Faster M operation for the swapper to be great again

[Martin Pieuchot]

Diff below uses two tricks to make uvm_pagermapin/out() faster and less
likely to fail in OOM situations.

These functions are used to map buffers when swapping pages in/out and
when faulting on mmaped files.  robert@ even measured a 75% improvement
when populating pages related to files that aren't yet in the buffer
cache.

The first trick is to use the direct map when available.  I'm doing this
for single pages but km_alloc(9) also does that for single segment...
uvm_io() only maps one page at a time for the moment so this should be
enough.

The second trick is to use pmap_kenter_pa() which doesn't fail and is
faster.

With this changes the "freeze" happening on my server when entering many
pages to swap in OOM situation is much shorter and the machine becomes
quickly responsive.

ok?

Index: uvm/uvm_pager.c
===================================================================
RCS file: /cvs/src/sys/uvm/uvm_pager.c,v
retrieving revision 1.81
diff -u -p -r1.81 uvm_pager.c
--- uvm/uvm_pager.c     28 Jun 2022 19:07:40 -0000      1.81
+++ uvm/uvm_pager.c     30 Jun 2022 13:34:46 -0000
@@ -258,6 +258,16 @@ uvm_pagermapin(struct vm_page **pps, int
        vsize_t size;
        struct vm_page *pp;
 
+#ifdef __HAVE_PMAP_DIRECT
+       /* use direct mappings for single page */
+       if (npages == 1) {
+               KASSERT(pps[0]);
+               KASSERT(pps[0]->pg_flags & PG_BUSY);
+               kva = pmap_map_direct(pps[0]);
+               return kva;
+       }
+#endif
+
        prot = PROT_READ;
        if (flags & UVMPAGER_MAPIN_READ)
                prot |= PROT_WRITE;
@@ -273,14 +283,7 @@ uvm_pagermapin(struct vm_page **pps, int
                pp = *pps++;
                KASSERT(pp);
                KASSERT(pp->pg_flags & PG_BUSY);
-               /* Allow pmap_enter to fail. */
-               if (pmap_enter(pmap_kernel(), cva, VM_PAGE_TO_PHYS(pp),
-                   prot, PMAP_WIRED | PMAP_CANFAIL | prot) != 0) {
-                       pmap_remove(pmap_kernel(), kva, cva);
-                       pmap_update(pmap_kernel());
-                       uvm_pseg_release(kva);
-                       return 0;
-               }
+               pmap_kenter_pa(cva, VM_PAGE_TO_PHYS(pp), prot);
        }
        pmap_update(pmap_kernel());
        return kva;
@@ -294,8 +297,15 @@ uvm_pagermapin(struct vm_page **pps, int
 void
 uvm_pagermapout(vaddr_t kva, int npages)
 {
+#ifdef __HAVE_PMAP_DIRECT
+       /* use direct mappings for single page */
+       if (npages == 1) {
+               pmap_unmap_direct(kva);
+               return;
+       }
+#endif
 
-       pmap_remove(pmap_kernel(), kva, kva + ((vsize_t)npages << PAGE_SHIFT));
+       pmap_kremove(kva, (vsize_t)npages << PAGE_SHIFT);
        pmap_update(pmap_kernel());
        uvm_pseg_release(kva);