quoted

Currently, zswap_cpu_comp_dead() calls crypto_free_acomp() while holding
the per-CPU acomp_ctx mutex. crypto_free_acomp() then holds scomp_lock
(through crypto_exit_scomp_ops_async()).

On the other hand, crypto_alloc_acomp_node() holds the scomp_lock
(through crypto_scomp_init_tfm()), and then allocates memory.
If the allocation results in reclaim, we may attempt to hold the per-CPU
acomp_ctx mutex.

The bug is in acomp.  crypto_free_acomp() should never have to wait for a memory
allocation.  That is what needs to be fixed.

crypto_free_acomp() does not explicitly wait for an allocation, but it
waits for scomp_lock (in crypto_exit_scomp_ops_async()), which may be
held while allocating memory from crypto_scomp_init_tfm().

Are you suggesting that crypto_exit_scomp_ops_async() should not be
holding scomp_lock?

But really the bounce buffering in acomp (which is what is causing this problem)
should not exist at all.  There is really no practical use case for it; it's
just there because of the Crypto API's insistence on shoehorning everything into
scatterlists for no reason...

I am assuming this about scomp_scratch logic, which is what we need to
hold the scomp_lock for, resulting in this problem.

If this is something that can be done right away I am fine with dropping
this patch for an alternative fix, although it may be nice to reduce the
lock critical section in zswap_cpu_comp_dead() to the bare minimum
anyway.