[v3,0/3] memcg/hugetlb: Rework memcg hugetlb charging

Message

Joshua Hahn Dec. 11, 2024, 8:39 p.m. UTC

This series cleans up memcg's hugetlb charging logic by deprecating the
current memcg hugetlb try-charge + {commit, cancel} logic present in
alloc_hugetlb_folio. A single function mem_cgroup_charge_hugetlb takes
its place instead. This makes the code more maintainable by simplifying
the error path and reduces memcg's footprint in hugetlb logic.

This patch introduces a few changes in the hugetlb folio allocation
error path:
(a) Instead of having multiple return points, we consolidate them to
    two: one for reaching the memcg limit or running out of memory
    (-ENOMEM) and one for hugetlb allocation fails / limit being
    reached (-ENOSPC).
(b) Previously, the memcg limit was checked before the folio is acquired,
    meaning the hugeTLB folio isn't acquired if the limit is reached.
    This patch performs the charging after the folio is reached, meaning
    if memcg's limit is reached, the acquired folio is freed right away.

This patch builds on two earlier patch series: [2] which adds memcg
hugeTLB counters, and [3] which deprecates charge moving and removes the
last references to mem_cgroup_cancel_charge. The request for this cleanup
can be found in [2].

Suggested-by: Shakeel Butt <[e-mail]>
Signed-off-by: Joshua Hahn <[e-mail]>

[1] https://lore.kernel.org/all/[e-mail]/
[2] https://lore.kernel.org/all/[e-mail]/
[3] https://lore.kernel.org/linux-mm/[e-mail]/

---
Changelog
v3:
  * Fixed build error where mem_cgroup_charge_hugetlb was defined inside
    an #ifdef CONFIG_MEMCG. It is now defined in the #else case as well.
  * memcg_accounts_hugetlb is no longer forward-declared, just
    statically defined earlier.
  * The #ifdef is moved outside memcg_accounts_hugetlb's definition.
  * lruvec_stat_mod_folio() is now called only except when memcg
    charging fails. That is, even when memcg is not supported or memcg
    does not account hugetlb, we still update the (global) stats.
    * For this to happen, memcg_charge_hugetlb no longer returns
      -EOPNOTSUPP, just returns 0 when it is not supported.
  * kernel-doc comment added for mem_cgroup_charge_hugetlb
  * s/deprecate/remove in commit messages for clarity.
  * Thank you Shakeel, Yosry, and SJ for your input!
v2:
  * Removed declaration of memcg_accounts_hugetlb from memcontrol.h
  * Moved second call to memcg_accounts_hugetlb from 2nd patch to 1st
  * Changed error behavior in alloc_hugetlb_folio: v1 included a bug
    that uncharged hugetlb_cgroup twice when memecg's limit was reached
  * mem_cgroup_charge_hugetlb no longer called with hugetlb_lock held
  * Moved mem_cgroup_hugetlb_{try, charge} deprecation to patch 3
  * mem_cgroup_charge_hugetlb always decrements memcg's refcount
  * Fully cleaned up mem_cgroup_{cancel,commit}_charge
  * Fixed typos

Joshua Hahn (3):
  memcg/hugetlb: Introduce memcg_accounts_hugetlb
  memcg/hugetlb: Introduce mem_cgroup_charge_hugetlb
  memcg/hugetlb: Remove memcg hugetlb try-commit-cancel protocol

 include/linux/memcontrol.h | 21 ++-------
 mm/hugetlb.c               | 35 ++++++---------
 mm/memcontrol.c            | 92 ++++++++++++++++----------------------
 3 files changed, 56 insertions(+), 92 deletions(-)