Skip to contentSkip to conversation

PMA phase 2 - per CPU address spaces

PATCH In review Peter Crosthwaite 4 participants 19 messages Aug 26, 2014 – Oct 21, 2015 Kbuild

lore
Share & tools
lore
#0 lore raw

cover PMA phase 2 - per CPU address spaces

Hi All,

This series sets up CPUs with configurable address spaces. This follows on from Edgars original work and moves towards removal of address_space_memory and support for arbitrary memory heirachies/layouts.

Fuller context in RFC:

http://lists.gnu.org/archive/html/qemu-devel/2014-06/msg00483.html

Follow up series' will add the rest of that functionality.

Regards,
Peter

Changed since RFC:
Limit scope to only CPU and Address Space plumbing Got compiling and tested in fully configured build.

Peter Crosthwaite (6):
memory: address_space_init: do nothing if no root region given
memory: AddressSpace: Implement ref counting
memory: Add address_space_init_shareable()
qom: Move cpu.o to obj-y.
qom/cpu: Add Memory Region Property
cpu: Delay address space init until realize

 Makefile.target       |  1 +
 exec.c                |  1 -
 include/exec/memory.h |  4 ++++
 include/qom/cpu.h     |  1 +
 memory.c              | 34 ++++++++++++++++++++++++++++++++++
 qom/Makefile.objs     |  2 +-
 qom/cpu.c             | 16 ++++++++++++++++
 7 files changed, 57 insertions(+), 2 deletions(-)
Signature · 1 line
2.1.0.1.g27b9230
#1 lore raw

PATCH 1/6 memory: address_space_init: do nothing if no root region given

Just ignore this case. This is needed for bus master device realize when the machine model doesn't connect an attachment. Then machine model may decide to not set a memory region for mastering yet the device will attempt to create itself an address space come realize time. Gracefully do nothing.

---

 memory.c | 4 ++++
 1 file changed, 4 insertions(+)
memory.c +4 -0
index 42317a2..e6014d9 100644
@@ -1925,6 +1925,10 @@ void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
         memory_init();
     }
 
+    if (!root) {
+        return;
+    }
+
     memory_region_transaction_begin();
     as->root = root;
     as->current_map = g_new(FlatView, 1);
Signature · 1 line
2.1.0.1.g27b9230
#2 lore raw

PATCH 2/6 memory: AddressSpace: Implement ref counting

In reply to Peter Crosthwaite

To allow for sharing of AddressSpaces between multiple masters.

---

 include/exec/memory.h | 1 +
 memory.c              | 5 +++++
 2 files changed, 6 insertions(+)
include/exec/memory.h +1 -0
index d165b27..18cb6b2 100644
@@ -207,6 +207,7 @@ struct AddressSpace {
     /* All fields are private. */
     char *name;
     MemoryRegion *root;
+    int ref_count;
     struct FlatView *current_map;
     int ioeventfd_nb;
     struct MemoryRegionIoeventfd *ioeventfds;
memory.c +5 -0
index e6014d9..27beef1 100644
@@ -1930,6 +1930,7 @@ void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
     }
 
     memory_region_transaction_begin();
+    as->ref_count = 1;
     as->root = root;
     as->current_map = g_new(FlatView, 1);
     flatview_init(as->current_map);
@@ -1946,6 +1947,10 @@ void address_space_destroy(AddressSpace *as)
 {
     MemoryListener *listener;
 
+    as->ref_count--;
+    if (as->ref_count) {
+        return;
+    }
     /* Flush out anything from MemoryListeners listening in on this */
     memory_region_transaction_begin();
     as->root = NULL;
Signature · 1 line
2.1.0.1.g27b9230
#4 lore raw

PATCH 4/6 qom: Move cpu.o to obj-y.

In reply to Peter Crosthwaite

It's not really a common object and this is needed to give it access to CONFIG_USER_ONLY definition. Move it to regular obj-y.

