hide struct start_info from PVH code
Subsystems (2)
cover hide struct start_info from PVH code
While working on xenstore-stubdom live update support I came across some really unpleasant code using struct start_info even when running in PVH mode.
Especially suspend/resume was doing some crazy stuff, including overwriting the initial struct hvm_start_info partially.
This series is hiding struct start_info from PVH code. This didn't result in any runtime error, but in future I'll need to access struct hvm_start_info some time after boot, so I need to get that fixed.
Tested especially with xl suspend/resume and xl save/restore in PV and in PVH mode.
Juergen Gross (6):
suspend: drop crazy pfn handling for PVH
setup: drop parameter from map_shared_info()
x86/pv: remove global start_info
console: drop parameter from get_console()
xenbus: drop parameter from get_xenbus()
x86/pv: make start_info_ptr PV-only
arch/arm/setup.c | 6 ++-- arch/x86/setup.c | 69 ++++++++++++-------------------------------- console.c | 10 +++---- hypervisor.c | 2 +- include/console.h | 2 +- include/hypervisor.h | 14 ++------- include/xenbus.h | 4 +-- xenbus.c | 13 ++++----- 8 files changed, 39 insertions(+), 81 deletions(-)
Signature · 1 line
2.43.0
PATCH 1/6 suspend: drop crazy pfn handling for PVH
Suspend for PVH is doing some crazy stuff with the Xenstore and console PFNs:
- In arch_pre_suspend() the PFNs are "translated" to MFNs, even if that translation is a nop.
- The result is stored via start_info_ptr, which points to a struct start_info only in case of PV mode.
- After resuming or a cancelled suspend, the PFNs are retrieved again and "translated".
- The "MFNs" are stored again via start_info_ptr, which is still not pointing at a struct start_info, and from where the data is never read again after that.
Drop all of that completely.
--- arch/x86/setup.c | 22 ---------------------- 1 file changed, 22 deletions(-)
arch/x86/setup.c +0 -22
@@ -227,17 +227,6 @@ void arch_pre_suspend(void)virt_to_pfn (mfn_to_virt (start_info_ptr->store_mfn)); start_info_ptr->console.domU.mfn =virt_to_pfn (mfn_to_virt (start_info_ptr->console.domU.mfn)); -#else -uint64_t store_v; -uint64_t console_v; - -if (hvm_get_parameter (HVM_PARAM_STORE_PFN , &store_v) ) -BUG (); - start_info_ptr->store_mfn = store_v; - -if (hvm_get_parameter (HVM_PARAM_CONSOLE_PFN , &console_v) ) -BUG (); - start_info_ptr->console.domU.mfn = console_v;#endif unmap_shared_info ();
@@ -262,17 +251,6 @@ void arch_post_suspend(int canceled) } else {memcpy (&start_info, start_info_ptr,sizeof (start_info_t)); } -#else -uint64_t store_v; -uint64_t console_v; - -if (hvm_get_parameter (HVM_PARAM_STORE_PFN , &store_v)) -BUG (); - start_info_ptr->store_mfn =pfn_to_mfn (store_v); - -if (hvm_get_parameter (HVM_PARAM_CONSOLE_PFN , &console_v)) -BUG (); - start_info_ptr->console.domU.mfn =pfn_to_mfn (console_v);#endif HYPERVISOR_shared_info =map_shared_info ((void *) start_info_ptr);
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 1/6] suspend: drop crazy pfn handling for PVH
Quoted 13 lines · Juergen Gross
Suspend for PVH is doing some crazy stuff with the Xenstore and console PFNs:
- In
arch_pre_suspend()the PFNs are "translated" to MFNs, even if that translation is a nop.
- The result is stored viastart_info_ptr, which points to a structstart_infoonly in case of PV mode.
- After resuming or a cancelled suspend, the PFNs are retrieved again and "translated".
- The "MFNs" are stored again viastart_info_ptr, which is still not pointing at a structstart_info, and from where the data is never read again after that.
Quoted 3 lines
Drop all of that completely.
Signed-off-by Juergen Gross <[email protected]>
start_info_ptr points at an hvm_start_info. Assuming I counted properly, accessing ->console/store is writing past the end of the struct.
