From: Helge Deller <deller@gmx.de>
To: matoro <matoro_mailinglist_kernel@matoro.tk>
Cc: Jens Axboe <axboe@kernel.dk>,
io-uring@vger.kernel.org, Linux Ia64 <linux-ia64@vger.kernel.org>,
glaubitz@physik.fu-berlin.de, Sam James <sam@gentoo.org>
Subject: Re: [PATCH 1/5] io_uring: Adjust mapping wrt architecture aliasing requirements
Date: Thu, 13 Jul 2023 09:27:28 +0200 [thread overview]
Message-ID: <ZK+nYIxe6zf2vYwH@ls3530> (raw)
In-Reply-To: <8bb091fb5fd00072842fe92b03abac9b@matoro.tk>
* matoro <matoro_mailinglist_kernel@matoro.tk>:
> On 2023-07-12 16:30, Helge Deller wrote:
> > On 7/12/23 21:05, Helge Deller wrote:
> > > On 7/12/23 19:28, matoro wrote:
> > > > On 2023-07-12 12:24, Helge Deller wrote:
> > > > > Hi Matoro,
> > > > >
> > > > > * matoro <matoro_mailinglist_kernel@matoro.tk>:
> > > > > > On 2023-03-14 13:16, Jens Axboe wrote:
> > > > > > > From: Helge Deller <deller@gmx.de>
> > > > > > >
> > > > > > > Some architectures have memory cache aliasing requirements (e.g. parisc)
> > > > > > > if memory is shared between userspace and kernel. This patch fixes the
> > > > > > > kernel to return an aliased address when asked by userspace via mmap().
> > > > > > >
> > > > > > > Signed-off-by: Helge Deller <deller@gmx.de>
> > > > > > > Signed-off-by: Jens Axboe <axboe@kernel.dk>
> > > > > > > ---
> > > > > > > io_uring/io_uring.c | 51 +++++++++++++++++++++++++++++++++++++++++++++
> > > > > > > 1 file changed, 51 insertions(+)
> > > > > > >
> > > > > > > diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
> > > > > > > index 722624b6d0dc..3adecebbac71 100644
> > > > > > > --- a/io_uring/io_uring.c
> > > > > > > +++ b/io_uring/io_uring.c
> > > > > > > @@ -72,6 +72,7 @@
> > > > > > > #include <linux/io_uring.h>
> > > > > > > #include <linux/audit.h>
> > > > > > > #include <linux/security.h>
> > > > > > > +#include <asm/shmparam.h>
> > > > > > >
> > > > > > > #define CREATE_TRACE_POINTS
> > > > > > > #include <trace/events/io_uring.h>
> > > > > > > @@ -3317,6 +3318,54 @@ static __cold int io_uring_mmap(struct file
> > > > > > > *file, struct vm_area_struct *vma)
> > > > > > > return remap_pfn_range(vma, vma->vm_start, pfn, sz,
> > > > > > > vma->vm_page_prot);
> > > > > > > }
> > > > > > >
> > > > > > > +static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
> > > > > > > + unsigned long addr, unsigned long len,
> > > > > > > + unsigned long pgoff, unsigned long flags)
> > > > > > > +{
> > > > > > > + const unsigned long mmap_end = arch_get_mmap_end(addr, len, flags);
> > > > > > > + struct vm_unmapped_area_info info;
> > > > > > > + void *ptr;
> > > > > > > +
> > > > > > > + /*
> > > > > > > + * Do not allow to map to user-provided address to avoid breaking the
> > > > > > > + * aliasing rules. Userspace is not able to guess the offset address
> > > > > > > of
> > > > > > > + * kernel kmalloc()ed memory area.
> > > > > > > + */
> > > > > > > + if (addr)
> > > > > > > + return -EINVAL;
> > > > > > > +
> > > > > > > + ptr = io_uring_validate_mmap_request(filp, pgoff, len);
> > > > > > > + if (IS_ERR(ptr))
> > > > > > > + return -ENOMEM;
> > > > > > > +
> > > > > > > + info.flags = VM_UNMAPPED_AREA_TOPDOWN;
> > > > > > > + info.length = len;
> > > > > > > + info.low_limit = max(PAGE_SIZE, mmap_min_addr);
> > > > > > > + info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
> > > > > > > +#ifdef SHM_COLOUR
> > > > > > > + info.align_mask = PAGE_MASK & (SHM_COLOUR - 1UL);
> > > > > > > +#else
> > > > > > > + info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
> > > > > > > +#endif
> > > > > > > + info.align_offset = (unsigned long) ptr;
> > > > > > > +
> > > > > > > + /*
> > > > > > > + * A failed mmap() very likely causes application failure,
> > > > > > > + * so fall back to the bottom-up function here. This scenario
> > > > > > > + * can happen with large stack limits and large mmap()
> > > > > > > + * allocations.
