qemu-devel
[Top][All Lists]
Advanced

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

Re: [Qemu-devel] [PATCH v3] linux-user: add support for MADV_DONTNEED


From: Simon Hausmann
Subject: Re: [Qemu-devel] [PATCH v3] linux-user: add support for MADV_DONTNEED
Date: Thu, 6 Sep 2018 12:11:57 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.0


On 9/6/18 11:34 AM, Laurent Vivier wrote:
Le 27/08/2018 à 10:40, Simon Hausmann a écrit :
Most flags to madvise() are just hints, so typically ignoring the
syscall and returning okay is fine. However applications exist that do
rely on MADV_DONTNEED behavior to guarantee that upon subsequent access
the mapping is refreshed from the backing file or zero for anonymous
mappings. The file backed mappings this is tricky - hence the original
comment - but for anonymous mappings it is safe to forward. We just need
to remember which those mappings are, which is now stored in the flags.

Cc: Riku Voipio <address@hidden>
Cc: Laurent Vivier <address@hidden>
Cc: Sami Nurmenniemi <address@hidden>
Signed-off-by: Simon Hausmann <address@hidden>

--
v2:
   - align start and len on host page size
v3:
   - align start and len to target page size as it may be bigger than
     host
   - use immediate return in do_syscall
   - forward MADV_DONTNEED advice only for anonymously mapped pages
   - remember which mappings are anonymous in the page flags and preserve
     those bits across mprotect calls.
---
  accel/tcg/translate-all.c |  8 +++++--
  bsd-user/mmap.c           |  8 +++----
  include/exec/cpu-all.h    | 11 ++++++++-
  linux-user/elfload.c      |  3 ++-
  linux-user/mmap.c         | 49 +++++++++++++++++++++++++++++++++++----
  linux-user/qemu.h         |  1 +
  linux-user/syscall.c      | 12 ++++------
  7 files changed, 72 insertions(+), 20 deletions(-)

diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c
index 898c3bb3d1..467fbd9aeb 100644
--- a/accel/tcg/translate-all.c
+++ b/accel/tcg/translate-all.c
@@ -2481,7 +2481,8 @@ int page_get_flags(target_ulong address)
  /* Modify the flags of a page and invalidate the code if necessary.
     The flag PAGE_WRITE_ORG is positioned automatically depending
     on PAGE_WRITE.  The mmap_lock should already be held.  */
-void page_set_flags(target_ulong start, target_ulong end, int flags)
+void page_set_flags(target_ulong start, target_ulong end, int flags,
+                    enum page_set_flags_mode mode)
Is it possible to not add the mode, and only use "flags" to set the new
flag? Using page_get_flags() to keep flags that needed to be kept?


Hm, the reason why I didn't do that is because page_set_flags operates

on a range while page_get_flags() gets the flags for only one page.

If you prefer not to have the mode, then I can see only two other options:


    (1) Assume that the flags are typically homogeneous and use something like

page_set_flags(start, end, page_get_flags(start) | whatever)


    (2) Change the call sites to not use the range setting ability of page_set_flags but

loop itself.


I don't like either, so I went for the mode. But I can change to any way you prefer.


  {
      target_ulong addr, len;
@@ -2513,7 +2514,10 @@ void page_set_flags(target_ulong start, target_ulong end, int flags)
              p->first_tb) {
              tb_invalidate_phys_page(addr, 0);
          }
-        p->flags = flags;
+        if (mode == PAGE_SET_ALL_FLAGS)
+            p->flags = flags;
+        else /* PAGE_SET_PROTECTION */
+            p->flags |= (p->flags & ~(PAGE_BITS - 1)) | (flags & PAGE_BITS);
      }
  }
diff --git a/bsd-user/mmap.c b/bsd-user/mmap.c
index 17f4cd80aa..4bfac81af4 100644
--- a/bsd-user/mmap.c
+++ b/bsd-user/mmap.c
@@ -125,7 +125,7 @@ int target_mprotect(abi_ulong start, abi_ulong len, int 
prot)
          if (ret != 0)
              goto error;
      }
-    page_set_flags(start, start + len, prot | PAGE_VALID);
+    page_set_flags(start, start + len, prot, PAGE_SET_PROTECTION);
Why do you remove the PAGE_VALID flag?


As the mode set PAGE_SET_PROTECTION, the page_set_flags call only affects the protection

bits and the existing PAGE_VALID is preserved. I believe PAGE_VALID must be set for that range.

If that assumption is wrong, then I could change page_set_flags to implicitly set PAGE_VALID if the

mode is to apply new protection flags. Would that be ok?


...
diff --git a/linux-user/elfload.c b/linux-user/elfload.c
index 8638612aec..c59cf4359c 100644
--- a/linux-user/elfload.c
+++ b/linux-user/elfload.c
@@ -1702,7 +1702,8 @@ static void zero_bss(abi_ulong elf_bss, abi_ulong 
last_bss, int prot)
/* Ensure that the bss page(s) are valid */
      if ((page_get_flags(last_bss-1) & prot) != prot) {
-        page_set_flags(elf_bss & TARGET_PAGE_MASK, last_bss, prot | 
PAGE_VALID);
+        page_set_flags(elf_bss & TARGET_PAGE_MASK, last_bss, prot,
+                       PAGE_SET_PROTECTION);
      }
PAGE_VALID?


Oh! You're right, here I mistakenly thought the pages are already marked valid, but

it appears that this is a new mapping, so I'll fix that to set PAGE_VALID.


if (host_start < host_map_start) {
diff --git a/linux-user/mmap.c b/linux-user/mmap.c
index 41e0983ce8..fff29ee04b 100644
--- a/linux-user/mmap.c
+++ b/linux-user/mmap.c
@@ -124,7 +124,7 @@ int target_mprotect(abi_ulong start, abi_ulong len, int 
prot)
          if (ret != 0)
              goto error;
      }
-    page_set_flags(start, start + len, prot | PAGE_VALID);
+    page_set_flags(start, start + len, prot, PAGE_SET_PROTECTION);
      mmap_unlock();
PAGE_VALID?

...


Same - potentially wrong - reasoning as with the other mprotect() wrapper.

+int target_madvise(abi_ulong start, abi_ulong len, int advice)
+{
+    abi_ulong end, addr;
+    int ret;
+
+    len = TARGET_PAGE_ALIGN(len);
+    start &= TARGET_PAGE_MASK;
+
+    if (!guest_range_valid(start, len)) {
+        errno = EINVAL;
+        return -1;
+    }
+
+    /* A straight passthrough may not be safe because qemu sometimes
+       turns private file-backed mappings into anonymous mappings.
+       Most flags are hints so we ignore them.
+       One exception is made for MADV_DONTNEED on anonymous mappings,
+       that applications may rely on to zero out pages. */
+    if (advice & MADV_DONTNEED) {
+        end = start + len;
+        for (addr = start; addr < end; addr += TARGET_PAGE_SIZE) {
+            if (page_get_flags(addr) & PAGE_MAP_ANONYMOUS) {
+                ret = madvise(g2h(addr), TARGET_PAGE_SIZE, MADV_DONTNEED);
these values must be aligned on host page size.

Sorry, I think I assumed from the earlier statement that the target page size may be bigger than the

host and that that would be sufficient for host alignment. I didn't consider that the reverse may also apply.

Would the following be correct?


    (1) Use target aligned start/len for guest_range_valid call

    (2) Use host aligned start/end/increment for the loop (including the page_get_flags call parameters).


Thanks for the feedback :)


Simon







reply via email to

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