- 07 Aug, 2014 40 commits
-
-
Rasmus Villemoes authored
The compiler can generate slightly smaller and simpler code when it knows that "nbits" is non-negative. Since no-one passes a negative bit-count, this shouldn't affect the semantics. Signed-off-by:
Rasmus Villemoes <linux@rasmusvillemoes.dk> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org>
-
Rasmus Villemoes authored
The compiler can generate slightly smaller and simpler code when it knows that "nbits" is non-negative. Since no-one passes a negative bit-count, this shouldn't affect the semantics. Signed-off-by:
-
Rasmus Villemoes authored
The compiler can generate slightly smaller and simpler code when it knows that "nbits" is non-negative. Since no-one passes a negative bit-count, this shouldn't affect the semantics. Signed-off-by:
-
Rasmus Villemoes authored
Many functions in lib/bitmap.c start with an expression such as lim = bits/BITS_PER_LONG. Since bits has type (signed) int, and since gcc cannot know that it is in fact non-negative, it generates worse code than it could. These patches, mostly consisting of changing various parameters to unsigned, gives a slight overall code reduction: add/remove: 1/1 grow/shrink: 8/16 up/down: 251/-414 (-163) function old new delta tick_device_uses_broadcast 335 425 +90 __irq_alloc_descs 498 554 +56 __bitmap_andnot 73 115 +42 __bitmap_and 70 101 +31 bitmap_weight - 11 +11 copy_hugetlb_page_range 752 762 +10 follow_hugetlb_page 846 854 +8 hugetlb_init 1415 1417 +2 hugetlb_nrpages_setup 130 131 +1 hugetlb_add_hstate 377 376 -1 bitmap_allocate_region 82 80 -2 select_task_rq_fair 2202 2191 -11 hweight_long 66 55 -11 __reg_op 230 219 -11 dm_stats_message 2849 2833 -16 bitmap_parselist 92 74 -18 __bitmap_weight 115 97 -18 __bitmap_subset 153 129 -24 __bitmap_full 128 104 -24 __bitmap_empty 120 96 -24 bitmap_set 179 149 -30 bitmap_clear 185 155 -30 __bitmap_equal 136 105 -31 __bitmap_intersects 148 108 -40 __bitmap_complement 109 67 -42 tick_device_setup_broadcast_func.isra 81 - -81 [The increases in __bitmap_and{,not} are due to bug fixes 17/18,18/18. No idea why bitmap_weight suddenly appears.] While 163 bytes treewide is insignificant, I believe the bitmap functions are often called with locks held, so saving even a few cycles might be worth it. While making these changes, I found a few other things that might be worth including. 16,17,18 are actual bug fixes. The rest shouldn't change the behaviour of any of the functions, provided no-one passed negative nbits values. If something should come up, it should be fairly bisectable. A few issues I thought about, but didn't know what to do with: * Many of the functions misbehave if nbits is compile-time 0; the out-of-line functions generally handle 0 correctly. bitmap_fill() is particularly bad, whether the 0 is known at compile time or not. It would probably be nice to add detection of at least compile-time 0 and handle that appropriately. * I didn't change __bitmap_shift_{left,right} to use unsigned because I want to fully understand why the algorithm works before making that change. However, AFAICT, they behave correctly for all (positive) shift amounts. This is not the case for the small_const_nbits versions. If for example nbits = n = BITS_PER_LONG, the shift operators turn into no-ops (at least on x86), so one get *dst = *src, whereas one would expect to get *dst=0. That difference in behaviour is somewhat annoying. This patch (of 18): The compiler can generate slightly smaller and simpler code when it knows that "nbits" is non-negative. Since no-one passes a negative bit-count, this shouldn't affect the semantics. Signed-off-by: -
Andrew Morton authored
Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk> Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com> Cc: Don Mullis <don.mullis@gmail.com> Cc: Dave Chinner <david@fromorbit.com> Signed-off-by:
-
Rasmus Villemoes authored
The helper merge_and_restore_back_links() makes sure to call the caller's cmp function during the final ->prev pointer fixup, so that the cmp function may call cond_resched(). However, if the cmp function does not call cond_resched() at all, this is entirely redundant. If it does, doing at least two function calls for every two pointer assignments is a bit excessive. This patch limits the calls to once for every 256 iterations. Signed-off-by:
