Deal with kmalloc vs ksize vs __alloc_size

[kees]

There needs to be a way to deal with __alloc_size not matching ksize().

https://lore.kernel.org/lkml/202202281516.19274C0@keescook/

Plan: remove side-effect from ksize() and refactor all users to either use krealloc() afterwards, or kmalloc_roundup_size() before.

in v6.1:

• [x] commit 05a940656e1e ("slab: Introduce kmalloc_size_roundup()")

in -next:

• [x] commit 36875a063b5e ("net: ipa: Proactively round up to kmalloc bucket size")
• [x] commit ab3f7828c979 ("openvswitch: Use kmalloc_size_roundup() to match ksize() usage")
• [x] commit d6dd508080a3 ("bnx2: Use kmalloc_size_roundup() to match ksize() usage")
• [x] commit 12d6c1d3a2ad ("skbuff: Proactively round up to kmalloc bucket size")
• [x] mempool: Do not use ksize() for poisoning https://lore.kernel.org/lkml/166693938482.29415.7034851115705424459.tip-bot2@tip-bot2/
• [x] dma-buf: Proactively round up to kmalloc bucket size https://lore.kernel.org/lkml/20221018090858.never.941-kees@kernel.org/
• [x] btrfs: send: Proactively round up to kmalloc bucket size https://lore.kernel.org/lkml/20220923202822.2667581-8-keescook@chromium.org/
• [x] coredump: Proactively round up to kmalloc bucket size https://lore.kernel.org/lkml/20221018090701.never.996-kees@kernel.org/
• [x] devres: Use kmalloc_size_roundup() to match ksize() usage https://lore.kernel.org/lkml/20221018090406.never.856-kees@kernel.org/
• [x] igb: Proactively round up to kmalloc bucket size https://lore.kernel.org/lkml/20221018092340.never.556-kees@kernel.org/
• [x] bpf: Use kmalloc_size_roundup() to match ksize() usage https://lore.kernel.org/linux-hardening/20221029024444.gonna.633-kees@kernel.org/

partially reviewed:

unreviewed:

finally:

• [x] mm: Make ksize() a reporting-only function https://lore.kernel.org/lkml/20221022180455.never.023-kees@kernel.org

[kees]

Idea from keithp: create a wrapper func to re-assign __alloc_size() as if we went through realloc(). For example:

static inline __alloc_size(2) void *__resize_kmalloc(void *objp, size)
{
    return objp;
}
#define ksize(objp) ({                          \
    /*                              \
     * Getting the actual allocation size means the __alloc_size    \
     * hints are no longer valid, and the compiler needs to     \
     * learn the new one.                       \
     */                             \
        size_t __new_kmalloc_size;                                  \
        __new_kmalloc_size = _ksize(objp);                              \
        objp = __resize_kmalloc(objp, __new_kmalloc_size);              \
    __new_kmalloc_size;                         \
})

[kees]

See https://github.com/ClangBuiltLinux/linux/issues/1599

[kees]

The wrapper won't work because GCC ignores __alloc_size on inlines. 😠

So, introduce kmalloc_size_roundup() and remove the side-effect from ksize().