PATCH 2/6 setup: drop parameter from map_shared_info()
In reply to Juergen GrossThe parameter of map_shared_info() is evaluated in PV mode only, and it is always the pointer to the initial struct start_info from boot time.
By setting start_info_ptr a little bit earlier the PV variant of map_shared_info() can just use start_info_ptr instead of a parameter.
--- arch/arm/setup.c | 2 +- arch/x86/setup.c | 12 ++++++------ hypervisor.c | 2 +- include/hypervisor.h | 2 +- 4 files changed, 9 insertions(+), 9 deletions(-)
arch/arm/setup.c +1 -1
@@ -38,7 +38,7 @@ void arch_init(void *dtb_pointer, uint32_t physical_offset) device_tree = dtb_pointer;/* Map shared_info page */ - HYPERVISOR_shared_info =map_shared_info (NULL ); + HYPERVISOR_shared_info =map_shared_info ();get_console (NULL );get_xenbus (NULL );
arch/x86/setup.c +6 -6
@@ -79,11 +79,10 @@ static inline void sse_init(void) {#ifdef CONFIG_PARAVIRT #define hpc_init () -shared_info_t *map_shared_info (void *p) +shared_info_t *map_shared_info (void ) {int rc; -start_info_t *si = p; -unsigned long pa = si->shared_info; +unsigned long pa = start_info_ptr->shared_info;if ( (rc =HYPERVISOR_update_va_mapping ((unsigned long )shared_info,__pte (pa |7 ),UVMF_INVLPG )) )
@@ -185,6 +184,8 @@ arch_init(void *par) {static char hello[] ="Bootstrapping...\n" ; + start_info_ptr = par; +hpc_init (); (void )HYPERVISOR_console_io (CONSOLEIO_write,strlen (hello), hello);
@@ -206,7 +207,7 @@ arch_init(void *par)get_cmdline (par);/* Grab the shared_info pointer and put it in a safe place. */ - HYPERVISOR_shared_info =map_shared_info (par); + HYPERVISOR_shared_info =map_shared_info ();/* print out some useful information */ print_start_of_day (par);
@@ -214,7 +215,6 @@ arch_init(void *par)#ifdef CONFIG_PARAVIRT memcpy (&start_info, par,sizeof (start_info));#endif - start_info_ptr = (start_info_t *)par;start_kernel (); }
@@ -253,7 +253,7 @@ void arch_post_suspend(int canceled) }#endif - HYPERVISOR_shared_info =map_shared_info ((void *) start_info_ptr); + HYPERVISOR_shared_info =map_shared_info ();#ifndef CONFIG_PARAVIRT xen_callback_vector ();#endif
hypervisor.c +1 -1
@@ -65,7 +65,7 @@ int hvm_set_parameter(int idx, uint64_t value)return HYPERVISOR_hvm_op (HVMOP_set_param, &xhv); } -shared_info_t *map_shared_info (void *p) +shared_info_t *map_shared_info (void ) {struct xen_add_to_physmap xatp;
include/hypervisor.h +1 -1
@@ -42,7 +42,7 @@ extern union start_info_union start_info_union;int hvm_get_parameter (int idx,uint64_t *value);int hvm_set_parameter (int idx,uint64_t value);#endif -shared_info_t *map_shared_info (void *p); +shared_info_t *map_shared_info (void );void unmap_shared_info (void );void force_evtchn_callback (void );void do_hypervisor_callback (struct pt_regs *regs);
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 2/6] setup: drop parameter from map_shared_info()
Quoted 15 lines · Juergen Gross
The parameter of
map_shared_info()is evaluated in PV mode only, and it is always the pointer to the initial structstart_infofrom boot time.By setting
start_info_ptra little bit earlier the PV variant ofmap_shared_info()can just usestart_info_ptrinstead of a parameter.Signed-off-by Juergen Gross <[email protected]>---
arch/x86/setup.c +0
-0
arch/x86/setup.c arch_init(void *par) · line 187
Quoted 4 lines
arch/x86/setup.c +1
-0@@ -185,6 +184,8 @@ arch_init(void *par) {static char hello[] ="Bootstrapping...\n" ; + start_info_ptr = par;
Should start_info_ptr only be set for CONFIG_PARAVIRT?
Oh, patch 6 does that. Strangely, patch 3 and 6 are not in my inbox.
