* [PATCH 01 of 12] xenalyze: Allow several summary handlers to register on a single vmexit
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
@ 2011-11-28 16:17 ` George Dunlap
2011-11-28 16:17 ` [PATCH 02 of 12] xenalyze: Reorganize cr trace handling George Dunlap
` (10 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:17 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
VMX tends to have several distinct kinds of handlers for a single
vmexit type; notably, EXCEPTION_NMI gets page faults, other traps,
emulations, and sometimes other kinds of traps as well; while for
SVM, each different trap has a different call type.
This patch introduces a signal-style callback mechanism that allows
several callbacks to be attached to a given exit reason. Each of these
callbacks will be called in the case of a summary event.
In order to distinguish different kinds of callbacks, the callback also
includes a value which is passed as-is to the function when it's called.
It's the size of a pointer so that it can be used as a pointer if necessary;
typically it will probably be a small integer value.
Each vmexit should only register a handler once with any given data value;
there is some syntactic sugar to make sure this happens.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 288573cb0f11 -r 2e85d4c4042e xenalyze.c
--- a/xenalyze.c Fri Oct 07 11:49:04 2011 +0100
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -1308,6 +1308,14 @@ struct pcpu_info;
#define RESYNCS_MAX 17
#define PF_XEN_FIXUP_UNSYNC_RESYNC_MAX 2
+struct hvm_data;
+
+struct hvm_summary_handler_node {
+ void (*handler)(struct hvm_data *, void* data);
+ void *data;
+ struct hvm_summary_handler_node *next;
+};
+
struct hvm_data {
/* Summary information */
int init;
@@ -1318,7 +1326,7 @@ struct hvm_data {
/* SVM / VMX compatibility. FIXME - should be global */
char ** exit_reason_name;
int exit_reason_max;
- void (* exit_reason_summary_handler[HVM_EXIT_REASON_MAX])(struct hvm_data *);
+ struct hvm_summary_handler_node *exit_reason_summary_handler_list[HVM_EXIT_REASON_MAX];
/* Information about particular exit reasons */
struct {
@@ -3011,23 +3019,58 @@ void hvm_short_summary(struct hvm_short_
}
}
-void hvm_set_summary_handler(struct hvm_data *h, void (*s)(struct hvm_data *h)) {
+/* Wrapper to try to make sure this is only called once per
+ * call site, rather than walking through the list each time */
+#define hvm_set_summary_handler(_h, _s, _d) \
+ do { \
+ static int done=0; \
+ int ret; \
+ if(!done) { \
+ if ((ret=__hvm_set_summary_handler(_h, _s, _d))) \
+ fprintf(stderr, "%s: hvm_set_summary_handler returned %d\n", \
+ __func__, ret); \
+ done=1; \
+ } \
+ } while(0)
+
+int __hvm_set_summary_handler(struct hvm_data *h, void (*s)(struct hvm_data *h, void*d), void*d) {
/* Set summary handler */
if(h->exit_reason < h->exit_reason_max)
{
- if(h->exit_reason_summary_handler[h->exit_reason])
+ struct hvm_summary_handler_node *p, **q;
+
+ /* Find the end of the list, checking to make sure there are no
+ * duplicates along the way */
+ q=&h->exit_reason_summary_handler_list[h->exit_reason];
+ p = *q;
+ while(p)
{
- if(h->exit_reason_summary_handler[h->exit_reason]!= s)
- fprintf(warn, "Strange, unexpected summary handler for exit_reason %s (%d!)\n",
- h->exit_reason_name[h->exit_reason], h->exit_reason);
- }
- else
- {
- h->exit_reason_summary_handler[h->exit_reason] = s;
- fprintf(warn, "Setting exit_reason %s (%d) summary handler\n",
- h->exit_reason_name[h->exit_reason], h->exit_reason);
- }
- }
+ if(p->handler == s && p->data == d)
+ {
+ fprintf(stderr, "%s: Unexpected duplicate handler %p,%p\n",
+ __func__, s, d);
+ error(ERR_STRICT, NULL);
+ return -EBUSY;
+ }
+ q=&p->next;
+ p=*q;
+ }
+
+ assert(p==NULL);
+
+ /* Insert the new handler */
+ p=malloc(sizeof(*p));
+ if (!p) {
+ fprintf(stderr, "%s: Malloc failed!\n", __func__);
+ error(ERR_SYSTEM, NULL);
+ }
+ p->handler=s;
+ p->data = d;
+ p->next=*q;
+ *q=p;
+ return 0;
+ }
+ return -EINVAL;
}
#define SIGN_EXTENDED_BITS (~((1ULL<<48)-1))
@@ -3040,7 +3083,7 @@ static inline int is_valid_addr64(unsign
return ((va & SIGN_EXTENDED_BITS) == 0);
}
-void hvm_pf_xen_summary(struct hvm_data *h) {
+void hvm_pf_xen_summary(struct hvm_data *h, void *d) {
int i,j, k;
for(i=0; i<PF_XEN_MAX; i++)
@@ -3314,7 +3357,7 @@ void hvm_pf_xen_postprocess(struct hvm_d
}
/* Set summary handler */
- hvm_set_summary_handler(h, hvm_pf_xen_summary);
+ hvm_set_summary_handler(h, hvm_pf_xen_summary, NULL);
}
if(opt.dump_cooked)
@@ -4061,7 +4104,7 @@ void cr3_dump_list(struct cr3_value_stru
free(qsort_array);
}
-void hvm_cr3_write_summary(struct hvm_data *h) {
+void hvm_cr3_write_summary(struct hvm_data *h, void *data) {
int j;
for(j=0; j<RESYNCS_MAX; j++)
@@ -4071,7 +4114,7 @@ void hvm_cr3_write_summary(struct hvm_da
" *[MAX] ");
}
-void hvm_cr_write_summary(struct hvm_data *h)
+void hvm_cr_write_summary(struct hvm_data *h, void *data)
{
int i;
@@ -4081,7 +4124,7 @@ void hvm_cr_write_summary(struct hvm_dat
" cr%d ", i);
switch(i) {
case 3:
- hvm_cr3_write_summary(h);
+ hvm_cr3_write_summary(h, NULL);
break;
}
}
@@ -4168,11 +4211,11 @@ void hvm_cr_write_postprocess(struct hvm
if ( opt.svm_mode ) {
/* For svm, only need a summary for cr3 */
if ( h->exit_reason == VMEXIT_CR3_WRITE )
- hvm_set_summary_handler(h, hvm_cr3_write_summary);
+ hvm_set_summary_handler(h, hvm_cr3_write_summary, NULL);
} else {
/* For vmx, real mode may cause EXNMI exits on cr accesses */
if ( h->exit_reason != EXIT_REASON_EXCEPTION_NMI )
- hvm_set_summary_handler(h, hvm_cr_write_summary);
+ hvm_set_summary_handler(h, hvm_cr_write_summary, NULL);
}
}
}
@@ -4229,7 +4272,7 @@ void hvm_cr_write_process(struct record_
}
/* msr_write */
-void hvm_msr_write_summary(struct hvm_data *h)
+void hvm_msr_write_summary(struct hvm_data *h, void *d)
{
}
@@ -4247,7 +4290,7 @@ void hvm_msr_write_postprocess(struct hv
}
/* Set summary handler */
- hvm_set_summary_handler(h, hvm_msr_write_summary);
+ hvm_set_summary_handler(h, hvm_msr_write_summary, NULL);
}
void hvm_msr_write_process(struct record_info *ri, struct hvm_data *h)
@@ -4274,7 +4317,7 @@ void hvm_msr_write_process(struct record
}
/* msr_read */
-void hvm_msr_read_summary(struct hvm_data *h)
+void hvm_msr_read_summary(struct hvm_data *h, void *d)
{
}
@@ -4292,7 +4335,7 @@ void hvm_msr_read_postprocess(struct hvm
}
/* Set summary handler */
- hvm_set_summary_handler(h, hvm_msr_read_summary);
+ hvm_set_summary_handler(h, hvm_msr_read_summary, NULL);
}
void hvm_msr_read_process(struct record_info *ri, struct hvm_data *h)
@@ -4318,7 +4361,7 @@ void hvm_msr_read_process(struct record_
h->post_process = hvm_msr_read_postprocess;
}
-void hvm_vmcall_summary(struct hvm_data *h)
+void hvm_vmcall_summary(struct hvm_data *h, void *d)
{
int i;
@@ -4343,6 +4386,7 @@ void hvm_vmcall_postprocess(struct hvm_d
else
update_summary(&h->summary.vmcall[HYPERCALL_MAX],
h->arc_cycles);
+ hvm_set_summary_handler(h, hvm_vmcall_summary, NULL);
}
}
@@ -4365,7 +4409,6 @@ void hvm_vmcall_process(struct record_in
}
if(opt.summary) {
- hvm_set_summary_handler(h, hvm_vmcall_summary);
h->inflight.vmcall.eax = r->eax;
h->post_process = hvm_vmcall_postprocess;
}
@@ -4391,7 +4434,7 @@ void hvm_inj_exc_process(struct record_i
}
-void hvm_intr_summary(struct hvm_data *h)
+void hvm_intr_summary(struct hvm_data *h, void *d)
