qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v5 16/50] multi-process: Synchronize remote memory


From: Jag Raman
Subject: Re: [PATCH v5 16/50] multi-process: Synchronize remote memory
Date: Wed, 4 Mar 2020 14:35:08 -0500
User-agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:68.0) Gecko/20100101 Thunderbird/68.5.0



On 3/4/2020 6:53 AM, Dr. David Alan Gilbert wrote:
* Jagannathan Raman (address@hidden) wrote:
Add memory-listener object which is used to keep the view of the RAM
in sync between QEMU and remote process.
A MemoryListener is registered for system-memory AddressSpace. The
listener sends SYNC_SYSMEM message to the remote process when memory
listener commits the changes to memory, the remote process receives
the message and processes it in the handler for SYNC_SYSMEM message.

TODO: No need to create object for remote memory listener.

Signed-off-by: Jagannathan Raman <address@hidden>
Signed-off-by: John G Johnson <address@hidden>
Signed-off-by: Elena Ufimtseva <address@hidden>
---
  Makefile.target                |   3 +
  hw/proxy/memory-sync.c         | 212 +++++++++++++++++++++++++++++++++++++++++
  hw/proxy/qemu-proxy.c          |   5 +
  include/hw/proxy/memory-sync.h |  37 +++++++
  include/hw/proxy/qemu-proxy.h  |   5 +
  remote/remote-main.c           |  11 +++
  6 files changed, 273 insertions(+)
  create mode 100644 hw/proxy/memory-sync.c
  create mode 100644 include/hw/proxy/memory-sync.h

diff --git a/Makefile.target b/Makefile.target
index cfd36c1..271d883 100644
--- a/Makefile.target
+++ b/Makefile.target
@@ -127,6 +127,9 @@ obj-$(CONFIG_TCG) += fpu/softfloat.o
  obj-y += target/$(TARGET_BASE_ARCH)/
  obj-y += disas.o
  obj-$(call notempty,$(TARGET_XML_FILES)) += gdbstub-xml.o
+ifeq ($(TARGET_NAME)-$(CONFIG_MPQEMU)-$(CONFIG_USER_ONLY), x86_64-y-)
+obj-$(CONFIG_MPQEMU) += hw/proxy/memory-sync.o
+endif
  LIBS := $(libs_cpu) $(LIBS)
obj-$(CONFIG_PLUGIN) += plugins/
diff --git a/hw/proxy/memory-sync.c b/hw/proxy/memory-sync.c
new file mode 100644
index 0000000..3edbb19
--- /dev/null
+++ b/hw/proxy/memory-sync.c
@@ -0,0 +1,212 @@
+/*
+ * Copyright © 2018, 2020 Oracle and/or its affiliates.
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ *
+ */
+
+#include <sys/types.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "qemu/osdep.h"
+#include "qemu/compiler.h"
+#include "qemu/int128.h"
+#include "qemu/range.h"
+#include "exec/memory.h"
+#include "exec/cpu-common.h"
+#include "cpu.h"
+#include "exec/ram_addr.h"
+#include "exec/address-spaces.h"
+#include "io/mpqemu-link.h"
+#include "hw/proxy/memory-sync.h"
+
+static const TypeInfo remote_mem_sync_type_info = {
+    .name          = TYPE_MEMORY_LISTENER,
+    .parent        = TYPE_OBJECT,
+    .instance_size = sizeof(RemoteMemSync),
+};
+
+static void remote_mem_sync_register_types(void)
+{
+    type_register_static(&remote_mem_sync_type_info);
+}
+
+type_init(remote_mem_sync_register_types)
+
+static void proxy_ml_begin(MemoryListener *listener)
+{
+    RemoteMemSync *sync = container_of(listener, RemoteMemSync, listener);
+    int mrs;
+
+    for (mrs = 0; mrs < sync->n_mr_sections; mrs++) {
+        memory_region_unref(sync->mr_sections[mrs].mr);
+    }
+
+    g_free(sync->mr_sections);
+    sync->mr_sections = NULL;
+    sync->n_mr_sections = 0;
+}
+
+static int get_fd_from_hostaddr(uint64_t host, ram_addr_t *offset)
+{
+    MemoryRegion *mr;
+    ram_addr_t off;
+
+    mr = memory_region_from_host((void *)(uintptr_t)host, &off);

Do you need to just check we found an 'mr' ?

