Skip to content
Closed
Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
29 commits
Select commit Hold shift + click to select a range
0a76c7e
Initial Large Page (2M) support for node.js
suresh-srinivas Mar 23, 2018
fa0e324
Add support for checking if explicitHugePages and transparentHugePage…
suresh-srinivas Mar 26, 2018
fdb45cd
Added License headers
suresh-srinivas Mar 30, 2018
b7791c2
Get rid of waring messages & code cleanup
suresh-srinivas Mar 30, 2018
6631eea
Protect the large pages under #ifdef NODE_ENABLE_LARGE_CODE_PAGES and…
suresh-srinivas Apr 4, 2018
843089c
Added configure option to enable huge pages
uttampawar Apr 9, 2018
077bc01
Added else clause to set node_use_large_pages=false by default
uttampawar Apr 9, 2018
7bdd9fc
Merge branch 'configure_fixes' into 2M-Pages-For-Code-16198
uttampawar Apr 9, 2018
b91de5a
Finished adding checks at appropriate places to handle possible error…
uttampawar May 8, 2018
7ef956a
Changed return type for two function from void to int
uttampawar May 9, 2018
b02e7a9
Fixed lint errors.
uttampawar May 24, 2018
2af82b1
Added one more condition check for verify the start address of newly …
uttampawar May 26, 2018
cae9285
Fixed syntax error due to wring data type. int64 to int64_t
uttampawar May 30, 2018
30114b6
Removed explictHugePages code
suresh-srinivas May 30, 2018
39e1f0d
Merge branch 'add-checks' into 2M-Pages-For-Code-16198
suresh-srinivas May 30, 2018
49cd0de
Merge branch '2M-Pages-For-Code-16198' of https://github.intel.com/DS…
suresh-srinivas May 30, 2018
2dd9e8c
Added Large Page Support
suresh-srinivas May 30, 2018
51d0f02
Merge branch 'intel-large-pages' of https://github.com/suresh-sriniva…
suresh-srinivas May 31, 2018
2f672ee
Update PR based on feedback
suresh-srinivas Jun 1, 2018
31504cc
Addressing the additional PR feedback
suresh-srinivas Jun 3, 2018
f998c58
Fix the gypi style issue
suresh-srinivas Jun 3, 2018
9f15cfc
Update with stylistic changes (eg char* instead of char *, IsLargePag…
suresh-srinivas Jun 4, 2018
d6de361
Add additional guard so large pages is only under Linux and target_ar…
suresh-srinivas Jun 4, 2018
29c7d13
Style fixes according to the Node C++ Style Guide
suresh-srinivas Jun 4, 2018
9828036
Eliminate ld.script and use implicit script
suresh-srinivas Jun 7, 2018
600cf54
Add #if defined(NODE_WANT_INTERNALS) && NODE_WANT_INTERNALS
suresh-srinivas Jun 7, 2018
bf259e2
Fix the test failures
suresh-srinivas Jun 7, 2018
f0a6dcb
Add detailed help message to configure --with-largepages
suresh-srinivas Jun 8, 2018
4610793
Fix style issues and created inline functions instead of macros for a…
suresh-srinivas Jun 12, 2018
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
Update PR based on feedback
  • Loading branch information
suresh-srinivas committed Jun 1, 2018
commit 2f672ee8fe4f9b21331ccaaa364ca9095d08db1b
6 changes: 2 additions & 4 deletions src/node.cc
Original file line number Diff line number Diff line change
Expand Up @@ -4379,9 +4379,8 @@ int Start(int argc, char** argv) {
CHECK_GT(argc, 0);

#ifdef NODE_ENABLE_LARGE_CODE_PAGES
if (node::largepages::isLargePagesEnabled()) {
if ((node::largepages::map_static_code_to_large_pages()) != 0) {
fprintf(stderr, "Mapping of static code to large pages failed.\n");
if (node::isLargePagesEnabled()) {
if ((node::map_static_code_to_large_pages()) != 0) {
fprintf(stderr, "Reverting to default page size\n");
}
}
Expand Down Expand Up @@ -4431,7 +4430,6 @@ int Start(int argc, char** argv) {
// will never be fully cleaned up.
v8_platform.Dispose();


delete[] exec_argv;
exec_argv = nullptr;

Expand Down
102 changes: 62 additions & 40 deletions src/node_large_page.cc
Original file line number Diff line number Diff line change
Expand Up @@ -35,65 +35,81 @@
#include <string>
#include <fstream>
#include <iostream>
#include <sstream>