> > > > > > > + */
> > > > > > > + addr = vm_unmapped_area(&info);
> > > > > > > + if (offset_in_page(addr)) {
> > > > > > > + info.flags = 0;
> > > > > > > + info.low_limit = TASK_UNMAPPED_BASE;
> > > > > > > + info.high_limit = mmap_end;
> > > > > > > + addr = vm_unmapped_area(&info);
> > > > > > > + }
> > > > > > > +
> > > > > > > + return addr;
> > > > > > > +}
> > > > > > > +
> > > > > > > #else /* !CONFIG_MMU */
> > > > > > >
> > > > > > > static int io_uring_mmap(struct file *file, struct vm_area_struct *vma)
> > > > > > > @@ -3529,6 +3578,8 @@ static const struct file_operations io_uring_fops
> > > > > > > = {
> > > > > > > #ifndef CONFIG_MMU
> > > > > > > .get_unmapped_area = io_uring_nommu_get_unmapped_area,
> > > > > > > .mmap_capabilities = io_uring_nommu_mmap_capabilities,
> > > > > > > +#else
> > > > > > > + .get_unmapped_area = io_uring_mmu_get_unmapped_area,
> > > > > > > #endif
> > > > > > > .poll = io_uring_poll,
> > > > > > > #ifdef CONFIG_PROC_FS
> > > > > >
> > > > > > Hi Jens, Helge - I've bisected a regression with
> > > > > > io_uring on ia64 to this
> > > > > > patch in 6.4. Unfortunately this breaks userspace
> > > > > > programs using io_uring,
> > > > > > the easiest one to test is cmake with an io_uring
> > > > > > enabled libuv (i.e., libuv
> > > > > > >= 1.45.0) which will hang.
> > > > > >
> > > > > > I am aware that ia64 is in a vulnerable place right now
> > > > > > which I why I am
> > > > > > keeping this spread limited. Since this clearly involves
> > > > > > architecture-specific changes for parisc,
> > > > >
> > > > > it isn't so much architecture-specific... (just one ifdef)
> > > > >
> > > > > > is there any chance of looking at
> > > > > > what is required to do the same for ia64? I looked at
> > > > > > 0ef36bd2b37815719e31a72d2beecc28ca8ecd26 ("parisc:
> > > > > > change value of SHMLBA
> > > > > > from 0x00400000 to PAGE_SIZE") and tried to replicate the SHMLBA ->
> > > > > > SHM_COLOUR change, but it made no difference.
> > > > > >
> > > > > > If hardware is necessary for testing, I can provide it,
> > > > > > including remote BMC
> > > > > > access for restarts/kernel debugging. Any takers?
> > > > >
> > > > > I won't have time to test myself, but maybe you could test?
> > > > >
> > > > > Basically we should try to find out why
> > > > > io_uring_mmu_get_unmapped_area()
> > > > > doesn't return valid addresses, while arch_get_unmapped_area()
> > > > > [in arch/ia64/kernel/sys_ia64.c] does.
> > > > >
> > > > > You could apply this patch first:
> > > > > It introduces a memory leak (as it requests memory twice),
> > > > > but maybe we
> > > > > get an idea?
> > > > > The ia64 arch_get_unmapped_area() searches for memory from bottom
> > > > > (flags=0), while io_uring function tries top-down first.
> > > > > Maybe that's
> > > > > the problem. And I don't understand the offset_in_page() check right
> > > > > now.
> > > > >
> > > > > diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
> > > > > index 3bca7a79efda..93b1964d2bbb 100644
> > > > > --- a/io_uring/io_uring.c
> > > > > +++ b/io_uring/io_uring.c
> > > > > @@ -3431,13 +3431,17 @@ static unsigned long
> > > > > io_uring_mmu_get_unmapped_area(struct file *filp,
> > > > > * can happen with large stack limits and large mmap()
> > > > > * allocations.
