Skip to contentSkip to conversation

Add MADV_COLLAPSE documentation

PATCH Superseded Zach OKeefe 3 participants 10 messages Oct 18, 2022 – Oct 21, 2022

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover Add MADV_COLLAPSE documentation

From: Zach O'Keefe <[email protected]>

v2 Forward

Changes from v1[1]:

- All patches have seen some reformatting and/or typo corrections (Alex Colomar).
- Patch 3 has had the commit description reworded to reflect that this commit is a "fix" rather than an "update" (Alex Colomar).
- Rebased on man-pages-6.01

[1] https://lore.kernel.org/linux-man/[email protected]/T/#m8e9e94ed52c99e7cf4969bd992492359c59a0faa

----------------------------------------------

Hey Alex / Michael,

This series adds MADV_COLLAPSE (expected for Linux 6.1) documentation to madvise(2) and process_madvise(2). A few prerequisite patches are included to fix up existing MADV_HUGEPAGE and process_madvise(2) documentation, as well as add some additional clarity for madvise(2) "advice" probing.

Series applies on man-pages-6.00, and I've attempted to use semantic newlines, though I can't claim I've made the right choices everywhere for long clauses.

Thanks,
Zach

Zach O'Keefe (4):
madvise.2: update THP file/shmem documentation for +5.4 madvise.2: document reliable probe for advice support process_madvise.2: CAP_SYS_ADMIN cleanup madvise.2: add documentation for MADV_COLLAPSE

 man2/madvise.2         | 133 +++++++++++++++++++++++++++++++++++++++--
 man2/process_madvise.2 |  13 +++-
 2 files changed, 140 insertions(+), 6 deletions(-)
Signature · 1 line
2.38.0.413.g74048e4d9e-goog
#1 lore raw

PATCH v2 1/4 madvise.2: update THP file/shmem documentation for +5.4

From: Zach O'Keefe <[email protected]>

Since Linux 5.4, Transparent Huge Pages now support both file-backed memory and shmem memory. Update MADV_HUGEPAGE advice description to reflect this.

Additionally, expand the description of requirements for memory to be considered eligible for THP: alignment / mapping requirements, VMA flags, prctl(2) settings, inode status, etc.

---
 man2/madvise.2 | 38 +++++++++++++++++++++++++++++++++++---
 1 file changed, 35 insertions(+), 3 deletions(-)
man2/madvise.2 +35 -3
index 81cce56af..64f788ace 100644
@@ -320,8 +320,6 @@ Enable Transparent Huge Pages (THP) for pages in the range specified by
 .I addr
 and
 .IR length .
-Currently, Transparent Huge Pages work only with private anonymous pages (see
-.BR mmap (2)).
 The kernel will regularly scan the areas marked as huge page candidates
 to replace them with huge pages.
 The kernel will also allocate huge pages directly when the region is
@@ -354,12 +352,46 @@ an access pattern that the developer knows in advance won't risk
 to increase the memory footprint of the application when transparent
 hugepages are enabled.
 .IP
+.\" commit 99cb0dbd47a15d395bf3faa78dc122bc5efe3fc0
+Since Linux 5.4,
+automatic scan of eligible areas and replacement by huge pages works with
+private anonymous pages (see
+.BR mmap (2)),
+shmem pages,
+and file-backed pages.
+For all memory types,
+memory may only be replaced by huge pages on hugepage-aligned boundaries.
+For file-mapped memory \(em including tmpfs (see
+.BR tmpfs (2))
+\(em the mapping must also be naturally hugepage-aligned within the file.
+Additionally,
+for file-backed,
+non-tmpfs memory,
+the file must not be open for write and the mapping must be executable.
+.IP
+The VMA must not be marked
+.BR VM_NOHUGEPAGE ,
+.BR VM_HUGETLB ,
+.BR VM_IO ,
+.BR VM_DONTEXPAND ,
+.BR VM_MIXEDMAP ,
+or
+.BR VM_PFNMAP ,
+nor can it be stack memory or backed by a DAX-enabled device
+(unless the DAX device is hot-plugged as System RAM).
+The process must also not have
+.B PR_SET_THP_DISABLE
+set (see
+.BR prctl (2) ).
+.IP
 The
 .B MADV_HUGEPAGE
 and
 .B MADV_NOHUGEPAGE
 operations are available only if the kernel was configured with
-.BR CONFIG_TRANSPARENT_HUGEPAGE .
+.B CONFIG_TRANSPARENT_HUGEPAGE
+and file/shmem memory is only supported if the kernel was configured with
+.BR CONFIG_READ_ONLY_THP_FOR_FS .
 .TP
 .BR MADV_NOHUGEPAGE " (since Linux 2.6.38)"
 Ensures that memory in the address range specified by
Signature · 1 line
2.38.0.413.g74048e4d9e-goog
#2 lore raw