---

 Makefile.target   | 1 +
 qom/Makefile.objs | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)
Makefile.target +1 -0
index 1e8d7ab..c91c8e5 100644
@@ -84,6 +84,7 @@ all: $(PROGS) stap
 # cpu emulator library
 obj-y = exec.o translate-all.o cpu-exec.o
 obj-y += tcg/tcg.o tcg/optimize.o
+obj-y += qom/
 obj-$(CONFIG_TCG_INTERPRETER) += tci.o
 obj-$(CONFIG_TCG_INTERPRETER) += disas/tci.o
 obj-y += fpu/softfloat.o
qom/Makefile.objs +1 -1
index 985003b..3604a4f 100644
@@ -1,3 +1,3 @@
 common-obj-y = object.o container.o qom-qobject.o
-common-obj-y += cpu.o
 common-obj-y += object_interfaces.o
+obj-y += cpu.o
Signature · 1 line
2.1.0.1.g27b9230
#10 lore raw

PATCH 5/6 qom/cpu: Add Memory Region Property

In reply to Peter Crosthwaite

Which is used to construct a per-CPU address space. The Address space will be created at realize time.

---

 include/qom/cpu.h |  1 +
 qom/cpu.c         | 11 +++++++++++
 2 files changed, 12 insertions(+)
include/qom/cpu.h +1 -0
index 1aafbf5..192d337 100644
@@ -242,6 +242,7 @@ struct CPUState {
     int64_t icount_extra;
     sigjmp_buf jmp_env;
 
+    MemoryRegion *mr;
     AddressSpace *as;
     MemoryListener *tcg_as_listener;
 
qom/cpu.c +11 -0
index b32dd0a..899071e 100644
@@ -25,6 +25,10 @@
 #include "qemu/log.h"
 #include "qemu/error-report.h"
 #include "sysemu/sysemu.h"
+#include "exec/memory.h"
+#include "qapi/visitor.h"
+#include "exec/address-spaces.h"
+#include "hw/qdev-properties.h"
 
 bool cpu_exists(int64_t id)
 {
@@ -314,6 +318,13 @@ static void cpu_common_initfn(Object *obj)
     CPUClass *cc = CPU_GET_CLASS(obj);
 
     cpu->gdb_num_regs = cpu->gdb_num_g_regs = cc->gdb_num_core_regs;
+#ifndef CONFIG_USER_ONLY
+    object_property_add_link(obj, "mr", TYPE_MEMORY_REGION,
+                             (Object **)&cpu->mr,
+                             qdev_prop_allow_set_link_before_realize,
+                             OBJ_PROP_LINK_UNREF_ON_RELEASE,
+                             &error_abort);
+#endif
 }
 
 static int64_t cpu_common_get_arch_id(CPUState *cpu)
Signature · 1 line
2.1.0.1.g27b9230
#14 lore raw

PATCH 6/6 cpu: Delay address space init until realize

In reply to Peter Crosthwaite

To allow it to be set by the user after init.

---

 exec.c    | 1 -
 qom/cpu.c | 5 +++++
 2 files changed, 5 insertions(+), 1 deletion(-)
exec.c +0 -1
index 5f9857c..15ae6b7 100644
@@ -492,7 +492,6 @@ void cpu_exec_init(CPUArchState *env)
     QTAILQ_INIT(&cpu->breakpoints);
     QTAILQ_INIT(&cpu->watchpoints);
 #ifndef CONFIG_USER_ONLY
-    cpu->as = &address_space_memory;
     cpu->thread_id = qemu_get_thread_id();
 #endif
     QTAILQ_INSERT_TAIL(&cpus, cpu, node);
qom/cpu.c +5 -0
index 899071e..f866d30 100644
@@ -305,6 +305,11 @@ static void cpu_common_realizefn(DeviceState *dev, Error **errp)
 {
     CPUState *cpu = CPU(dev);
 
+#ifndef CONFIG_USER_ONLY
+    cpu->as = cpu->mr ? address_space_init_shareable(cpu->mr, NULL)
+                      : &address_space_memory;
+#endif
+
     if (dev->hotplugged) {
         cpu_synchronize_post_init(cpu);
         notifier_list_notify(&cpu_added_notifiers, dev);
Signature · 1 line
2.1.0.1.g27b9230
#11 lore raw

Re: [Qemu-devel] [PATCH v1 5/6] qom/cpu: Add Memory Region Property

In reply to Peter Crosthwaite

Il 26/08/2014 02:59, Peter Crosthwaite ha scritto:

Quoted 38 lines

Which is used to construct a per-CPU address space. The Address space will be created at realize time.