{
int i;
@@ -4483,7 +4526,7 @@ void hvm_intr_process(struct hvm_data *h
if(opt.summary_info) {
if(opt.summary)
- hvm_set_summary_handler(h, hvm_intr_summary);
+ hvm_set_summary_handler(h, hvm_intr_summary, NULL);
if(vec < EXTERNAL_INTERRUPT_MAX)
h->summary.extint[vec]++;
@@ -5321,6 +5364,8 @@ void hvm_summary(struct hvm_data *h) {
printf("Exit reasons:\n");
for(i=0; i<h->exit_reason_max; i++) {
+ struct hvm_summary_handler_node *p;
+
if ( h->exit_reason_name[i] )
PRINT_SUMMARY(h->summary.exit_reason[i],
" %-20s ", h->exit_reason_name[i]);
@@ -5328,8 +5373,12 @@ void hvm_summary(struct hvm_data *h) {
PRINT_SUMMARY(h->summary.exit_reason[i],
" %20d ", i);
- if(h->exit_reason_summary_handler[i])
- h->exit_reason_summary_handler[i](h);
+ p=h->exit_reason_summary_handler_list[i];
+ while(p)
+ {
+ p->handler(h, p->data);
+ p=p->next;
+ }
}
printf("Guest interrupt counts:\n");
@@ -6238,7 +6287,7 @@ void shadow_process(struct pcpu_info *p)
if(sevt.minor <= PF_XEN_LAST_FAULT) {
h->inflight.pf_xen.pf_case = sevt.minor;
if(opt.summary) {
- hvm_set_summary_handler(h, hvm_pf_xen_summary);
+ hvm_set_summary_handler(h, hvm_pf_xen_summary, NULL);
}
}
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 02 of 12] xenalyze: Reorganize cr trace handling
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
2011-11-28 16:17 ` [PATCH 01 of 12] xenalyze: Allow several summary handlers to register on a single vmexit George Dunlap
@ 2011-11-28 16:17 ` George Dunlap
2011-11-28 16:18 ` [PATCH 03 of 12] xenalyze: Function-ize setting of h->post_process George Dunlap
` (9 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:17 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Reorganize the CR trace handling to take advantage of the
new summary handling. In particular, SVM has an exit per
CR register, while VMX has one exit for all CR accesses.
This allows them to coexist.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 2e85d4c4042e -r 6650e492be5d xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -4104,7 +4104,7 @@ void cr3_dump_list(struct cr3_value_stru
free(qsort_array);
}
-void hvm_cr3_write_summary(struct hvm_data *h, void *data) {
+void hvm_cr3_write_summary(struct hvm_data *h) {
int j;
for(j=0; j<RESYNCS_MAX; j++)
@@ -4116,18 +4116,12 @@ void hvm_cr3_write_summary(struct hvm_da
void hvm_cr_write_summary(struct hvm_data *h, void *data)
{
- int i;
-
- for(i=0; i<CR_MAX; i++)
- {
- PRINT_SUMMARY(h->summary.cr_write[i],
- " cr%d ", i);
- switch(i) {
- case 3:
- hvm_cr3_write_summary(h, NULL);
- break;
- }
- }
+ int cr=(int)data;
+
+ PRINT_SUMMARY(h->summary.cr_write[cr],
+ " cr%d ", cr);
+ if ( cr==3 )
+ hvm_cr3_write_summary(h);
}
void hvm_cr_write_postprocess(struct hvm_data *h)
@@ -4208,20 +4202,39 @@ void hvm_cr_write_postprocess(struct hvm
/* FIXME - deal with cr_read_summary */
if(h->exit_reason < h->exit_reason_max)
{
- if ( opt.svm_mode ) {
- /* For svm, only need a summary for cr3 */
- if ( h->exit_reason == VMEXIT_CR3_WRITE )
- hvm_set_summary_handler(h, hvm_cr3_write_summary, NULL);
- } else {
- /* For vmx, real mode may cause EXNMI exits on cr accesses */
- if ( h->exit_reason != EXIT_REASON_EXCEPTION_NMI )
- hvm_set_summary_handler(h, hvm_cr_write_summary, NULL);
+ /* Want a different "set" for each cr */
+ switch(h->inflight.cr_write.cr)
+ {
+#define case_cr(_x) \
+ case (_x): \
+ hvm_set_summary_handler(h, hvm_cr_write_summary, (void *)(_x)); \
+ break
+ case_cr(0);
+ case_cr(1);
+ case_cr(2);
+ case_cr(3);
+ case_cr(4);
+ case_cr(5);
+ case_cr(6);
+ case_cr(7);
+ case_cr(8);
+ case_cr(9);
+ case_cr(10);
+ case_cr(11);
+ case_cr(12);
+ case_cr(13);
+ case_cr(14);
+ case_cr(15);
+#undef case_cr
+ default:
+ fprintf(stderr, "Unexpected cr: %d\n", h->inflight.cr_write.cr);
+ error(ERR_SANITY, NULL);
+ break;
}
}
}
void hvm_exception_nmi_generic_postprocess(struct hvm_data *h);
-
void hvm_cr_write_process(struct record_info *ri, struct hvm_data *h)
{
union {
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 03 of 12] xenalyze: Function-ize setting of h->post_process
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
2011-11-28 16:17 ` [PATCH 01 of 12] xenalyze: Allow several summary handlers to register on a single vmexit George Dunlap
2011-11-28 16:17 ` [PATCH 02 of 12] xenalyze: Reorganize cr trace handling George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 04 of 12] xenalyze: Introduce generic summary functionality George Dunlap
` (8 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Introduce hvm_set_postprocess, so that we can make changes
to the setting and warning all in one place.
Special-case hvm_exception_nmi_generic_postprocess for now
to avoid regression; will get rid of it in the next c/s
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 6650e492be5d -r 47d436ca14d1 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -3073,6 +3073,20 @@ int __hvm_set_summary_handler(struct hvm
return -EINVAL;
}
+void hvm_exception_nmi_generic_postprocess(struct hvm_data *h);
+
+static int hvm_set_postprocess(struct hvm_data *h, void (*s)(struct hvm_data *h))
+{
+ if ( h->post_process == NULL
+ || h->post_process == hvm_exception_nmi_generic_postprocess )
+ {
+ h->post_process = s;
+ return 0;
+ }
+ else
+ return 1;
+}
+
#define SIGN_EXTENDED_BITS (~((1ULL<<48)-1))
#define HIGH_BIT(_v) ((_v) & (1ULL<<47))
static inline int is_valid_addr64(unsigned long long va)
@@ -3416,7 +3430,8 @@ void hvm_pf_xen_process(struct record_in
e->pt_index[4]);
}
- h->post_process = hvm_pf_xen_postprocess;
+ if ( hvm_set_postprocess(h, hvm_pf_xen_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
char * hvm_vlapic_icr_dest_shorthand_name[4] = {
@@ -3703,10 +3718,9 @@ void hvm_mmio_assist_process(struct reco
if((e->gpa & 0xfffff000) == 0xfee00000)
hvm_vlapic_handler(h);
- /* Catch MMIOs that don't go through the shadow code */
- if ( h->post_process == NULL )
- h->post_process = hvm_mmio_assist_postprocess;
-
+ /* Catch MMIOs that don't go through the shadow code; tolerate
+ * failures to set (probably shadow_mmio) */
+ hvm_set_postprocess(h, hvm_mmio_assist_postprocess);
}
void hvm_inj_virq_process(struct record_info *ri, struct hvm_data *h) {
@@ -3889,10 +3903,12 @@ void hvm_io_assist_process(struct record
if(mevt.write) {
h->inflight.io.is_write = 1;
- h->post_process = hvm_io_write_postprocess;
+ if ( hvm_set_postprocess(h, hvm_io_write_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
} else {
h->inflight.io.is_write = 0;
- h->post_process = hvm_io_read_postprocess;
+ if ( hvm_set_postprocess(h, hvm_io_read_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
if(opt.dump_all)
@@ -4234,7 +4250,6 @@ void hvm_cr_write_postprocess(struct hvm
}
}
-void hvm_exception_nmi_generic_postprocess(struct hvm_data *h);
void hvm_cr_write_process(struct record_info *ri, struct hvm_data *h)
{
union {
@@ -4258,13 +4273,11 @@ void hvm_cr_write_process(struct record_
h->inflight.cr_write.val = val = r->x32.val;
}
- /* In real mode, cr accesses may cause EXNMI vmexits */
- if ( !h->post_process
- || (!opt.svm_mode
- && h->post_process == hvm_exception_nmi_generic_postprocess) )
- h->post_process = hvm_cr_write_postprocess;
- else
- fprintf(warn, "Strange, h->postprocess already set!\n");
+ /* In vmx, in real mode, cr accesses may cause EXNMI vmexits.
