physmem: fd-based shared memory

Create MAP_SHARED RAMBlocks by mmap'ing a file descriptor rather than using
MAP_ANON, so the memory can be accessed in another process by passing and
mmap'ing the fd.  This will allow CPR to support memory-backend-ram and
memory-backend-shm objects, provided the user creates them with share=on.

Use memfd_create if available because it has no constraints.  If not, use
POSIX shm_open.  However, allocation on the opened fd may fail if the shm
mount size is too small, even if the system has free memory, so for backwards
compatibility fall back to qemu_anon_ram_alloc/MAP_ANON on failure.

For backwards compatibility on Windows, always use MAP_ANON.  share=on has
no purpose there, but the syntax is accepted, and must continue to work.

Lastly, quietly fall back to MAP_ANON if the system does not support
qemu_ram_alloc_from_fd.

Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/1736967650-129648-5-git-send-email-steven.sistare@oracle.com
Signed-off-by: Fabiano Rosas <farosas@suse.de>

system/physmem.c

@@ -48,6 +48,7 @@
 #include "qemu/qemu-print.h"
 #include "qemu/log.h"
 #include "qemu/memalign.h"
+#include "qemu/memfd.h"
 #include "exec/memory.h"
 #include "exec/ioport.h"
 #include "system/dma.h"
@@ -1948,6 +1949,7 @@ RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, ram_addr_t max_size,
                                  bool grow,
                                  Error **errp)
 {
+    ERRP_GUARD();
     RAMBlock *new_block;
     Error *local_err = NULL;
     int64_t file_size, file_align;
@@ -2068,6 +2070,25 @@ RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
 }
 #endif
 
+#ifdef CONFIG_POSIX
+/*
+ * Create MAP_SHARED RAMBlocks by mmap'ing a file descriptor, so it can be
+ * shared with another process if CPR is being used.  Use memfd if available
+ * because it has no size limits, else use POSIX shm.
+ */
+static int qemu_ram_get_shared_fd(const char *name, Error **errp)
+{
+    int fd;
+
+    if (qemu_memfd_check(0)) {
+        fd = qemu_memfd_create(name, 0, 0, 0, 0, errp);
+    } else {
+        fd = qemu_shm_alloc(0, errp);
+    }
+    return fd;
+}
+#endif
+
 static
 RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
                                   qemu_ram_resize_cb resized,
@@ -2081,6 +2102,41 @@ RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
     assert((ram_flags & ~(RAM_SHARED | RAM_RESIZEABLE | RAM_PREALLOC |
                           RAM_NORESERVE | RAM_GUEST_MEMFD)) == 0);
     assert(!host ^ (ram_flags & RAM_PREALLOC));
+    assert(max_size >= size);
+
+#ifdef CONFIG_POSIX         /* ignore RAM_SHARED for Windows */
+    if (!host) {
+        if (ram_flags & RAM_SHARED) {
+            const char *name = memory_region_name(mr);
+            int fd = qemu_ram_get_shared_fd(name, errp);
+
+            if (fd < 0) {
+                return NULL;
+            }
+
+            /* Use same alignment as qemu_anon_ram_alloc */
+            mr->align = QEMU_VMALLOC_ALIGN;
+
+            /*
+             * This can fail if the shm mount size is too small, or alloc from
+             * fd is not supported, but previous QEMU versions that called
+             * qemu_anon_ram_alloc for anonymous shared memory could have
+             * succeeded.  Quietly fail and fall back.
+             */
+            new_block = qemu_ram_alloc_from_fd(size, max_size, resized, mr,
+                                               ram_flags, fd, 0, false, NULL);
+            if (new_block) {
+                trace_qemu_ram_alloc_shared(name, new_block->used_length,
+                                            new_block->max_length, fd,
+                                            new_block->host);
+                return new_block;
+            }
+
+            close(fd);
+            /* fall back to anon allocation */
+        }
+    }
+#endif
 
     align = qemu_real_host_page_size();
     align = MAX(align, TARGET_PAGE_SIZE);
@@ -2092,7 +2148,6 @@ RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
     new_block->resized = resized;
     new_block->used_length = size;
     new_block->max_length = max_size;
-    assert(max_size >= size);
     new_block->fd = -1;
     new_block->guest_memfd = -1;
     new_block->page_size = qemu_real_host_page_size();

system/trace-events

@@ -33,6 +33,7 @@ address_space_map(void *as, uint64_t addr, uint64_t len, bool is_write, uint32_t
 find_ram_offset(uint64_t size, uint64_t offset) "size: 0x%" PRIx64 " @ 0x%" PRIx64
 find_ram_offset_loop(uint64_t size, uint64_t candidate, uint64_t offset, uint64_t next, uint64_t mingap) "trying size: 0x%" PRIx64 " @ 0x%" PRIx64 ", offset: 0x%" PRIx64" next: 0x%" PRIx64 " mingap: 0x%" PRIx64
 ram_block_discard_range(const char *rbname, void *hva, size_t length, bool need_madvise, bool need_fallocate, int ret) "%s@%p + 0x%zx: madvise: %d fallocate: %d ret: %d"
+qemu_ram_alloc_shared(const char *name, size_t size, size_t max_size, int fd, void *host) "%s size %zu max_size %zu fd %d host %p"
 
 # cpus.c
 vm_stop_flush_all(int ret) "ret %d"

util/memfd.c

@@ -194,17 +194,27 @@ bool qemu_memfd_alloc_check(void)
 /**
  * qemu_memfd_check():
  *
- * Check if host supports memfd.
+ * Check if host supports memfd.  Cache the answer for the common case flags=0.
  */
 bool qemu_memfd_check(unsigned int flags)
 {
 #ifdef CONFIG_LINUX
-    int mfd = memfd_create("test", flags | MFD_CLOEXEC);
+    int mfd;
+    static int memfd_check = MEMFD_TODO;
+
+    if (!flags && memfd_check != MEMFD_TODO) {
+        return memfd_check;
+    }
 
+    mfd = memfd_create("test", flags | MFD_CLOEXEC);
     if (mfd >= 0) {
         close(mfd);
-        return true;
     }
+    if (!flags) {
+        memfd_check = (mfd >= 0) ? MEMFD_OK : MEMFD_KO;
+    }
+    return (mfd >= 0);
+
 #endif
 
     return false;