Date: Thu, 17 Apr 2014 09:14:52 +0000 (UTC) From: "Cherry G. Mathew" <cherry@FreeBSD.org> To: src-committers@freebsd.org, svn-src-projects@freebsd.org Subject: svn commit: r264583 - projects/amd64_xen_pv/sys/amd64/xen Message-ID: <201404170914.s3H9Eq2v047387@svn.freebsd.org>
next in thread | raw e-mail | index | archive | help
Author: cherry Date: Thu Apr 17 09:14:52 2014 New Revision: 264583 URL: http://svnweb.freebsd.org/changeset/base/264583 Log: Enable mmu_map managed amd64 maps to be destroyed after use. This involves testing that the backing pages are no longer in use, and delinking them in the page hierarchy, before they are freed back to the kernel. This is the last supporting patch for what would have been the mmu_map(9) MI api, which I have now abandoned in favour of the DMAP based native approach to page table management. This will make the amd64/xen/pmap.c code much easier to maintain. WARNING: This KPI will be deprecated! Approved by: gibbs (implicit) Modified: projects/amd64_xen_pv/sys/amd64/xen/mmu_map.c projects/amd64_xen_pv/sys/amd64/xen/mmu_map.h Modified: projects/amd64_xen_pv/sys/amd64/xen/mmu_map.c ============================================================================== --- projects/amd64_xen_pv/sys/amd64/xen/mmu_map.c Thu Apr 17 06:40:11 2014 (r264582) +++ projects/amd64_xen_pv/sys/amd64/xen/mmu_map.c Thu Apr 17 09:14:52 2014 (r264583) @@ -188,6 +188,126 @@ struct mmu_map_index { * Make optional on DEBUG */ }; +/* Scan a page table for valid entries. Return true if any found */ +/* XXX: this is terribly slow - we need a queue/flush model to queue up frees and "flush" them all out together. */ +/* XXX; support huge pages */ +static bool +pdpt_empty(pdp_entry_t *pdpt) +{ + int i; + + KASSERT(pdpt != NULL, + ("Invalid pdpt\n")); + + for (i = 0;i < NPDPEPG;i++) { + if (pdpt[i] & PG_V) { + return false; + } + } + + return true; +} + +static bool +pdt_empty(pd_entry_t *pdt) +{ + int i; + + KASSERT(pdt != NULL, + ("Invalid pdt\n")); + + for (i = 0;i < NPDEPG;i++) { + if (pdt[i] & PG_V) { + return false; + } + } + + return true; +} + +static bool +pt_empty(pt_entry_t *pt) +{ + + int i; + + KASSERT(pt != NULL, + ("Invalid pt\n")); + + for (i = 0;i < NPTEPG;i++) { + if (pt[i] & PG_V) { + return false; + } + } + + return true; +} + +/* Free a table and update its entry in the hierarchy */ +static void +mmu_map_pdpt_free(struct mmu_map_index *pti, uintptr_t va) +{ + vm_paddr_t pml4tep_ma; + pml4_entry_t *pml4tep; + + KASSERT(pti->sanity == SANE, ("%s: pti insane!", __func__)); + + pml4tep = &pti->pml4t[pml4t_index(va)]; + + pml4tep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pml4tep) +); + xen_queue_pt_update(pml4tep_ma, 0); + xen_flush_queue(); + + /* The PDPT is empty. Free it and zero the + * pointer + */ + //if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pdpt); + pti->pdpt = NULL; +} + +static void +mmu_map_pdt_free(struct mmu_map_index *pti, uintptr_t va) +{ + vm_paddr_t pdptep_ma; + pdp_entry_t *pdptep; + + KASSERT(pti->sanity == SANE, ("%s: pti insane!", __func__)); + + pdptep = &pti->pdpt[pdpt_index(va)]; + + pdptep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pdptep)); + xen_queue_pt_update(pdptep_ma, 0); + xen_flush_queue(); + + /* The PDT is empty. Free it and zero the + * pointer + */ + // if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pdt); + pti->pdt = NULL; +} + +static void +mmu_map_pt_free(struct mmu_map_index *pti, uintptr_t va) +{ + vm_paddr_t pdtep_ma; + pd_entry_t *pdtep; + + KASSERT(pti->sanity == SANE, ("%s: pti insane!", __func__)); + + pdtep = &pti->pdt[pdt_index(va)]; + + /* Zap the backing PDT entry */ + pdtep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pdtep)); + xen_queue_pt_update(pdtep_ma, 0); + xen_flush_queue(); + + /* The PT is empty. Free it and zero the + * pointer */ + //if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pt); + pti->pt = NULL; +} + size_t mmu_map_t_size(void) { @@ -222,8 +342,7 @@ void mmu_map_t_fini(void *addr) pti->sanity = 0; if (mb->free != NULL) { - /* XXX: go through PT hierarchy and free + unmap - * unused tables */ + /* XXX */ } } @@ -312,10 +431,12 @@ mmu_map_inspect_va(struct pmap *pm, void return true; } - +#include <sys/proc.h> bool mmu_map_hold_va(struct pmap *pm, void *addr, uintptr_t va) { + bool alloced = false; + KASSERT(addr != NULL && pm != NULL, ("NULL arg(s) given")); struct mmu_map_index *pti = addr; @@ -324,7 +445,6 @@ mmu_map_hold_va(struct pmap *pm, void *a /* Reset pti */ pti->pml4t = pti->pdpt = pti->pdt = pti->pt = 0; - bool alloced = false; /* Did we have to alloc backing pages ? */ vm_paddr_t pt; pti->pml4t = pmap_get_pml4t(pm); @@ -416,12 +536,10 @@ mmu_map_release_va(struct pmap *pm, void if (pti->pdt != NULL) { /* Zap pdte */ - pd_entry_t *pdtep; - vm_paddr_t pdtep_ma; - - pdtep = &pti->pdt[pdt_index(va)]; - if (pti->pt == NULL) { + pd_entry_t *pdtep; + pdtep = &pti->pdt[pdt_index(va)]; + KASSERT(*pdtep == 0, ("%s(%d): mmu state machine out of sync!