| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <CAKEwX=P=frZmEXm26uTBN05gqLXoL-Shgk5P=EsMpYR16dW-sw@mail.gmail.com>
Date: Tue, 19 Dec 2023 10:43:21 -0800
From: Nhat Pham <nphamcs@...il.com>
To: Chris Li <chrisl@...nel.org>
Cc: Chengming Zhou <zhouchengming@...edance.com>, Seth Jennings <sjenning@...hat.com>,
Yosry Ahmed <yosryahmed@...gle.com>, Vitaly Wool <vitaly.wool@...sulko.com>,
Dan Streetman <ddstreet@...e.org>, Johannes Weiner <hannes@...xchg.org>,
Andrew Morton <akpm@...ux-foundation.org>, linux-kernel@...r.kernel.org,
linux-mm@...ck.org
Subject: Re: [PATCH v3 6/6] mm/zswap: directly use percpu mutex and buffer in load/store
On Tue, Dec 19, 2023 at 5:29 AM Chris Li <chrisl@...nel.org> wrote:
>
> Hi Chengming and Yosry,
>
> On Mon, Dec 18, 2023 at 3:50 AM Chengming Zhou
> <zhouchengming@...edance.com> wrote:
> >
> > Since the introduce of reusing the dstmem in the load path, it seems
> > confusing that we are now using acomp_ctx->dstmem and acomp_ctx->mutex
> > now for purposes other than what the naming suggests.
> >
> > Yosry suggested removing these two fields from acomp_ctx, and directly
> > using zswap_dstmem and zswap_mutex in both the load and store paths,
> > rename them, and add proper comments above their definitions that they
> > are for generic percpu buffering on the load and store paths.
> >
> > So this patch remove dstmem and mutex from acomp_ctx, and rename the
> > zswap_dstmem to zswap_buffer, using the percpu mutex and buffer on
> > the load and store paths.
>
> Sorry joining this discussion late.
>
> I get the rename of "dstmem" to "buffer". Because the buffer is used
> for both load and store as well. What I don't get is that, why do we
> move it out of the acomp_ctx struct. Now we have 3 per cpu entry:
> buffer, mutex and acomp_ctx. I think we should do the reverse, fold
> this three per cpu entry into one struct the acomp_ctx. Each per_cpu
> load() has a sequence of dance around the cpu id and disable preempt
> etc, while each of the struct member load is just a plan memory load.
> It seems to me it would be more optimal to combine this three per cpu
> entry into acomp_ctx. Just do the per cpu for the acomp_ctx once.
I agree with Chris. From a practicality POV, what Chris says here
makes sense. From a semantic POV, this buffer is only used in
(de)compression contexts - be it in store, load, or writeback - so it
belonging to the orignal struct still makes sense to me. Why separate
it out, without any benefits. Just rename the old field buffer or
zswap_buffer and call it a day? It will be a smaller patch too!
>
> >
> > Suggested-by: Yosry Ahmed <yosryahmed@...gle.com>
> > Signed-off-by: Chengming Zhou <zhouchengming@...edance.com>
> > ---
> > mm/zswap.c | 69 +++++++++++++++++++++++++++++++++-----------------------------
> > 1 file changed, 37 insertions(+), 32 deletions(-)
> >
> > diff --git a/mm/zswap.c b/mm/zswap.c
> > index 2c349fd88904..71bdcd552e5b 100644
> > --- a/mm/zswap.c
> > +++ b/mm/zswap.c
> > @@ -166,8 +166,6 @@ struct crypto_acomp_ctx {
> > struct crypto_acomp *acomp;
> > struct acomp_req *req;
> > struct crypto_wait wait;
> > - u8 *dstmem;
> > - struct mutex *mutex;
> > };
> >
> > /*
> > @@ -694,7 +692,7 @@ static void zswap_alloc_shrinker(struct zswap_pool *pool)
> > /*********************************
> > * per-cpu code
> > **********************************/
> > -static DEFINE_PER_CPU(u8 *, zswap_dstmem);
> > +static DEFINE_PER_CPU(u8 *, zswap_buffer);
> > /*
> > * If users dynamically change the zpool type and compressor at runtime, i.e.
> > * zswap is running, zswap can have more than one zpool on one cpu, but they
> > @@ -702,39 +700,39 @@ static DEFINE_PER_CPU(u8 *, zswap_dstmem);
> > */
> > static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
> >
> > -static int zswap_dstmem_prepare(unsigned int cpu)
> > +static int zswap_buffer_prepare(unsigned int cpu)
> > {
> > struct mutex *mutex;
> > - u8 *dst;
> > + u8 *buf;
> >
> > - dst = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
> > - if (!dst)
> > + buf = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
> > + if (!buf)
> > return -ENOMEM;
> >
> > mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
> > if (!mutex) {
> > - kfree(dst);
> > + kfree(buf);
> > return -ENOMEM;
> > }
> >
> > mutex_init(mutex);
> > - per_cpu(zswap_dstmem, cpu) = dst;
> > + per_cpu(zswap_buffer, cpu) = buf;
> > per_cpu(zswap_mutex, cpu) = mutex;
> > return 0;
> > }
> >
> > -static int zswap_dstmem_dead(unsigned int cpu)
> > +static int zswap_buffer_dead(unsigned int cpu)
> > {
> > struct mutex *mutex;
> > - u8 *dst;
> > + u8 *buf;
> >
> > mutex = per_cpu(zswap_mutex, cpu);
> > kfree(mutex);
> > per_cpu(zswap_mutex, cpu) = NULL;
> >
> > - dst = per_cpu(zswap_dstmem, cpu);
> > - kfree(dst);
> > - per_cpu(zswap_dstmem, cpu) = NULL;
> > + buf = per_cpu(zswap_buffer, cpu);
> > + kfree(buf);
> > + per_cpu(zswap_buffer, cpu) = NULL;
> >
> > return 0;
> > }
> > @@ -772,9 +770,6 @@ static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
> > acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
> > crypto_req_done, &acomp_ctx->wait);
> >
> > - acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
> > - acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
> > -
> > return 0;
> > }
> >
> > @@ -1397,15 +1392,21 @@ static void __zswap_load(struct zswap_entry *entry, struct page *page)
> > struct zpool *zpool = zswap_find_zpool(entry);
> > struct scatterlist input, output;
> > struct crypto_acomp_ctx *acomp_ctx;
> > - u8 *src;
> > + u8 *src, *buf;
> > + int cpu;
> > + struct mutex *mutex;
> >
> > - acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
> > - mutex_lock(acomp_ctx->mutex);
> > + cpu = raw_smp_processor_id();
> > + mutex = per_cpu(zswap_mutex, cpu);
> First per cpu call.