Regards,
Jason
Quoted 4 lines
arch/x86/setup.c +1
-0+hpc_init (); (void )HYPERVISOR_console_io (CONSOLEIO_write,strlen (hello), hello);
PATCH 3/6 x86/pv: remove global start_info
In reply to Juergen GrossAdd an extern declaration for start_info_ptr when running in PV mode and use that instead of the copy held in start_info. This allows to drop start_info.
--- arch/x86/setup.c | 17 +---------------- console.c | 2 +- include/hypervisor.h | 12 ++---------- xenbus.c | 2 +- 4 files changed, 5 insertions(+), 28 deletions(-)
arch/x86/setup.c +1 -16
@@ -35,17 +35,8 @@#include <xen/arch-x86/hvm/start_info.h> #include <xen/hvm/params.h> -#ifdef CONFIG_PARAVIRT -/* -* This structure contains start-of-day info, such as pagetable base pointer, -* address of the shared_info structure, and things like that. -*/ -union start_info_union start_info_union; -EXPORT_SYMBOL (start_info_union); -#endif -/* -* This pointer holds a reference to the copy of the start_info struct. +* This pointer holds a reference to the start_info struct. */ start_info_t *start_info_ptr;
@@ -212,10 +203,6 @@ arch_init(void *par)/* print out some useful information */ print_start_of_day (par); -#ifdef CONFIG_PARAVIRT -memcpy (&start_info, par,sizeof (start_info)); -#endif -start_kernel (); }
@@ -248,8 +235,6 @@ void arch_post_suspend(int canceled)if (canceled) { start_info_ptr->store_mfn =pfn_to_mfn (start_info_ptr->store_mfn); start_info_ptr->console.domU.mfn =pfn_to_mfn (start_info_ptr->console.domU.mfn); - }else { -memcpy (&start_info, start_info_ptr,sizeof (start_info_t)); }#endif
console.c +1 -1
@@ -411,7 +411,7 @@ void xencons_ring_resume(struct consfront_dev *dev)if ( dev ) {#if CONFIG_PARAVIRT -get_console (&start_info); +get_console (start_info_ptr);#else get_console (0 );#endif
include/hypervisor.h +2 -10
@@ -28,16 +28,8 @@/* hypervisor.c */ #ifdef CONFIG_PARAVIRT -/* -* a placeholder for the start of day information passed up from the hypervisor -*/ -union start_info_union -{ -start_info_t start_info; -char padding[512 ]; -}; -extern union start_info_union start_info_union; -#define start_info (start_info_union.start_info) +/* A pointer to the start of day information passed up from the hypervisor. */ +extern start_info_t *start_info_ptr;#else int hvm_get_parameter (int idx,uint64_t *value);int hvm_set_parameter (int idx,uint64_t value);
xenbus.c +1 -1
@@ -469,7 +469,7 @@ void resume_xenbus(int canceled)struct xsd_sockmsg *rep;#ifdef CONFIG_PARAVIRT -get_xenbus (&start_info); +get_xenbus (start_info_ptr);#else get_xenbus (0 );#endif
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 3/6] x86/pv: remove global start_info
Quoted 5 lines · Juergen Gross
Add an extern declaration for
start_info_ptrwhen running in PV mode and use that instead of the copy held instart_info. This allows to dropstart_info.Signed-off-by Juergen Gross <[email protected]>
PATCH 4/6 console: drop parameter from get_console()
In reply to Juergen GrossThe parameter of get_console() is used only when running in PV mode, and it always has the same value.
Drop the parameter and use start_info_ptr directly in the PV variant of get_console().
