|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] RE: [Xen-ia64-devel] RFC: ptc.ga implementation for SMP-g
Hi Tristan,
Sorry for late response.
Let's first make sure our understanding of ptc.ga is correct.
Assume there are two LPs on one box, we call them LP1 and LP2,
And rr7=0x107 on LP1, rr7=0x207 on LP2,
And below code segment is executed at LP1
movl R2=0xe000000000000000;
mov r3=16<<2;
ptc.ga r2,r3;
The result on LP1 is, all tlb entries with rid=0x107 which are overlapped
with (0xe000000000000000~0xe000000000010000) are purged.
The result on LP2 should be same with that on LP1.
But according to this patch,
The result on LP2 is, all tlb entries with rid=0x207 which are overlapped
with (0xe000000000000000~0xe000000000010000) are purged.
This is not the desirable result of ptc.ga.
IMO, when using IPI to emulate ptc.ga, the related rid should be passed to
other LPs as a parameter.
Thanks,
Anthony
>From: Tristan
>Gingold
>Sent: 2006?4?3? 21:38
>To: xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
>Subject: [Xen-ia64-devel] RFC: ptc.ga implementation for SMP-g
>
>Hi,
>
>after the comments, here is my updated patch for ptc.ga
>Please comment it.
>
>With this patch, the page_flags are always written atomically. Ptc only clear
>it. This eliminates itc and ptc conflicts.
>
>The other conflict is use. This is within ia64_page_fault, between
>vcpu_translate and vcpu_itc_no_srlz. This part of code is protected by a
>flag + counter: At entry the flag is set and the counter increment, at exit
>the flag is reset. Ptc.ga waits if the flag is set and retries if the
>counter has changes.
>
>Tristan.
>
>diff -r ddc279c91502 xen/arch/ia64/vmx/vmmu.c
>--- a/xen/arch/ia64/vmx/vmmu.c Fri Mar 31 21:04:16 2006
>+++ b/xen/arch/ia64/vmx/vmmu.c Mon Apr 3 11:20:57 2006
>@@ -459,7 +459,7 @@
> va = PAGEALIGN(ifa, ps);
> index = vtr_find_overlap(vcpu, va, ps, DSIDE_TLB);
> if (index>=0) {
>- vcpu->arch.dtrs[index].p=0;
>+ vcpu->arch.dtrs[index].pte.p=0;
> index = vtr_find_overlap(vcpu, va, ps, DSIDE_TLB);
> }
> hcb = vmx_vcpu_get_vtlb(vcpu);
>@@ -476,7 +476,7 @@
> va = PAGEALIGN(ifa, ps);
> index = vtr_find_overlap(vcpu, va, ps, ISIDE_TLB);
> if (index>=0) {
>- vcpu->arch.itrs[index].p=0;
>+ vcpu->arch.itrs[index].pte.p=0;
> index = vtr_find_overlap(vcpu, va, ps, ISIDE_TLB);
> }
> hcb = vmx_vcpu_get_vtlb(vcpu);
>diff -r ddc279c91502 xen/arch/ia64/xen/process.c
>--- a/xen/arch/ia64/xen/process.c Fri Mar 31 21:04:16 2006
>+++ b/xen/arch/ia64/xen/process.c Mon Apr 3 11:20:57 2006
>@@ -290,12 +290,22 @@
> return;
> }
>
>+ /* Bit 0 must not be set. */
>+ BUG_ON (PSCBX(current, tlb_inuse) & 1);
>+ /* Set bit 0 and increment counter (from bit 1). */
>+ current->arch.tlb_inuse += 3;
>+
> fault = vcpu_translate(current,address,is_data,0,&pteval,&itir,&iha);
> if (fault == IA64_NO_FAULT) {
> pteval = translate_domain_pte(pteval,address,itir);
>
> vcpu_itc_no_srlz(current,is_data?2:1,address,pteval,-1UL,(itir>>2)&0x3
>f);
>+ /* Clear in_use flag. */
>+ current->arch.tlb_inuse--;
> return;
> }
>+ /* Clear in_use flag. */
>+ current->arch.tlb_inuse--;
>+
> if (!user_mode (regs)) {
