On Wed, Dec 27, 2023 at 5:16 PM Chengming Zhou
<zhouchengming@xxxxxxxxxxxxx> wrote:
>
> On 2023/12/27 15:01, Barry Song wrote:
> > On Wed, Dec 27, 2023 at 7:38 PM Chengming Zhou
> > <zhouchengming@xxxxxxxxxxxxx> wrote:
> >>
> >> On 2023/12/27 14:25, Barry Song wrote:
> >>> On Wed, Dec 27, 2023 at 4:51 PM Chengming Zhou
> >>> <zhouchengming@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>> On 2023/12/27 08:23, Nhat Pham wrote:
> >>>>> On Tue, Dec 26, 2023 at 3:30 PM Chris Li <chrisl@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> Again, sorry I was looking at the decompression side rather than the
> >>>>>> compression side. The compression side does not even offer a safe
> >>>>>> version of the compression function.
> >>>>>> That seems to be dangerous. It seems for now we should make the zswap
> >>>>>> roll back to 2 page buffer until we have a safe way to do compression
> >>>>>> without overwriting the output buffers.
> >>>>>
> >>>>> Unfortunately, I think this is the way - at least until we rework the
> >>>>> crypto/compression API (if that's even possible?).
> >>>>> I still think the 2 page buffer is dumb, but it is what it is :(
> >>>>
> >>>> Hi,
> >>>>
> >>>> I think it's a bug in `scomp_acomp_comp_decomp()`, which doesn't use
> >>>> the caller passed "src" and "dst" scatterlist. Instead, it uses its own
> >>>> per-cpu "scomp_scratch", which have 128KB src and dst.
> >>>>
> >>>> When compression done, it uses the output req->dlen to copy scomp_scratch->dst
> >>>> to our dstmem, which has only one page now, so this problem happened.
> >>>>
> >>>> I still don't know why the alg->compress(src, slen, dst, &dlen) doesn't
> >>>> check the dlen? It seems an obvious bug, right?
> >>>>
> >>>> As for this problem in `scomp_acomp_comp_decomp()`, this patch below
> >>>> should fix it. I will set up a few tests to check later.
> >>>>
> >>>> Thanks!
> >>>>
> >>>> diff --git a/crypto/scompress.c b/crypto/scompress.c
> >>>> index 442a82c9de7d..e654a120ae5a 100644
> >>>> --- a/crypto/scompress.c
> >>>> +++ b/crypto/scompress.c
> >>>> @@ -117,6 +117,7 @@ static int scomp_acomp_comp_decomp(struct acomp_req *req, int dir)
> >>>> struct crypto_scomp *scomp = *tfm_ctx;
> >>>> void **ctx = acomp_request_ctx(req);
> >>>> struct scomp_scratch *scratch;
> >>>> + unsigned int dlen;
> >>>> int ret;
> >>>>
> >>>> if (!req->src || !req->slen || req->slen > SCOMP_SCRATCH_SIZE)
> >>>> @@ -128,6 +129,8 @@ static int scomp_acomp_comp_decomp(struct acomp_req *req, int dir)
> >>>> if (!req->dlen || req->dlen > SCOMP_SCRATCH_SIZE)
> >>>> req->dlen = SCOMP_SCRATCH_SIZE;
> >>>>
> >>>> + dlen = req->dlen;
> >>>> +
> >>>> scratch = raw_cpu_ptr(&scomp_scratch);
> >>>> spin_lock(&scratch->lock);
> >>>>
> >>>> @@ -145,6 +148,9 @@ static int scomp_acomp_comp_decomp(struct acomp_req *req, int dir)
> >>>> ret = -ENOMEM;
> >>>> goto out;
> >>>> }
> >>>> + } else if (req->dlen > dlen) {
> >>>> + ret = -ENOMEM;
> >>>> + goto out;
> >>>> }
> >>>
> >>> This can't fix the problem as crypto_scomp_compress() has written overflow data.
> >>
> >> No, crypto_scomp_compress() writes to its own scomp_scratch->dst memory, then copy
> >> to our dstmem.
> >>
> >>>
> >>> BTW, in many cases, hardware-accelerators drivers/crypto can do compression and
> >>> decompression by off-loading CPU;
> >>> we won't have a chance to let hardware check the dst buffer size. so
> >>> giving the dst buffer
> >>> with enough length to the hardware's dma engine is the right way. I
> >>> mean, we shouldn't
> >>> change dst from 2pages to 1page.
> >>
> >> But how do we know 2 pages is enough for any compression algorithm?
> >>
> >
> > There is no guarrette 2 pages is enough.
> >
> > We can invent our own compression algorithm, in our algorithm, we can
> > add a lot of metadata, for example, longer than 4KB when the source data
> > is uncompress-able. then dst can be larger than 2 * PAGE_SIZE. but this
> > is not the case :-) This kind of algorithm may never succeed.
> >
> > For those in-tree algorithms, we have a WORST_COMPR_FACTOR. in
> > ubifs, zram and zswap, we all use "2" as the worst comp factor.
>
> Thanks for your explanation! Maybe it's best for us to return to 2 pages
> if no other people's comments. And this really need more documentation :-)
I agree. we need some doc.
besides, i actually think we can skip zswap frontend if
over-compression is really
happening.
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -1318,7 +1318,7 @@ bool zswap_store(struct folio *folio)
if (zpool_malloc_support_movable(zpool))
gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
ret = zpool_malloc(zpool, dlen, gfp, &handle);
- if (ret == -ENOSPC) {
+ if (ret == -ENOSPC || dlen > PAGE_SIZE) {
zswap_reject_compress_poor++;
goto put_dstmem;
}
since we are not saving but wasting space in this corner case?
> since there is no any comment or check in the acomp compress interface.
>
> /*
> * @src: Source Data
> * @dst: Destination data
> * @slen: Size of the input buffer
> * @dlen: Size of the output buffer and number of bytes produced
> * @flags: Internal flags
> * @__ctx: Start of private context data
> */
> struct acomp_req {
> struct crypto_async_request base;
> struct scatterlist *src;
> struct scatterlist *dst;
> unsigned int slen;
> unsigned int dlen;
> u32 flags;
> void *__ctx[] CRYPTO_MINALIGN_ATTR;
> };
>
> >
> > /*
> > * Some compressors, like LZO, may end up with more data then the input buffer.
> > * So UBIFS always allocates larger output buffer, to be sure the compressor
> > * will not corrupt memory in case of worst case compression.
> > */
> > #define WORST_COMPR_FACTOR 2
> >
> > #ifdef CONFIG_FS_ENCRYPTION
> > #define UBIFS_CIPHER_BLOCK_SIZE FSCRYPT_CONTENTS_ALIGNMENT
> > #else
> > #define UBIFS_CIPHER_BLOCK_SIZE 0
> > #endif
> >
> > /*
> > * How much memory is needed for a buffer where we compress a data node.
> > */
> > #define COMPRESSED_DATA_NODE_BUF_SZ \
> > (UBIFS_DATA_NODE_SZ + UBIFS_BLOCK_SIZE * WORST_COMPR_FACTOR)
> >
> >
> > For years, we have never seen 2 pages that can be a problem. but 1
> > page is definitely
> > not enough, I remember I once saw many cases where accelerators' dmaengine
> > can write more than 1 page.
> >
> >> Thanks.
> >>
> >>>
> >>>> scatterwalk_map_and_copy(scratch->dst, req->dst, 0, req->dlen,
> >>>> 1);
> >>>
> >>>
>
Thanks
Barry
