-
Eric B Munson authored
mlock() allows a user to control page out of program memory, but this comes at the cost of faulting in the entire mapping when it is allocated. For large mappings where the entire area is not necessary this is not ideal. Instead of forcing all locked pages to be present when they are allocated, this set creates a middle ground. Pages are marked to be placed on the unevictable LRU (locked) when they are first used, but they are not faulted in by the mlock call. This series introduces a new mlock() system call that takes a flags argument along with the start address and size. This flags argument gives the caller the ability to request memory be locked in the traditional way, or to be locked after the page is faulted in. A new MCL flag is added to mirror the lock on fault behavior from mlock() in mlockall(). There are two main use cases that this set covers. The first is the security focussed mlock case. A buffer is needed that cannot be written to swap. The maximum size is known, but on average the memory used is significantly less than this maximum. With lock on fault, the buffer is guaranteed to never be paged out without consuming the maximum size every time such a buffer is created. The second use case is focussed on performance. Portions of a large file are needed and we want to keep the used portions in memory once accessed. This is the case for large graphical models where the path through the graph is not known until run time. The entire graph is unlikely to be used in a given invocation, but once a node has been used it needs to stay resident for further processing. Given these constraints we have a number of options. We can potentially waste a large amount of memory by mlocking the entire region (this can also cause a significant stall at startup as the entire file is read in). We can mlock every page as we access them without tracking if the page is already resident but this introduces large overhead for each access. The third option is mapping the entire region with PROT_NONE and using a signal handler for SIGSEGV to mprotect(PROT_READ) and mlock() the needed page. Doing this page at a time adds a significant performance penalty. Batching can be used to mitigate this overhead, but in order to safely avoid trying to mprotect pages outside of the mapping, the boundaries of each mapping to be used in this way must be tracked and available to the signal handler. This is precisely what the mm system in the kernel should already be doing. For mlock(MLOCK_ONFAULT) the user is charged against RLIMIT_MEMLOCK as if mlock(MLOCK_LOCKED) or mmap(MAP_LOCKED) was used, so when the VMA is created not when the pages are faulted in. For mlockall(MCL_ONFAULT) the user is charged as if MCL_FUTURE was used. This decision was made to keep the accounting checks out of the page fault path. To illustrate the benefit of this set I wrote a test program that mmaps a 5 GB file filled with random data and then makes 15,000,000 accesses to random addresses in that mapping. The test program was run 20 times for each setup. Results are reported for two program portions, setup and execution. The setup phase is calling mmap and optionally mlock on the entire region. For most experiments this is trivial, but it highlights the cost of faulting in the entire region. Results are averages across the 20 runs in milliseconds. mmap with mlock(MLOCK_LOCKED) on entire range: Setup avg: 8228.666 Processing avg: 8274.257 mmap with mlock(MLOCK_LOCKED) before each access: Setup avg: 0.113 Processing avg: 90993.552 mmap with PROT_NONE and signal handler and batch size of 1 page: With the default value in max_map_count, this gets ENOMEM as I attempt to change the permissions, after upping the sysctl significantly I get: Setup avg: 0.058 Processing avg: 69488.073 mmap with PROT_NONE and signal handler and batch size of 8 pages: Setup avg: 0.068 Processing avg: 38204.116 mmap with PROT_NONE and signal handler and batch size of 16 pages: Setup avg: 0.044 Processing avg: 29671.180 mmap with mlock(MLOCK_ONFAULT) on entire range: Setup avg: 0.189 Processing avg: 17904.899 The signal handler in the batch cases faulted in memory in two steps to avoid having to know the start and end of the faulting mapping. The first step covers the page that caused the fault as we know that it will be possible to lock. The second step speculatively tries to mlock and mprotect the batch size - 1 pages that follow. There may be a clever way to avoid this without having the program track each mapping to be covered by this handeler in a globally accessible structure, but I could not find it. It should be noted that with a large enough batch size this two step fault handler can still cause the program to crash if it reaches far beyond the end of the mapping. These results show that if the developer knows that a majority of the mapping will be used, it is better to try and fault it in at once, otherwise mlock(MLOCK_ONFAULT) is significantly faster. The performance cost of these patches are minimal on the two benchmarks I have tested (stream and kernbench). The following are the average values across 20 runs of stream and 10 runs of kernbench after a warmup run whose results were discarded. Avg throughput in MB/s from stream using 1000000 element arrays Test 4.2-rc1 4.2-rc1+lock-on-fault Copy: 10,566.5 10,421 Scale: 10,685 10,503.5 Add: 12,044.1 11,814.2 Triad: 12,064.8 11,846.3 Kernbench optimal load 4.2-rc1 4.2-rc1+lock-on-fault Elapsed Time 78.453 78.991 User Time 64.2395 65.2355 System Time 9.7335 9.7085 Context Switches 22211.5 22412.1 Sleeps 14965.3 14956.1 This patch (of 6): Extending the mlock system call is very difficult because it currently does not take a flags argument. A later patch in this set will extend mlock to support a middle ground between pages that are locked and faulted in immediately and unlocked pages. To pave the way for the new system call, the code needs some reorganization so that all the actual entry point handles is checking input and translating to VMA flags. Signed-off-by:Eric B Munson <emunson@akamai.com> Acked-by:
1aab92ec