Re: [PATCH v1] memory tier: consolidate the initialization of memory tiers

["Ho-Ren Chuang" <horen.chuang@linux.dev>]

June 24, 2024 at 1:27 AM, "Huang, Ying" <ying.huang@intel.com> wrote:

Hi Huang, Ying,

Thanks for your feedback. Replies inlined.

> 
> Hi, Jack,
> 
> Thanks for patch!
> 
> "Ho-Ren (Jack) Chuang" <horen.chuang@linux.dev> writes:
> 
> > 
> > If we simply move the set_node_memory_tier() from memory_tier_init() to
> > 
> >  late_initcall(), it will result in HMAT not registering the
> > 
> >  mt_adistance_algorithm callback function, because set_node_memory_tier()
> > 
> >  is not performed during the memory tiering initialization phase,
> > 
> >  leading to a lack of correct default_dram information.
> > 
> >  Therefore, we introduced a nodemask to pass the information of the
> > 
> >  default DRAM nodes. The reason for not choosing to reuse
> > 
> >  default_dram_type->nodes is that it is not clean enough. So in the end,
> > 
> >  we use a __initdata variable, which is a variable that is released once
> > 
> >  initialization is complete, including both CPU and memory nodes for HMAT
> > 
> >  to iterate through.
> > 
> >  Besides, since default_dram_type may be checked/used during the
> > 
> >  initialization process of HMAT and drivers, it is better to keep the
> > 
> >  allocation of default_dram_type in memory_tier_init().
> > 
> >  Signed-off-by: Ho-Ren (Jack) Chuang <horenchuang@bytedance.com>
> > 
> 
> Suggested-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
> 

Thank you for your help with the input. Will add it in the v2.

> > 
> > ---
> > 
> >  Hi all,
> > 
> >  The current memory tier initialization process is distributed across two
> > 
> >  different functions, memory_tier_init() and memory_tier_late_init(). This
> > 
> >  design is hard to maintain. Thus, this patch is proposed to reduce the
> > 
> >  possible code paths by consolidating different initialization patches into one.
> > 
> >  The earlier discussion with Jonathan and Ying is listed here:
> > 
> >  https://lore.kernel.org/lkml/20240405150244.00004b49@Huawei.com/
> > 
> >  If we want to put these two initializations together, they must be placed
> > 
> >  together in the later function. Because only at that time, the HMAT information
> > 
> >  will be ready, adist between nodes can be calculated, and memory tiering can be
> > 
> >  established based on the adist. So we position the initialization at
> > 
> >  memory_tier_init() to the memory_tier_late_init() call.
> > 
> >  Moreover, it's natural to keep memory_tier initialization in drivers at
> > 
> >  device_initcall() level.
> > 
> >  This patchset is based on commits cf93be18fa1b and a72a30af550c:
> > 
> >  [0/2] https://lkml.kernel.org/r/20240405000707.2670063-1-horenchuang@bytedance.com
> > 
> >  [1/2] https://lkml.kernel.org/r/20240405000707.2670063-2-horenchuang@bytedance.com
> > 
> >  [1/2] https://lkml.kernel.org/r/20240405000707.2670063-3-horenchuang@bytedance.com
> > 
> 
> It appears that you should switch the parts before and after "---".
> 
> This is the real patch description, as pointed out by Andrew too.
> 

Thank you for the suggestion. I plan to write the real patch description in
the cover letter in the next version to avoid any misunderstanding.

> > 
> > Thanks,
> > 
> >  Ho-Ren (Jack) Chuang
> > 
> >  drivers/acpi/numa/hmat.c | 4 ++-
> > 
> >  include/linux/memory-tiers.h | 6 ++++
> > 
> >  mm/memory-tiers.c | 70 ++++++++++++++++++------------------
> > 
> >  3 files changed, 44 insertions(+), 36 deletions(-)
> > 
> >  diff --git a/drivers/acpi/numa/hmat.c b/drivers/acpi/numa/hmat.c
> > 
> >  index 2c8ccc91ebe6..31a77a3324a8 100644
> > 
> >  --- a/drivers/acpi/numa/hmat.c
> > 
> >  +++ b/drivers/acpi/numa/hmat.c
> > 
> >  @@ -939,11 +939,13 @@ static int hmat_set_default_dram_perf(void)
> > 
> >  int nid, pxm;
> > 
> >  struct memory_target *target;
> > 
> >  struct access_coordinate *attrs;
> > 
> >  + nodemask_t default_dram_nodes;
> > 
> >  
> > 
> >  if (!default_dram_type)
> > 
> >  return -EIO;
> > 
> >  
> > 
> >  - for_each_node_mask(nid, default_dram_type->nodes) {
> > 
> >  + default_dram_nodes = mt_get_default_dram_nodemask();
> > 
> >  + for_each_node_mask(nid, default_dram_nodes) {
> > 
> 
> We don't need 'default_dram_type' in the function actually. It appears
> 
> that we can hide it in memory-tiers.c now?
> 

Do you mean to remove the "if (!default_dram_type) return -EIO;" here?
If so, I agree, it's not used anymore here.