\n", __func__, __LINE__)); } else { @@ -443,33 +561,23 @@ mmu_map_release_va(struct pmap *pm, void return; } - /* We can free the PT only after the PDT entry is zapped */ - if (memcchr(pti->pt, 0, PAGE_SIZE) == NULL) { - /* Zap the backing PDT entry */ - pdtep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pdtep)); - xen_queue_pt_update(pdtep_ma, 0); - xen_flush_queue(); - - /* The PT is empty. Free it and zero the - * pointer */ - if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pt); - pti->pt = NULL; + if (pt_empty(pti->pt)) { + mmu_map_pt_free(pti, va); +#ifdef notyet + pm->pm_stats.resident_count--; +#endif } - } KASSERT(pti->pdpt != 0, ("Invalid pdpt\n")); } if (pti->pdpt != NULL) { /* Zap pdpte */ - - pdp_entry_t *pdptep; - vm_paddr_t pdptep_ma; - - pdptep = &pti->pdpt[pdpt_index(va)]; - if (pti->pdt == NULL) { + pdp_entry_t *pdptep; + pdptep = &pti->pdpt[pdpt_index(va)]; + KASSERT(*pdptep == 0, ("%s(%d): mmu state machine out of sync!\n", __func__, __LINE__)); } @@ -491,28 +599,20 @@ mmu_map_release_va(struct pmap *pm, void return; } - /* We can free the PDT only after the PDPT entry is zapped */ - if (memcchr(pti->pdt, 0, PAGE_SIZE) == NULL) { - pdptep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pdptep)); - xen_queue_pt_update(pdptep_ma, 0); - xen_flush_queue(); - - /* The PDT is empty. Free it and zero the - * pointer - */ - if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pdt); - pti->pdt = NULL; + if (pdt_empty(pti->pdt)) { + mmu_map_pdt_free(pti, va); +#ifdef notyet + pm->pm_stats.resident_count--; +#endif } KASSERT(pti->pml4t != 0, ("Invalid pml4t\n")); } if (pti->pml4t != NULL) { /* Zap pml4te */ - pml4_entry_t *pml4tep; - vm_paddr_t pml4tep_ma; - - pml4tep = &pti->pml4t[pml4t_index(va)]; - if (pti->pdpt == NULL) { + pml4_entry_t *pml4tep; + pml4tep = &pti->pml4t[pml4t_index(va)]; + KASSERT(*pml4tep == 0, ("%s(%d): mmu state machine out of sync!\n", __func__, __LINE__)); } @@ -534,17 +634,12 @@ mmu_map_release_va(struct pmap *pm, void return; } - if (memcchr(pti->pdpt, 0, PAGE_SIZE) == NULL) { - pml4tep_ma = xpmap_ptom(pti->ptmb.vtop((uintptr_t)pml4tep) -); - xen_queue_pt_update(pml4tep_ma, 0); - xen_flush_queue(); + if (pdpt_empty(pti->pdpt)) { + mmu_map_pdpt_free(pti, va); +#ifdef notyet + pm->pm_stats.resident_count--; +#endif - /* The PDPT is empty. Free it and zero the - * pointer - */ - if (pti->ptmb.free) pti->ptmb.free((uintptr_t)pti->pdpt); - pti->pdpt = NULL; } } @@ -553,5 +648,6 @@ mmu_map_release_va(struct pmap *pm, void * higher level aliasing issues across pmaps and vcpus that * can't be addressed here. */ + } Modified: projects/amd64_xen_pv/sys/amd64/xen/mmu_map.h ============================================================================== --- projects/amd64_xen_pv/sys/amd64/xen/mmu_map.h Thu Apr 17 06:40:11 2014 (r264582) +++ projects/amd64_xen_pv/sys/amd64/xen/mmu_map.h Thu Apr 17 09:14:52 2014 (r264583) @@ -133,10 +133,18 @@ bool mmu_map_inspect_va(struct pmap *, m */ bool mmu_map_hold_va(struct pmap *, mmu_map_t, vm_offset_t); -/* Optionally release resources after tear down of a va->pa mapping */ +/* Optionally release resources needed to "inspect" a va->pa mapping + * Note: This function is advisory in nature. The va->pa mapping + * itself may *NOT* be modified. + */ void mmu_map_release_va(struct pmap *, mmu_map_t, vm_offset_t); /* + * Release all resources allocated via mmu_map_hold_va(). + */ +void mmu_map_destroy(struct pmap *); + +/* * Machine dependant "view" into the page table hierarchy FSM. * On amd64, there are four tables that are consulted for a va->pa * translation. This information may be extracted by the MD functions
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201404170914.s3H9Eq2v047387>