OK, we'll add this check.


+    if (offset) {
+        *offset = off;
+    }
+
+    return memory_region_get_fd(mr);
+}
+
+static bool proxy_mrs_can_merge(uint64_t host, uint64_t prev_host, size_t size)
+{
+    bool merge;
+    int fd1, fd2;
+
+    fd1 = get_fd_from_hostaddr(host, NULL);
+
+    fd2 = get_fd_from_hostaddr(prev_host, NULL);
+
+    merge = (fd1 == fd2);
+
+    merge &= ((prev_host + size) == host);

It's interesting; I think the vhost code checks that the two mr's are
the same where you are checking for the same underlying fd - but I think
that's OK.
(I wonder if we need to check offset's within the fd's match up when
they're merged - since you added that offset feature in an earlier
patch?
That would also need checking in vhost_region_add_section)

If the fds are the same, and the subsequent check ((prev_host + size) ==
host) passes, then I believe the offsets would match as well.


+    return merge;
+}
+
+static void proxy_ml_region_addnop(MemoryListener *listener,
+                                   MemoryRegionSection *section)
+{
+    RemoteMemSync *sync = container_of(listener, RemoteMemSync, listener);
+    bool need_add = true;
+    uint64_t mrs_size, mrs_gpa, mrs_page;
+    uintptr_t mrs_host;
+    RAMBlock *mrs_rb;
+    MemoryRegionSection *prev_sec;
+
+    if (!(memory_region_is_ram(section->mr) &&
+          !memory_region_is_rom(section->mr))) {
+        return;
+    }
+
+    mrs_rb = section->mr->ram_block;
+    mrs_page = (uint64_t)qemu_ram_pagesize(mrs_rb);
+    mrs_size = int128_get64(section->size);
+    mrs_gpa = section->offset_within_address_space;
+    mrs_host = (uintptr_t)memory_region_get_ram_ptr(section->mr) +
+               section->offset_within_region;
+
+    if (get_fd_from_hostaddr(mrs_host, NULL) <= 0) {
+        return;
+    }
+
+    mrs_host = mrs_host & ~(mrs_page - 1);
+    mrs_gpa = mrs_gpa & ~(mrs_page - 1);
+    mrs_size = ROUND_UP(mrs_size, mrs_page);

OK, just note the more complex code in vhost_region_add_section for page
aligning regions that are needed for postcopy; I think that would be the
same if you wanted to do postcopy with remote processes.

Since mmap requires the addresses to be aligned with a page boundry, we
added these checks. We are essentially doing the same with alignage as
compared with vhost user. So we should be compliant with postcopy I
believe.


+    if (sync->n_mr_sections) {
+        prev_sec = sync->mr_sections + (sync->n_mr_sections - 1);
+        uint64_t prev_gpa_start = prev_sec->offset_within_address_space;
+        uint64_t prev_size = int128_get64(prev_sec->size);
+        uint64_t prev_gpa_end   = range_get_last(prev_gpa_start, prev_size);
+        uint64_t prev_host_start =
+            (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr) +
+            prev_sec->offset_within_region;
+        uint64_t prev_host_end = range_get_last(prev_host_start, prev_size);
+
+        if (mrs_gpa <= (prev_gpa_end + 1)) {
+            if (mrs_gpa < prev_gpa_start) {
+                assert(0);
+            }

g_assert(mrs_gpa < prev_gpa_start);

Thank you, we'll update the above check with the version you proposed.