+ * Account them under that heading; otherwise, complain */
+ if ( hvm_set_postprocess(h, hvm_cr_write_postprocess) )
+ fprintf(warn, "%s: Strange, h->postprocess already set!\n",
+ __func__);
if(opt.dump_all)
{
@@ -4326,7 +4339,8 @@ void hvm_msr_write_process(struct record
r->addr, r->val);
}
- h->post_process = hvm_msr_write_postprocess;
+ if ( hvm_set_postprocess(h, hvm_msr_write_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
/* msr_read */
@@ -4371,7 +4385,8 @@ void hvm_msr_read_process(struct record_
r->addr, r->val);
}
- h->post_process = hvm_msr_read_postprocess;
+ if ( hvm_set_postprocess(h, hvm_msr_read_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void hvm_vmcall_summary(struct hvm_data *h, void *d)
@@ -4421,10 +4436,10 @@ void hvm_vmcall_process(struct record_in
r->eax);
}
- if(opt.summary) {
- h->inflight.vmcall.eax = r->eax;
- h->post_process = hvm_vmcall_postprocess;
- }
+ h->inflight.vmcall.eax = r->eax;
+
+ if ( hvm_set_postprocess(h, hvm_vmcall_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void hvm_inj_exc_process(struct record_info *ri, struct hvm_data *h)
@@ -5640,8 +5655,8 @@ void shadow_emulate_process(struct recor
flag_string(e), e->flags,
e->pt_level, e->corresponding_va);
- h->post_process = shadow_emulate_postprocess;
-
+ if ( hvm_set_postprocess(h, shadow_emulate_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
struct shadow_emulate_other {
@@ -5776,7 +5791,8 @@ void shadow_unsync_process(struct record
e->pt_level,
e->corresponding_va);
- h->post_process = shadow_unsync_postprocess;
+ if ( hvm_set_postprocess(h, shadow_unsync_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
#endif
@@ -5802,7 +5818,8 @@ void shadow_emulate_other_process(struct
e->gfn,
e->va);
- h->post_process = shadow_fault_generic_postprocess;
+ if ( hvm_set_postprocess(h, shadow_fault_generic_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void shadow_fixup_postprocess(struct hvm_data *h)
@@ -5962,7 +5979,8 @@ void shadow_fixup_process(struct record_
flag_string(e));
}
- h->post_process = shadow_fixup_postprocess;
+ if ( hvm_set_postprocess(h, shadow_fixup_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void shadow_mmio_postprocess(struct hvm_data *h)
@@ -6058,7 +6076,8 @@ void shadow_mmio_process(struct record_i
(e->pf_case==PF_XEN_FAST_MMIO)?"fast ":"",
e->va);
- h->post_process = shadow_mmio_postprocess;
+ if ( hvm_set_postprocess(h, shadow_mmio_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void shadow_propagate_postprocess(struct hvm_data *h)
@@ -6150,7 +6169,8 @@ void shadow_propagate_process(struct rec
e->va, e->gl1e, e->flags,
flag_string(e));
- h->post_process = shadow_propagate_postprocess;
+ if ( hvm_set_postprocess(h, shadow_propagate_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void shadow_fault_generic_dump(unsigned int event, uint32_t *d, char *prefix,
@@ -6216,7 +6236,8 @@ void shadow_fault_generic_process(struct
"]", ri->dump_header);
h->inflight.pf_xen.pf_case = sevt.minor;
- h->post_process = shadow_fault_generic_postprocess;
+ if ( hvm_set_postprocess(h, shadow_fault_generic_postprocess) )
+ fprintf(warn, "%s: Strange, postprocess already set\n", __func__);
}
void shadow_resync_process(struct record_info *ri, struct hvm_data *h)
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 04 of 12] xenalyze: Introduce generic summary functionality
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (2 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 03 of 12] xenalyze: Function-ize setting of h->post_process George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 05 of 12] xenalyze: Handle new hvm_event traces George Dunlap
` (7 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Allow generic processing of hvm traces to generate summary
information, connected to specific exit reasons.
This makes hvm_generic_postprocess the default handler (which
means it can be overriden in hvm_set_postprocess()), and
also replaces hvm_exception_nmi_generic_postprocess with the
hvm_generic_postprocess.
Also add a handler for exits without an HVM trace, and make
a header for hvm_pf_xen_summary().
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 47d436ca14d1 -r 844117c5a51e xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -953,6 +953,7 @@ char * hvm_trap_name[HVM_TRAP_MAX] = {
enum {
+ HVM_EVENT_HANDLER_NONE = 0,
HVM_EVENT_HANDLER_PF_XEN = 1,
HVM_EVENT_HANDLER_PF_INJECT,
HVM_EVENT_HANDLER_INJ_EXC,
@@ -984,7 +985,7 @@ enum {
HVM_EVENT_HANDLER_MAX
};
char * hvm_event_handler_name[HVM_EVENT_HANDLER_MAX] = {
- NULL,
+ "(no handler)",
"pf_xen",
"pf_inject",
"inj_exc",
@@ -1343,6 +1344,7 @@ struct hvm_data {
struct event_cycle_summary cr_write[CR_MAX];
struct event_cycle_summary cr3_write_resyncs[RESYNCS_MAX+1];
struct event_cycle_summary vmcall[HYPERCALL_MAX+1];
+ struct event_cycle_summary generic[HVM_EVENT_HANDLER_MAX];
struct hvm_gi_struct {
int count;
struct cycle_summary runtime[GUEST_INTERRUPT_CASE_MAX];
@@ -3073,12 +3075,12 @@ int __hvm_set_summary_handler(struct hvm
return -EINVAL;
}
-void hvm_exception_nmi_generic_postprocess(struct hvm_data *h);
+void hvm_generic_postprocess(struct hvm_data *h);
static int hvm_set_postprocess(struct hvm_data *h, void (*s)(struct hvm_data *h))
{
if ( h->post_process == NULL
- || h->post_process == hvm_exception_nmi_generic_postprocess )
+ || h->post_process == hvm_generic_postprocess )
{
h->post_process = s;
return 0;
@@ -3099,26 +3101,35 @@ static inline int is_valid_addr64(unsign
void hvm_pf_xen_summary(struct hvm_data *h, void *d) {
int i,j, k;
-
+
+ printf(" page_fault\n");
for(i=0; i<PF_XEN_MAX; i++)
{
- PRINT_SUMMARY(h->summary.pf_xen[i],
- " %-25s ", pf_xen_name[i]);
+ if( pf_xen_name[i] )
+ {
+ PRINT_SUMMARY(h->summary.pf_xen[i],
+ " %-25s ", pf_xen_name[i]);
+ }
+ else
+ {
+ PRINT_SUMMARY(h->summary.pf_xen[i],
+ " [%23d] ", i);
+ }
switch(i){
case PF_XEN_NON_EMULATE:
for(j=0; j<PF_XEN_NON_EMUL_MAX; j++)
PRINT_SUMMARY(h->summary.pf_xen_non_emul[j],
- " *%-13s ", pf_xen_non_emul_name[j]);
+ " *%-13s ", pf_xen_non_emul_name[j]);
break;
case PF_XEN_EMULATE:
for(j=0; j<PF_XEN_EMUL_MAX; j++) {
PRINT_SUMMARY(h->summary.pf_xen_emul[j],
- " *%-13s ", pf_xen_emul_name[j]);
+ " *%-13s ", pf_xen_emul_name[j]);
if(j == PF_XEN_EMUL_EARLY_UNSHADOW) {
int k;
for(k=0; k<5; k++) {
PRINT_SUMMARY(h->summary.pf_xen_emul_early_unshadow[k],
- " +[%d] ", k);
+ " +[%d] ", k);
}
}
}
@@ -3126,14 +3137,14 @@ void hvm_pf_xen_summary(struct hvm_data
case PF_XEN_FIXUP:
for(j=0; j<PF_XEN_FIXUP_MAX; j++) {
PRINT_SUMMARY(h->summary.pf_xen_fixup[j],
- " *%-13s ", pf_xen_fixup_name[j]);
+ " *%-13s ", pf_xen_fixup_name[j]);
if(j == PF_XEN_FIXUP_UNSYNC ) {
for(k=0; k<PF_XEN_FIXUP_UNSYNC_RESYNC_MAX; k++) {