-
Rasmus Villemoes authored
There is no reason to maintain the list structure while freeing the debug elements. Aside from the redundant pointer manipulations, it is also inefficient from a locality-of-reference viewpoint, since they are visited in a random order (wrt. the order they were allocated). Furthermore, if we jumped to exit: after detecting list corruption, it is actually dangerous. So just free the elements in the order they were allocated, using the backing array elts. Allocate that using kcalloc(), so that if allocation of one of the debug element fails, we just end up calling kfree(NULL) for the trailing elements. Minor details: Use sizeof(*elts) instead of sizeof(void *), and return err immediately when allocation of elts fails, to avoid introducing another label just before the final return statement. Signed-off-by:
-
Rasmus Villemoes authored
Add a check to make sure that the prev pointer of the list head points to the last element on the list. Signed-off-by:
-
Rasmus Villemoes authored
Signed-off-by:
-
Joe Perches authored
It's been nearly 3 years now since commit 55036ba7 ("lib: rename pack_hex_byte() to hex_byte_pack()") so it's time to remove this deprecated and unused static inline. Signed-off-by:
Joe Perches <joe@perches.com> Signed-off-by:
-
Fabian Frederick authored
Use kernel.h definition. Signed-off-by:
Fabian Frederick <fabf@skynet.be> Cc: Alexey Dobriyan <adobriyan@gmail.com> Signed-off-by:
-
Mathias Krause authored
Complement commit 68aecfb9 ("lib/string_helpers.c: make arrays static") by making the arrays const -- not only pointing to const strings. This moves them out of the data section to the r/o data section: text data bss dec hex filename 1150 176 0 1326 52e lib/string_helpers.old.o 1326 0 0 1326 52e lib/string_helpers.new.o Signed-off-by:
Mathias Krause <minipli@googlemail.com> Cc: James Bottomley <James.Bottomley@HansenPartnership.com> Signed-off-by:
-
Gui Hecheng authored
For modern filesystems such as btrfs, t/p/e size level operations are common. add size unit t/p/e parsing to memparse Signed-off-by:
Gui Hecheng <guihc.fnst@cn.fujitsu.com> Acked-by:
David Rientjes <rientjes@google.com> Reviewed-by:
Satoru Takeuchi <takeuchi_satoru@jp.fujitsu.com> Signed-off-by:
-
George Spelvin authored
The function may be useful for other drivers, so export it. (Suggested by Tejun Heo.) Note that I inverted the return value of glob_match; returning true on match seemed to make more sense. Signed-off-by:
George Spelvin <linux@horizon.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by:
-
George Spelvin authored
This was useful during development, and is retained for future regression testing. GCC appears to have no way to place string literals in a particular section; adding __initconst to a char pointer leaves the string itself in the default string section, where it will not be thrown away after module load. Thus all string constants are kept in explicitly declared and named arrays. Sorry this makes printk a bit harder to read. At least the tests are more compact. Signed-off-by:
-
George Spelvin authored
This is a helper function from drivers/ata/libata_core.c, where it is used to blacklist particular device models. It's being moved to lib/ so other drivers may use it for the same purpose. This implementation in non-recursive, so is safe for the kernel stack. [akpm@linux-foundation.org: fix sparse warning] Signed-off-by:
-
Sergey Senozhatsky authored
Cleanup unused `if 0'-ed functions, which have been dead since 2006 (commits 87c2ce3b ("lib/zlib*: cleanups") by Adrian Bunk and 4f3865fb ("zlib_inflate: Upgrade library code to a recent version") by Richard Purdie): - zlib_deflateSetDictionary - zlib_deflateParams - zlib_deflateCopy - zlib_inflateSync - zlib_syncsearch - zlib_inflateSetDictionary - zlib_inflatePrime Signed-off-by:
Sergey Senozhatsky <sergey.senozhatsky@gmail.com> Signed-off-by:
-
Ken Helias authored