---

 include/qom/cpu.h |  1 +
 qom/cpu.c         | 11 +++++++++++
 2 files changed, 12 insertions(+)
include/qom/cpu.h +1 -0
index 1aafbf5..192d337 100644
@@ -242,6 +242,7 @@ struct CPUState {
     int64_t icount_extra;
     sigjmp_buf jmp_env;
 
+    MemoryRegion *mr;
     AddressSpace *as;
     MemoryListener *tcg_as_listener;
 
qom/cpu.c +4 -0
index b32dd0a..899071e 100644
@@ -25,6 +25,10 @@
 #include "qemu/log.h"
 #include "qemu/error-report.h"
 #include "sysemu/sysemu.h"
+#include "exec/memory.h"
+#include "qapi/visitor.h"
+#include "exec/address-spaces.h"
+#include "hw/qdev-properties.h"
 
 bool cpu_exists(int64_t id)
 {

qom/cpu.c static void cpu_common_initfn(Object *obj) · line 322

Quoted 5 lines
qom/cpu.c +2 -0
@@ -314,6 +318,13 @@ static void cpu_common_initfn(Object *obj)
     CPUClass *cc = CPU_GET_CLASS(obj);
 
     cpu->gdb_num_regs = cpu->gdb_num_g_regs = cc->gdb_num_core_regs;
+#ifndef CONFIG_USER_ONLY
+    object_property_add_link(obj, "mr", TYPE_MEMORY_REGION,

Please use a clearer name such as "memory".

Paolo

Quoted 9 lines
qom/cpu.c +5 -0
+                             (Object **)&cpu->mr,
+                             qdev_prop_allow_set_link_before_realize,
+                             OBJ_PROP_LINK_UNREF_ON_RELEASE,
+                             &error_abort);
+#endif
 }
 
 static int64_t cpu_common_get_arch_id(CPUState *cpu)
#12 lore raw
In reply to Peter Crosthwaite
Quoted 22 lines · Peter Crosthwaite

Hi All,

This series sets up CPUs with configurable address spaces. This follows on from Edgars original work and moves towards removal of address_space_memory and support for arbitrary memory heirachies/layouts.

Fuller context in RFC:

http://lists.gnu.org/archive/html/qemu-devel/2014-06/msg00483.html

Follow up series' will add the rest of that functionality.

Regards,
Peter

Changed since RFC:
Limit scope to only CPU and Address Space plumbing Got compiling and tested in fully configured build.

Quoted 7 lines

Peter Crosthwaite (6):
memory: address_space_init: do nothing if no root region given
memory: AddressSpace: Implement ref counting
memory: Add address_space_init_shareable()
qom: Move cpu.o to obj-y.
qom/cpu: Add Memory Region Property
cpu: Delay address space init until realize

It looks like patch 3/6 never reached the list: care to resend the series?

thanks
-- PMM

#5 lore raw

Re: [Qemu-devel] [PATCH v1 4/6] qom: Move cpu.o to obj-y.

In reply to Peter Crosthwaite
Quoted 1 line · Peter Crosthwaite

It's not really a common object

In what sense isn't it a common object? It's only compiled once, rather than per-target. We should avoid moving object files from compiled-once to compiled-per-target if we can...

thanks
-- PMM

#6 lore raw

Re: [Qemu-devel] [PATCH v1 4/6] qom: Move cpu.o to obj-y.

Il 01/09/2014 19:43, Peter Maydell ha scritto:

Quoted 7 lines · Peter Crosthwaite
Quoted 1 line

It's not really a common object

In what sense isn't it a common object? It's only compiled once, rather than per-target. We should avoid moving object files from compiled-once to compiled-per-target if we can...

Should we add a softmmu_enabled() function similar to kvm_enabled(), that returns true for system emulation and false for user-mode emulation?

