- 22 Feb, 2014 40 commits
-
-
H. Peter Anvin authored
commit 03bbd596 upstream. If SMAP support is not compiled into the kernel, don't enable SMAP in CR4 -- in fact, we should clear it, because the kernel doesn't contain the proper STAC/CLAC instructions for SMAP support. Found by Fengguang Wu's test system. Reported-by:
Fengguang Wu <fengguang.wu@intel.com> Link: http://lkml.kernel.org/r/20140213124550.GA30497@localhostSigned-off-by:
H. Peter Anvin <hpa@linux.intel.com> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-
Beomho Seo authored
commit bef44abc upstream. This effects the reported scale of the raw values, and thus userspace applications that use this value. One micro tesla equal 0.01 gauss. So I have fixed calculation formula And add RAW_TO_GAUSS macro. ASA is in the range of 0 to 255. If multiply 0.003, calculation result(in_magn_[*]_scale) is always 0. So multiply 3000 and return and IIO_VAL_INT_PLUS_MICRO. As a result, read_raw call back function return accurate scale value. Signed-off-by:
Beomho Seo <beomho.seo@samsung.com> Signed-off-by:
Jonathan Cameron <jic23@kernel.org> Signed-off-by:
-
Marcus Folkesson authored
commit c76782d1 upstream. This is necessary since timestamp is calculated as the last element in iio_compute_scan_bytes(). Without this fix any userspace code reading the layout of the buffer via sysfs will incorrectly interpret the data leading some nasty corruption. Signed-off-by:
Marcus Folkesson <marcus.folkesson@gmail.com> Acked-by:
Lars-Peter Clausen <lars@metafoo.de> Signed-off-by:
-
Guenter Roeck authored
commit 55b40d37 upstream. In kernel version 3.13, devm_regulator_get() may return no error if a regulator is undeclared. regulator_get_voltage() will return -EINVAL if this happens. This causes the driver to fail loading if the vref regulator is not declared. Since vref is optional, call devm_regulator_get_optional instead. Signed-off-by:
Guenter Roeck <linux@roeck-us.net> Signed-off-by:
-
Hartmut Knaack authored
commit 38408d05 upstream. Only free an IRQ in error_free_irq, if it has been requested previously. Signed-off-by:
Hartmut Knaack <knaack.h@gmx.de> Acked-by:
-
H Hartley Sweeten authored
commit 1e85c1ea upstream. The last value written to a analog output channel is cached in the private data of this driver for readback. Currently, the wrong value is cached in the (*insn_write) functions. The current code stores the data[n] value for readback afer the loop has written all the values. At this time 'n' points past the end of the data array. Fix the functions by using a local variable to hold the data being written to the analog output channel. This variable is then used after the loop is complete to store the readback value. The current value is retrieved before the loop in case no values are actually written.. Signed-off-by:
H Hartley Sweeten <hsweeten@visionengravers.com> Reviewed-by:
Ian Abbott <abbotti@mev.co.uk> Signed-off-by:
-
Larry Finger authored
commit 08951f10 upstream. There is a typo in the device list that interchanges the vendor and product codes for one of the entries. This exchange was determined by noticing that the vendor code is 0x07b8 for Abocom at http://www.linux-usb.org/usb.ids. Signed-off-by:
Larry Finger <Larry.Finger@lwfinger.net> Signed-off-by:
-
Cédric Dufour authored
commit 8b9e418c upstream. The changes introduced in commit 4b1a25f0 ("fix build when CONFIG_UIDGID_STRICT_TYPE_CHECKS is on") got the UID check the wrong way around, leading to "Permission denied" when a regular user attempts to retrieve his quota (lfs quota -u ...) but allowing him to retrieve other users quota. Full details at: https://jira.hpdd.intel.com/browse/LU-4530 Cc: Peng Tao <tao.peng@emc.com> Signed-off-by:
Cédric Dufour <cedric.dufour@idiap.ch> Signed-off-by:
-
Bjørn Mork authored
commit f948dcf9 upstream. This device was mentioned in an OpenWRT forum. Seems to have a "standard" Sierra Wireless ifnumber to function layout: 0: qcdm 2: nmea 3: modem 8: qmi 9: storage Signed-off-by:
Bjørn Mork <bjorn@mork.no> Signed-off-by:
-
Petr Písař authored