PATCH v2 2/4 madvise.2: document reliable probe for advice support

In reply to Zach OKeefe

From: Zach O'Keefe <[email protected]>

EINVAL is an overloaded error code for madvise(2) and it's not clear under what context it means "advice is not valid" vs another error.

Explicitly document that madvise(0, 0, advice) can reliably be used to probe for kernel support for "advice", returning zero iff "advice" is supported by the kernel.

---
 man2/madvise.2 | 5 +++++
 1 file changed, 5 insertions(+)
man2/madvise.2 +5 -0
index 64f788ace..df3413cc8 100644
@@ -790,6 +790,11 @@ that are not mapped, the Linux version of
 ignores them and applies the call to the rest (but returns
 .B ENOMEM
 from the system call, as it should).
+.PP
+.IR madvise(0,\ 0,\ advice)
+will return zero iff
+.I advice
+is supported by the kernel and can be relied on to probe for support.
 .\" .SH HISTORY
 .\" The
 .\" .BR madvise ()
Signature · 1 line
2.38.0.413.g74048e4d9e-goog
#3 lore raw

PATCH v2 3/4 process_madvise.2: CAP_SYS_ADMIN cleanup

In reply to Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

Correct this in the man-page for process_madvise(2).

Fixes a144f458b ("process_madvise.2: Document process_madvise(2)")
Cc Suren Baghdasaryan <[email protected]>
Cc Minchan Kim <[email protected]>
---
 man2/process_madvise.2 | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
man2/process_madvise.2 +2 -1
index 6208206e4..7bee1a098 100644
@@ -113,7 +113,8 @@ check (see
 in addition,
 because of the performance implications of applying the advice,
 the caller must have the
-.B CAP_SYS_ADMIN
+.\" commit 96cfe2c0fd23ea7c2368d14f769d287e7ae1082e
+.B CAP_SYS_NICE
 capability.
 .SH RETURN VALUE
 On success,
Signature · 1 line
2.38.0.413.g74048e4d9e-goog
Reviewed-by on v2 3/4
#4 lore raw

Re: [PATCH man-pages v2 3/4] process_madvise.2: CAP_SYS_ADMIN cleanup

Quoted 16 lines · Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

Quoted 1 line

Correct this in the man-page for process_madvise(2).

Thanks for fixing my mistake!

Quoted 5 lines
Fixes a144f458b ("process_madvise.2: Document process_madvise(2)")
Cc Suren Baghdasaryan <[email protected]>
Cc Minchan Kim <[email protected]>
Quoted 21 lines
---
 man2/process_madvise.2 | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
man2/process_madvise.2 +2 -2
index 6208206e4..7bee1a098 100644
@@ -113,7 +113,8 @@ check (see
 in addition,
 because of the performance implications of applying the advice,
 the caller must have the
-.B CAP_SYS_ADMIN
+.\" commit 96cfe2c0fd23ea7c2368d14f769d287e7ae1082e
+.B CAP_SYS_NICE
 capability.
 .SH RETURN VALUE
 On success,
--

2.38.0.413.g74048e4d9e-goog

#5 lore raw

Re: [PATCH man-pages v2 3/4] process_madvise.2: CAP_SYS_ADMIN cleanup

In reply to Zach OKeefe

Hi Zach!