> > > > > */
> > > > > +/* compare to arch_get_unmapped_area() in
> > > > > arch/ia64/kernel/sys_ia64.c */
> > > > > addr = vm_unmapped_area(&info);
> > > > > - if (offset_in_page(addr)) {
> > > > > +printk("io_uring_mmu_get_unmapped_area() address 1 is:
> > > > > %px\n", addr);
> > > > > + addr = NULL;
> > > > > + if (!addr) {
> > > > > info.flags = 0;
> > > > > info.low_limit = TASK_UNMAPPED_BASE;
> > > > > info.high_limit = mmap_end;
> > > > > addr = vm_unmapped_area(&info);
> > > > > }
> > > > > +printk("io_uring_mmu_get_unmapped_area() returns address
> > > > > %px\n", addr);
> > > > >
> > > > > return addr;
> > > > > }
> > > > >
> > > > >
> > > > > Another option is to disable the call to
> > > > > io_uring_nommu_get_unmapped_area())
> > > > > with the next patch. Maybe you could add printks() to ia64's
> > > > > arch_get_unmapped_area()
> > > > > and check what it returns there?
> > > > >
> > > > > @@ -3654,6 +3658,8 @@ static const struct file_operations
> > > > > io_uring_fops = {
> > > > > #ifndef CONFIG_MMU
> > > > > .get_unmapped_area = io_uring_nommu_get_unmapped_area,
> > > > > .mmap_capabilities = io_uring_nommu_mmap_capabilities,
> > > > > +#elif 0 /* IS_ENABLED(CONFIG_IA64) */
> > > > > + .get_unmapped_area = NULL,
> > > > > #else
> > > > > .get_unmapped_area = io_uring_mmu_get_unmapped_area,
> > > > > #endif
> > > > >
> > > > > Helge
> > > >
> > > > Thanks Helge. Sample output from that first patch:
> > > >
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > address 1 is: 1ffffffffff40000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > returns address 2000000001e40000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > address 1 is: 1ffffffffff20000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > returns address 2000000001f20000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > address 1 is: 1ffffffffff30000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > returns address 2000000001f30000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > address 1 is: 1ffffffffff90000
> > > > [Wed Jul 12 13:09:50 2023] io_uring_mmu_get_unmapped_area()
> > > > returns address 2000000001f90000
> > > >
> > > > This pattern seems to be pretty stable, I tried instead just
> > > > directly returning the result of a call to
> > > > arch_get_unmapped_area() at the end of the function and it seems
> > > > similar:
> > > >
> > > > [Wed Jul 12 13:27:07 2023] io_uring_mmu_get_unmapped_area()
> > > > would return address 1ffffffffffd0000
> > > > [Wed Jul 12 13:27:07 2023] but arch_get_unmapped_area() would
> > > > return address 2000000001f00000
> > > > [Wed Jul 12 13:27:07 2023] io_uring_mmu_get_unmapped_area()
> > > > would return address 1ffffffffff00000
> > > > [Wed Jul 12 13:27:07 2023] but arch_get_unmapped_area() would
> > > > return address 1ffffffffff00000
> > > > [Wed Jul 12 13:27:07 2023] io_uring_mmu_get_unmapped_area()
> > > > would return address 1fffffffffe20000
> > > > [Wed Jul 12 13:27:07 2023] but arch_get_unmapped_area() would
> > > > return address 2000000002000000
> > > > [Wed Jul 12 13:27:07 2023] io_uring_mmu_get_unmapped_area()
> > > > would return address 1fffffffffe30000
> > > > [Wed Jul 12 13:27:07 2023] but arch_get_unmapped_area() would
> > > > return address 2000000002100000
> > > >
> > > > Is that enough of a clue to go on?
> > >
> > > SHMLBA on ia64 is 0x100000:
> > > arch/ia64/include/asm/shmparam.h:#define SHMLBA (1024*1024)
> > > but the values returned by io_uring_mmu_get_unmapped_area() does not
> > > fullfill this.
> > >
> > > So, probably ia64's SHMLBA isn't pulled in correctly in
> > > io_uring/io_uring.c.
> > > Check value of this line:
> > > info.align_mask = PAGE_MASK & (SHMLBA - 1UL);
> > >
> > > You could also add
> > > #define SHM_COLOUR 0x100000
> > > in front of the
> > > #ifdef SHM_COLOUR
> > > (define SHM_COLOUR in io_uring/kbuf.c too).
> >
> > What is the value of PAGE_SIZE and "ptr" on your machine?