> > 
> > pxm = node_to_pxm(nid);
> > 
> >  target = find_mem_target(pxm);
> > 
> >  if (!target)
> > 
> >  diff --git a/include/linux/memory-tiers.h b/include/linux/memory-tiers.h
> > 
> >  index 0d70788558f4..1567db7bd40e 100644
> > 
> >  --- a/include/linux/memory-tiers.h
> > 
> >  +++ b/include/linux/memory-tiers.h
> > 
> >  @@ -51,6 +51,7 @@ int mt_perf_to_adistance(struct access_coordinate *perf, int *adist);
> > 
> >  struct memory_dev_type *mt_find_alloc_memory_type(int adist,
> > 
> >  struct list_head *memory_types);
> > 
> >  void mt_put_memory_types(struct list_head *memory_types);
> > 
> >  +nodemask_t mt_get_default_dram_nodemask(void);
> > 
> >  #ifdef CONFIG_MIGRATION
> > 
> >  int next_demotion_node(int node);
> > 
> >  void node_get_allowed_targets(pg_data_t *pgdat, nodemask_t *targets);
> > 
> >  @@ -149,5 +150,10 @@ static inline struct memory_dev_type *mt_find_alloc_memory_type(int adist,
> > 
> >  static inline void mt_put_memory_types(struct list_head *memory_types)
> > 
> >  {
> > 
> >  }
> > 
> >  +
> > 
> >  +static inline nodemask_t mt_get_default_dram_nodemask(void)
> > 
> >  +{
> > 
> >  + return NODE_MASK_NONE;
> > 
> >  +}
> > 
> >  #endif /* CONFIG_NUMA */
> > 
> >  #endif /* _LINUX_MEMORY_TIERS_H */
> > 
> >  diff --git a/mm/memory-tiers.c b/mm/memory-tiers.c
> > 
> >  index 6632102bd5c9..7d4b7f53dd8f 100644
> > 
> >  --- a/mm/memory-tiers.c
> > 
> >  +++ b/mm/memory-tiers.c
> > 
> >  @@ -43,6 +43,7 @@ static LIST_HEAD(memory_tiers);
> > 
> >  static LIST_HEAD(default_memory_types);
> > 
> >  static struct node_memory_type_map node_memory_types[MAX_NUMNODES];
> > 
> >  struct memory_dev_type *default_dram_type;
> > 
> >  +static nodemask_t default_dram_nodes __initdata = NODE_MASK_NONE;
> > 
> >  
> > 
> >  static const struct bus_type memory_tier_subsys = {
> > 
> >  .name = "memory_tiering",
> > 
> >  @@ -125,6 +126,11 @@ static inline struct memory_tier *to_memory_tier(struct device *device)
> > 
> >  return container_of(device, struct memory_tier, dev);
> > 
> >  }
> > 
> >  
> > 
> >  +nodemask_t __init mt_get_default_dram_nodemask(void)
> > 
> >  +{
> > 
> >  + return default_dram_nodes;
> > 
> >  +}
> > 
> >  +
> > 
> 
> Why not just expose 'default_dram_nodes'?
> 

I was thinking encapsulating it should be more systematic/structural.
Do you think exposing it is better?

> > 
> > static __always_inline nodemask_t get_memtier_nodemask(struct memory_tier *memtier)
> > 
> >  {
> > 
> >  nodemask_t nodes = NODE_MASK_NONE;
> > 
> >  @@ -671,27 +677,38 @@ EXPORT_SYMBOL_GPL(mt_put_memory_types);
> > 
> >  
> > 
> >  /*
> > 
> >  * This is invoked via `late_initcall()` to initialize memory tiers for
> > 
> >  - * CPU-less memory nodes after driver initialization, which is
> > 
> >  - * expected to provide `adistance` algorithms.
> > 
> >  + * memory nodes, both with and without CPUs. After the initialization of
> > 
> >  + * firmware and devices, adistance algorithms are expected to be provided.
> > 
> >  */
> > 
> >  static int __init memory_tier_late_init(void)
> > 
> >  {
> > 
> >  int nid;
> > 
> >  + struct memory_tier *memtier;
> > 
> >  
> > 
> >  guard(mutex)(&memory_tier_lock);
> > 
> >  + /*
> > 
> >  + * Look at all the existing and uninitialized N_MEMORY nodes and
> > 
> >  + * add them to default memory tier or to a tier if we already have
> > 
> >  + * memory types assigned.
> > 
> >  + */
> > 
> >  for_each_node_state(nid, N_MEMORY) {
> > 
> 
> During the function run, the node may change between N_MEMORY and
> 
> !N_MEMORY in theory. So, it appears necessary to get/put_online_mems()
> 
> in the function?
> 

Thanks for the catch. I will add get/put_online_mems().