PRINT_SUMMARY(h->summary.pf_xen_fixup_unsync_resync[k],
- " +[%3d] ", k);
+ " +[%3d] ", k);
}
PRINT_SUMMARY(h->summary.pf_xen_fixup_unsync_resync[k],
- " +[max] ");
+ " +[max] ");
}
}
break;
@@ -4552,6 +4563,10 @@ void hvm_intr_process(struct hvm_data *h
update_eip(&h->v->d->interrupt_eip_list, rip, 0, 0, NULL);
}
+ /* Disable generic postprocessing */
+ /* FIXME: Do the summary stuff in a post-processor */
+ h->post_process = NULL;
+
if(opt.summary_info) {
if(opt.summary)
hvm_set_summary_handler(h, hvm_intr_summary, NULL);
@@ -4631,25 +4646,6 @@ void hvm_pf_inject_process(struct record
}
}
-void hvm_exception_nmi_generic_postprocess(struct hvm_data *h)
-{
- static int warned = 0;
- if(opt.dump_cooked)
- {
- printf(" %s exnmi no_handler\n",
- h->dump_header);
- }
-
- if(opt.summary_info)
- update_summary(&h->summary.pf_xen[PF_XEN_NO_HANDLER],
- h->arc_cycles);
-
- if(!warned) {
- warned++;
- fprintf(warn, "Strange, no handler for exnmi!\n");
- }
-}
-
void hvm_npf_process(struct record_info *ri, struct hvm_data *h)
{
struct {
@@ -4694,31 +4690,86 @@ void hvm_rdtsc_process(struct record_inf
h->last_rdtsc = r->tsc;
}
+void hvm_generic_summary(struct hvm_data *h, void *data)
+{
+ int evt = (int)data;
+
+ assert(evt < HVM_EVENT_HANDLER_MAX);
+
+ PRINT_SUMMARY(h->summary.generic[evt],
+ " %s ", hvm_event_handler_name[evt]);
+
+}
+
void hvm_generic_postprocess(struct hvm_data *h)
{
+ int evt = 0;
+ static unsigned registered[HVM_EVENT_HANDLER_MAX] = { 0 };
+
+ if ( h->inflight.generic.event )
+ evt = (h->inflight.generic.event - TRC_HVM_HANDLER) & ~TRC_64_FLAG;
+ else {
+ static unsigned warned[HVM_EXIT_REASON_MAX] = { 0 };
+ /* Some exits we don't expect a handler; just return */
+ if(opt.svm_mode)
+ {
+ }
+ else
+ {
+ switch(h->exit_reason)
+ {
+ /* These just need us to go through the return path */
+ case EXIT_REASON_PENDING_INTERRUPT:
+ case EXIT_REASON_TPR_BELOW_THRESHOLD:
+ /* Not much to log now; may need later */
+ case EXIT_REASON_WBINVD:
+ return;
+ default:
+ break;
+ }
+ }
+ if ( !warned[h->exit_reason] )
+ {
+ /* If we aren't a known exception, warn and log results */
+ fprintf(warn, "%s: Strange, exit %x missing a handler\n",
+ __func__, h->exit_reason);
+ warned[h->exit_reason]=1;
+ }
+ }
+
+ if ( evt > HVM_EVENT_HANDLER_MAX || evt < 0)
+ {
+ fprintf(warn, "%s: invalid hvm event %x\n",
+ __func__, h->inflight.generic.event);
+ error(ERR_RECORD, NULL);
+ return;
+ }
+
if(opt.summary_info) {
+ update_summary(&h->summary.generic[evt],
+ h->arc_cycles);
+
+ /* NB that h->exit_reason may be 0, so we offset by 1 */
+ if ( registered[evt] )
+ {
+ static unsigned warned[HVM_EXIT_REASON_MAX] = { 0 };
+ if ( registered[evt] != h->exit_reason+1 && !warned[h->exit_reason])
+ {
+ fprintf(warn, "%s: HVM evt %x in %x and %x!\n",
+ __func__, evt, registered[evt]-1, h->exit_reason);
+ warned[h->exit_reason]=1;
+ }
+ }
+ else
+ {
+ int ret;
+ if((ret=__hvm_set_summary_handler(h, hvm_generic_summary, (void *)evt)))
+ fprintf(stderr, "%s: hvm_set_summary_handler returned %d\n",
+ __func__, ret);
+ registered[evt]=h->exit_reason+1;
+ }
/* HLT checked at hvm_vmexit_close() */
}
-
- if(opt.dump_cooked)
- {
- int i, evt = h->event_handler;
-
- if(evt < HVM_EVENT_HANDLER_MAX)
- printf(" %s %s [",
- h->dump_header,
- hvm_event_handler_name[evt]);
- else
- printf(" %s %d [",
- h->dump_header,
- evt);
-
- for(i=0; i<4; i++) {
- printf(" %x", h->inflight.generic.d[i]);
- }
-
- printf(" ]\n");
- }
}
void hvm_generic_dump(struct record_info *ri, char * prefix)
@@ -4878,18 +4929,15 @@ needs_vmexit:
case TRC_HVM_CLTS:
case TRC_HVM_LMSW:
case TRC_HVM_LMSW64:
- default:
- if ( h->post_process != NULL )
- fprintf(warn, "Strange, h->postprocess already set!\n");
- /* Guest NMI should expect h->postprocess to be set.
- FIXME: Handle NMI specially. */
+ case TRC_HVM_NMI:
case TRC_HVM_CR_READ:
case TRC_HVM_CR_READ64:
- case TRC_HVM_NMI:
- /* FIXME: Really should check to make sure post_process is NULL or the exnmi default... */
+ default:
+ if ( h->post_process != hvm_generic_postprocess )
+ fprintf(warn, "%s: Strange, h->postprocess set!\n",
+ __func__);
h->inflight.generic.event = ri->event;
bcopy(h->d, h->inflight.generic.d, sizeof(unsigned int) * 4);
- h->post_process = hvm_generic_postprocess;
if(opt.dump_all)
{
hvm_generic_dump(ri, "]");
@@ -5168,18 +5216,8 @@ void hvm_vmexit_process(struct record_in
h->wrmap_bf = 0;
h->short_summary_done = 0;
- h->post_process = NULL;
- if(!opt.svm_mode)
- {
- switch(h->exit_reason)
- {
- case EXIT_REASON_EXCEPTION_NMI:
- h->post_process = hvm_exception_nmi_generic_postprocess;
- break;
- default:
- ;
- }
- }
+ h->post_process = hvm_generic_postprocess;
+ h->inflight.generic.event = 0;
{
struct time_struct t;
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 05 of 12] xenalyze: Handle new hvm_event traces
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (3 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 04 of 12] xenalyze: Introduce generic summary functionality George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 06 of 12] xenalyze: Relocate pio and mmio enumaration structs to their own sub-struct George Dunlap
` (6 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 844117c5a51e -r 463ac7003722 trace.h
--- a/trace.h Mon Nov 28 16:16:23 2011 +0000
+++ b/trace.h Mon Nov 28 16:16:23 2011 +0000
@@ -162,6 +162,9 @@
#define TRC_HVM_RDTSC (TRC_HVM_HANDLER + 0x1a)
#define TRC_HVM_INTR_WINDOW (TRC_HVM_HANDLER + 0x20)
#define TRC_HVM_NPF (TRC_HVM_HANDLER + 0x21)
+#define TRC_HVM_REALMODE_EMULATE (TRC_HVM_HANDLER + 0x22)
+#define TRC_HVM_TRAP (TRC_HVM_HANDLER + 0x23)
+#define TRC_HVM_TRAP_DEBUG (TRC_HVM_HANDLER + 0x24)
#define TRC_HVM_IOPORT_WRITE (TRC_HVM_HANDLER + 0x216)
#define TRC_HVM_IOMEM_WRITE (TRC_HVM_HANDLER + 0x217)
diff -r 844117c5a51e -r 463ac7003722 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -982,6 +982,9 @@ enum {
HVM_EVENT_RDTSC,
HVM_EVENT_INTR_WINDOW=0x20, /* Oops... skipped 0x1b-1f */
HVM_EVENT_NPF,
+ HVM_EVENT_REALMODE_EMULATE,
+ HVM_EVENT_TRAP,
+ HVM_EVENT_TRAP_DEBUG,
HVM_EVENT_HANDLER_MAX
};
char * hvm_event_handler_name[HVM_EVENT_HANDLER_MAX] = {
@@ -1014,6 +1017,9 @@ char * hvm_event_handler_name[HVM_EVENT_
"rdtsc",
[HVM_EVENT_INTR_WINDOW]="intr_window",
"npf",
+ "realmode_emulate",
+ "trap",
+ "trap_debug",
};
enum {
@@ -4930,6 +4936,9 @@ needs_vmexit:
case TRC_HVM_LMSW:
case TRC_HVM_LMSW64:
case TRC_HVM_NMI:
+ case TRC_HVM_REALMODE_EMULATE:
+ case TRC_HVM_TRAP:
+ case TRC_HVM_TRAP_DEBUG:
case TRC_HVM_CR_READ:
case TRC_HVM_CR_READ64:
default:
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 06 of 12] xenalyze: Relocate pio and mmio enumaration structs to their own sub-struct