commit 0930b095 upstream. \E[3J console code (secure clear screen) needs to update_screen(vc) in order to write-through blanks into off-screen video memory. This has been removed accidentally in 3.6 by: commit 81732c3b Author: Jean-François Moine <moinejf@free.fr> Date: Thu Sep 6 19:24:13 2012 +0200 tty vt: Fix line garbage in virtual console on command line edition Signed-off-by:
Petr Písař <petr.pisar@atlas.cz> Signed-off-by:
-
Mika Kuoppala authored
commit 1d2cb9a5 upstream. Each invocation of va_copy() must be matched by a corresponding invocation of va_end() in the same function. This regression has been introduced in commit e29bb4eb Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Sep 20 10:20:59 2013 +0100 drm/i915: Use a temporary va_list for two-pass string handling Signed-off-by:
Mika Kuoppala <mika.kuoppala@intel.com> Reviewed-by:
Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:
Daniel Vetter <daniel.vetter@ffwll.ch> Signed-off-by:
-
Christian König authored
commit b927e1c2 upstream. Otherwise decoding isn't really useable. bug: https://bugs.freedesktop.org/show_bug.cgi?id=71448Signed-off-by:
Christian König <christian.koenig@amd.com> Signed-off-by:
Alex Deucher <alexander.deucher@amd.com> Signed-off-by:
-
Alex Deucher authored
commit 858a41c8 upstream. Otherwise decoding isn't really useable. Signed-off-by:
-
Lars Poeschel authored
commit 3ac06b90 upstream. 3GPP TS 07.10 states in section 5.4.6.3.7: "The length byte contains the value 2 or 3 ... depending on the break signal." The break byte is optional and if it is sent, the length is 3. In fact the driver was not able to work with modems that send this break byte in their modem status control message. If the modem just sends the break byte if it is really set, then weird things might happen. The code for deconding the modem status to the internal linux presentation in gsm_process_modem has already a big comment about this 2 or 3 byte length thing and it is already able to decode the brk, but the code calling the gsm_process_modem function in gsm_control_modem does not encode it and hand it over the right way. This patch fixes this. Without this fix if the modem sends the brk byte in it's modem status control message the driver will hang when opening a muxed channel. Signed-off-by:
Lars Poeschel <poeschel@lemonage.de> Signed-off-by:
-
NeilBrown authored
commit 2ec197db upstream. If an NFS client attempts to get a lock (using NLM) and the lock is not available, the server will remember the request and when the lock becomes available it will send a GRANT request to the client to provide the lock. If the client already held an adjacent lock, the GRANT callback will report the union of the existing and new locks, which can confuse the client. This happens because __posix_lock_file (called by vfs_lock_file) updates the passed-in file_lock structure when adjacent or over-lapping locks are found. To avoid this problem we take a copy of the two fields that can be changed (fl_start and fl_end) before the call and restore them afterwards. An alternate would be to allocate a 'struct file_lock', initialise it, use locks_copy_lock() to take a copy, then locks_release_private() after the vfs_lock_file() call. But that is a lot more work. Reported-by:
Olaf Kirch <okir@suse.com> Signed-off-by:
NeilBrown <neilb@suse.de> Signed-off-by:
J. Bruce Fields <bfields@redhat.com> Signed-off-by:
-
Doug Anderson authored
commit d3d89c46 upstream. The ntc thermistor code was doing math whose temporary result might have overflowed 32-bits. We need some casts in there to make it safe. In one example I found: - pullup_uV: 1800000 - result of iio_read_channel_raw: 3226 - 1800000 * 3226 => 0x15a1cbc80 Signed-off-by:
Doug Anderson <dianders@chromium.org> Signed-off-by:
-
Paul Bolle authored
commit 5bbb2ae3 upstream. bind_get() checks the device number it is called with. It uses MAX_RAW_MINORS for the upper bound. But MAX_RAW_MINORS is set at compile time while the actual number of raw devices can be set at runtime. This means the test can either be too strict or too lenient. And if the test ends up being too lenient bind_get() might try to access memory beyond what was allocated for "raw_devices". So check against the runtime value (max_raw_minors) in this function. Signed-off-by:
Paul Bolle <pebolle@tiscali.nl> Acked-by:
Jan Kara <jack@suse.cz> Signed-off-by:
-
Qipan Li authored
commit fb78b811 upstream. commit 8b9ade9f coming from Viresh Kumar "tty: serial: sirfsoc: drop uart_port->lock before calling tty_flip_buffer_push()" broke sirfsoc uart driver by knic: [ 5.129122] BUG: spinlock already unlocked on CPU#0, ip6tables/1331 [ 5.132554] lock: sirfsoc_uart_ports+0x4/0x8a0, .magic: dead4ead, .owner: <none>/-1, .owner_cpu: -1 [ 5.141651] CPU: 0 PID: 1331 Comm: ip6tables Tainted: G W O 3.10.16 #3 [ 5.148866] [<c0013528>] (unwind_backtrace+0x0/0xe0) from [<c0010e70>] (show_stack+0x10/0x14) [ 5.157362] [<c0010e70>] (show_stack+0x10/0x14) from [<c01a5e68>] (do_raw_spin_unlock+0x40/0xc8) [ 5.166125] [<c01a5e68>] (do_raw_spin_unlock+0x40/0xc8) from [<c03ff8b4>] (_raw_spin_unlock+0x8/0x40) [ 5.175322] [<c03ff8b4>] (_raw_spin_unlock+0x8/0x40) from [<c0203fcc>] (sirfsoc_uart_pio_rx_chars+0xa4/0xc0) [ 5.185120] [<c0203fcc>] (sirfsoc_uart_pio_rx_chars+0xa4/0xc0) from [<c0204fb8>] (sirfsoc_rx_tmo_process_tl+0xdc/0x1e0) [ 5.195875] [<c0204fb8>] (sirfsoc_rx_tmo_process_tl+0xdc/0x1e0) from [<c0024b50>] (tasklet_action+0x8c/0xec) [ 5.205673] [<c0024b50>] (tasklet_action+0x8c/0xec) from [<c00242a8>] (__do_softirq+0xec/0x1d4) [ 5.214347] [<c00242a8>] (__do_softirq+0xec/0x1d4) from [<c0024428>] (do_softirq+0x48/0x54) [ 5.222674] [<c0024428>] (do_softirq+0x48/0x54) from [<c0024690>] (irq_exit+0x74/0xc0) [ 5.230573] [<c0024690>] (irq_exit+0x74/0xc0) from [<c000e1e8>] (handle_IRQ+0x6c/0x90) [ 5.238465] [<c000e1e8>] (handle_IRQ+0x6c/0x90) from [<c000d500>] (__irq_svc+0x40/0x70) [ 5.246446] [<c000d500>] (__irq_svc+0x40/0x70) from [<c0092e7c>] (mark_page_accessed+0xc/0x68) [ 5.255034] [<c0092e7c>] (mark_page_accessed+0xc/0x68) from [<c00a2a4c>] (unmap_single_vma+0x3bc/0x550) [ 5.264402] [<c00a2a4c>] (unmap_single_vma+0x3bc/0x550) from [<c00a3b4c>] (unmap_vmas+0x44/0x54) [ 5.273164] [<c00a3b4c>] (unmap_vmas+0x44/0x54) from [<c00a81a8>] (exit_mmap+0xc4/0x1e0) [ 5.281233] [<c00a81a8>] (exit_mmap+0xc4/0x1e0) from [<c001bb78>] (mmput+0x3c/0xdc) [ 5.288868] [<c001bb78>] (mmput+0x3c/0xdc) from [<c0021b0c>] (do_exit+0x30c/0x828) [ 5.296413] [<c0021b0c>] (do_exit+0x30c/0x828) from [<c0022dac>] (do_group_exit+0x4c/0xb0) [ 5.304653] [<c0022dac>] (do_group_exit+0x4c/0xb0) from [<c0022e20>] (__wake_up_parent+0x0/0x18) Root cause: the commit dropped uart_port->lock before calling tty_flip_buffer_push(), but in sirfsoc-uart, sirfsoc_uart_pio_rx_chars() can be called by sirfsoc_rx_tmo_process_tl(). here uart_port->lock has not been taken yet. so that caused unpaired lock/unlock. Solution: This patch is doing a quick fix for that, it adds spin_lock/unlock(&port->lock) protect to sirfsoc_uart_pio_rx_chars() in sirfsoc_rx_tmo_process_tl() to keep spin_lock/unlock in pair. Signed-off-by:
Qipan Li <Qipan.Li@csr.com> Signed-off-by:
Barry Song <Baohua.Song@csr.com> Signed-off-by:
-
Axel Lin authored
commit f7db1588 upstream. Otherwise, spi_setup() fails with unsupported mode bits message. Signed-off-by:
Axel Lin <axel.lin@ingics.com> Signed-off-by:
Mark Brown <broonie@linaro.org> Signed-off-by:
-
Kleber Sacilotto de Souza authored
commit 14e2abb7 upstream. On IBM pseries systems the device_type device-tree property of a PCIe bridge contains the string "pciex". The of_bus_pci_match() function was looking only for "pci" on this property, so in such cases the bus matching code was falling back to the default bus, causing problems on functions that should be using "assigned-addresses" for region address translation. This patch fixes the problem by also looking for "pciex" on the PCI bus match function. v2: added comment Signed-off-by:
Kleber Sacilotto de Souza <klebers@linux.vnet.ibm.com> Acked-by:
Grant Likely <grant.likely@linaro.org> Signed-off-by:
Rob Herring <robh@kernel.org> Signed-off-by:
-
Emmanuel Grumbach authored
commit 8e2a866e upstream. Not doing so will let BT kill our probe requests leading to failures in scan. Reviewed-by:
Johannes Berg <johannes.berg@intel.com> Signed-off-by:
Emmanuel Grumbach <emmanuel.grumbach@intel.com> Signed-off-by:
-
Emmanuel Grumbach authored