Paolo

#7 lore raw

Re: [Qemu-devel] [PATCH v1 4/6] qom: Move cpu.o to obj-y.

Quoted 4 lines · Paolo Bonzini

Il 01/09/2014 19:43, Peter Maydell ha scritto:

Quoted 3 lines · Peter Crosthwaite
Quoted 1 line

It's not really a common object

Quoted 8 lines
Quoted 4 lines

In what sense isn't it a common object? It's only compiled once, rather than per-target. We should avoid moving object files from compiled-once to compiled-per-target if we can...

Should we add a softmmu_enabled() function similar to kvm_enabled(), that returns true for system emulation and false for user-mode emulation?

Doesn't really help me in this case, as I need conditional compliation due to use of the maybe-exists &address_space_memory. I would also need to add get_system_address_space() which return NULL for user-mode as well.

Regards,
Peter

Quoted 3 lines

Paolo

#8 lore raw

Re: [Qemu-devel] [PATCH v1 4/6] qom: Move cpu.o to obj-y.

In reply to Peter Maydell
Quoted 8 lines · Peter Maydell
Quoted 1 line · Peter Crosthwaite

It's not really a common object

In what sense isn't it a common object? It's only compiled once, rather than per-target. We should avoid moving object files from compiled-once to compiled-per-target if we can...

Yeh, I introduce the variation with address_space_memory reliance. CPU needs to see it for the default AR. It doesn't exist in user-mode though.

Regards,
Peter

Quoted 3 lines

thanks
-- PMM

#9 lore raw

Re: [Qemu-devel] [PATCH v1 4/6] qom: Move cpu.o to obj-y.

Il 02/09/2014 00:56, Peter Crosthwaite ha scritto:

Quoted 10 lines · Peter Maydell
Quoted 8 lines · Peter Crosthwaite
Quoted 1 line

It's not really a common object

In what sense isn't it a common object? It's only compiled once, rather than per-target. We should avoid moving object files from compiled-once to compiled-per-target if we can...

Quoted 3 lines

Yeh, I introduce the variation with address_space_memory reliance. CPU needs to see it for the default AR. It doesn't exist in user-mode though.

cpus.c already contains softmmu-specific of cpu.c. Perhaps you can add something in there (and in *-user/main.c).

Paolo

#3 lore raw

PATCH 3/6 memory: Add address_space_init_shareable()

In reply to Peter Crosthwaite

This will either create a new AS or return a pointer to an already existing equivalent one. Both name and root mr must match.

The motivation is to reuse address spaces as much as possible. It's going to be quite common that bus masters out in device land have pointers to the same memory region for their mastering yet each will need to create its own address space. Let the memory API implement sharing for them.

Aside from the perf optimisations, this should reduce the amount of redundant output on info mtree as well.

Thee returned value will be malloced, but the malloc will be automatically freed when the AS runs out of refs.

---
We could change the equivalency test only match mr to support device specific naming of these shared ASes. The singleton AS can ultimately only have one name however. So perhaps some strcatting each time a new sharer is added to the share. That or first-in-best-dressed.

Changed since v1:
Implement ref counting and garbage collection.

 include/exec/memory.h |  3 +++
 memory.c              | 25 +++++++++++++++++++++++++
 2 files changed, 28 insertions(+)
include/exec/memory.h +3 -0
index 18cb6b2..cbd7336 100644
@@ -208,6 +208,7 @@ struct AddressSpace {
     char *name;
     MemoryRegion *root;
     int ref_count;
+    bool malloced;
     struct FlatView *current_map;
     int ioeventfd_nb;
     struct MemoryRegionIoeventfd *ioeventfds;
@@ -977,6 +978,8 @@ void mtree_info(fprintf_function mon_printf, void *f);
  */
 void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name);
 
+AddressSpace *address_space_init_shareable(MemoryRegion *root,
+                                           const char *name);
 
 /**
  * address_space_destroy: destroy an address space
memory.c +25 -0
index 27beef1..08ab918 100644
@@ -1932,6 +1932,7 @@ void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
     memory_region_transaction_begin();
     as->ref_count = 1;
     as->root = root;
+    as->malloced = false;
     as->current_map = g_new(FlatView, 1);
     flatview_init(as->current_map);
     as->ioeventfd_nb = 0;
@@ -1943,6 +1944,27 @@ void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
     memory_region_transaction_commit();
 }
 