> /* The fault occurs inside Xen. */
> if (!ia64_done_with_exception(regs)) {
>diff -r ddc279c91502 xen/arch/ia64/xen/vcpu.c
>--- a/xen/arch/ia64/xen/vcpu.c Fri Mar 31 21:04:16 2006
>+++ b/xen/arch/ia64/xen/vcpu.c Mon Apr 3 11:20:57 2006
>@@ -1276,7 +1276,7 @@
> // FIXME: also need to check && (!trp->key || vcpu_pkr_match(trp->key))
> static inline int vcpu_match_tr_entry(TR_ENTRY *trp, UINT64 ifa, UINT64 rid)
> {
>- return trp->p && trp->rid == rid
>+ return trp->pte.p && trp->rid == rid
> && ifa >= trp->vadr
> && ifa <= (trp->vadr + (1L << trp->ps) - 1);
> }
>@@ -1328,7 +1328,7 @@
> if (vcpu_quick_region_check(vcpu->arch.dtr_regions,address)) {
> for (trp = vcpu->arch.dtrs, i = NDTRS; i; i--, trp++) {
> if (vcpu_match_tr_entry(trp,address,rid)) {
>- *pteval = trp->page_flags;
>+ *pteval = trp->pte.val;
> *itir = trp->itir;
> tr_translate_count++;
> return IA64_NO_FAULT;
>@@ -1341,7 +1341,7 @@
> if (vcpu_quick_region_check(vcpu->arch.itr_regions,address)) {
> for (trp = vcpu->arch.itrs, i = NITRS; i; i--, trp++) {
> if (vcpu_match_tr_entry(trp,address,rid)) {
>- *pteval = trp->page_flags;
>+ *pteval = trp->pte.val;
> *itir = trp->itir;
> tr_translate_count++;
> return IA64_NO_FAULT;
>@@ -1354,7 +1354,7 @@
> // FIXME?: check dtlb for inst accesses too, else bad things happen?
> trp = &vcpu->arch.dtlb;
> if (/* is_data && */ vcpu_match_tr_entry(trp,address,rid)) {
>- if (vcpu->domain==dom0 && !in_tpa) *pteval = trp->page_flags;
>+ if (vcpu->domain==dom0 && !in_tpa) *pteval = trp->pte.val;
> else *pteval = vcpu->arch.dtlb_pte;
> *itir = trp->itir;
> dtlb_translate_count++;
>@@ -1691,24 +1691,27 @@
>
> static inline void vcpu_purge_tr_entry(TR_ENTRY *trp)
> {
>- trp->p = 0;
>+ trp->pte.val = 0;
> }
>
> static void vcpu_set_tr_entry(TR_ENTRY *trp, UINT64 pte, UINT64 itir, UINT64
>ifa)
> {
> UINT64 ps;
>+ union pte_flags new_pte;
>
> trp->itir = itir;
> trp->rid = VCPU(current,rrs[ifa>>61]) & RR_RID_MASK;
>- trp->p = 1;
> ps = trp->ps;
>- trp->page_flags = pte;
>- if (trp->pl < 2) trp->pl = 2;
>+ new_pte.val = pte;
>+ if (new_pte.pl < 2) new_pte.pl = 2;
> trp->vadr = ifa & ~0xfff;
> if (ps > 12) { // "ignore" relevant low-order bits
>- trp->ppn &= ~((1UL<<(ps-12))-1);
>+ new_pte.ppn &= ~((1UL<<(ps-12))-1);
> trp->vadr &= ~((1UL<<ps)-1);
> }
>+
>+ /* Atomic write. */
>+ trp->pte.val = new_pte.val;
> }
>
> IA64FAULT vcpu_itr_d(VCPU *vcpu, UINT64 slot, UINT64 pte,
>@@ -1877,12 +1880,15 @@
> struct ptc_ga_args {
> unsigned long vadr;
> unsigned long addr_range;
>+ struct vcpu *vcpu;
> };
>
> static void ptc_ga_remote_func (void *varg)
> {
> struct ptc_ga_args *args = (struct ptc_ga_args *)varg;
>- vhpt_flush_address (args->vadr, args->addr_range);
>+
>+ if (args->vcpu->processor == smp_processor_id ())
>+ vhpt_flush_address (args->vadr, args->addr_range);
> }
> #endif
>
>@@ -1900,6 +1906,7 @@
>
> args.vadr = vadr;
> args.addr_range = addr_range;
>+ args.vcpu = vcpu;
>
> /* This method is very conservative and should be optimized:
> - maybe IPI calls can be avoided,
>@@ -1909,22 +1916,40 @@
> */
> for_each_vcpu (d, v) {
> if (v != vcpu) {
>- /* Purge tc entry.
