On 2024/12/3 09:01, Sridhar, Kanchana P wrote:
Hi Chengming, Yosry,
-----Original Message-----
From: Yosry Ahmed <yosryahmed@xxxxxxxxxx>
Sent: Monday, December 2, 2024 11:33 AM
To: Chengming Zhou <chengming.zhou@xxxxxxxxx>
Cc: Sridhar, Kanchana P <kanchana.p.sridhar@xxxxxxxxx>; linux-
kernel@xxxxxxxxxxxxxxx; linux-mm@xxxxxxxxx; hannes@xxxxxxxxxxx;
nphamcs@xxxxxxxxx; usamaarif642@xxxxxxxxx; ryan.roberts@xxxxxxx;
21cnbao@xxxxxxxxx; akpm@xxxxxxxxxxxxxxxxxxxx; Feghali, Wajdi K
<wajdi.k.feghali@xxxxxxxxx>; Gopal, Vinodh <vinodh.gopal@xxxxxxxxx>
Subject: Re: [PATCH v1 2/2] mm: zswap: zswap_store_pages() simplifications
for batching.
On Wed, Nov 27, 2024 at 11:00 PM Chengming Zhou
<chengming.zhou@xxxxxxxxx> wrote:
On 2024/11/28 06:53, Kanchana P Sridhar wrote:
In order to set up zswap_store_pages() to enable a clean batching
implementation in [1], this patch implements the following changes:
1) Addition of zswap_alloc_entries() which will allocate zswap entries for
all pages in the specified range for the folio, upfront. If this fails,
we return an error status to zswap_store().
2) Addition of zswap_compress_pages() that calls zswap_compress() for
each
page, and returns false if any zswap_compress() fails, so
zswap_store_page() can cleanup resources allocated and return an
error
status to zswap_store().
3) A "store_pages_failed" label that is a catch-all for all failure points
in zswap_store_pages(). This facilitates cleaner error handling within
zswap_store_pages(), which will become important for IAA compress
batching in [1].
[1]: https://patchwork.kernel.org/project/linux-mm/list/?series=911935
Signed-off-by: Kanchana P Sridhar <kanchana.p.sridhar@xxxxxxxxx>
---
mm/zswap.c | 93 +++++++++++++++++++++++++++++++++++++++++----
---------
1 file changed, 71 insertions(+), 22 deletions(-)
diff --git a/mm/zswap.c b/mm/zswap.c
index b09d1023e775..db80c66e2205 100644
--- a/mm/zswap.c
+++ b/mm/zswap.c
@@ -1409,9 +1409,56 @@ static void shrink_worker(struct work_struct
*w)
* main API
**********************************/
+static bool zswap_compress_pages(struct page *pages[],
+ struct zswap_entry *entries[],
+ u8 nr_pages,
+ struct zswap_pool *pool)
+{
+ u8 i;
+
+ for (i = 0; i < nr_pages; ++i) {
+ if (!zswap_compress(pages[i], entries[i], pool))
+ return false;
+ }
+
+ return true;
+}
How about introducing a `zswap_compress_folio()` interface which
can be used by `zswap_store()`?
```
zswap_store()
nr_pages = folio_nr_pages(folio)
entries = zswap_alloc_entries(nr_pages)
ret = zswap_compress_folio(folio, entries, pool)
// store entries into xarray and LRU list
```
And this version `zswap_compress_folio()` is very simple for now:
```
zswap_compress_folio()
nr_pages = folio_nr_pages(folio)
for (index = 0; index < nr_pages; ++index) {
struct page *page = folio_page(folio, index);
if (!zswap_compress(page, &entries[index], pool))
return false;
}
return true;
```
This can be easily extended to support your "batched" version.
Then the old `zswap_store_page()` could be removed.
The good point is simplicity, that we don't need to slice folio
into multiple batches, then repeat the common operations for each
batch, like preparing entries, storing into xarray and LRU list...
+1
Thanks for the code review comments. One question though: would
it make sense to trade-off the memory footprint cost with the code
simplification? For instance, lets say we want to store a 64k folio.
We would allocate memory for 16 zswap entries, and lets say one of
the compressions fails, we would deallocate memory for all 16 zswap
entries. Could this lead to zswap_entry kmem_cache starvation and
subsequent zswap_store() failures in multiple processes scenarios?
Ah, I get your consideration. But it's the unlikely case, right?
If the case you mentioned above happens a lot, I think yes, we should
optimize its memory footprint to avoid allocation and deallocation.
On the other hand, we should consider a folio would be compressed
successfully in most cases. So we have to allocate all entries
eventually.
Based on your consideration, I think your way is ok too, although
I think the patch 2/2 should be dropped, since it hides pages loop
in smaller functions, as Yosry mentioned too.
In other words, allocating entries in smaller batches -- more specifically,
only the compress batchsize -- seems to strike a balance in terms of
memory footprint, while mitigating the starvation aspect, and possibly
also helping latency (allocating a large # of zswap entries and potentially
deallocating, could impact latency).
If we consider the likely case (compress successfully), the whole
latency should be better, right? Since we can bulk allocate all
entries at first, and bulk insert to xarray and LRU at last.
If we agree with the merits of processing a large folio in smaller batches:
this in turn requires we store the smaller batches of entries in the
xarray/LRU before moving to the next batch. Which means all the
zswap_store() ops need to be done for a batch before moving to the next
batch.
Both way is ok for me based on your memory footprint consideration
above.
Thanks.