> > 
> > - /*
> > 
> >  - * Some device drivers may have initialized memory tiers
> > 
> >  - * between `memory_tier_init()` and `memory_tier_late_init()`,
> > 
> >  - * potentially bringing online memory nodes and
> > 
> >  - * configuring memory tiers. Exclude them here.
> > 
> >  - */
> > 
> >  - if (node_memory_types[nid].memtype)
> > 
> >  - continue;
> > 
> >  + if (!node_state(nid, N_CPU))
> > 
> 
> Why? I think that we should "continue" here even if node_state(nid,
> 
> N_CPU).
> 

Do you mean no matter node_state(nid, N_CPU) or !node_state(nid, N_CPU),
as long as if (node_memory_types[nid].memtype) is true, we
should "continue"?

I think you are right, at this moment, we only care if the
node_memory_types[nid].memtype is set or not.
If not, we should set it here. If yes, we should continue.
If my understanding is correct, I will fix it in the v2.

> > 
> > + /*
> > 
> >  + * Some device drivers may have initialized
> > 
> >  + * memory tiers, potentially bringing memory nodes
> > 
> >  + * online and configuring memory tiers.
> > 
> >  + * Exclude them here.
> > 
> >  + */
> > 
> >  + if (node_memory_types[nid].memtype)
> > 
> >  + continue;
> > 
> >  
> > 
> >  - set_node_memory_tier(nid);
> > 
> >  + memtier = set_node_memory_tier(nid);
> > 
> >  + if (IS_ERR(memtier))
> > 
> >  + /*
> > 
> >  + * Continue with memtiers we are able to setup.
> > 
> >  + */
> > 
> >  + break;
> > 
> >  }
> > 
> >  -
> > 
> >  establish_demotion_targets();
> > 
> >  
> > 
> >  return 0;
> > 
> >  @@ -876,7 +893,6 @@ static int __meminit memtier_hotplug_callback(struct notifier_block *self,
> > 
> >  static int __init memory_tier_init(void)
> > 
> >  {
> > 
> >  int ret, node;
> > 
> >  - struct memory_tier *memtier;
> > 
> >  
> > 
> >  ret = subsys_virtual_register(&memory_tier_subsys, NULL);
> > 
> >  if (ret)
> > 
> >  @@ -887,7 +903,8 @@ static int __init memory_tier_init(void)
> > 
> >  GFP_KERNEL);
> > 
> >  WARN_ON(!node_demotion);
> > 
> >  #endif
> > 
> >  - mutex_lock(&memory_tier_lock);
> > 
> >  +
> > 
> >  + guard(mutex)(&memory_tier_lock);
> > 
> >  /*
> > 
> >  * For now we can have 4 faster memory tiers with smaller adistance
> > 
> >  * than default DRAM tier.
> > 
> >  @@ -898,28 +915,11 @@ static int __init memory_tier_init(void)
> > 
> >  panic("%s() failed to allocate default DRAM tier\n", __func__);
> > 
> >  
> > 
> >  /*
> > 
> >  - * Look at all the existing N_MEMORY nodes and add them to
> > 
> >  - * default memory tier or to a tier if we already have memory
> > 
> >  - * types assigned.
> > 
> >  + * Record nodes with memory and CPU to set default DRAM performance.
> > 
> >  */
> > 
> 
> For one line comments, we can use
> 
>  /* Record nodes with memory and CPU to set default DRAM performance. */
> 

Thank you for the guidance. Will fix in the v2.

> > 
> > - for_each_node_state(node, N_MEMORY) {
> > 
> >  - if (!node_state(node, N_CPU))
> > 
> >  - /*
> > 
> >  - * Defer memory tier initialization on
> > 
> >  - * CPUless numa nodes. These will be initialized
> > 
> >  - * after firmware and devices are initialized.
> > 
> >  - */
> > 
> >  - continue;
> > 
> >  -
> > 
> >  - memtier = set_node_memory_tier(node);
> > 
> >  - if (IS_ERR(memtier))
> > 
> >  - /*
> > 
> >  - * Continue with memtiers we are able to setup
> > 
> >  - */
> > 
> >  - break;
> > 
> >  - }
> > 
> >  - establish_demotion_targets();
> > 
> >  - mutex_unlock(&memory_tier_lock);
> > 
> >  + for_each_node_state(node, N_MEMORY)
> > 
> >  + if (node_state(node, N_CPU))
> > 
> >  + node_set(node, default_dram_nodes);
> > 
> 
> Why not use
> 
>  nodes_andnot(default_dram_nodes, node_states[N_MEMORY], node_states[N_CPU]);
> 

Instead of using nodes_andnot(), should nodes_and() be correct? because we wanna
record the nodes that are both N_MEMORY and N_CPU.

> > 
> > hotplug_memory_notifier(memtier_hotplug_callback, MEMTIER_HOTPLUG_PRI);
> > 
> >  return 0;
> > 
> 
> --
> 
> Best Regards,
> 
> Huang, Ying
>