On Sun, Mar 3, 2024 at 6:42 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>
> Hi, Jack,
>
> "Ho-Ren (Jack) Chuang" <horenchuang@xxxxxxxxxxxxx> writes:
>
> > * Introduce `mt_init_with_hmat()`
> > We defer memory tier initialization for those CPUless NUMA nodes
> > until acquiring HMAT info. `mt_init_with_hmat()` is introduced to
> > post-create CPUless memory tiers after obtaining HMAT info.
> > It iterates through each CPUless memory node, creating memory tiers if
> > necessary. Finally, it calculates demotion tables again at the end.
> >
> > * Introduce `hmat_find_alloc_memory_type()`
> > Find or allocate a memory type in the `hmat_memory_types` list.
> >
> > * Make `set_node_memory_tier()` more generic
> > This function can also be used for setting other memory types for a node.
> > To do so, a new argument is added to specify a memory type.
> >
> > * Handle cases where there is no HMAT when creating memory tiers
> > If no HMAT is specified, it falls back to using `default_dram_type`.
> >
> > * Change adist calculation code to use another new lock, mt_perf_lock.
> > Iterating through CPUlist nodes requires holding the `memory_tier_lock`.
> > However, `mt_calc_adistance()` will end up trying to acquire the same lock,
> > leading to a potential deadlock. Therefore, we propose introducing a
> > standalone `mt_perf_lock` to protect `default_dram_perf`. This approach not
> > only avoids deadlock but also prevents holding a large lock simultaneously.
>
> The patch description is used to described why we need the change, and
> how we do that, but not what we do. People can tell what is done from
> the code itself.
>
Got it. Thanks. Will rewrite it after the code is finalized.
> > Signed-off-by: Ho-Ren (Jack) Chuang <horenchuang@xxxxxxxxxxxxx>
> > Signed-off-by: Hao Xiang <hao.xiang@xxxxxxxxxxxxx>
> > ---
> > drivers/acpi/numa/hmat.c | 3 ++
> > include/linux/memory-tiers.h | 6 +++
> > mm/memory-tiers.c | 76 ++++++++++++++++++++++++++++++++----
> > 3 files changed, 77 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/acpi/numa/hmat.c b/drivers/acpi/numa/hmat.c
> > index d6b85f0f6082..9f57338b3cb5 100644
> > --- a/drivers/acpi/numa/hmat.c
> > +++ b/drivers/acpi/numa/hmat.c
> > @@ -1038,6 +1038,9 @@ static __init int hmat_init(void)
> > if (!hmat_set_default_dram_perf())
> > register_mt_adistance_algorithm(&hmat_adist_nb);
> >
> > + /* Post-create CPUless memory tiers after getting HMAT info */
> > + mt_init_with_hmat();
> > +
> > return 0;
> > out_put:
> > hmat_free_structures();
> > diff --git a/include/linux/memory-tiers.h b/include/linux/memory-tiers.h
> > index 69e781900082..2f845e90c033 100644
> > --- a/include/linux/memory-tiers.h
> > +++ b/include/linux/memory-tiers.h
> > @@ -48,6 +48,7 @@ int mt_calc_adistance(int node, int *adist);
> > int mt_set_default_dram_perf(int nid, struct access_coordinate *perf,
> > const char *source);
> > int mt_perf_to_adistance(struct access_coordinate *perf, int *adist);
> > +void mt_init_with_hmat(void);
>
> HMAT isn't universally available. It's a driver in fact. So, don't put
> driver specific code in general code.
>
Please see below regarding "move code to hmat.c"
> > #ifdef CONFIG_MIGRATION
> > int next_demotion_node(int node);
> > void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets);
> > @@ -136,5 +137,10 @@ static inline int mt_perf_to_adistance(struct access_coordinate *perf, int *adis
> > {
> > return -EIO;
> > }
> > +
> > +static inline void mt_init_with_hmat(void)
> > +{
> > +
> > +}
> > #endif /* CONFIG_NUMA */
> > #endif /* _LINUX_MEMORY_TIERS_H */
> > diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
> > index 0537664620e5..7a0a579b3deb 100644
> > --- a/mm/memory-tiers.c
> > +++ b/mm/memory-tiers.c
> > @@ -35,7 +35,9 @@ struct node_memory_type_map {
> > };
> >
> > static DEFINE_MUTEX(memory_tier_lock);
> > +static DEFINE_MUTEX(mt_perf_lock);
> > static LIST_HEAD(memory_tiers);
> > +static LIST_HEAD(hmat_memory_types);
> > static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
> > struct memory_dev_type *default_dram_type;
> >
> > @@ -502,7 +504,7 @@ static inline void __init_node_memory_type(int node, struct memory_dev_type *mem
> > }
> > }
> >
> > -static struct memory_tier *set_node_memory_tier(int node)
> > +static struct memory_tier *set_node_memory_tier(int node, struct memory_dev_type *new_memtype)
>
> No. memory_dev_type are passed to the function via node_memory_types[node].memtype.
>
Got it. Will mimic the way kmem.c does. Thanks.
> > {
> > struct memory_tier *memtier;
> > struct memory_dev_type *memtype;
> > @@ -514,7 +516,7 @@ static struct memory_tier *set_node_memory_tier(int node)
> > if (!node_state(node, N_MEMORY))
> > return ERR_PTR(-EINVAL);
> >
> > - __init_node_memory_type(node, default_dram_type);
> > + __init_node_memory_type(node, new_memtype);
> >
> > memtype = node_memory_types[node].memtype;
> > node_set(node, memtype->nodes);
> > @@ -623,6 +625,56 @@ void clear_node_memory_type(int node, struct memory_dev_type *memtype)
> > }
> > EXPORT_SYMBOL_GPL(clear_node_memory_type);
> >
> > +static struct memory_dev_type *hmat_find_alloc_memory_type(int adist)
>
> Similar function existed in drivers/dax/kmem.c. Please abstract them
> and move them here.