> > + mutex_lock(mutex);
> > +
> > + acomp_ctx = per_cpu_ptr(entry->pool->acomp_ctx, cpu);
> Second per cpu
> >
> > src = zpool_map_handle(zpool, entry->handle, ZPOOL_MM_RO);
> > if (!zpool_can_sleep_mapped(zpool)) {
> > - memcpy(acomp_ctx->dstmem, src, entry->length);
> > - src = acomp_ctx->dstmem;
> > + buf = per_cpu(zswap_buffer, cpu);
>
> Here is the function that does the third per_cpu call. I think doing
> one per_cpu and the rest of them load from the context is more
> optimal.
> Conceptually it is cleaner as well. It is clear what this mutex is
> supposed to protect. It is protecting the compression context struct.
> Move it out as per cpu, it is less clear what is the protection scope
> of the mutex.
>
> What am I missing?
>
> Chris
>
>
> > + memcpy(buf, src, entry->length);
> > + src = buf;
> > zpool_unmap_handle(zpool, entry->handle);
> > }
> >
> > @@ -1415,7 +1416,7 @@ static void __zswap_load(struct zswap_entry *entry, struct page *page)
> > acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
> > BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
> > BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
> > - mutex_unlock(acomp_ctx->mutex);
> > + mutex_unlock(mutex);
> >
> > if (zpool_can_sleep_mapped(zpool))
> > zpool_unmap_handle(zpool, entry->handle);
> > @@ -1551,6 +1552,8 @@ bool zswap_store(struct folio *folio)
> > u8 *src, *dst;
> > gfp_t gfp;
> > int ret;
> > + int cpu;
> > + struct mutex *mutex;
> >
> > VM_WARN_ON_ONCE(!folio_test_locked(folio));
> > VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
> > @@ -1636,11 +1639,13 @@ bool zswap_store(struct folio *folio)
> > }
> >
> > /* compress */
> > - acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
> > + cpu = raw_smp_processor_id();
> > + mutex = per_cpu(zswap_mutex, cpu);
> > + mutex_lock(mutex);
> >
> > - mutex_lock(acomp_ctx->mutex);
> > + acomp_ctx = per_cpu_ptr(entry->pool->acomp_ctx, cpu);
> > + dst = per_cpu(zswap_buffer, cpu);
> >
> > - dst = acomp_ctx->dstmem;
> > sg_init_table(&input, 1);
> > sg_set_page(&input, page, PAGE_SIZE, 0);
> >
> > @@ -1683,7 +1688,7 @@ bool zswap_store(struct folio *folio)
> > buf = zpool_map_handle(zpool, handle, ZPOOL_MM_WO);
> > memcpy(buf, dst, dlen);
> > zpool_unmap_handle(zpool, handle);
> > - mutex_unlock(acomp_ctx->mutex);
> > + mutex_unlock(mutex);
> >
> > /* populate entry */
> > entry->swpentry = swp_entry(type, offset);
> > @@ -1726,7 +1731,7 @@ bool zswap_store(struct folio *folio)
> > return true;
> >
> > put_dstmem:
> > - mutex_unlock(acomp_ctx->mutex);
> > + mutex_unlock(mutex);
> > put_pool:
> > zswap_pool_put(entry->pool);
> > freepage:
> > @@ -1902,10 +1907,10 @@ static int zswap_setup(void)
> > }
> >
> > ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
> > - zswap_dstmem_prepare, zswap_dstmem_dead);
> > + zswap_buffer_prepare, zswap_buffer_dead);
> > if (ret) {
> > - pr_err("dstmem alloc failed\n");
> > - goto dstmem_fail;
> > + pr_err("buffer alloc failed\n");
> > + goto buffer_fail;
> > }
> >
> > ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
> > @@ -1940,7 +1945,7 @@ static int zswap_setup(void)
> > zswap_pool_destroy(pool);
> > hp_fail:
> > cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
> > -dstmem_fail:
> > +buffer_fail:
> > kmem_cache_destroy(zswap_entry_cache);
> > cache_fail:
> > /* if built-in, we aren't unloaded on failure; don't allow use */
> >
> > --
> > b4 0.10.1
Powered by blists - more mailing lists