+            if ((section->mr == prev_sec->mr) &&
+                proxy_mrs_can_merge(mrs_host, prev_host_start,
+                                    (mrs_gpa - prev_gpa_start))) {
+                uint64_t max_end = MAX(prev_host_end, mrs_host + mrs_size);
+                need_add = false;
+                prev_sec->offset_within_address_space =
+                    MIN(prev_gpa_start, mrs_gpa);
+                prev_sec->offset_within_region =
+                    MIN(prev_host_start, mrs_host) -
+                    (uintptr_t)memory_region_get_ram_ptr(prev_sec->mr);
+                prev_sec->size = int128_make64(max_end - MIN(prev_host_start,
+                                                             mrs_host));
+            }
+        }
+    }
+
+    if (need_add) {
+        ++sync->n_mr_sections;
+        sync->mr_sections = g_renew(MemoryRegionSection, sync->mr_sections,
+                                    sync->n_mr_sections);
+        sync->mr_sections[sync->n_mr_sections - 1] = *section;
+        sync->mr_sections[sync->n_mr_sections - 1].fv = NULL;
+        memory_region_ref(section->mr);
+    }

I'd add some tracing in this function; it's a nightmare to debug when it
does something unexpected.

Thank you, we'll add the tracing.


+}
+
+static void proxy_ml_commit(MemoryListener *listener)
+{
+    RemoteMemSync *sync = container_of(listener, RemoteMemSync, listener);
+    MPQemuMsg msg;
+    MemoryRegionSection section;
+    ram_addr_t offset;
+    uintptr_t host_addr;
+    int region;
+
+    memset(&msg, 0, sizeof(MPQemuMsg));
+
+    msg.cmd = SYNC_SYSMEM;
+    msg.bytestream = 0;
+    msg.num_fds = sync->n_mr_sections;
+    msg.size = sizeof(msg.data1);
+    assert(msg.num_fds <= REMOTE_MAX_FDS);
+
+    for (region = 0; region < sync->n_mr_sections; region++) {
+        section = sync->mr_sections[region];
+        msg.data1.sync_sysmem.gpas[region] =
+            section.offset_within_address_space;
+        msg.data1.sync_sysmem.sizes[region] = int128_get64(section.size);
+        host_addr = (uintptr_t)memory_region_get_ram_ptr(section.mr) +
+                    section.offset_within_region;
+        msg.fds[region] = get_fd_from_hostaddr(host_addr, &offset);

Since you already have section.mr you cna use memory_region_get_fd.

OK.