> > For 4k page size I get:
> > SHMLBA -1 -> FFFFF
> > PAGE_MASK -> FFFFFFFFF000
> > so,
> > info.align_mask = PAGE_MASK & (SHMLBA - 1UL) = 0xFF000;
> > You could try to set nfo.align_mask = 0xfffff;
> >
> > Helge
>
> Using 64KiB (65536) PAGE_SIZE here. 64-bit pointers.
>
> Tried both #define SHM_COLOUR 0x100000, as well and info.align_mask =
> 0xFFFFF, but both of them made the problem change from 100% reproducible, to
> intermittent.
>
> After inspecting the ouput I observed that it hangs only when the first
> allocation returns an address below 0x2000000000000000, and the second
> returns an address above it. When both addresses are above it, it does not
> hang. Examples:
>
> When it works:
> $ cmake --version
> cmake version 3.26.4
>
> CMake suite maintained and supported by Kitware (kitware.com/cmake).
> $ dmesg --color=always -T | tail -n 4
> [Wed Jul 12 20:32:37 2023] io_uring_mmu_get_unmapped_area() would return
> address 1fffffffffe20000
> [Wed Jul 12 20:32:37 2023] but arch_get_unmapped_area() would return address
> 2000000002000000
> [Wed Jul 12 20:32:37 2023] io_uring_mmu_get_unmapped_area() would return
> address 1fffffffffe50000
> [Wed Jul 12 20:32:37 2023] but arch_get_unmapped_area() would return address
> 2000000002100000
>
>
> When it hangs:
> $ cmake --version
> cmake version 3.26.4
>
> CMake suite maintained and supported by Kitware (kitware.com/cmake).
> ^C
> $ dmesg --color=always -T | tail -n 4
> [Wed Jul 12 20:33:12 2023] io_uring_mmu_get_unmapped_area() would return
> address 1ffffffffff00000
> [Wed Jul 12 20:33:12 2023] but arch_get_unmapped_area() would return address
> 1ffffffffff00000
> [Wed Jul 12 20:33:12 2023] io_uring_mmu_get_unmapped_area() would return
> address 1fffffffffe60000
> [Wed Jul 12 20:33:12 2023] but arch_get_unmapped_area() would return address
> 2000000001f00000
>
> Is io_uring_mmu_get_unmapped_area supported to always return addresses above
> 0x2000000000000000?
Yes, with the patch below.
> Any reason why it is not doing so sometimes?
It depends on the parameters for vm_unmapped_area(). Specifically
info.flags=0.
Try this patch:
diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
index 3bca7a79efda..b259794ab53b 100644
--- a/io_uring/io_uring.c
+++ b/io_uring/io_uring.c
@@ -3429,10 +3429,13 @@ static unsigned long io_uring_mmu_get_unmapped_area(struct file *filp,
* A failed mmap() very likely causes application failure,
* so fall back to the bottom-up function here. This scenario
* can happen with large stack limits and large mmap()
- * allocations.
+ * allocations. Use bottom-up on IA64 for correct aliasing.
*/
- addr = vm_unmapped_area(&info);
- if (offset_in_page(addr)) {
+ if (IS_ENABLED(CONFIG_IA64))
+ addr = NULL;
+ else
+ addr = vm_unmapped_area(&info);
+ if (!addr) {
info.flags = 0;
info.low_limit = TASK_UNMAPPED_BASE;
info.high_limit = mmap_end;
Helge
next prev parent reply other threads:[~2023-07-13 7:28 UTC|newest]
Thread overview: 11+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20230314171641.10542-1-axboe@kernel.dk>
[not found] ` <20230314171641.10542-2-axboe@kernel.dk>
2023-07-12 4:43 ` [PATCH 1/5] io_uring: Adjust mapping wrt architecture aliasing requirements matoro
2023-07-12 16:24 ` Helge Deller
2023-07-12 17:28 ` matoro
2023-07-12 19:05 ` Helge Deller
2023-07-12 20:30 ` Helge Deller
2023-07-13 0:35 ` matoro
2023-07-13 7:27 ` Helge Deller [this message]
2023-07-13 23:57 ` matoro
2023-07-16 6:54 ` Helge Deller
2023-07-16 18:03 ` matoro
2023-07-16 20:54 ` Helge Deller
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=ZK+nYIxe6zf2vYwH@ls3530 \
--to=deller@gmx.de \
--cc=axboe@kernel.dk \
--cc=glaubitz@physik.fu-berlin.de \
--cc=io-uring@vger.kernel.org \
--cc=linux-ia64@vger.kernel.org \
--cc=matoro_mailinglist_kernel@matoro.tk \
--cc=sam@gentoo.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).