commit b900a87b upstream. This can be useful to be able to spot the firmware version from the error reports without needing to fetch it from another place. Signed-off-by:
-
Emmanuel Grumbach authored
commit c5128654 upstream. The driver wasn't reading the NVM properly. While this didn't lead to any issue until now, it seems that there is an old version of the NVM in the wild. In this version, the A band channels appear to be valid but the SKU capabilities (another field of the NVM) says that A band isn't supported at all. With this specific version of the NVM, the driver would think that A band is supported while the HW / firmware don't. This leads to asserts. Reviewed-by:
-
Geert Uytterhoeven authored
commit 1f802f82 upstream. This reverts commit e120cc0d. It causes a NULL pointer dereference with drivers using the generic spi_transfer_one_message(), which always calls spi_finalize_current_message(), which zeroes master->cur_msg. Drivers implementing transfer_one_message() theirselves must always call spi_finalize_current_message(), even if the transfer failed: * @transfer_one_message: the subsystem calls the driver to transfer a single * message while queuing transfers that arrive in the meantime. When the * driver is finished with this message, it must call * spi_finalize_current_message() so the subsystem can issue the next * transfer Signed-off-by:
Geert Uytterhoeven <geert+renesas@linux-m68k.org> Signed-off-by:
-
Pontus Fuchs authored
commit f12cb289 upstream. When the netlink skb is exhausted split_start is left set. In the subsequent retry, with a larger buffer, the dump is continued from the failing point instead of from the beginning. This was causing my rt28xx based USB dongle to now show up when running "iw list" with an old iw version without split dump support. Fixes: 3713b4e3 ("nl80211: allow splitting wiphy information in dumps") Signed-off-by:
Pontus Fuchs <pontus.fuchs@gmail.com> [avoid the entire workaround when state->split is set] Signed-off-by:
-
Martin Schwidefsky authored
commit 8d7f6690 upstream. The kernel currently crashes with a low-address-protection exception if a user space process executes an instruction that tries to use the linkage stack. Set the base-ASTE origin and the subspace-ASTE origin of the dispatchable-unit-control-table to point to a dummy ASTE. Set up control register 15 to point to an empty linkage stack with no room left. A user space process with a linkage stack instruction will still crash but with a different exception which is correctly translated to a segmentation fault instead of a kernel oops. Signed-off-by:
Martin Schwidefsky <schwidefsky@de.ibm.com> Signed-off-by:
-
Michael Holzheu authored
commit d7736ff5 upstream. Dumps created by kdump or zfcpdump can contain invalid memory holes when dumping z/VM systems that have memory pressure. For example: # zgetdump -i /proc/vmcore. Memory map: 0000000000000000 - 0000000000bfffff (12 MB) 0000000000e00000 - 00000000014fffff (7 MB) 000000000bd00000 - 00000000f3bfffff (3711 MB) The memory detection function find_memory_chunks() issues tprot to find valid memory chunks. In case of CMM it can happen that pages are marked as unstable via set_page_unstable() in arch_free_page(). If z/VM has released that pages, tprot returns -EFAULT and indicates a memory hole. So fix this and switch off CMM in case of kdump or zfcpdump. Signed-off-by:
Michael Holzheu <holzheu@linux.vnet.ibm.com> Signed-off-by:
-
Oleksij Rempel authored
commit 4fcfc744 upstream. Raw id and FW id should be switched. Tested-by:
Oleksij Rempel <linux@rempel-privat.de> Signed-off-by:
John W. Linville <linville@tuxdriver.com> Signed-off-by:
-
Sujith Manoharan authored
commit 8298383c upstream. Even though we make sure PowerSave is not enabled by default by disabling the flag, WIPHY_FLAG_PS_ON_BY_DEFAULT on init, PS could be enabled by userspace based on various factors like battery usage etc. Since PS in ath9k is just broken and has been untested for years, remove support for it, but allow a user to explicitly enable it using a module parameter. Signed-off-by:
Sujith Manoharan <c_manoha@qca.qualcomm.com> Signed-off-by:
-
Oleksij Rempel authored
