On 06/11/24(Wed) 15:58, mpi@grenadille.net wrote:
> Diff below greatly improves the responsiveness of the page daemon for
> 64bit archs with a low/high memory split.  The improvement comes from
> a more precise calculation of how many low pages have to be freed.  As
> a result the amount of pages written to swap is decreased by ~50% in my
> tests and my arm64 machine becomes responsive during heavy swapping.
> 
> The diff includes:
> 
> - Use a global "struct uvm_pmalloc" to notify failed nowait allocations
>   in order to look at the managed lists.  The current algorithm does not
>   call uvmpd_scan() if there have been only nowait allocations.
> 
> - Skip calling the shrinkers and grabbing some locks if the page daemon
>   is awoken to rebalance the active/inactive lists.
> 
> - Do not bother releasing high pages if all we are interested in are low
>   pages
> 
> - Try to deactivate low pages first only if we are not short on swap
>   slots

4th diff:

Use the same mechanism to notify failed nowait & wait allocs to the page daemon.

As a side effect the page daemon now consider releasing inactive pages when
a nowait allocation for low pages failed.
---
 sys/uvm/uvm_pdaemon.c   | 25 ++++++++++++-------------
 sys/uvm/uvm_pmemrange.c | 11 +++++++----
 2 files changed, 19 insertions(+), 17 deletions(-)

diff --git a/sys/uvm/uvm_pdaemon.c b/sys/uvm/uvm_pdaemon.c
index 8af0ff181a6..614209393e2 100644
--- a/sys/uvm/uvm_pdaemon.c
+++ b/sys/uvm/uvm_pdaemon.c
@@ -194,7 +194,7 @@ uvmpd_tune(void)
  * recover at least some memory in the most restricted region (assumed
  * to be dma_constraint).
  */
-volatile int uvm_nowait_failed;
+struct uvm_pmalloc nowait_pma;
 
 static inline int
 uvmpd_pma_done(struct uvm_pmalloc *pma)
@@ -219,11 +219,19 @@ uvm_pageout(void *arg)
 	(void) spl0();
 	uvmpd_tune();
 
+	/*
+	 * XXX realistically, this is what our nowait callers probably
+	 * care about.
+	 */
+	nowait_pma.pm_constraint = dma_constraint;
+	nowait_pma.pm_size = (16 << PAGE_SHIFT); /* XXX */
+	nowait_pma.pm_flags = 0;
+
 	for (;;) {
 		long size;
 
 		uvm_lock_fpageq();
-		if (!uvm_nowait_failed && TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
+		if (TAILQ_EMPTY(&uvm.pmr_control.allocs)) {
 			msleep_nsec(&uvm.pagedaemon, &uvm.fpageqlock, PVM,
 			    "pgdaemon", INFSLP);
 			uvmexp.pdwoke++;
@@ -233,15 +241,7 @@ uvm_pageout(void *arg)
 			pma->pm_flags |= UVM_PMA_BUSY;
 			constraint = pma->pm_constraint;
 		} else {
-			if (uvm_nowait_failed) {
-				/*
-				 * XXX realistically, this is what our
-				 * nowait callers probably care about
-				 */
-				constraint = dma_constraint;
-				uvm_nowait_failed = 0;
-			} else
-				constraint = no_constraint;
+			constraint = no_constraint;
 		}
 		/* How many pages do we need to free during this round? */
 		shortage = uvmexp.freetarg - uvmexp.free + BUFPAGES_DEFICIT;
@@ -303,8 +303,7 @@ uvm_pageout(void *arg)
 			pma->pm_flags &= ~UVM_PMA_BUSY;
 			if (pma->pm_flags & UVM_PMA_FREED) {
 				pma->pm_flags &= ~UVM_PMA_LINKED;
-				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma,
-				    pmq);
+				TAILQ_REMOVE(&uvm.pmr_control.allocs, pma, pmq);
 				wakeup(pma);
 			}
 		}
diff --git a/sys/uvm/uvm_pmemrange.c b/sys/uvm/uvm_pmemrange.c
index 287308db735..207f2d31a79 100644
--- a/sys/uvm/uvm_pmemrange.c
+++ b/sys/uvm/uvm_pmemrange.c
@@ -841,7 +841,7 @@ uvm_pmr_extract_range(struct uvm_pmemrange *pmr, struct vm_page *pg,
  * recover at least some memory in the most restricted region (assumed
  * to be dma_constraint).
  */
-extern volatile int uvm_nowait_failed;
+extern struct uvm_pmalloc nowait_pma;
 
 /*
  * Acquire a number of pages.
@@ -1190,9 +1190,12 @@ fail:
 		    flags & UVM_PLA_FAILOK) == 0)
 			goto retry;
 		KASSERT(flags & UVM_PLA_FAILOK);
-	} else {
-		if (!(flags & UVM_PLA_NOWAKE)) {
-			uvm_nowait_failed = 1;
+	} else if (!(flags & UVM_PLA_NOWAKE)) {
+		struct uvm_pmalloc *pma = &nowait_pma;
+
+		if (!(nowait_pma.pm_flags & UVM_PMA_LINKED)) {
+			nowait_pma.pm_flags = UVM_PMA_LINKED;
+			TAILQ_INSERT_TAIL(&uvm.pmr_control.allocs, pma, pmq);
 			wakeup(&uvm.pagedaemon);
 		}
 	}
-- 
2.46.1