// The functions in this file map the text segment of node into 2M pages.
// The algorithm is quite simple
// The algorithm is simple
// 1. Find the text region of node binary in memory
// Examine the /proc/self/maps to determine the currently mapped text
// region and obtain the start and end
// Modify the start to point to the very beginning of node text segment
// (from variable nodetext setup in ld.script)
// Align the address of start and end to Large Page Boundaries
//
// 2. Move the text region to large pages
// Map a new area and copy the original code there
// Use mmap using the start address with MAP_FIXED so we get exactly the
// same virtual address
// Use madvise with MADV_HUGE_PAGE to use Anonymous 2M Pages
// If successful copy the code there and unmap the original region.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

worthwhile to expand the algorithm a bit more - helps later maintainance.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure that sounds good. I will add more details.

extern char __executable_start;
extern char __etext;
extern char __nodetext;

namespace node {
namespace largepages {
#define ALIGN(x, a) (((x) + (a) - 1) & ~((a) - 1))
#define PAGE_ALIGN_UP(x, a) ALIGN(x, a)
#define PAGE_ALIGN_DOWN(x, a) ((x) & ~((a) - 1))
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

style nit: these could all be inline functions, right?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks. Yes fixed them to be inline functions.


struct TextRegion {
void * from;
void * to;
int totalHugePages;
int64_t offset;
bool found_text_region;
void * from;
void * to;
int total_hugepages;
bool found_text_region;
};

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

how many of these fields are used?
can we follow a consistent variable naming convention? (camel casing vs snake casing etc.) I suggest look around in node.cc for the prevailing nomenclature.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure, will look and make it consistent.

static void printSystemError(int error) {
fprintf(stderr, "Hugepages WARNING: %s\n", strerror(error));
return;
}

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am against printing anything - if this whole exercise works or fails, let it be transparent to outside.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It is easy to trip up on some OS configuration/settings. At the beginning as this is getting tested on more systems, we wanted to be able to provide some messages for users so we can also fix those problems.

// The format of the maps file is the following
// address perms offset dev inode pathname
// 00400000-00452000 r-xp 00000000 08:02 173521 /usr/bin/dbus-daemon

static struct TextRegion find_node_text_region() {
FILE *f;
int64_t start, end, offset, inode;
char perm[5], dev[6], name[256];
int ret;
const size_t hugePageSize = 2L * 1024 * 1024;
std::ifstream ifs;
std::string map_line;
std::string permission;
char dash;
int64_t start, end;
const size_t hps = 2L * 1024 * 1024;
struct TextRegion nregion;

nregion.found_text_region = false;

f = fopen("/proc/self/maps", "r");
ret = fscanf(f, "%lx-%lx %4s %lx %5s %ld %s\n",
&start, &end, perm, &offset, dev, &inode, name);
fclose(f);
ifs.open("/proc/self/maps");
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Assumes /proc is mounted. Not necessarily true.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok, added checks.

std::getline(ifs, map_line);
std::istringstream iss(map_line);
ifs.close();

iss >> std::hex >> start;
iss >> dash;
iss >> std::hex >> end;
iss >> permission;

if (ret == 7 &&
perm[0] == 'r' && perm[1] == '-' && perm[2] == 'x') {
// Checking if the region is from node binary and executable
if (permission.compare("r-xp") == 0) {
start = (unsigned int64_t) &__nodetext;
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Don't use C style casts, use proper C++ casts (in this case reinterpret_cast<uint64_t>.)

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok fixed that. in fact took out that line and use nodetext directly.

char *from = reinterpret_cast<char *>PAGE_ALIGN_UP(start, hugePageSize);
char *to = reinterpret_cast<char *>PAGE_ALIGN_DOWN(end, hugePageSize);
char *from = reinterpret_cast<char *>PAGE_ALIGN_UP(start, hps);
char *to = reinterpret_cast<char *>PAGE_ALIGN_DOWN(end, hps);

if (from < to) {
size_t size = (intptr_t)to - (intptr_t)from;
size_t size = to - from;
nregion.found_text_region = true;
nregion.from = from;
nregion.to = to;
nregion.offset = offset;
nregion.totalHugePages = size/hugePageSize;
return nregion;
nregion.total_hugepages = size / hps;
}
}

Expand All @@ -105,7 +121,8 @@ struct TextRegion {

ifs.open("/sys/kernel/mm/transparent_hugepage/enabled");
if (!ifs) {
fprintf(stderr, "WARNING: Couldn't check hugepages support\n");
fprintf(stderr, "Could not open file: " \
"/sys/kernel/mm/transparent_hugepage/enabled\n");
return false;
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can we follow a consistent return value convention? (returning null vs. -1 vs errno) I suggest look arounf in other APIs for the prevailing convention.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure

}

Expand All @@ -127,13 +144,15 @@ struct TextRegion {
return ret_status;
}

// Moving the text region to large pages. We need to be very careful.
// a) This function itself should not be moved.
// We use a gcc option to put it outside the ".text" section
// b) This function should not call any function(s) that might be moved.
// 1. We map a new area and copy the original code there
// 2. We mmap using HUGE_TLB
// 3. If successful we copy the code there and unmap the original region.
// Moving the text region to large pages. We need to be very careful.
// a) This function itself should not be moved.
// We use a gcc option to put it outside the ".text" section
// b) This function should not call any function(s) that might be moved.
// 1. map a new area and copy the original code there
// 2. mmap using the start address with MAP_FIXED so we get exactly
// the same virtual address
// 3. madvise with MADV_HUGE_PAGE
// 3. If successful copy the code there and unmap the original region
int
__attribute__((__section__(".eh_frame")))
__attribute__((__aligned__(2 * 1024 * 1024)))
Expand All @@ -156,6 +175,10 @@ struct TextRegion {

memcpy(nmem, r.from, size);

// We already know the original page is r-xp
// (PROT_READ, PROT_EXEC, MAP_PRIVATE)
// We want PROT_WRITE because we are writing into it.
// We want it at the fixed address and we use MAP_FIXED.
tmem = mmap(start, size,
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what are the chances you get what you want at the same location? start is already mapped, right? how about passing a nullptr here and get the huge pages wherever available?

Copy link
Copy Markdown
Contributor Author

@suresh-srinivas suresh-srinivas Jun 1, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If addr is not NULL, then the kernel takes it as a hint about where to place the mapping. Since we are using MAP_FIXED it is not a hint but a requirement. We are relying on the new mapping being at the same virtual address so we dont have to do any fix up of the code. Otherwise we will have to fixup the branches, offsets etc.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok - I was thinking on the lines on mremap , but looking at the man page for that and your implementation, I guess it is one and the same.

PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1 , 0);
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what is the rationale behind predefining the memory attributes? ideally we want to inherit the attribute from the old small pages?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We do want to be able to write to this so PROT_WRITE is needed. We are using MAP_FIXED to place the mapping at exactly that address.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Updated the code comments
We already know the original page is r-xp (PROT_READ, PROT_EXEC, MAP_PRIVATE)
We want PROT_WRITE because we are writing into it.
We want it at the fixed address and we use MAP_FIXED.
We are using MAP_ANONYMOUS so it's contents are initialized to zero and it is also not backed by any file.

Expand All @@ -166,7 +189,7 @@ struct TextRegion {
}

if (tmem != start) {
fprintf(stderr, "Unable to allocate hugepages.n");
fprintf(stderr, "Unable to allocate hugepages\n");
munmap(nmem, size);
munmap(tmem, size);
return -1;
Expand Down Expand Up @@ -213,14 +236,14 @@ struct TextRegion {

// This is the primary API called from main
int map_static_code_to_large_pages() {
struct TextRegion n = find_node_text_region();
if (n.found_text_region == false) {
fprintf(stderr, "Hugepages WARNING: failed to map static code\n");
struct TextRegion r = find_node_text_region();
if (r.found_text_region == false) {
fprintf(stderr, "Hugepages WARNING: failed to find text region \n");
return -1;
}

if (n.to <= reinterpret_cast<void *> (&move_text_region_to_large_pages))
return move_text_region_to_large_pages(n);
if (r.to <= reinterpret_cast<void *> (&move_text_region_to_large_pages))
return move_text_region_to_large_pages(r);

return -1;
}
Expand All @@ -229,5 +252,4 @@ struct TextRegion {
return isTransparentHugePagesEnabled();
}

} // namespace largepages
} // namespace node
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I admit I kinda checked out after the first 100 lines. This file has so many style errors, can you fix those first, please?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you be specific? I ran the make -s lint and fixed all the ones pointed to by that.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The linter doesn't catch everything but to name three things:

  1. Inconsistent indentation.
  2. Star leans left in pointers (void* p, not void *p or void * p.)
  3. Casing: IsLargePagesEnabled, not isLargePagesEnabled.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks I have updated the PR with 2) and 3). I am using standard indentation of 2 spaces.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@suresh-srinivas There’s https://github.com/nodejs/node/blob/master/CPP_STYLE_GUIDE.md that would help a lot with this. Also, comparing with other code in our codebase if you’re unsure. :)

(In particular, we don’t indent code inside namespaces)

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@addaleax thanks much. I will look at this and fix styles accordingly.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@addaleax @bnoordhuis @gireeshpunathil fixed the style errors. Let me know if you see anything else.

2 changes: 0 additions & 2 deletions src/node_large_page.h
Original file line number Diff line number Diff line change
Expand Up @@ -24,10 +24,8 @@
#define SRC_NODE_LARGE_PAGE_H_

namespace node {
Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should be enclosed in

#if defined(NODE_WANT_INTERNALS) && NODE_WANT_INTERNALS

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks I fixed it.

namespace largepages {
bool isLargePagesEnabled();
int map_static_code_to_large_pages();
} // namespace largepages
} // namespace node

#endif // SRC_NODE_LARGE_PAGE_H_
Expand Down