--- arch/arm/setup.c | 2 +- arch/x86/setup.c | 2 +- console.c | 10 +++++----- include/console.h | 2 +- 4 files changed, 8 insertions(+), 8 deletions(-)
arch/arm/setup.c +1 -1
@@ -40,7 +40,7 @@ void arch_init(void *dtb_pointer, uint32_t physical_offset)/* Map shared_info page */ HYPERVISOR_shared_info =map_shared_info (); -get_console (NULL ); +get_console ();get_xenbus (NULL );gic_init ();
arch/x86/setup.c +1 -1
@@ -193,7 +193,7 @@ arch_init(void *par)/* WARN: don't do printk before here, it uses information from shared_info. Use xprintk instead. */ -get_console (par); +get_console ();get_xenbus (par);get_cmdline (par);
console.c +5 -5
@@ -197,15 +197,15 @@ uint32_t console_evtchn;static struct consfront_dev *resume_xen_console (struct consfront_dev *dev);#ifdef CONFIG_PARAVIRT -void get_console (void *p) +void get_console (void ) { -start_info_t *si = p; +start_info_t *si = start_info_ptr; console_ring =mfn_to_virt (si->console.domU.mfn); console_evtchn = si->console.domU.evtchn; }#else -void get_console (void *p) +void get_console (void ) {uint64_t v =-1 ;
@@ -411,9 +411,9 @@ void xencons_ring_resume(struct consfront_dev *dev)if ( dev ) {#if CONFIG_PARAVIRT -get_console (start_info_ptr); +get_console ();#else -get_console (0 ); +get_console ();#endif resume_xen_console (dev); }
include/console.h +1 -1
@@ -76,7 +76,7 @@ void xprintk(const char *fmt, ...) __attribute__ ((format (printf, 1, 2)));void xencons_rx (char *buf,unsigned len,struct pt_regs *regs);void xencons_tx (void ); -void get_console (void *p); +void get_console (void );void init_console (void );void console_print (struct consfront_dev *dev,const char *data,int length);void fini_consfront (struct consfront_dev *dev);
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 4/6] console: drop parameter from get_console()
Quoted 7 lines · Juergen Gross
The parameter of
get_console()is used only when running in PV mode, and it always has the same value.Drop the parameter and use
start_info_ptrdirectly in the PV variant ofget_console().Signed-off-by Juergen Gross <[email protected]>
PATCH 5/6 xenbus: drop parameter from get_xenbus()
In reply to Juergen GrossThe parameter of get_xenbus() is used only when running in PV mode, and it always has the same value.
Drop the parameter and use start_info_ptr directly in the PV variant of get_xenbus().
--- arch/arm/setup.c | 2 +- arch/x86/setup.c | 2 +- include/xenbus.h | 4 ++-- xenbus.c | 13 +++++-------- 4 files changed, 9 insertions(+), 12 deletions(-)
arch/arm/setup.c +1 -1
@@ -41,7 +41,7 @@ void arch_init(void *dtb_pointer, uint32_t physical_offset) HYPERVISOR_shared_info =map_shared_info ();get_console (); -get_xenbus (NULL ); +get_xenbus ();gic_init ();
arch/x86/setup.c +1 -1
@@ -194,7 +194,7 @@ arch_init(void *par)/* WARN: don't do printk before here, it uses information from shared_info. Use xprintk instead. */ get_console (); -get_xenbus (par); +get_xenbus ();get_cmdline (par);/* Grab the shared_info pointer and put it in a safe place. */
include/xenbus.h +2 -2
@@ -12,7 +12,7 @@ extern struct xenstore_domain_interface *xenstore_buf;/* Initialize the XenBus system. */ void init_xenbus (void ); -void get_xenbus (void *p); +void get_xenbus (void );#else #define xenbus_evtchn ~0#define xenstore_buf NULL
@@ -20,7 +20,7 @@ void get_xenbus(void *p);static inline void init_xenbus (void ) { } -static inline void get_xenbus (void *p) +static inline void get_xenbus (void ) { }#endif
xenbus.c +5 -8
@@ -74,15 +74,15 @@ uint32_t xenbus_evtchn;EXPORT_SYMBOL (xenbus_evtchn);#ifdef CONFIG_PARAVIRT -void get_xenbus (void *p) +void get_xenbus (void ) { -start_info_t *si = p; +start_info_t *si = start_info_ptr; xenbus_evtchn = si->store_evtchn; xenstore_buf =mfn_to_virt (si->store_mfn); }#else -void get_xenbus (void *p) +void get_xenbus (void ) {uint64_t v;
@@ -468,11 +468,8 @@ void resume_xenbus(int canceled)struct write_req req[2 ];struct xsd_sockmsg *rep; -#ifdef CONFIG_PARAVIRT -get_xenbus (start_info_ptr); -#else -get_xenbus (0 ); -#endif +get_xenbus (); +unmask_evtchn (xenbus_evtchn);if ( !canceled )
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 5/6] xenbus: drop parameter from get_xenbus()
Quoted 7 lines · Juergen Gross
The parameter of
get_xenbus()is used only when running in PV mode, and it always has the same value.Drop the parameter and use
start_info_ptrdirectly in the PV variant ofget_xenbus().Signed-off-by Juergen Gross <[email protected]>
PATCH 6/6 x86/pv: make start_info_ptr PV-only
In reply to Juergen GrossHide start_info_ptr via CONFIG_PARAVIRT in order to avoid accesses to it form not PV specific code.