+AddressSpace *address_space_init_shareable(MemoryRegion *root, const char *name)
+{
+    AddressSpace *as;
+
+    if (!root) {
+        return NULL;
+    }
+
+    QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
+        if (root == as->root && !strcmp(name ? name : "anonymous", as->name)) {
+            as->ref_count++;
+            return as;
+        }
+    }
+
+    as = g_malloc0(sizeof *as);
+    address_space_init(as, root, name);
+    as->malloced = true;
+    return as;
+}
+
 void address_space_destroy(AddressSpace *as)
 {
     MemoryListener *listener;
@@ -1965,6 +1987,9 @@ void address_space_destroy(AddressSpace *as)
     flatview_unref(as->current_map);
     g_free(as->name);
     g_free(as->ioeventfds);
+    if (as->malloced) {
+        g_free(as);
+    }
 }
 
 bool io_mem_read(MemoryRegion *mr, hwaddr addr, uint64_t *pval, unsigned size)
Signature · 1 line
2.1.0.1.g27b9230
#13 lore raw
In reply to Peter Maydell
Quoted 31 lines · Peter Maydell
Quoted 22 lines · Peter Crosthwaite

Hi All,

This series sets up CPUs with configurable address spaces. This follows on from Edgars original work and moves towards removal of address_space_memory and support for arbitrary memory heirachies/layouts.

Fuller context in RFC:

http://lists.gnu.org/archive/html/qemu-devel/2014-06/msg00483.html

Follow up series' will add the rest of that functionality.

Regards,
Peter

Changed since RFC:
Limit scope to only CPU and Address Space plumbing Got compiling and tested in fully configured build.

Quoted 7 lines

Peter Crosthwaite (6):
memory: address_space_init: do nothing if no root region given
memory: AddressSpace: Implement ref counting
memory: Add address_space_init_shareable()
qom: Move cpu.o to obj-y.
qom/cpu: Add Memory Region Property
cpu: Delay address space init until realize

Quoted 3 lines

It looks like patch 3/6 never reached the list: care to resend the series?

Still have the original handy so sent the missing single. Should thread properly.

I'll action Paolo's comments and respin V2 shortly.

Regards,
Peter

Quoted 3 lines

thanks
-- PMM

#15 lore raw
In reply to Peter Crosthwaite
Quoted 19 lines · Peter Crosthwaite

This series sets up CPUs with configurable address spaces. This follows on from Edgars original work and moves towards removal of address_space_memory and support for arbitrary memory heirachies/layouts.

Fuller context in RFC:

http://lists.gnu.org/archive/html/qemu-devel/2014-06/msg00483.html

Follow up series' will add the rest of that functionality.

Regards,
Peter

Changed since RFC:
Limit scope to only CPU and Address Space plumbing Got compiling and tested in fully configured build.

Quoted 7 lines

Peter Crosthwaite (6):
memory: address_space_init: do nothing if no root region given
memory: AddressSpace: Implement ref counting
memory: Add address_space_init_shareable()
qom: Move cpu.o to obj-y.
qom/cpu: Add Memory Region Property
cpu: Delay address space init until realize

Just to let you know, I'm taking some of these patches into a series I'm working on for multiple address-space support (for ARM trustzone). Basically I'm taking patches 2 and 3, plus a variant of patch 5 that avoids the need to make cpu.o an obj-y object. Then I have some more patches of my own which add an extra QOM property for ARM CPUs for the secure-memory address space.

Alpha-quality work-in-progress patches currently at
https://git.linaro.org/people/peter.maydell/qemu-arm.git multi-ases web view:
https://git.linaro.org/people/peter.maydell/qemu-arm.git/shortlog/refs/heads/multi-ases

I'm not completely certain about the AS reference-counting code right now; will have to come back and think harder about it later.

thanks
-- PMM

#16 lore raw
Quoted 29 lines · Peter Maydell
Quoted 19 lines · Peter Crosthwaite

This series sets up CPUs with configurable address spaces. This follows on from Edgars original work and moves towards removal of address_space_memory and support for arbitrary memory heirachies/layouts.