Thanks for the work! :-)

Quoted 6 lines · Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

Quoted 8 lines

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

Those two commits are several versions apart:

alx@asus5775:~/src/linux/linux$ git describe --contains ecb8ac8b1f14
v5.10-rc1~87^2~14
alx@asus5775:~/src/linux/linux$ git describe --contains 96cfe2c0fd23
v5.12-rc3~12^2~9

If I understand the paragraph above, from 5.10 to 5.12 the capability required was CAP_SYS_ADMIN?

Cheers,

Alex

Quoted 25 lines

Correct this in the man-page for process_madvise(2).

Fixes a144f458b ("process_madvise.2: Document process_madvise(2)")
Cc Suren Baghdasaryan <[email protected]>
Cc Minchan Kim <[email protected]>
---
  man2/process_madvise.2 | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
man2/process_madvise.2 +2 -1
index 6208206e4..7bee1a098 100644
@@ -113,7 +113,8 @@ check (see
  in addition,
  because of the performance implications of applying the advice,
  the caller must have the
-.B CAP_SYS_ADMIN
+.\" commit 96cfe2c0fd23ea7c2368d14f769d287e7ae1082e
+.B CAP_SYS_NICE
  capability.
  .SH RETURN VALUE
  On success,
Signature · 1 line
<http://www.alejandro-colomar.es/>
#6 lore raw

Re: [PATCH man-pages v2 3/4] process_madvise.2: CAP_SYS_ADMIN cleanup

Quoted 15 lines · Alejandro Colomar
Quoted 14 lines · Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

[...]

Quoted 1 line

If I understand the paragraph above, from 5.10 to 5.12 the capability required was CAP_SYS_ADMIN?

Or was it CAP_SYS_PTRACE?

Signature · 1 line
<http://www.alejandro-colomar.es/>
#7 lore raw

Re: [PATCH man-pages v2 3/4] process_madvise.2: CAP_SYS_ADMIN cleanup

Hey Alex!

Thanks for taking the time to review!

Quoted 18 lines · Alejandro Colomar
Quoted 7 lines · Alejandro Colomar
Quoted 6 lines · Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

Quoted 8 lines
Quoted 8 lines

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

Quoted 6 lines

[...]

Quoted 2 lines

If I understand the paragraph above, from 5.10 to 5.12 the capability required was CAP_SYS_ADMIN?

Or was it CAP_SYS_PTRACE?

Starting in 5.10, there was no CAP_* capability requirement - only PTRACE_MODE_ATTACH_FSCREDS (aka PTRACE_MODE_ATTACH | PTRACE_MODE_REALCREDS). Now, my understanding of the algorithm employed for ptrace access mode checking isn't to be trusted, but AFAIK, a caller having CAP_SYS_PTRACE in the target's user namespace (directly or transitively) isn't required to pass this (though it makes it easier). ptrace(2) has an overview of the algorithm.

Starting in 5.12, CAP_SYS_NICE was added as a requirement, and the ptrace algorithm used changed to PTRACE_MODE_READ.

If you think recording the differences in kernel versions in the man-page is important, let me know and I can amend this patch.

Thanks,
Zcah

Quoted 2 lines

--
<http://www.alejandro-colomar.es/>

#8 lore raw

Re: [PATCH man-pages v2 3/4] process_madvise.2: CAP_SYS_ADMIN cleanup

Hey Zach,

Quoted 39 lines · Zach O'Keefe

Hey Alex!

Thanks for taking the time to review!

Quoted 18 lines · Alejandro Colomar
Quoted 7 lines · Alejandro Colomar
Quoted 6 lines · Zach OKeefe

From: Zach O'Keefe <[email protected]>

The initial commit of process_madvise(2) to man-pages project included an error, indicating that CAP_SYS_ADMIN capability was required when, in fact, CAP_SYS_NICE was the required capability.