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (4 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 05 of 12] xenalyze: Handle new hvm_event traces George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 07 of 12] xenalyze: Handle MMIO records from different vmexits George Dunlap
` (5 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
In preparation for the next patch, which does some MMIO reorganizing.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 463ac7003722 -r 910605f7ade3 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -1361,7 +1361,9 @@ struct hvm_data {
/* IPI Latency */
struct event_cycle_summary ipi_latency;
int ipi_count[256];
- struct io_address *mmio, *pio;
+ struct {
+ struct io_address *mmio, *pio;
+ } io;
} summary;
/* In-flight accumulation information */
@@ -3674,7 +3676,7 @@ void enumerate_mmio(struct hvm_data *h)
struct pf_xen_extra *e = &h->inflight.pf_xen;
if ( e->mmio_data_valid )
- update_io_address(&h->summary.mmio, e->gpa, e->mmio_is_write, h->arc_cycles, e->va);
+ update_io_address(&h->summary.io.mmio, e->gpa, e->mmio_is_write, h->arc_cycles, e->va);
}
void hvm_mmio_assist_postprocess(struct hvm_data *h)
@@ -3874,7 +3876,7 @@ void hvm_io_write_postprocess(struct hvm
h->inflight.io.val);
}
if(opt.with_pio_enumeration)
- update_io_address(&h->summary.pio, h->inflight.io.port, 1, h->arc_cycles, 0);
+ update_io_address(&h->summary.io.pio, h->inflight.io.port, 1, h->arc_cycles, 0);
}
void hvm_io_read_postprocess(struct hvm_data *h)
@@ -3886,7 +3888,7 @@ void hvm_io_read_postprocess(struct hvm_
h->inflight.io.val);
}
if(opt.with_pio_enumeration)
- update_io_address(&h->summary.pio, h->inflight.io.port, 0, h->arc_cycles, 0);
+ update_io_address(&h->summary.io.pio, h->inflight.io.port, 0, h->arc_cycles, 0);
if(opt.scatterplot_io && h->inflight.io.port == opt.scatterplot_io_port)
scatterplot_vs_time(h->exit_tsc, P.now - h->exit_tsc);
}
@@ -5492,8 +5494,8 @@ void hvm_summary(struct hvm_data *h) {
if(h->summary.ipi_count[i])
printf(" [%3d] %10d\n",
i, h->summary.ipi_count[i]);
- hvm_io_address_summary(h->summary.pio, "IO address summary:");
- hvm_io_address_summary(h->summary.mmio, "MMIO address summary:");
+ hvm_io_address_summary(h->summary.io.pio, "IO address summary:");
+ hvm_io_address_summary(h->summary.io.mmio, "MMIO address summary:");
}
/* ---- Shadow records ---- */
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 07 of 12] xenalyze: Handle MMIO records from different vmexits
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (5 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 06 of 12] xenalyze: Relocate pio and mmio enumaration structs to their own sub-struct George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 08 of 12] xenalyze: Add option to skip vga range in MMIO enumeration George Dunlap
` (4 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
handle_mmio() is called in Xen as a generic way to do emulation;
in VMX, this at least includes not only EXCEPTION_NMI (page faults
in shadow mode) and NPF (nested page fault in HAP mode), but also
APIC_ACCESS and sometimes even CR_ACCESS.
This patch allows xenalyze to process and summarize mmio records
for different vmexits separately.
The change that causes the biggest number of line changes is pulling
the MMIO information out of pf_xen_extra and in its own struct, which
is kept separate from the inflight union.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 910605f7ade3 -r c9f583c65e07 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -1252,6 +1252,20 @@ char * pf_xen_name[PF_XEN_MAX] = {
#define CORR_VA_INVALID (0ULL-1)
+enum {
+ NONPF_MMIO_APIC,
+ NONPF_MMIO_NPF,
+ NONPF_MMIO_UNKNOWN,
+ NONPF_MMIO_MAX
+};
+
+struct mmio_info {
+ unsigned long long gpa;
+ unsigned long long va; /* Filled only by shadow */
+ unsigned data;
+ unsigned data_valid:1, is_write:1;
+};
+
struct pf_xen_extra {
unsigned long long va;
union {
@@ -1302,9 +1316,7 @@ struct pf_xen_extra {
/* Flags */
unsigned corr_valid:1,
corr_is_kernel:1,
- va_is_kernel:1,
- mmio_data_valid:1,
- mmio_is_write:1;
+ va_is_kernel:1;
};
struct pcpu_info;
@@ -1351,6 +1363,7 @@ struct hvm_data {
struct event_cycle_summary cr3_write_resyncs[RESYNCS_MAX+1];
struct event_cycle_summary vmcall[HYPERCALL_MAX+1];
struct event_cycle_summary generic[HVM_EVENT_HANDLER_MAX];
+ struct event_cycle_summary mmio[NONPF_MMIO_MAX];
struct hvm_gi_struct {
int count;
struct cycle_summary runtime[GUEST_INTERRUPT_CASE_MAX];
@@ -1367,33 +1380,37 @@ struct hvm_data {
} summary;
/* In-flight accumulation information */
- union {
- struct {
- unsigned port:31,
- is_write:1;
- unsigned int val;
- } io;
- struct pf_xen_extra pf_xen;
- struct {
- unsigned cr;
- unsigned long long val;
- int repromote;
- } cr_write;
- struct {
- unsigned addr;
- unsigned long long val;
- } msr;
- struct {
- unsigned int event;
- uint32_t d[4];
- } generic;
- struct {
- unsigned eax;
- } vmcall;
- struct {
- unsigned vec;
- } intr;
- } inflight;
+ struct {
+ union {
+ struct {
+ unsigned port:31,
+ is_write:1;
+ unsigned int val;
+ } io;
+ struct pf_xen_extra pf_xen;
+ struct {
+ unsigned cr;
+ unsigned long long val;
+ int repromote;
+ } cr_write;
+ struct {
+ unsigned addr;
+ unsigned long long val;
+ } msr;
+ struct {
+ unsigned int event;
+ uint32_t d[4];
+ } generic;
+ struct {
+ unsigned eax;
+ } vmcall;
+ struct {
+ unsigned vec;
+ } intr;
+ };
+ /* MMIO gets its separate area, since many exits may use it */
+ struct mmio_info mmio;
+ }inflight;
int resyncs;
void (*post_process)(struct hvm_data *);
tsc_t exit_tsc, arc_cycles, entry_tsc;
@@ -3309,6 +3326,7 @@ void pf_preprocess(struct pf_xen_extra *
void hvm_pf_xen_preprocess(unsigned event, struct hvm_data *h) {
struct pf_xen_extra *e = &h->inflight.pf_xen;
+ struct mmio_info *m = &h->inflight.mmio;
struct hvm_pf_xen_record *r = (typeof(r))h->d;
if(event == TRC_HVM_PF_XEN64)
@@ -3325,7 +3343,7 @@ void hvm_pf_xen_preprocess(unsigned even
e->error_code = r->x32.error_code;
}
- if(e->mmio_data_valid)
+ if(m->data_valid)
e->pf_case = PF_XEN_MMIO;
else
{
@@ -3363,8 +3381,11 @@ void hvm_pf_xen_postprocess(struct hvm_d
struct pf_xen_extra *e = &h->inflight.pf_xen;
if(opt.summary_info) {
- update_summary(&h->summary.pf_xen[e->pf_case],
- h->arc_cycles);
+ if(e->pf_case)
+ update_summary(&h->summary.pf_xen[e->pf_case],
+ h->arc_cycles);
+ else
+ fprintf(warn, "Strange, pf_case 0!\n");
switch(e->pf_case)
{
case PF_XEN_EMULATE:
@@ -3530,7 +3551,7 @@ struct outstanding_ipi *find_vec(struct
void hvm_vlapic_icr_handler(struct hvm_data *h)
{
- struct pf_xen_extra *e = &h->inflight.pf_xen;
+ struct mmio_info *m = &h->inflight.mmio;
union {
unsigned int val;
struct {
@@ -3544,7 +3565,7 @@ void hvm_vlapic_icr_handler(struct hvm_d
_res2:2,
dest_shorthand:2;
};
- } icr = { .val = e->data };
+ } icr = { .val = m->data };