The name was modified from hlist_add_after() to hlist_add_behind() when adjusting the order of arguments to match the one with klist_add_after(). This is necessary to break old code when it would use it the wrong way. Make klist follow this naming scheme for consistency. Signed-off-by:
Ken Helias <kenhelias@firemail.de> Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> Cc: Christoph Hellwig <hch@infradead.org> Cc: Hugh Dickins <hughd@google.com> Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com> Signed-off-by:
-
Ken Helias authored
All other add functions for lists have the new item as first argument and the position where it is added as second argument. This was changed for no good reason in this function and makes using it unnecessary confusing. The name was changed to hlist_add_behind() to cause unconverted code to generate a compile error instead of using the wrong parameter order. [akpm@linux-foundation.org: coding-style fixes] Signed-off-by:
-
Ken Helias authored
The argument names for hlist_add_after() are poorly chosen because they look the same as the ones for hlist_add_before() but have to be used differently. hlist_add_after_rcu() has made a better choice. Signed-off-by:
-
Neil Zhang authored
Fix coccinelle warnings. Signed-off-by:
Neil Zhang <zhangwm@marvell.com> Signed-off-by:
-
Jan Kara authored
We need interrupts disabled when calling console_trylock_for_printk() only so that cpu id we pass to can_use_console() remains valid (for other things console_sem provides all the exclusion we need and deadlocks on console_sem due to interrupts are impossible because we use down_trylock()). However if we are rescheduled, we are guaranteed to run on an online cpu so we can easily just get the cpu id in can_use_console(). We can lose a bit of performance when we enable interrupts in vprintk_emit() and then disable them again in console_unlock() but OTOH it can somewhat reduce interrupt latency caused by console_unlock(). We differ from (reverted) commit 939f04be in that we avoid calling console_unlock() from vprintk_emit() with lockdep enabled as that has unveiled quite some bugs leading to system freezes during boot (e.g. https://lkml.org/lkml/2014/5/30/242, https://lkml.org/lkml/2014/6/28/521). Signed-off-by:
Jan Kara <jack@suse.cz> Tested-by:
Andreas Bombe <aeb@debian.org> Signed-off-by:
-
Alex Elder authored
Some small cleanups to kernel/printk/printk.c. None of them should cause any change in behavior. - When CONFIG_PRINTK is defined, parenthesize the value of LOG_LINE_MAX. - When CONFIG_PRINTK is *not* defined, there is an extra LOG_LINE_MAX definition; delete it. - Pull an assignment out of a conditional expression in console_setup(). - Use isdigit() in console_setup() rather than open coding it. - In update_console_cmdline(), drop a NUL-termination assignment; the strlcpy() call that precedes it guarantees it's not needed. - Simplify some logic in printk_timed_ratelimit(). Signed-off-by:Alex Elder <elder@linaro.org> Reviewed-by:
Petr Mladek <pmladek@suse.cz> Cc: Andi Kleen <ak@linux.intel.com> Cc: Borislav Petkov <bp@suse.de> Cc: Jan Kara <jack@suse.cz> Cc: John Stultz <john.stultz@linaro.org> Cc: Steven Rostedt <rostedt@goodmis.org> Signed-off-by:
-
Alex Elder authored
Use the IS_ENABLED() macro rather than #ifdef blocks to set certain global values. Signed-off-by:
Borislav Petkov <bp@suse.de> Reviewed-by:
-
Alex Elder authored
Fix a few comments that don't accurately describe their corresponding code. It also fixes some minor typographical errors. Signed-off-by:
-
Alex Elder authored
Commit a8fe19eb ("kernel/printk: use symbolic defines for console loglevels") makes consistent use of symbolic values for printk() log levels. The naming scheme used is different from the one used for DEFAULT_MESSAGE_LOGLEVEL though. Change that symbol name to be MESSAGE_LOGLEVEL_DEFAULT for consistency. And because the value of that symbol comes from a similarly-named config option, rename CONFIG_DEFAULT_MESSAGE_LOGLEVEL as well. Signed-off-by:
-
Alex Elder authored