Quoted 8 lines
Quoted 8 lines

The initial commit of process_madvise(2) to Linux, commit ecb8ac8b1f14 ("mm/madvise: introduce process_madvise() syscall: an external memory hinting API"), relied on PTRACE_MODE_ATTACH_FSCREDS (see ptrace(2)), but was amended by commit 96cfe2c0fd23 ("mm/madvise: replace ptrace attach requirement for process_madvise") which replaced this with a combination of PTRACE_MODE_READ and CAP_SYS_NICE (PTRACE_MODE_READ to prevent leaking ASLR metadata and CAP_SYS_NICE for influencing process performance).

Quoted 6 lines

[...]

Quoted 2 lines

If I understand the paragraph above, from 5.10 to 5.12 the capability required was CAP_SYS_ADMIN?

Or was it CAP_SYS_PTRACE?

Starting in 5.10, there was no CAP_* capability requirement - only PTRACE_MODE_ATTACH_FSCREDS (aka PTRACE_MODE_ATTACH | PTRACE_MODE_REALCREDS). Now, my understanding of the algorithm employed for ptrace access mode checking isn't to be trusted, but AFAIK, a caller having CAP_SYS_PTRACE in the target's user namespace (directly or transitively) isn't required to pass this (though it makes it easier). ptrace(2) has an overview of the algorithm.

Quoted 2 lines

Starting in 5.12, CAP_SYS_NICE was added as a requirement, and the ptrace algorithm used changed to PTRACE_MODE_READ.

Understood.

Quoted 3 lines

If you think recording the differences in kernel versions in the man-page is important, let me know and I can amend this patch.

Yes; since it was live during 2 versions, I think we should at least mention it.
A couple of lines in NOTES might be enough.

Thanks,

Alex

Quoted 6 lines

Thanks,
Zcah

Quoted 2 lines

--
<http://www.alejandro-colomar.es/>

Signature · 1 line
<http://www.alejandro-colomar.es/>
#9 lore raw

PATCH v2 4/4 madvise.2: add documentation for MADV_COLLAPSE

In reply to Zach OKeefe

From: Zach O'Keefe <[email protected]>

Linux 6.1 introduced MADV_COLLAPSE in upstream commit 7d8faaf15545 ("mm/madvise: introduce MADV_COLLAPSE sync hugepage collapse") and upstream commit 34488399fa08 ("mm/madvise: add file and shmem support to MADV_COLLAPSE"). Update the man-pages for madvise(2) and process_madvise(2).