Fuller context in RFC:

http://lists.gnu.org/archive/html/qemu-devel/2014-06/msg00483.html

Follow up series' will add the rest of that functionality.

Regards,
Peter

Changed since RFC:
Limit scope to only CPU and Address Space plumbing Got compiling and tested in fully configured build.

Quoted 7 lines

Peter Crosthwaite (6):
memory: address_space_init: do nothing if no root region given
memory: AddressSpace: Implement ref counting
memory: Add address_space_init_shareable()
qom: Move cpu.o to obj-y.
qom/cpu: Add Memory Region Property
cpu: Delay address space init until realize

Quoted 8 lines

Just to let you know, I'm taking some of these patches into a series I'm working on for multiple address-space support (for ARM trustzone). Basically I'm taking patches 2 and 3, plus a variant of patch 5 that avoids the need to make cpu.o an obj-y object. Then I have some more patches of my own which add an extra QOM property for ARM CPUs for the secure-memory address space.

Cool. That is largely orthogonal to the goal of the referenced RFC (link is still above in quoted text) which was thinking more about the machine models. See the PetaLogix patches at the end. With the core work going through in this project, the groundwork for arbitrary machine AS hierarchy is there and we should think about deprecating address_space_memory for new SoCs and DMA capable devices.

This might also help Marcin and Christian who were working with some heterogenous archs and the memory spaces came up on both accounts.

Is info mtree well behaved? I remember needing some changes to mtree due to verbosity when there were multiple ases and complex aliases at play. Code is in the Xilinx tree, let me know if you need some digging.

Quoted 9 lines

Alpha-quality work-in-progress patches currently at
https://git.linaro.org/people/peter.maydell/qemu-arm.git multi-ases web view:
https://git.linaro.org/people/peter.maydell/qemu-arm.git/shortlog/refs/heads/multi-ases

I'm not completely certain about the AS reference-counting code right now; will have to come back and think harder about it later.

I did once attempt to QOMify address spaces themselves which would implement this on core layers but I think we decided against that IIRC (and just use MRs as the QOMified handle for ases).

Regards,
Peter

Quoted 2 lines

thanks
-- PMM

#17 lore raw
Quoted 10 lines · Peter Crosthwaite
Quoted 8 lines · Peter Maydell

Just to let you know, I'm taking some of these patches into a series I'm working on for multiple address-space support (for ARM trustzone). Basically I'm taking patches 2 and 3, plus a variant of patch 5 that avoids the need to make cpu.o an obj-y object. Then I have some more patches of my own which add an extra QOM property for ARM CPUs for the secure-memory address space.

Quoted 6 lines

Cool. That is largely orthogonal to the goal of the referenced RFC (link is still above in quoted text) which was thinking more about the machine models. See the PetaLogix patches at the end. With the core work going through in this project, the groundwork for arbitrary machine AS hierarchy is there and we should think about deprecating address_space_memory for new SoCs and DMA capable devices.

Yes. We should probably also try to be consistent about what QOM property names we use for handing the MR to bus masters. I went for 'memory' and 'secure-memory', but maybe there's a better scheme?

Quoted 7 lines

This might also help Marcin and Christian who were working with some heterogenous archs and the memory spaces came up on both accounts.

Is info mtree well behaved? I remember needing some changes to mtree due to verbosity when there were multiple ases and complex aliases at play. Code is in the Xilinx tree, let me know if you need some digging.

I haven't done any testing beyond "haven't actually broken booting of something that doesn't use the second AS" yet :-) I'll check info mtree.

The other area I need to think about is KVM support -- KVM doesn't really like having different ASes between CPUs (IIRC you could do it but the performance would be bad). The uses I have in mind would be ok with this limitation but maybe we need an assertion somewhere to avoid people hitting it by mistake.