Thank you!
--
Jag


+        msg.data1.sync_sysmem.offsets[region] = offset;
+    }
+    mpqemu_msg_send(&msg, sync->mpqemu_link->com);
+}
+
+void deconfigure_memory_sync(RemoteMemSync *sync)
+{
+    memory_listener_unregister(&sync->listener);
+}
+
+/*
+ * TODO: Memory Sync need not be instantianted once per every proxy device.
+ *       All remote devices are going to get the exact same updates at the
+ *       same time. It therefore makes sense to have a broadcast model.
+ *
+ *       Broadcast model would involve running the MemorySync object in a
+ *       thread. MemorySync would contain a list of mpqemu-link objects
+ *       that need notification. proxy_ml_commit() could send the same
+ *       message to all the links at the same time.
+ */
+void configure_memory_sync(RemoteMemSync *sync, MPQemuLinkState *mpqemu_link)
+{
+    sync->n_mr_sections = 0;
+    sync->mr_sections = NULL;
+
+    sync->mpqemu_link = mpqemu_link;
+
+    sync->listener.begin = proxy_ml_begin;
+    sync->listener.commit = proxy_ml_commit;
+    sync->listener.region_add = proxy_ml_region_addnop;
+    sync->listener.region_nop = proxy_ml_region_addnop;
+    sync->listener.priority = 10;
+
+    memory_listener_register(&sync->listener, &address_space_memory);
+}
diff --git a/hw/proxy/qemu-proxy.c b/hw/proxy/qemu-proxy.c
index b17d9bb..d3a9d38 100644
--- a/hw/proxy/qemu-proxy.c
+++ b/hw/proxy/qemu-proxy.c
@@ -16,6 +16,8 @@
  #include "qapi/qmp/qjson.h"
  #include "qapi/qmp/qstring.h"
  #include "hw/proxy/qemu-proxy.h"
+#include "hw/proxy/memory-sync.h"
+#include "qom/object.h"
static void pci_proxy_dev_realize(PCIDevice *dev, Error **errp); @@ -243,6 +245,8 @@ static void init_proxy(PCIDevice *dev, char *command, char *exec_name, mpqemu_init_channel(pdev->mpqemu_link, &pdev->mpqemu_link->com,
                          pdev->socket);
+
+    configure_memory_sync(pdev->sync, pdev->mpqemu_link);
  }
static void pci_proxy_dev_realize(PCIDevice *device, Error **errp)
@@ -261,6 +265,7 @@ static void pci_proxy_dev_realize(PCIDevice *device, Error 
**errp)
      dev->set_proxy_sock = set_proxy_sock;
      dev->get_proxy_sock = get_proxy_sock;
      dev->init_proxy = init_proxy;
+    dev->sync = REMOTE_MEM_SYNC(object_new(TYPE_MEMORY_LISTENER));
  }
static void send_bar_access_msg(PCIProxyDev *dev, MemoryRegion *mr,
diff --git a/include/hw/proxy/memory-sync.h b/include/hw/proxy/memory-sync.h
new file mode 100644
index 0000000..d8329c9
--- /dev/null
+++ b/include/hw/proxy/memory-sync.h
@@ -0,0 +1,37 @@
+/*
+ * Copyright © 2018, 2020 Oracle and/or its affiliates.
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ *
+ */
+
+#ifndef MEMORY_SYNC_H
+#define MEMORY_SYNC_H
+
+#include <sys/types.h>
+
+#include "qemu/osdep.h"
+#include "qom/object.h"
+#include "exec/memory.h"
+#include "io/mpqemu-link.h"
+
+#define TYPE_MEMORY_LISTENER "memory-listener"
+#define REMOTE_MEM_SYNC(obj) \
+            OBJECT_CHECK(RemoteMemSync, (obj), TYPE_MEMORY_LISTENER)
+
+typedef struct RemoteMemSync {
+    Object obj;
+
+    MemoryListener listener;
+
+    int n_mr_sections;
+    MemoryRegionSection *mr_sections;
+
+    MPQemuLinkState *mpqemu_link;
+} RemoteMemSync;
+
+void configure_memory_sync(RemoteMemSync *sync, MPQemuLinkState *mpqemu_link);
+void deconfigure_memory_sync(RemoteMemSync *sync);
+
+#endif
diff --git a/include/hw/proxy/qemu-proxy.h b/include/hw/proxy/qemu-proxy.h
index 44e370e..c93ffe3 100644
--- a/include/hw/proxy/qemu-proxy.h
+++ b/include/hw/proxy/qemu-proxy.h
@@ -10,6 +10,7 @@
  #define QEMU_PROXY_H
#include "io/mpqemu-link.h"
+#include "hw/proxy/memory-sync.h"
#define TYPE_PCI_PROXY_DEV "pci-proxy-dev" @@ -37,8 +38,12 @@ extern const MemoryRegionOps proxy_default_ops;
  struct PCIProxyDev {
      PCIDevice parent_dev;
+ int n_mr_sections;
+    MemoryRegionSection *mr_sections;
+
      MPQemuLinkState *mpqemu_link;
+ RemoteMemSync *sync;
      pid_t remote_pid;
      int socket;
diff --git a/remote/remote-main.c b/remote/remote-main.c
index acd8daf..9512a3b 100644
--- a/remote/remote-main.c
+++ b/remote/remote-main.c
@@ -34,6 +34,7 @@
  #include "block/block.h"
  #include "exec/ramlist.h"
  #include "exec/memattrs.h"
+#include "exec/address-spaces.h"
static MPQemuLinkState *mpqemu_link;
  PCIDevice *remote_pci_dev;
@@ -161,6 +162,16 @@ static void process_msg(GIOCondition cond, MPQemuChannel 
*chan)
              goto finalize_loop;
          }
          break;
+    case SYNC_SYSMEM:
+        /*
+         * TODO: ensure no active DMA is happening when
+         * sysmem is being updated

In practice this turns out to be hard!

Dave

+         */
+        remote_sysmem_reconfig(msg, &err);
+        if (err) {
+            goto finalize_loop;
+        }
+        break;
      default:
          error_setg(&err, "Unknown command");
          goto finalize_loop;
--
1.8.3.1

--
Dr. David Alan Gilbert / address@hidden / Manchester, UK




reply via email to

[Prev in Thread] Current Thread [Next in Thread]