void ipi_send(struct vcpu_data *ov, int vec)
{
@@ -3590,7 +3611,7 @@ void hvm_vlapic_icr_handler(struct hvm_d
o->count);
}
- if(e->mmio_is_write) {
+ if(m->is_write) {
if(opt.dump_all || opt.dump_cooked) {
printf(" [vla] d%dv%d icr vec %d %s\n",
h->v->d->did, h->v->vid,
@@ -3657,9 +3678,9 @@ void hvm_vlapic_eoi_handler(struct hvm_d
void hvm_vlapic_handler(struct hvm_data *h)
{
- struct pf_xen_extra *e = &h->inflight.pf_xen;
-
- switch(e->gpa) {
+ struct mmio_info *m = &h->inflight.mmio;
+
+ switch(m->gpa) {
case 0xfee00300:
hvm_vlapic_icr_handler(h);
break;
@@ -3673,14 +3694,56 @@ void hvm_vlapic_handler(struct hvm_data
/* Also called by shadow_mmio_postprocess */
void enumerate_mmio(struct hvm_data *h)
{
- struct pf_xen_extra *e = &h->inflight.pf_xen;
-
- if ( e->mmio_data_valid )
- update_io_address(&h->summary.io.mmio, e->gpa, e->mmio_is_write, h->arc_cycles, e->va);
+ struct mmio_info *m = &h->inflight.mmio;
+
+ if ( m->data_valid )
+ update_io_address(&h->summary.io.mmio, m->gpa, m->is_write, h->arc_cycles, m->va);
+}
+
+void hvm_mmio_summary(struct hvm_data *h, void *data)
+{
+ int reason=(int)data;
+
+ PRINT_SUMMARY(h->summary.mmio[reason],
+ " mmio ");
}
void hvm_mmio_assist_postprocess(struct hvm_data *h)
{
+ int reason;
+
+ switch(h->exit_reason)
+ {
+ case VMEXIT_NPF:
+ case EXIT_REASON_EPT_VIOLATION:
+ reason=NONPF_MMIO_NPF;
+ hvm_set_summary_handler(h, hvm_mmio_summary, (void *)reason);
+ break;
+ case EXIT_REASON_APIC_ACCESS:
+ reason=NONPF_MMIO_APIC;
+ hvm_set_summary_handler(h, hvm_mmio_summary, (void *)reason);
+ break;
+ default:
+ {
+ static int warned = 0;
+ if (!warned)
+ {
+ fprintf(stderr, "%s: Strange, MMIO with unexpected exit reason %d\n",
+ __func__, h->exit_reason);
+ warned=1;
+ }
+ reason=NONPF_MMIO_UNKNOWN;
+ hvm_set_summary_handler(h, hvm_mmio_summary, (void *)reason);
+ break;
+ }
+ }
+
+ if(opt.summary_info)
+ {
+ update_summary(&h->summary.mmio[reason],
+ h->arc_cycles);
+ }
+
if ( opt.with_mmio_enumeration )
enumerate_mmio(h);
}
@@ -3688,8 +3751,7 @@ void hvm_mmio_assist_postprocess(struct
#define HVM_IO_ASSIST_WRITE 0x200
void hvm_mmio_assist_process(struct record_info *ri, struct hvm_data *h)
{
- struct pf_xen_extra *e = &h->inflight.pf_xen;
-
+ struct mmio_info *m = &h->inflight.mmio;
union {
struct {
unsigned int gpa;
@@ -3711,30 +3773,32 @@ void hvm_mmio_assist_process(struct reco
} mevt = { .event = ri->event };
if(mevt.x64) {
- e->gpa = r->x64.gpa;
- e->data = r->x64.data;
+ m->gpa = r->x64.gpa;
+ m->data = r->x64.data;
if(ri->extra_words*(sizeof(unsigned int))==sizeof(r->x64))
- e->mmio_data_valid=1;
+ m->data_valid=1;
} else {
- e->gpa = r->x32.gpa;
- e->data = r->x32.data;
+ m->gpa = r->x32.gpa;
+ m->data = r->x32.data;
if(ri->extra_words*(sizeof(unsigned int))==sizeof(r->x32))
- e->mmio_data_valid=1;
- }
-
- e->mmio_is_write = mevt.write;
+ m->data_valid=1;
+ }
+
+ m->is_write = mevt.write;
if(opt.dump_all)
{
- if(e->mmio_data_valid)
- printf("]%s mmio_assist %c gpa %llx data %x\n", h->dump_header,
- mevt.write?'w':'r', e->gpa, e->data);
+ if(m->data_valid)
+ printf("]%s mmio_assist %c gpa %llx data %x\n",
+ h->dump_header,
+ mevt.write?'w':'r',
+ m->gpa, m->data);
else
printf("]%s mmio_assist %c gpa %llx (no data)\n", h->dump_header,
- mevt.write?'w':'r', e->gpa);
- }
-
- if((e->gpa & 0xfffff000) == 0xfee00000)
+ mevt.write?'w':'r', m->gpa);
+ }
+
+ if((m->gpa & 0xfffff000) == 0xfee00000)
hvm_vlapic_handler(h);
/* Catch MMIOs that don't go through the shadow code; tolerate
@@ -5857,6 +5921,7 @@ void shadow_emulate_other_process(struct
e->gfn = r.gfn;
e->va = r.va;
+ e->pf_case = sevt.minor;
pf_preprocess(e, h->v->guest_paging_levels);
@@ -6035,10 +6100,15 @@ void shadow_fixup_process(struct record_
void shadow_mmio_postprocess(struct hvm_data *h)
{
struct pf_xen_extra *e = &h->inflight.pf_xen;
+ struct mmio_info *m = &h->inflight.mmio;
if ( opt.summary_info )
{
- update_summary(&h->summary.pf_xen[e->pf_case], h->arc_cycles);
+ if(e->pf_case)
+ update_summary(&h->summary.pf_xen[e->pf_case],
+ h->arc_cycles);
+ else
+ fprintf(warn, "Strange, pf_case 0!\n");
hvm_update_short_summary(h, HVM_SHORT_SUMMARY_MMIO);
}
@@ -6048,23 +6118,23 @@ void shadow_mmio_postprocess(struct hvm_
if ( opt.dump_cooked )
{
- if(e->mmio_data_valid)
+ if(m->data_valid)
printf(" %s %smmio %s va %llx eip %llx%s gpa %llx data %x\n",
h->dump_header,
(e->pf_case==PF_XEN_FAST_MMIO)?"fast ":"",
- e->mmio_is_write?"write":"read",
+ m->is_write?"write":"read",
e->va,
h->rip, find_symbol(h->rip),
- e->gpa,
- e->data);
+ m->gpa,
+ m->data);
else
printf(" %s %smmio %s va %llx eip %llx%s gpa %llx (no data)\n",
h->dump_header,
(e->pf_case==PF_XEN_FAST_MMIO)?"fast ":"",
- e->mmio_is_write?"write":"read",
- e->va,
+ m->is_write?"write":"read",
+ m->va,
h->rip, find_symbol(h->rip),
- e->gpa);
+ m->gpa);
}
}
@@ -6072,6 +6142,7 @@ void shadow_mmio_postprocess(struct hvm_
void shadow_mmio_process(struct record_info *ri, struct hvm_data *h)
{
struct pf_xen_extra *e = &h->inflight.pf_xen;
+ struct mmio_info *m = &h->inflight.mmio;
union {
/* for PAE, guest_l1e may be 64 while guest_va may be 32;
so put it first for alignment sake. */
@@ -6104,7 +6175,7 @@ void shadow_mmio_process(struct record_i
error(ERR_RECORD, ri);
return;
}
- e->va = r->gpl2.va;
+ e->va = m->va = r->gpl2.va;
break;
case 4:
if(sizeof(r->gpl4) != ri->extra_words * 4)
@@ -6115,7 +6186,7 @@ void shadow_mmio_process(struct record_i
error(ERR_RECORD, ri);
return;
}
- e->va = r->gpl4.va;
+ e->va = m->va = r->gpl4.va;
break;
}
@@ -6135,7 +6206,11 @@ void shadow_propagate_postprocess(struct
if ( opt.summary_info )
{
- update_summary(&h->summary.pf_xen[e->pf_case], h->arc_cycles);
+ if(e->pf_case)
+ update_summary(&h->summary.pf_xen[e->pf_case],
+ h->arc_cycles);
+ else
+ fprintf(warn, "Strange, pf_case 0!\n");
hvm_update_short_summary(h, HVM_SHORT_SUMMARY_PROPAGATE);
}
@@ -6255,20 +6330,20 @@ void shadow_fault_generic_dump(unsigned
void shadow_fault_generic_postprocess(struct hvm_data *h)
{
- union shadow_event sevt = { .event = h->inflight.generic.event };
+ struct pf_xen_extra *e = &h->inflight.pf_xen;
+ if ( e->pf_case < PF_XEN_NOT_SHADOW || e->pf_case > PF_XEN_LAST_FAULT )
+ {
+ fprintf(warn, "%s: Strange, unexpected case %d\n",
+ __func__, e->pf_case);
+ return;
+ }
+
if(opt.summary_info) {
- update_summary(&h->summary.pf_xen[sevt.minor],
- h->arc_cycles);
+ update_summary(&h->summary.pf_xen[e->pf_case],
+ h->arc_cycles);
hvm_update_short_summary(h, HVM_SHORT_SUMMARY_PROPAGATE);
}
-
- if(opt.dump_cooked)
- {
- shadow_fault_generic_dump(h->inflight.generic.event,
- h->inflight.generic.d,
- " ", h->dump_header);
- }
}
void shadow_fault_generic_process(struct record_info *ri, struct hvm_data *h)