>- Can we do this directly ? Well, this is just a
>- single atomic write. */
>- vcpu_purge_tr_entry(&PSCBX(v,dtlb));
>- vcpu_purge_tr_entry(&PSCBX(v,itlb));
>+ unsigned int count;
>+ int proc;
>+
>+ do {
>+ /* Wait until the tlb is not used. */
>+ while ((count = PSCBX(vcpu, tlb_inuse)) & 1)
>+ cpu_relax ();
>+
>+ /* Purge tc entries. */
>+ vcpu_purge_tr_entry(&PSCBX(vcpu,dtlb));
>+ vcpu_purge_tr_entry(&PSCBX(vcpu,itlb));
>+
>+ /* Loop until the entry is not used. */
>+ } while (count != PSCBX(vcpu, tlb_inuse));
>+
>+ /* Here, the vcpu tlb is cleared.
>+ However the mapping is still present in the VHPT
>+ and the tc. */
> #ifdef VHPT_GLOBAL
>- /* Flush VHPT on remote processors.
>- FIXME: invalidate directly the entries? */
>- smp_call_function_single
>- (v->processor, &ptc_ga_remote_func,
>- &args, 0, 1);
>+ /* Flush VHPT on remote processors. */
>+ do {
>+ proc = v->processor;
>+ smp_call_function_single
>+ (v->processor, &ptc_ga_remote_func,
>+ &args, 0, 1);
>+ /* Try again if VCPU has migrated. */
>+ } while (proc != v->processor);
> #endif
> }
> }
> #endif
>
>+ /* No needs to watch tlb_inuse for local processor, since it is
>+ executing the ptc.ga. */
> #ifdef VHPT_GLOBAL
> vhpt_flush_address(vadr,addr_range);
> #endif
>diff -r ddc279c91502 xen/include/asm-ia64/domain.h
>--- a/xen/include/asm-ia64/domain.h Fri Mar 31 21:04:16 2006
>+++ b/xen/include/asm-ia64/domain.h Mon Apr 3 11:20:57 2006
>@@ -44,13 +44,19 @@
> offsetof(vcpu_info_t, evtchn_upcall_mask))
>
> struct arch_vcpu {
>-#if 1
> TR_ENTRY itrs[NITRS];
> TR_ENTRY dtrs[NDTRS];
> TR_ENTRY itlb;
> TR_ENTRY dtlb;
>- unsigned int itr_regions;
>- unsigned int dtr_regions;
>+
>+ /* Bit is set for regions having an itr/dtr. */
>+ unsigned short itr_regions;
>+ unsigned short dtr_regions;
>+
>+ /* Bit 0 means the itlb/dtlb is in use.
>+ bits 31-1 are a generation counter. There is only one writers. */
>+ volatile unsigned int tlb_inuse;
>+
> unsigned long itlb_pte;
> unsigned long dtlb_pte;
> unsigned long irr[4];
>@@ -61,7 +67,7 @@
> unsigned long domain_itm;
> unsigned long domain_itm_last;
> unsigned long xen_itm;
>-#endif
>+
> mapped_regs_t *privregs; /* save the state of vcpu */
> unsigned long metaphysical_rr0; // from arch_domain (so is
> pinned)
> unsigned long metaphysical_rr4; // from arch_domain (so is
> pinned)
>diff -r ddc279c91502 xen/include/asm-ia64/tlb.h
>--- a/xen/include/asm-ia64/tlb.h Fri Mar 31 21:04:16 2006
>+++ b/xen/include/asm-ia64/tlb.h Mon Apr 3 11:20:57 2006
>@@ -4,23 +4,24 @@
> #define NITRS 8
> #define NDTRS 8
>
>+union pte_flags {
>+ struct {
>+ unsigned long p : 1; // 0
>+ unsigned long : 1; // 1
>+ unsigned long ma : 3; // 2-4
>+ unsigned long a : 1; // 5
>+ unsigned long d : 1; // 6
>+ unsigned long pl : 2; // 7-8
>+ unsigned long ar : 3; // 9-11
>+ unsigned long ppn : 38; // 12-49
>+ unsigned long : 2; // 50-51
>+ unsigned long ed : 1; // 52
>+ };
>+ unsigned long val;
>+};
>+
> typedef struct {
>- union {
>- struct {
>- unsigned long p : 1; // 0
>- unsigned long : 1; // 1
>- unsigned long ma : 3; // 2-4
>- unsigned long a : 1; // 5
>- unsigned long d : 1; // 6
>- unsigned long pl : 2; // 7-8
>- unsigned long ar : 3; // 9-11
>- unsigned long ppn : 38; // 12-49
>- unsigned long : 2; // 50-51
>- unsigned long ed : 1; // 52
>- };
>- unsigned long page_flags;
>- };
>-
>+ volatile union pte_flags pte;
> union {
> struct {
> unsigned long : 2; // 0-1
_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel
|
 |
Lists.xenproject.org is hosted with RackSpace, monitoring our |