In do_syslog() there's a path used by kmsg_poll() and kmsg_read() that only needs to know whether there's any data available to read (and not its size). These callers only check for non-zero return. As a shortcut, do_syslog() returns the difference between what has been logged and what has been "seen." The comments say that the "count of records" should be returned but it's not. Instead it returns (log_next_idx - syslog_idx), which is a difference between buffer offsets--and the result could be negative. The behavior is the same (it'll be zero or not in the same cases), but the count of records is more meaningful and it matches what the comments say. So change the code to return that. Signed-off-by:
-
Luis R. Rodriguez authored
The default size of the ring buffer is too small for machines with a large amount of CPUs under heavy load. What ends up happening when debugging is the ring buffer overlaps and chews up old messages making debugging impossible unless the size is passed as a kernel parameter. An idle system upon boot up will on average spew out only about one or two extra lines but where this really matters is on heavy load and that will vary widely depending on the system and environment. There are mechanisms to help increase the kernel ring buffer for tracing through debugfs, and those interfaces even allow growing the kernel ring buffer per CPU. We also have a static value which can be passed upon boot. Relying on debugfs however is not ideal for production, and relying on the value passed upon bootup is can only used *after* an issue has creeped up. Instead of being reactive this adds a proactive measure which lets you scale the amount of contributions you'd expect to the kernel ring buffer under load by each CPU in the worst case scenario. We use num_possible_cpus() to avoid complexities which could be introduced by dynamically changing the ring buffer size at run time, num_possible_cpus() lets us use the upper limit on possible number of CPUs therefore avoiding having to deal with hotplugging CPUs on and off. This introduces the kernel configuration option LOG_CPU_MAX_BUF_SHIFT which is used to specify the maximum amount of contributions to the kernel ring buffer in the worst case before the kernel ring buffer flips over, the size is specified as a power of 2. The total amount of contributions made by each CPU must be greater than half of the default kernel ring buffer size (1 << LOG_BUF_SHIFT bytes) in order to trigger an increase upon bootup. The kernel ring buffer is increased to the next power of two that would fit the required minimum kernel ring buffer size plus the additional CPU contribution. For example if LOG_BUF_SHIFT is 18 (256 KB) you'd require at least 128 KB contributions by other CPUs in order to trigger an increase of the kernel ring buffer. With a LOG_CPU_BUF_SHIFT of 12 (4 KB) you'd require at least anything over > 64 possible CPUs to trigger an increase. If you had 128 possible CPUs the amount of minimum required kernel ring buffer bumps to: ((1 << 18) + ((128 - 1) * (1 << 12))) / 1024 = 764 KB Since we require the ring buffer to be a power of two the new required size would be 1024 KB. This CPU contributions are ignored when the "log_buf_len" kernel parameter is used as it forces the exact size of the ring buffer to an expected power of two value. [pmladek@suse.cz: fix build] Signed-off-by:
Luis R. Rodriguez <mcgrof@suse.com> Signed-off-by:
Davidlohr Bueso <davidlohr@hp.com> Tested-by:
-
Luis R. Rodriguez authored
Signed-off-by:
-
Luis R. Rodriguez authored
In practice the power of 2 practice of the size of the kernel ring buffer remains purely historical but not a requirement, specially now that we have LOG_ALIGN and use it for both static and dynamic allocations. It could have helped with implicit alignment back in the days given the even the dynamically sized ring buffer was guaranteed to be aligned so long as CONFIG_LOG_BUF_SHIFT was set to produce a __LOG_BUF_LEN which is architecture aligned, since log_buf_len=n would be allowed only if it was > __LOG_BUF_LEN and we always ended up rounding the log_buf_len=n to the next power of 2 with roundup_pow_of_two(), any multiple of 2 then should be also architecture aligned. These assumptions of course relied heavily on CONFIG_LOG_BUF_SHIFT producing an aligned value but users can always change this. We now have precise alignment requirements set for the log buffer size for both static and dynamic allocations, but lets upkeep the old practice of using powers of 2 for its size to help with easy expected scalable values and the allocators for dynamic allocations. We'll reuse this later so move this into a helper. Signed-off-by:
-
Luis R. Rodriguez authored
We have to consider alignment for the ring buffer both for the default static size, and then also for when an dynamic allocation is made when the log_buf_len=n kernel parameter is passed to set the size specifically to a size larger than the default size set by the architecture through CONFIG_LOG_BUF_SHIFT. The default static kernel ring buffer can be aligned properly if architectures set CONFIG_LOG_BUF_SHIFT properly, we provide ranges for the size though so even if CONFIG_LOG_BUF_SHIFT has a sensible aligned value it can be reduced to a non aligned value. Commit 6ebb017d ("printk: Fix alignment of buf causing crash on ARM EABI") by Andrew Lunn ensures the static buffer is always aligned and the decision of alignment is done by the compiler by using __alignof__(struct log). When log_buf_len=n is used we allocate the ring buffer dynamically. Dynamic allocation varies, for the early allocation called before setup_arch() memblock_virt_alloc() requests a page aligment and for the default kernel allocation memblock_virt_alloc_nopanic() requests no special alignment, which in turn ends up aligning the allocation to SMP_CACHE_BYTES, which is L1 cache aligned. Since we already have the required alignment for the kernel ring buffer though we can do better and request explicit alignment for LOG_ALIGN. This does that to be safe and make dynamic allocation alignment explicit. Signed-off-by:
-
Geoff Levand authored
Signed-off-by:
Geoff Levand <geoff@infradead.org> Signed-off-by:
-
Joe Perches authored
The DEFINE_SIMPLE_ATTRIBUTE macro should not end in a ; Fix the one use in the kernel tree that did not have a semicolon. Signed-off-by:
Guenter Roeck <linux@roeck-us.net> Acked-by:
Luca Tettamanti <kronos.it@gmail.com> Signed-off-by:
-
Jiri Kosina authored
We have been chasing a memory corruption bug, which turned out to be caused by very old gcc (4.3.4), which happily turned conditional load into a non-conditional one, and that broke correctness (the condition was met only if lock was held) and corrupted memory. This particular problem with that particular code did not happen when never gccs were used. I've brought this up with our gcc folks, as I wanted to make sure that this can't really happen again, and it turns out it actually can. Quoting Martin Jambor <mjambor@suse.cz>: "More current GCCs are more careful when it comes to replacing a conditional load with a non-conditional one, most notably they check that a store happens in each iteration of _a_ loop but they assume loops are executed. They also perform a simple check whether the store cannot trap which currently passes only for non-const variables. A simple testcase demonstrating it on an x86_64 is for example the following: $ cat cond_store.c int g_1 = 1; int g_2[1024] __attribute__((section ("safe_section"), aligned (4096))); int c = 4; int __attribute__ ((noinline)) foo (void) { int l; for (l = 0; (l != 4); l++) { if (g_1) return l; for (g_2[0] = 0; (g_2[0] >= 26); ++g_2[0]) ; } return 2; } int main (int argc, char* argv[]) { if (mprotect (g_2, sizeof(g_2), PROT_READ) == -1) { int e = errno; error (e, e, "mprotect error %i", e); } foo (); __builtin_printf("OK\n"); return 0; } /* EOF */ $ ~/gcc/trunk/inst/bin/gcc cond_store.c -O2 --param allow-store-data-races=0 $ ./a.out OK $ ~/gcc/trunk/inst/bin/gcc cond_store.c -O2 --param allow-store-data-races=1 $ ./a.out Segmentation fault The testcase fails the same at least with 4.9, 4.8 and 4.7. Therefore I would suggest building kernels with this parameter set to zero. I also agree with Jikos that the default should be changed for -O2. I have run most of the SPEC 2k6 CPU benchmarks (gamess and dealII failed, at -O2, not sure why) compiled with and without this option and did not see any real difference between respective