Re: [RFC PATCH v0] mm/vmscan: Add readahead LRU to improve readahead file page reclamation efficiency

On 2025/9/17 0:33, Yuanchu Xie wrote:

On Tue, Sep 16, 2025 at 2:22 AM Lei Liu [off-list ref] wrote:

quoted

...

2. Solution Proposal
Introduce a Readahead LRU to track pages brought in via readahead. During
memory reclamation, prioritize scanning this LRU to reclaim pages that
have not been accessed recently. For pages in the Readahead LRU that are
accessed, move them back to the inactive_file LRU to await subsequent
reclamation.

I'm unsure this is the right solution though, given all users would
have this readahead LRU on and we don't have performance numbers
besides application startup here.
My impression is that readahead behavior is highly dependent on the
hardware, the workload, and the desired behavior, so making the
readahead{-adjacent} behavior more amenable to tuning seems like the
right direction.

Maybe relevant discussions: https://lwn.net/Articles/897786/

I only skimmed the code but noticed a few things:

quoted

diff --git a/fs/proc/meminfo.c b/fs/proc/meminfo.c
index a458f1e112fd..4f3f031134fd 100644
--- a/fs/proc/meminfo.c
+++ b/fs/proc/meminfo.c

@@ -71,6 +71,7 @@ static int meminfo_proc_show(struct seq_file *m, void *v)
         show_val_kb(m, "Inactive(anon): ", pages[LRU_INACTIVE_ANON]);
         show_val_kb(m, "Active(file):   ", pages[LRU_ACTIVE_FILE]);
         show_val_kb(m, "Inactive(file): ", pages[LRU_INACTIVE_FILE]);
+       show_val_kb(m, "ReadAhead(file):",

I notice both readahead and read ahead in this patch. Stick to the
conventional one (readahead).

quoted

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 8d3fa3a91ce4..57dac828aa4f 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h

@@ -127,6 +127,7 @@ enum pageflags {
  #ifdef CONFIG_ARCH_USES_PG_ARCH_3
         PG_arch_3,
  #endif
+       PG_readahead_lru,

More pageflags...

b/include/trace/events/mmflags.h

quoted

index aa441f593e9a..2dbc1701e838 100644

--- a/include/trace/events/mmflags.h
+++ b/include/trace/events/mmflags.h

@@ -159,7 +159,8 @@ TRACE_DEFINE_ENUM(___GFP_LAST_BIT);
         DEF_PAGEFLAG_NAME(reclaim),                                     \
         DEF_PAGEFLAG_NAME(swapbacked),                                  \
         DEF_PAGEFLAG_NAME(unevictable),                                 \
-       DEF_PAGEFLAG_NAME(dropbehind)                                   \
+       DEF_PAGEFLAG_NAME(dropbehind),                                  \
+       DEF_PAGEFLAG_NAME(readahead_lru)                                \
  IF_HAVE_PG_MLOCK(mlocked)                                              \
  IF_HAVE_PG_HWPOISON(hwpoison)                                          \
  IF_HAVE_PG_IDLE(idle)                                                  \

@@ -309,6 +310,7 @@ IF_HAVE_VM_DROPPABLE(VM_DROPPABLE,  "droppable"     )               \
                 EM (LRU_ACTIVE_ANON, "active_anon") \
                 EM (LRU_INACTIVE_FILE, "inactive_file") \
                 EM (LRU_ACTIVE_FILE, "active_file") \
+               EM(LRU_READ_AHEAD_FILE, "readahead_file") \

Likewise, inconsistent naming.

quoted

diff --git a/mm/migrate.c b/mm/migrate.c
index 9e5ef39ce73a..0feab4d89d47 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c

@@ -760,6 +760,8 @@ void folio_migrate_flags(struct folio *newfolio, struct folio *folio)
                 folio_set_workingset(newfolio);
         if (folio_test_checked(folio))
                 folio_set_checked(newfolio);
+       if (folio_test_readahead_lru(folio))
+               folio_set_readahead_lru(folio);

newfolio

Understood—I'll revise accordingly.

quoted

  /*

@@ -5800,6 +5837,87 @@ static void lru_gen_shrink_node(struct pglist_data *pgdat, struct scan_control *