--- arch/x86/setup.c | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-)
arch/x86/setup.c +11 -5
@@ -35,11 +35,6 @@#include <xen/arch-x86/hvm/start_info.h> #include <xen/hvm/params.h> -/* -* This pointer holds a reference to the start_info struct. -*/ -start_info_t *start_info_ptr; -/* * Shared page for communicating with the hypervisor. * Events flags go here, for example.
@@ -68,6 +63,11 @@ static inline void sse_init(void) {#endif #ifdef CONFIG_PARAVIRT +/* +* This pointer holds a reference to the start_info struct. +*/ +start_info_t *start_info_ptr; +#define hpc_init ()shared_info_t *map_shared_info (void )
@@ -175,7 +175,9 @@ arch_init(void *par) {static char hello[] ="Bootstrapping...\n" ; +#ifdef CONFIG_PARAVIRT start_info_ptr = par; +#endif hpc_init (); (void )HYPERVISOR_console_io (CONSOLEIO_write,strlen (hello), hello);
@@ -226,7 +228,11 @@ int arch_suspend(void)* This hypercall returns 1 if the suspend * was cancelled and 0 if resuming in a new domain */ +#ifdef CONFIG_PARAVIRT return HYPERVISOR_suspend (virt_to_mfn (start_info_ptr)); +#else +return HYPERVISOR_suspend (0 ); +#endif }void arch_post_suspend (int canceled)
Signature · 1 line
2.43.0
Re: [MINI-OS PATCH 6/6] x86/pv: make start_info_ptr PV-only
Quoted 4 lines · Juergen Gross
Hide
start_info_ptrviaCONFIG_PARAVIRTin order to avoid accesses to it form not PV specific code.Signed-off-by Juergen Gross <[email protected]>
Quoted 14 lines · Juergen Gross
While working on xenstore-stubdom live update support I came across some really unpleasant code using struct
start_infoeven when running in PVH mode.Especially suspend/resume was doing some crazy stuff, including overwriting the initial struct
hvm_start_infopartially.This series is hiding struct
start_infofrom PVH code. This didn't result in any runtime error, but in future I'll need to access structhvm_start_infosome time after boot, so I need to get that fixed.Tested especially with xl suspend/resume and xl save/restore in PV and in PVH mode.
Quoted 7 lines
Juergen Gross (6):
suspend: drop crazy pfn handling for PVH
setup: drop parameter frommap_shared_info()
x86/pv: remove globalstart_info
console: drop parameter fromget_console()
xenbus: drop parameter fromget_xenbus()
x86/pv: makestart_info_ptrPV-only
Before I even try: Does this series interfere in any way with the kexec one, and hence may need re-basing on top of that (which I've pushed a minute ago)?
Jan
Quoted 23 lines · Jan Beulich
Quoted 14 lines · Juergen Gross
While working on xenstore-stubdom live update support I came across some really unpleasant code using struct
start_infoeven when running in PVH mode.Especially suspend/resume was doing some crazy stuff, including overwriting the initial struct
hvm_start_infopartially.This series is hiding struct
start_infofrom PVH code. This didn't result in any runtime error, but in future I'll need to access structhvm_start_infosome time after boot, so I need to get that fixed.Tested especially with xl suspend/resume and xl save/restore in PV and in PVH mode.
Quoted 7 lines
Juergen Gross (6):
suspend: drop crazy pfn handling for PVH
setup: drop parameter frommap_shared_info()
x86/pv: remove globalstart_info
console: drop parameter fromget_console()
xenbus: drop parameter fromget_xenbus()
x86/pv: makestart_info_ptrPV-only
Quoted 2 lines
Before I even try: Does this series interfere in any way with the kexec one, and hence may need re-basing on top of that (which I've pushed a minute ago)?
I'm not aware of any possible conflict.
Juergen