run-times" Hopefully the default will be changed in newer gccs, but let's force it for kernel builds so that we are on a safe side even when older gcc are used. The code in question was out-of-tree printk-in-NMI (yeah, surprise suprise, once again) patch written by Petr Mladek, let me quote his comment from our internal bugzilla: "I have spent few days investigating inconsistent state of kernel ring buffer. It went out that it was caused by speculative store generated by gcc-4.3.4. The problem is in assembly generated for make_free_space(). The functions is called the following way: + vprintk_emit(); + log = MAIN_LOG; // with logbuf_lock or log = NMI_LOG; // with nmi_logbuf_lock cont_add(log, ...); + cont_flush(log, ...); + log_store(log, ...); + log_make_free_space(log, ...); If called with log = NMI_LOG then only nmi_log_* global variables are safe to modify but the generated code does store also into (main_)log_* global variables: <log_make_free_space>: 55 push %rbp 89 f6 mov %esi,%esi 48 8b 05 03 99 51 01 mov 0x1519903(%rip),%rax # ffffffff82620868 <nmi_log_next_id> 44 8b 1d ec 98 51 01 mov 0x15198ec(%rip),%r11d # ffffffff82620858 <log_next_idx> 8b 35 36 60 14 01 mov 0x1146036(%rip),%esi # ffffffff8224cfa8 <log_buf_len> 44 8b 35 33 60 14 01 mov 0x1146033(%rip),%r14d # ffffffff8224cfac <nmi_log_buf_len> 4c 8b 2d d0 98 51 01 mov 0x15198d0(%rip),%r13 # ffffffff82620850 <log_next_seq> 4c 8b 25 11 61 14 01 mov 0x1146111(%rip),%r12 # ffffffff8224d098 <log_buf> 49 89 c2 mov %rax,%r10 48 21 c2 and %rax,%rdx 48 8b 1d 0c 99 55 01 mov 0x155990c(%rip),%rbx # ffffffff826608a0 <nmi_log_buf> 49 c1 ea 20 shr $0x20,%r10 48 89 55 d0 mov %rdx,-0x30(%rbp) 44 29 de sub %r11d,%esi 45 29 d6 sub %r10d,%r14d 4c 8b 0d 97 98 51 01 mov 0x1519897(%rip),%r9 # ffffffff82620840 <log_first_seq> eb 7e jmp ffffffff81107029 <log_make_free_space+0xe9> [...] 85 ff test %edi,%edi # edi = 1 for NMI_LOG 4c 89 e8 mov %r13,%rax 4c 89 ca mov %r9,%rdx 74 0a je ffffffff8110703d <log_make_free_space+0xfd> 8b 15 27 98 51 01 mov 0x1519827(%rip),%edx # ffffffff82620860 <nmi_log_first_id> 48 8b 45 d0 mov -0x30(%rbp),%rax 48 39 c2 cmp %rax,%rdx # end of loop 0f 84 da 00 00 00 je ffffffff81107120 <log_make_free_space+0x1e0> [...] 85 ff test %edi,%edi # edi = 1 for NMI_LOG 4c 89 0d 17 97 51 01 mov %r9,0x1519717(%rip) # ffffffff82620840 <log_first_seq> ^^^^^^^^^^^^^^^^^^^^^^^^^^ KABOOOM 74 35 je ffffffff81107160 <log_make_free_space+0x220> It stores log_first_seq when edi == NMI_LOG. This instructions are used also when edi == MAIN_LOG but the store is done speculatively before the condition is decided. It is unsafe because we do not have "logbuf_lock" in NMI context and some other process migh modify "log_first_seq" in parallel" I believe that the best course of action is both - building kernel (and anything multi-threaded, I guess) with that optimization turned off - persuade gcc folks to change the default for future releases Signed-off-by:Jiri Kosina <jkosina@suse.cz> Cc: Martin Jambor <mjambor@suse.cz> Cc: Petr Mladek <pmladek@suse.cz> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Marek Polacek <polacek@redhat.com> Cc: Jakub Jelinek <jakub@redhat.com> Cc: Steven Noonan <steven@uplinklabs.net> Cc: Richard Biener <richard.guenther@gmail.com> Cc: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by:
-
Dan Streetman authored
Change zswap to use the zpool api instead of directly using zbud. Add a boot-time param to allow selecting which zpool implementation to use, with zbud as the default. Signed-off-by:
Dan Streetman <ddstreet@ieee.org> Tested-by:
Seth Jennings <sjennings@variantweb.net> Cc: Weijie Yang <weijie.yang@samsung.com> Cc: Minchan Kim <minchan@kernel.org> Cc: Nitin Gupta <ngupta@vflare.org> Signed-off-by:
-
Dan Streetman authored
Update zbud and zsmalloc to implement the zpool api. [fengguang.wu@intel.com: make functions static] Signed-off-by:
-
Dan Streetman authored
Add zpool api. zpool provides an interface for memory storage, typically of compressed memory. Users can select what backend to use; currently the only implementations are zbud, a low density implementation with up to two compressed pages per storage page, and zsmalloc, a higher density implementation with multiple compressed pages per storage page. Signed-off-by:
-
Dan Streetman authored
Change the type of the zbud_alloc() size param from unsigned int to size_t. Technically, this should not make any difference, as the zbud implementation already restricts the size to well within either type's limits; but as zsmalloc (and kmalloc) use size_t, and zpool will use size_t, this brings the size parameter type in line with zsmalloc/zpool. Signed-off-by:
-
Weijie Yang authored
Currently, we use a rwlock tb_lock to protect concurrent access to the whole zram meta table. However, according to the actual access model, there is only a small chance for upper user to access the same table[index], so the current lock granularity is too big. The idea of optimization is to change the lock granularity from whole meta table to per table entry (table -> table[index]), so that we can protect concurrent access to the same table[index], meanwhile allow the maximum concurrency. With this in mind, several kinds of locks which could be used as a per-entry lock were tested and compared: Test environment: x86-64 Intel Core2 Q8400, system memory 4GB, Ubuntu 12.04, kernel v3.15.0-rc3 as base, zram with 4 max_comp_streams LZO. iozone test: iozone -t 4 -R -r 16K -s 200M -I +Z (1GB zram with ext4 filesystem, take the average of 10 tests, KB/s) Test base CAS spinlock rwlock bit_spinlock ------------------------------------------------------------------- Initial write 1381094 1425435 1422860 1423075 1421521 Rewrite 1529479 1641199 1668762 1672855 1654910 Read 8468009 11324979 11305569 11117273 10997202 Re-read 8467476 11260914 11248059 11145336 10906486 Reverse Read 6821393 8106334 8282174 8279195 8109186 Stride read 7191093 8994306 9153982 8961224 9004434 Random read 7156353 8957932 9167098 8980465 8940476 Mixed workload 4172747 5680814 5927825 5489578 5972253 Random write 1483044 1605588 1594329 1600453 1596010 Pwrite 1276644 1303108 1311612 1314228 1300960 Pread 4324337 4632869 4618386 4457870 4500166 To enhance the possibility of access the same table[index] concurrently, set zram a small disksize(10MB) and let threads run with large loop count. fio test: fio --bs=32k --randrepeat=1 --randseed=100 --refill_buffers --scramble_buffers=1 --direct=1 --loops=3000 --numjobs=4 --filename=/dev/zram0 --name=seq-write --rw=write --stonewall --name=seq-read --rw=read --stonewall --name=seq-readwrite --rw=rw --stonewall --name=rand-readwrite --rw=randrw --stonewall (10MB zram raw block device, take the average of 10 tests, KB/s) Test base CAS spinlock rwlock bit_spinlock ------------------------------------------------------------- seq-write 933789 999357 1003298 995961 1001958 seq-read 5634130 6577930 6380861 6243912 6230006 seq-rw 1405687 1638117 1640256 1633903 1634459 rand-rw 1386119 1614664 1617211 1609267 1612471 All the optimization methods show a higher performance than the base, however, it is hard to say which method is the most appropriate. On the other hand, zram is mostly used on small embedded system, so we don't want to increase any memory footprint. This patch pick the bit_spinlock method, pack object size and page_flag into an unsigned long table.value, so as to not increase any memory overhead on both 32-bit and 64-bit system. On the third hand, even though different kinds of locks have different performances, we can ignore this difference, because: if zram is used as zram swapfile, the swap subsystem can prevent concurrent access to the same swapslot; if zram is used as zram-blk for set up filesystem on it, the upper filesystem and the page cache also prevent concurrent access of the same block mostly. So we can ignore the different performances among locks. Acked-by:Weijie Yang <weijie.yang@samsung.com> Signed-off-by:
Minchan Kim <minchan@kernel.org> Cc: Jerome Marchand <jmarchan@redhat.com> Cc: Nitin Gupta <ngupta@vflare.org> Signed-off-by:
-
Minchan Kim authored
Some architectures (eg, hexagon and PowerPC) could use PAGE_SHIFT of 16 or more. In these cases u16 is not sufficiently large to represent a compressed page's size so use size_t. Signed-off-by:
-