---
 man2/madvise.2         | 90 +++++++++++++++++++++++++++++++++++++++++-
 man2/process_madvise.2 | 10 +++++
 2 files changed, 98 insertions(+), 2 deletions(-)
man2/madvise.2 +88 -2
index df3413cc8..b03fc731d 100644
@@ -385,9 +385,10 @@ set (see
 .BR prctl (2) ).
 .IP
 The
-.B MADV_HUGEPAGE
+.BR MADV_HUGEPAGE ,
+.BR MADV_NOHUGEPAGE ,
 and
-.B MADV_NOHUGEPAGE
+.B MADV_COLLAPSE
 operations are available only if the kernel was configured with
 .B CONFIG_TRANSPARENT_HUGEPAGE
 and file/shmem memory is only supported if the kernel was configured with
@@ -400,6 +401,81 @@ and
 .I length
 will not be backed by transparent hugepages.
 .TP
+.BR MADV_COLLAPSE " (since Linux 6.1)"
+.\" commit 7d8faaf155454f8798ec56404faca29a82689c77
+.\" commit 34488399fa08faaf664743fa54b271eb6f9e1321
+Perform a best-effort synchronous collapse of the native pages mapped by the
+memory range into Transparent Huge Pages (THPs).
+.B MADV_COLLAPSE
+operates on the current state of memory of the calling process and makes no
+persistent changes or guarantees on how pages will be mapped,
+constructed,
+or faulted in the future.
+.IP
+.B MADV_COLLAPSE
+supports private anonymous pages (see
+.BR mmap (2)),
+shmem pages,
+and file-backed pages.
+See
+.B MADV_HUGEPAGE
+for general information on memory requirements for THP.
+If the range provided spans multiple VMAs,
+the semantics of the collapse over each VMA is independent from the others.
+If collapse of a given huge page-aligned/sized region fails,
+the operation may continue to attempt collapsing the remainder of the
+specified memory.
+.B MADV_COLLAPSE
+will automatically clamp the provided range to be hugepage-aligned.
+.IP
+All non-resident pages covered by the range will first be
+swapped/faulted-in,
+before being copied onto a freshly allocated hugepage.
+If the native pages compose the same PTE-mapped hugepage,
+and are suitably aligned,
+allocation of a new hugepage may be elided and collapse may happen
+in-place.
+Unmapped pages will have their data directly initialized to 0 in the new
+hugepage.
+However,
+for every eligible hugepage-aligned/sized region to be collapsed,
+at least one page must currently be backed by physical memory.
+.IP
+.BR MADV_COLLAPSE
+is independent of any sysfs
+(see
+.BR sysfs (5))
+setting under
+.IR /sys/kernel/mm/transparent_hugepage ,
+both in terms of determining THP eligibility,
+and allocation semantics.
+See Linux kernel source file
+.I Documentation/admin\-guide/mm/transhuge.rst
+for more information.
+.BR MADV_COLLAPSE
+also ignores
+.B huge=
+tmpfs mount when operating on tmpfs files.
+Allocation for the new hugepage may enter direct reclaim and/or compaction,
+regardless of VMA flags
+(though
+.BR VM_NOHUGEPAGE
+is still respected).
+.IP
+When the system has multiple NUMA nodes,
+the hugepage will be allocated from the node providing the most native
+pages.
+.IP
+If all hugepage-sized/aligned regions covered by the provided range were
+either successfully collapsed,
+or were already PMD-mapped THPs,
+this operation will be deemed successful.
+Note that this doesn’t guarantee anything about other possible mappings of
+the memory.
+Also note that many failures might have occurred since the operation may
+continue to collapse in the event collapse of a single hugepage-sized/aligned
+region fails.
+.TP
 .BR MADV_DONTDUMP " (since Linux 3.4)"
 .\" commit 909af768e88867016f427264ae39d27a57b6a8ed
 .\" commit accb61fe7bb0f5c2a4102239e4981650f9048519
@@ -619,6 +695,11 @@ A kernel resource was temporarily unavailable.
 .B EBADF
 The map exists, but the area maps something that isn't a file.
 .TP
+.B EBUSY
+(for
+.BR MADV_COLLAPSE )
+Could not charge hugepage to cgroup: cgroup limit exceeded.
+.TP
 .B EFAULT
 .I advice
 is
@@ -716,6 +797,11 @@ maximum resident set size.
 Not enough memory: paging in failed.
 .TP
 .B ENOMEM
+(for
+.BR MADV_COLLAPSE )
+Not enough memory: could not allocate hugepage.
+.TP
+.B ENOMEM
 Addresses in the specified range are not currently
 mapped, or are outside the address space of the process.
 .TP
man2/process_madvise.2 +10 -0
index 7bee1a098..900210106 100644
@@ -73,6 +73,10 @@ argument is one of the following values:
 See
 .BR madvise (2).
 .TP
+.B MADV_COLLAPSE
+See
+.BR madvise (2).
+.TP
 .B MADV_PAGEOUT
 See
 .BR madvise (2).
@@ -170,6 +174,12 @@ The caller does not have permission to access the address space of the process
 .TP
 .B ESRCH
 The target process does not exist (i.e., it has terminated and been waited on).
+.PP
+See
+.BR madvise (2)
+for
+.IR advice -specific
+errors.
 .SH VERSIONS
 This system call first appeared in Linux 5.10.
 .\" commit ecb8ac8b1f146915aa6b96449b66dd48984caacc
Signature · 1 line
2.38.0.413.g74048e4d9e-goog