Quoted 4 lines
Quoted 4 lines

Alpha-quality work-in-progress patches currently at
https://git.linaro.org/people/peter.maydell/qemu-arm.git multi-ases web view:
https://git.linaro.org/people/peter.maydell/qemu-arm.git/shortlog/refs/heads/multi-ases

(I see I messed up some of the patch authorship data; will fix.)

Quoted 8 lines
Quoted 4 lines

I'm not completely certain about the AS reference-counting code right now; will have to come back and think harder about it later.

I did once attempt to QOMify address spaces themselves which would implement this on core layers but I think we decided against that IIRC (and just use MRs as the QOMified handle for ases).

Why does address_space_init_shareable() insist on the name string matching as well as the MR, by the way? That seems like a recipe for having less sharing than we could for no particularly useful reason...

thanks
-- PMM

#18 lore raw
Quoted 19 lines · Peter Maydell
Quoted 10 lines · Peter Crosthwaite
Quoted 8 lines · Peter Maydell

Just to let you know, I'm taking some of these patches into a series I'm working on for multiple address-space support (for ARM trustzone). Basically I'm taking patches 2 and 3, plus a variant of patch 5 that avoids the need to make cpu.o an obj-y object. Then I have some more patches of my own which add an extra QOM property for ARM CPUs for the secure-memory address space.

Quoted 6 lines

Cool. That is largely orthogonal to the goal of the referenced RFC (link is still above in quoted text) which was thinking more about the machine models. See the PetaLogix patches at the end. With the core work going through in this project, the groundwork for arbitrary machine AS hierarchy is there and we should think about deprecating address_space_memory for new SoCs and DMA capable devices.

Quoted 5 lines

Yes. We should probably also try to be consistent about what QOM property names we use for handing the MR to bus masters. I went for 'memory' and 'secure-memory', but maybe there's a better scheme?

Sometimes the TRM gives a unique symbolic name. E.g. your average Xilinx soft-IP TRM will name the bus master 'M_AXI" or something like that. We should follow that if it exists. Food for thought, here is a diagram of a Microblaze processor with five well-named bus masters with this kind of naming scheme:

https://reference.digilentinc.com/_media/vivado:mig_37.jpg?w=500&tok=cca517

I'm guessing there is nothing in ARM ARM for CPUs though?

In the absence of a TRM name and for devices with only one bus master, I used "dma".

Quoted 30 lines
Quoted 7 lines

This might also help Marcin and Christian who were working with some heterogenous archs and the memory spaces came up on both accounts.

Is info mtree well behaved? I remember needing some changes to mtree due to verbosity when there were multiple ases and complex aliases at play. Code is in the Xilinx tree, let me know if you need some digging.

I haven't done any testing beyond "haven't actually broken booting of something that doesn't use the second AS" yet :-) I'll check info mtree.

The other area I need to think about is KVM support -- KVM doesn't really like having different ASes between CPUs (IIRC you could do it but the performance would be bad). The uses I have in mind would be ok with this limitation but maybe we need an assertion somewhere to avoid people hitting it by mistake.

Quoted 4 lines
Quoted 4 lines

Alpha-quality work-in-progress patches currently at
https://git.linaro.org/people/peter.maydell/qemu-arm.git multi-ases web view:
https://git.linaro.org/people/peter.maydell/qemu-arm.git/shortlog/refs/heads/multi-ases

(I see I messed up some of the patch authorship data; will fix.)

Quoted 5 lines
Quoted 4 lines

I'm not completely certain about the AS reference-counting code right now; will have to come back and think harder about it later.

Quoted 9 lines
Quoted 3 lines

I did once attempt to QOMify address spaces themselves which would implement this on core layers but I think we decided against that IIRC (and just use MRs as the QOMified handle for ases).

Why does address_space_init_shareable() insist on the name string matching as well as the MR, by the way? That seems like a recipe for having less sharing than we could for no particularly useful reason...

I think it was conservatism on my part to keep alive the possibility of two different address-spaces sharing a root MR but with different props on the AS struct level. If no such use case exists then just the root MR match should be good enough. I guess put another way, is we can do this if address spaces themselves are guaranteed to be stateless.

Regards,
Peter

Quoted 2 lines

thanks
-- PMM