  #endif /* CONFIG_LRU_GEN */

+static unsigned long shrink_read_ahead_list(unsigned long nr_to_scan,
+                                           unsigned long nr_to_reclaim,
+                                           struct lruvec *lruvec,
+                                           struct scan_control *sc)
+{
+       LIST_HEAD(l_hold);
+       LIST_HEAD(l_reclaim);
+       LIST_HEAD(l_inactive);
+       unsigned long nr_scanned = 0;
+       unsigned long nr_taken = 0;
+       unsigned long nr_reclaimed = 0;
+       unsigned long vm_flags;
+       enum vm_event_item item;
+       struct pglist_data *pgdat = lruvec_pgdat(lruvec);
+       struct reclaim_stat stat = { 0 };
+
+       lru_add_drain();
+
+       spin_lock_irq(&lruvec->lru_lock);
+       nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold, &nr_scanned,
+                                     sc, LRU_READ_AHEAD_FILE);
+
+       __count_vm_events(PGSCAN_READAHEAD_FILE, nr_scanned);
+       __mod_node_page_state(pgdat, NR_ISOLATED_FILE, nr_taken);
+       item = PGSCAN_KSWAPD + reclaimer_offset(sc);
+       if (!cgroup_reclaim(sc))
+               __count_vm_events(item, nr_scanned);
+       count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned);
+       __count_vm_events(PGSCAN_FILE, nr_scanned);
+       spin_unlock_irq(&lruvec->lru_lock);
+
+       if (nr_taken == 0)
+               return 0;
+
+       while (!list_empty(&l_hold)) {
+               struct folio *folio;
+
+               cond_resched();
+               folio = lru_to_folio(&l_hold);
+               list_del(&folio->lru);
+               folio_clear_readahead_lru(folio);
+
+               if (folio_referenced(folio, 0, sc->target_mem_cgroup, &vm_flags)) {
+                       list_add(&folio->lru, &l_inactive);
+                       continue;
+               }
+               folio_clear_active(folio);
+               list_add(&folio->lru, &l_reclaim);
+       }
+
+       nr_reclaimed = shrink_folio_list(&l_reclaim, pgdat, sc, &stat, true,
+                                        lruvec_memcg(lruvec));
+
+       list_splice(&l_reclaim, &l_inactive);
+
+       spin_lock_irq(&lruvec->lru_lock);
+       move_folios_to_lru(lruvec, &l_inactive);
+       __mod_node_page_state(pgdat, NR_ISOLATED_FILE, -nr_taken);
+
+       __count_vm_events(PGSTEAL_READAHEAD_FILE, nr_reclaimed);
+       item = PGSTEAL_KSWAPD + reclaimer_offset(sc);
+       if (!cgroup_reclaim(sc))
+               __count_vm_events(item, nr_reclaimed);
+       count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed);
+       __count_vm_events(PGSTEAL_FILE, nr_reclaimed);
+       spin_unlock_irq(&lruvec->lru_lock);

I see the idea is that readahead pages should be scanned before the
rest of inactive file. I wonder if this is achievable without adding
another LRU.


Thanks,
Yuanchu

Hi， Yuanchu

Thank you for your valuable feedback!

1.We initially considered keeping readahead pages in the system's 
existing inactive/active LRUs without adding a dedicated LRU. However, 
this approach may lead to inefficient reclamation of readahead pages.

Reason: When scanning the inactive LRU, processing readahead pages can 
be frequently interrupted by non-readahead pages (e.g., shared/accessed 
pages). The reference checks for these non-readahead pages incur 
significant overhead, slowing down the scanning and reclamation of 
readahead pages.
Thus, isolating readahead pages in a readahead LRU allows more targeted 
reclamation, significantly accelerating scanning and recycling efficiency.

2.That said, this solution does raise valid concerns. As you rightly 
pointed out, enabling this LRU globally may not align with all users' 
needs since not every scenario requires it.

3.For now, this remains a preliminary solution. The primary goal of this 
RFC is to highlight the issue of excessive readahead overhead and gather 
community insights for better alternatives.

We are actively exploring solutions without adding a new LRU for future 
iterations.

Best regards,
Lei Liu