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 08 of 12] xenalyze: Add option to skip vga range in MMIO enumeration
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (6 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 07 of 12] xenalyze: Handle MMIO records from different vmexits George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 09 of 12] xenalyze: Sort cr3 enumerated values by start time George Dunlap
` (3 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Emulations to the VGA range during boot can be extremely numerous,
sometimes causing out-of-memory errors. Add an option to
skip the VGA range (0xa0000-0xbffff) during enumeration, and
enable it by default.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r c9f583c65e07 -r 3c6da152c844 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -48,6 +48,15 @@ struct array_struct {
int count;
};
+#define warn_once(_x...) \
+ do { \
+ static int _w=1; \
+ if ( _w ) { \
+ _w=0; \
+ fprintf(warn, ##_x); \
+ } \
+ } while(0) \
+
/* -- Global variables -- */
struct {
int fd;
@@ -160,6 +169,7 @@ struct {
with_mmio_enumeration:1,
with_interrupt_eip_enumeration:1,
show_default_domain_summary:1,
+ mmio_enumeration_skip_vga:1,
progress:1,
svm_mode:1,
summary:1,
@@ -234,6 +244,7 @@ struct {
.with_mmio_enumeration = 0,
.with_interrupt_eip_enumeration = 0,
.show_default_domain_summary = 0,
+ .mmio_enumeration_skip_vga = 1,
.progress = 0,
.svm_mode = 0,
.summary = 0,
@@ -3692,10 +3703,21 @@ void hvm_vlapic_handler(struct hvm_data
}
/* Also called by shadow_mmio_postprocess */
+#define MMIO_VGA_START (0xa0000)
+#define MMIO_VGA_END (0xbffff)
void enumerate_mmio(struct hvm_data *h)
{
struct mmio_info *m = &h->inflight.mmio;
+ /* Skip vga area */
+ if ( opt.mmio_enumeration_skip_vga
+ && m->gpa >= MMIO_VGA_START
+ && m->gpa < MMIO_VGA_END)
+ {
+ warn_once("WARNING: Not enumerationg MMIO in VGA range. Use --mmio-enumeration-skip-vga=0 to override.\n");
+ return;
+ }
+
if ( m->data_valid )
update_io_address(&h->summary.io.mmio, m->gpa, m->is_write, h->arc_cycles, m->va);
}
@@ -9464,6 +9486,7 @@ enum {
OPT_INTERVAL_DOMAIN_GRANT_MAPS,
/* Summary info */
OPT_SHOW_DEFAULT_DOMAIN_SUMMARY,
+ OPT_MMIO_ENUMERATION_SKIP_VGA,
OPT_SAMPLE_SIZE,
OPT_REPORT_PCPU,
/* Guest info */
@@ -9649,6 +9672,14 @@ error_t cmd_parser(int key, char *arg, s
argp_usage(state);
break;
}
+ case OPT_MMIO_ENUMERATION_SKIP_VGA:
+ {
+ char * inval;
+ opt.mmio_enumeration_skip_vga = (int)strtol(arg, &inval, 0);
+ if( inval == arg )
+ argp_usage(state);
+ break;
+ }
case OPT_SCATTERPLOT_INTERRUPT_EIP:
{
char * inval;
@@ -10182,6 +10213,12 @@ const struct argp_option cmd_opts[] = {
.group = OPT_GROUP_SUMMARY,
.doc = "Show default domain information on summary", },
+ { .name = "mmio-enumeration-skip-vga",
+ .key = OPT_MMIO_ENUMERATION_SKIP_VGA,
+ .arg = "[0|1]",
+ .group = OPT_GROUP_SUMMARY,
+ .doc = "Control whether we enumerate MMIO accesses to the VGA area, which can be extremly high during boot. Default: 0", },
+
{ .name = "sample-size",
.key = OPT_SAMPLE_SIZE,
.arg = "size",
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 09 of 12] xenalyze: Sort cr3 enumerated values by start time
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (7 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 08 of 12] xenalyze: Add option to skip vga range in MMIO enumeration George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 10 of 12] xenalyze: Enable more cr3 output George Dunlap
` (2 subsequent siblings)
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Add a new sort method to sort by start time, rather than by total
time, and make it the default.
Really should make it command-line switchable, but I'll leave that
improvement until later.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 3c6da152c844 -r 46c366d71c74 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -4144,7 +4144,7 @@ void cr3_dump_list(struct cr3_value_stru
struct cr3_value_struct **qsort_array;
int i, N=0;
- int cr3_compare(const void *_a, const void *_b) {
+ int cr3_compare_total(const void *_a, const void *_b) {
struct cr3_value_struct *a=*(typeof(&a))_a;
struct cr3_value_struct *b=*(typeof(&a))_b;
@@ -4161,6 +4161,18 @@ void cr3_dump_list(struct cr3_value_stru
return -1;
}
+ int cr3_compare_start(const void *_a, const void *_b) {
+ struct cr3_value_struct *a=*(typeof(&a))_a;
+ struct cr3_value_struct *b=*(typeof(&a))_b;
+
+ if(a->first_time > b->first_time)
+ return 1;
+ else if(b->first_time == a->first_time)
+ return 0;
+ else
+ return -1;
+ }
+
if(!head)
return;
@@ -4180,7 +4192,7 @@ void cr3_dump_list(struct cr3_value_stru
/* Sort the array by time */
qsort(qsort_array, N, sizeof(struct eip_list_struct *),
- cr3_compare);
+ cr3_compare_start);
/* WARNING: don't use N after this point unless you copy this variable */
#if 0
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 10 of 12] xenalyze: Enable more cr3 output
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (8 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 09 of 12] xenalyze: Sort cr3 enumerated values by start time George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 11 of 12] xenalyze: Optimize pcpu_string George Dunlap
2011-11-28 16:18 ` [PATCH 12 of 12] xenalyze: Make max_active_pcpu calculation smarter George Dunlap
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
Some of the cr3 ouptut was #if 0'd out. Put it back in.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r 46c366d71c74 -r dab5ba28630a xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -4220,8 +4220,7 @@ void cr3_dump_list(struct cr3_value_stru
print_cycle_percent_summary(&p->hv_time, p->run_time, desc);
hvm_short_summary(&p->hvm, p->run_time, " + ");
-#if 0
- printf(" Seen: %4lu.%09lu-%4lu.%09lu switch %d flush %d\n",
+ printf(" Seen: %4u.%09u-%4u.%09u switch %d flush %d\n",
first.s, first.ns,
last.s, last.ns,
p->switch_count, p->flush_count);
@@ -4231,7 +4230,6 @@ void cr3_dump_list(struct cr3_value_stru
p->destroy.switch_count,
p->destroy.fixup_user,
p->destroy.emulate_corr_user);
-#endif
}
free(qsort_array);
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 11 of 12] xenalyze: Optimize pcpu_string
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (9 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 10 of 12] xenalyze: Enable more cr3 output George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
2011-11-28 16:18 ` [PATCH 12 of 12] xenalyze: Make max_active_pcpu calculation smarter George Dunlap
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
pcpu_string() is called on every line of the dump output,
and the basic loop is O(N) with the number of pcpus. There
was an optimization that would be O(1), but it had a bug so that
it never actually went into effect.
