On Fri, Feb 16, 2024 at 5:07 PM Chengming Zhou
<zhouchengming@xxxxxxxxxxxxx> wrote:
>
> On 2024/2/16 16:23, Yosry Ahmed wrote:
> > On Fri, Feb 16, 2024 at 04:05:39PM +1300, Barry Song wrote:
> >> From: Barry Song <v-songbaohua@xxxxxxxx>
> >>
> >> My commit fc8580edbaa6 ("mm: zsmalloc: return -ENOSPC rather than -EINVAL
> >> in zs_malloc while size is too large") wanted to depend on zs_malloc's
> >> returned ENOSPC to distinguish the case that compressed data is larger
> >> than the original data from normal compression cases. The commit, for
> >> sure, was correct and worked as expected but the code wouldn't run to
> >> there after commit 744e1885922a ("crypto: scomp - fix req->dst buffer
> >> overflow") as Chengming's this patch makes zswap_store() goto out
> >> immediately after the special compression case happens. So there is
> >> no chance to execute zs_malloc() now. We need to fix the count right
> >> after compressions return ENOSPC.
> >>
> >> Fixes: fc8580edbaa6 ("mm: zsmalloc: return -ENOSPC rather than -EINVAL in zs_malloc while size is too large")
> >
> > I don't see how this is a fix for that commit. Commit fc8580edbaa6 made
> > sure zsmalloc returns a correct errno when the compressed size is too
> > large. The fact that zswap stores were failing before calling into
> > zsmalloc and not reporting the error correctly in debug counters is not
> > that commits fault.
> >
Hi Yosry, Chengming,
Thanks for your quick responses.
> > I think the proper fixes should be 744e1885922a if it introduced the
> > first scenario where -ENOSPC can be returned from scomp without handling
> > it properly in zswap. If -ENOSPC was a possible return value before
> > that, then it should be cb61dad80fdc ("zswap: export compression failure
> > stats"), where the counter was introduced.
>
> Right, 744e1885922a maybe a better fixes target.
I agree 744e1885922a is a better fixes target.
>
> >
> >> Cc: Chengming Zhou <zhouchengming@xxxxxxxxxxxxx>
> >> Cc: Nhat Pham <nphamcs@xxxxxxxxx>
> >> Cc: Sergey Senozhatsky <senozhatsky@xxxxxxxxxxxx>
> >> Signed-off-by: Barry Song <v-songbaohua@xxxxxxxx>
> >> ---
> >> mm/zswap.c | 5 ++++-
> >> 1 file changed, 4 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/mm/zswap.c b/mm/zswap.c
> >> index 6319d2281020..9a21dbe8c056 100644
> >> --- a/mm/zswap.c
> >> +++ b/mm/zswap.c
> >> @@ -1627,7 +1627,10 @@ bool zswap_store(struct folio *folio)
> >> dlen = acomp_ctx->req->dlen;
> >>
> >> if (ret) {
> >> - zswap_reject_compress_fail++;
> >> + if (ret == -ENOSPC)
> >> + zswap_reject_compress_poor++;
> >> + else
> >> + zswap_reject_compress_fail++;
> >
> > With this diff, we have four locations in zswap_store() where we
> > increment zswap_reject_compress_{poor/fail}.
> >
> > How about the following instead?A
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index 62fe307521c93..3a7e8ba7f6116 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -1059,24 +1059,16 @@ static bool zswap_compress(struct folio *folio, struct zswap_entry *entry)
> > */
> > ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
> > dlen = acomp_ctx->req->dlen;
> > - if (ret) {
> > - zswap_reject_compress_fail++;
> > + if (ret)
> > goto unlock;
> > - }
> >
> > zpool = zswap_find_zpool(entry);
> > gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
> > if (zpool_malloc_support_movable(zpool))
> > gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
> > ret = zpool_malloc(zpool, dlen, gfp, &handle);
> > - if (ret == -ENOSPC) {
> > - zswap_reject_compress_poor++;
> > - goto unlock;
> > - }
> > - if (ret) {
> > - zswap_reject_alloc_fail++;
> > + if (ret)
> > goto unlock;
> > - }
> >
> > buf = zpool_map_handle(zpool, handle, ZPOOL_MM_WO);
> > memcpy(buf, dst, dlen);
> > @@ -1086,6 +1078,10 @@ static bool zswap_compress(struct folio *folio, struct zswap_entry *entry)
> > entry->length = dlen;
> >
> > unlock:
> > + if (ret == -ENOSPC)
> > + zswap_reject_compress_poor++;
> > + else if (ret)
> > + zswap_reject_alloc_fail++;
>
> Here have two cases: zswap_reject_compress_fail, zswap_reject_alloc_fail.
Is it safe to differentiate these two cases by checking ret == -ENOMEM ?
otherwise, it seems the original patch still makes more sense?
>
> > mutex_unlock(&acomp_ctx->mutex);
> > return ret == 0;
> > }
> >
> >> goto put_dstmem;
> >> }
> >>
> >> --
> >> 2.34.1
> >>
Thanks
Barry