>
Got it. Will try. Thanks.
> > +{
> > + bool found = false;
> > + struct memory_dev_type *mtype;
> > +
> > + list_for_each_entry(mtype, &hmat_memory_types, list) {
> > + if (mtype->adistance == adist) {
> > + found = true;
> > + break;
> > + }
> > + }
> > + if (!found) {
> > + mtype = alloc_memory_type(adist);
> > + if (!IS_ERR(mtype))
> > + list_add(&mtype->list, &hmat_memory_types);
> > + }
> > + return mtype;
> > +}
> > +
> > +static void mt_create_with_hmat(int node)
> > +{
> > + struct memory_dev_type *mtype = NULL;
> > + int adist = MEMTIER_ADISTANCE_DRAM;
> > +
> > + mt_calc_adistance(node, &adist);
> > + if (adist != MEMTIER_ADISTANCE_DRAM) {
> > + mtype = hmat_find_alloc_memory_type(adist);
> > + if (IS_ERR(mtype))
> > + pr_err("%s() failed to allocate a tier\n", __func__);
> > + } else {
> > + mtype = default_dram_type;
> > + }
> > +
> > + set_node_memory_tier(node, mtype);
> > +}
> > +
> > +void mt_init_with_hmat(void)
> > +{
> > + int nid;
> > +
> > + mutex_lock(&memory_tier_lock);
> > + for_each_node_state(nid, N_MEMORY)
> > + if (!node_state(nid, N_CPU))
> > + mt_create_with_hmat(nid);
> > +
> > + establish_demotion_targets();
> > + mutex_unlock(&memory_tier_lock);
> > +}
> > +EXPORT_SYMBOL_GPL(mt_init_with_hmat);
> > +
>
> I guess that we can put most hmat related code above in hmat.c.
>
To put the heat-related code to hmat.c I will need to export some
static functions in memory-tiers.c, like set_node_memory_tier() and
establish_demotion_targets(). Is that ok?
> > static void dump_hmem_attrs(struct access_coordinate *coord, const char *prefix)
> > {
> > pr_info(
> > @@ -636,7 +688,7 @@ int mt_set_default_dram_perf(int nid, struct access_coordinate *perf,
> > {
> > int rc = 0;
> >
> > - mutex_lock(&memory_tier_lock);
> > + mutex_lock(&mt_perf_lock);
> > if (default_dram_perf_error) {
> > rc = -EIO;
> > goto out;
> > @@ -684,7 +736,7 @@ int mt_set_default_dram_perf(int nid, struct access_coordinate *perf,
> > }
> >
> > out:
> > - mutex_unlock(&memory_tier_lock);
> > + mutex_unlock(&mt_perf_lock);
> > return rc;
> > }
> >
> > @@ -700,7 +752,7 @@ int mt_perf_to_adistance(struct access_coordinate *perf, int *adist)
> > perf->read_bandwidth + perf->write_bandwidth == 0)
> > return -EINVAL;
> >
> > - mutex_lock(&memory_tier_lock);
> > + mutex_lock(&mt_perf_lock);
> > /*
> > * The abstract distance of a memory node is in direct proportion to
> > * its memory latency (read + write) and inversely proportional to its
> > @@ -713,7 +765,7 @@ int mt_perf_to_adistance(struct access_coordinate *perf, int *adist)
> > (default_dram_perf.read_latency + default_dram_perf.write_latency) *
> > (default_dram_perf.read_bandwidth + default_dram_perf.write_bandwidth) /
> > (perf->read_bandwidth + perf->write_bandwidth);
> > - mutex_unlock(&memory_tier_lock);
> > + mutex_unlock(&mt_perf_lock);
> >
> > return 0;
> > }
> > @@ -797,7 +849,7 @@ static int __meminit memtier_hotplug_callback(struct notifier_block *self,
> > break;
> > case MEM_ONLINE:
> > mutex_lock(&memory_tier_lock);
> > - memtier = set_node_memory_tier(arg->status_change_nid);
> > + memtier = set_node_memory_tier(arg->status_change_nid, default_dram_type);
> > if (!IS_ERR(memtier))
> > establish_demotion_targets();
> > mutex_unlock(&memory_tier_lock);
> > @@ -836,7 +888,15 @@ static int __init memory_tier_init(void)
> > * types assigned.
> > */
> > for_each_node_state(node, N_MEMORY) {
> > - memtier = set_node_memory_tier(node);
> > + if (!node_state(node, N_CPU))
> > + /*
> > + * Defer memory tier initialization on CPUless numa nodes.
> > + * These will be initialized when HMAT information is
> > + * available.
> > + */
> > + continue;
> > +
> > + memtier = set_node_memory_tier(node, default_dram_type);
>
> On system with HMAT, how to fall back CPU-less node to
> default_dram_type? I found your description, but I don't find it in code.
>
I assume you meant without HMAT, if so,
because if no HMAT, adist will not be updated in mt_calc_adistance():
+ int adist = MEMTIER_ADISTANCE_DRAM;
+
+ mt_calc_adistance(node, &adist);
+ if (adist != MEMTIER_ADISTANCE_DRAM) {
…
+ } else {
+ mtype = default_dram_type;
+ }
+
+ set_node_memory_tier(node, mtype);
> > if (IS_ERR(memtier))
> > /*
> > * Continue with memtiers we are able to setup
>
> --
> Best Regards,
> Huang, Ying
--
Best regards,
Ho-Ren (Jack) Chuang
莊賀任