Put the optimization into effect, so that only the bits that need
to change are changed.
This means, however, that when other things change, like power
states or lost records, that other parts of the string need to
be redrawn. Put in hooks to redraw when these kinds of things
change.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r dab5ba28630a -r b7d88a8858b7 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -1942,6 +1942,7 @@ struct {
/* Function prototypes */
char * pcpu_string(int pcpu);
+void pcpu_string_draw(struct pcpu_info *p);
void process_generic(struct record_info *ri);
void dump_generic(FILE *f, struct record_info *ri);
ssize_t __read_record(int fd, struct trace_record *rec, loff_t offset);
@@ -6970,6 +6971,7 @@ void vcpu_prev_update(struct pcpu_info *
set:
pcpu_runstate_update(p, tsc);
p->current = NULL;
+ pcpu_string_draw(p);
runstate_update(prev, new_runstate, tsc);
}
@@ -7051,6 +7053,7 @@ void vcpu_next_update(struct pcpu_info *
next->p = p;
p->current = next;
+ pcpu_string_draw(p);
p->time.tsc = tsc;
p->lost_record.seen_valid_schedule = 1;
}
@@ -7089,6 +7092,7 @@ void vcpu_start(struct pcpu_info *p, str
runstate_update(v, RUNSTATE_RUNNING, p->first_tsc);
p->time.tsc = p->first_tsc;
p->current = v;
+ pcpu_string_draw(p);
v->p = p;
}
@@ -7524,6 +7528,7 @@ void sched_default_vcpu_activate(struct
v->runstate.state = RUNSTATE_RUNNING;
p->current = v;
+ pcpu_string_draw(p);
}
void sched_default_domain_init(void)
@@ -7825,7 +7830,10 @@ void pm_process(struct pcpu_info *p) {
ri->dump_header,
ri->d[0]);
if ( ri->d[0] <= CSTATE_MAX )
+ {
p->power_state=ri->d[0];
+ pcpu_string_draw(p);
+ }
break;
case TRC_PM_IDLE_EXIT:
if (opt.dump_all )
@@ -7837,6 +7845,7 @@ void pm_process(struct pcpu_info *p) {
printf("Strange, pm_idle_end %d, power_state %d!\n",
ri->d[0], p->power_state);
p->power_state = 0;
+ pcpu_string_draw(p);
break;
default:
if(opt.dump_all || opt.dump_cooked) {
@@ -8437,6 +8446,7 @@ void process_lost_records(struct pcpu_in
p->lost_record.active = 1;
p->lost_record.tsc = first_tsc;
+ pcpu_string_draw(p);
{
/* Any vcpu which is not actively running may be scheduled on the
@@ -8520,6 +8530,7 @@ void process_lost_records_end(struct pcp
p->lost_record.active = 0;
+ pcpu_string_draw(p);
P.lost_cpus--;
if(P.lost_cpus < 0) {
fprintf(warn, "ERROR: lost_cpus fell below 0 for pcpu %d!\n",
@@ -9230,38 +9241,59 @@ ssize_t read_record(int fd, struct pcpu_
return ri->size;
}
-/* WARNING not thread-safe */
+/*
+ * This funciton gets called for every record when doing dump. Try to
+ * make it efficient by changing the minimum amount from the last
+ * call. Do this by:
+ * - Keeping track of the last pcpu called, so we can just set that to -
+ * - Keeping track of how many pcpus we've "drawn", and only "drawing" new ones
+ * - Updating the current one
+ *
+ * FIXME: Need to deal with pcpu states changing...
+ *
+ * WARNING not thread-safe
+ */
+
+char __pcpu_string[MAX_CPUS+1] = { 0 };
+void pcpu_string_draw(struct pcpu_info *p)
+{
+ char *s = __pcpu_string;
+ int i=p->pid;
+
+ if(p->lost_record.active)
+ s[i]='l';
+ else if (!p->current)
+ s[i]=' ';
+ else if (p->current->d->did == DEFAULT_DOMAIN)
+ s[i]='.';
+ else if (p->current->d->did == IDLE_DOMAIN)
+ {
+ if ( opt.dump_show_power_states )
+ s[i]=p->power_state+'0';
+ else
+ s[i]='-';
+ }
+ else
+ s[i]='|';
+}
+
char * pcpu_string(int pcpu)
{
- static char s[MAX_CPUS+1] = { 0 };
+ char *s = __pcpu_string;
static int max_active_pcpu=-1, last_pcpu=-1;
assert(P.max_active_pcpu < MAX_CPUS);
assert(pcpu <= P.max_active_pcpu);
if(last_pcpu >= 0)
- s[last_pcpu]='-';
+ pcpu_string_draw(P.pcpu+last_pcpu);
if(P.max_active_pcpu > max_active_pcpu)
{
int i;
- for(i=max_active_pcpu + 1; i<= P.max_active_pcpu; i++) {
- if(P.pcpu[i].lost_record.active)
- s[i]='l';
- else if (!P.pcpu[i].current)
- s[i]=' ';
- else if (P.pcpu[i].current->d->did == DEFAULT_DOMAIN)
- s[i]='.';
- else if (P.pcpu[i].current->d->did == IDLE_DOMAIN)
- {
- if ( opt.dump_show_power_states )
- s[i]=P.pcpu[i].power_state+'0';
- else
- s[i]='-';
- }
- else
- s[i]='|';
- }
+ for(i=max_active_pcpu + 1; i<= P.max_active_pcpu; i++)
+ pcpu_string_draw(P.pcpu+i);
+ max_active_pcpu=P.max_active_pcpu;
}
s[pcpu]='x';
^ permalink raw reply [flat|nested] 13+ messages in thread* [PATCH 12 of 12] xenalyze: Make max_active_pcpu calculation smarter
2011-11-28 16:17 [PATCH 00 of 12] xenalyze: Lots of updates George Dunlap
` (10 preceding siblings ...)
2011-11-28 16:18 ` [PATCH 11 of 12] xenalyze: Optimize pcpu_string George Dunlap
@ 2011-11-28 16:18 ` George Dunlap
11 siblings, 0 replies; 13+ messages in thread
From: George Dunlap @ 2011-11-28 16:18 UTC (permalink / raw)
To: xen-devel; +Cc: george.dunlap
oprofile reports that during summary runs, over 30% of the
time processing is spent in choose_next_record. As a first
step towards eliminating the loop found in that function,
take out the calculation of P.max_active_pcpu, and make it
smarter.
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
diff -r b7d88a8858b7 -r 223e8ad4c755 xenalyze.c
--- a/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
+++ b/xenalyze.c Mon Nov 28 16:16:23 2011 +0000
@@ -8675,6 +8675,21 @@ void deactivate_pcpu(struct pcpu_info *p
lose_vcpu(p->current, p->last_tsc);
}
p->active = 0;
+ if ( p->pid == P.max_active_pcpu )
+ {
+ int i, max_active_pcpu = -1;
+ for(i=0; i<=P.max_active_pcpu; i++)
+ {
+ if(!P.pcpu[i].active)
+ continue;
+
+ max_active_pcpu = i;
+ }
+ P.max_active_pcpu = max_active_pcpu;
+ fprintf(warn, "%s: Setting max_active_pcpu to %d\n",
+ __func__, max_active_pcpu);
+ }
+
}
/* Helper function to process tsc-related record info */
@@ -9304,17 +9319,20 @@ char * pcpu_string(int pcpu)
struct pcpu_info * choose_next_record(void)
{
- int i, max_active_pcpu = -1;
+ int i;
struct pcpu_info * p, *min_p=NULL;
loff_t min_offset = 0;
+ /* Need to:
+ * - find the pcpu with the lowest order_tsc
+ * - Find the lowest file offset
+ */
for(i=0; i<=P.max_active_pcpu; i++)
{
p = P.pcpu+i;
if(!p->active)
continue;
- max_active_pcpu = i;
if(!min_p || p->order_tsc < min_p->order_tsc)
min_p = p;
@@ -9322,13 +9340,11 @@ struct pcpu_info * choose_next_record(vo
min_offset = p->file_offset;
}
- P.max_active_pcpu = max_active_pcpu;
-
if(opt.progress && min_offset >= G.progress.update_offset)
progress_update(min_offset);
/* If there are active pcpus, make sure we chose one */
- assert(min_p || (max_active_pcpu==-1));
+ assert(min_p || (P.max_active_pcpu==-1));
return min_p;
}
^ permalink raw reply [flat|nested] 13+ messages in thread