commit 6bca610d upstream. It is a copy/paste of patch provided by Sujith for ath9k. "Even though we make sure PowerSave is not enabled by default by disabling the flag, WIPHY_FLAG_PS_ON_BY_DEFAULT on init, PS could be enabled by userspace based on various factors like battery usage etc. Since PS in ath9k is just broken and has been untested for years, remove support for it, but allow a user to explicitly enable it using a module parameter." Signed-off-by:
-
Stanislaw Gruszka authored
commit 2fa4cb90 upstream. sta_rc_update() callback must be atomic, hence we can not take mutexes or do other operations, which can sleep in ath9k_htc_sta_rc_update(). I think we can just return from ath9k_htc_sta_rc_update(), if it is called without IEEE80211_RC_SUPP_RATES_CHANGED bit. That will help with scheduling while atomic bug for most cases (except mesh and IBSS modes). For mesh and IBSS I do not see other solution like creating additional workqueue, because sending firmware command require us to sleep, but this can be done in additional patch. Patch partially fixes bug: https://bugzilla.redhat.com/show_bug.cgi?id=990955Signed-off-by:
Stanislaw Gruszka <sgruszka@redhat.com> Signed-off-by:
-
Johannes Berg authored
commit 338f977f upstream. The "new" fragmentation code (since my rewrite almost 5 years ago) erroneously sets skb->len rather than using skb_trim() to adjust the length of the first fragment after copying out all the others. This leaves the skb tail pointer pointing to after where the data originally ended, and thus causes the encryption MIC to be written at that point, rather than where it belongs: immediately after the data. The impact of this is that if software encryption is done, then a) encryption doesn't work for the first fragment, the connection becomes unusable as the first fragment will never be properly verified at the receiver, the MIC is practically guaranteed to be wrong b) we leak up to 8 bytes of plaintext (!) of the packet out into the air This is only mitigated by the fact that many devices are capable of doing encryption in hardware, in which case this can't happen as the tail pointer is irrelevant in that case. Additionally, fragmentation is not used very frequently and would normally have to be configured manually. Fix this by using skb_trim() properly. Fixes: 2de8e0d9 ("mac80211: rewrite fragmentation") Reported-by:
Jouni Malinen <j@w1.fi> Signed-off-by:
-
Emmanuel Grumbach authored
commit 0297ea17 upstream. When the driver cannot start the AP or when the assignement of the beacon goes wrong, we need to unassign the vif. Signed-off-by:
-
Eliad Peller authored
commit 2f617435 upstream. ieee80211_start_roc_work() might add a new roc to existing roc, and tell cfg80211 it has already started. However, this might happen before the roc cookie was set, resulting in REMAIN_ON_CHANNEL (started) event with null cookie. Consequently, it can make wpa_supplicant go out of sync. Fix it by setting the roc cookie earlier. Signed-off-by:
Eliad Peller <eliad@wizery.com> Signed-off-by:
-
Steve French authored
commit 83e3bc23 upstream. The get/set ACL xattr support for CIFS ACLs attempts to send old cifs dialect protocol requests even when mounted with SMB2 or later dialects. Sending cifs requests on an smb2 session causes problems - the server drops the session due to the illegal request. This patch makes CIFS ACL operations protocol specific to fix that. Attempting to query/set CIFS ACLs for SMB2 will now return EOPNOTSUPP (until we add worker routines for sending query ACL requests via SMB2) instead of sending invalid (cifs) requests. A separate followon patch will be needed to fix cifs_acl_to_fattr (which takes a cifs specific u16 fid so can't be abstracted to work with SMB2 until that is changed) and will be needed to fix mount problems when "cifsacl" is specified on mount with e.g. vers=2.1 Signed-off-by:
Steve French <smfrench@gmail.com> Reviewed-by:
Shirish Pargaonkar <spargaonkar@suse.com> Signed-off-by:
-
Steve French authored
commit d979f3b0 upstream. Changeset 666753c3 added protocol operations for get/setxattr to avoid calling cifs operations on smb2/smb3 mounts for xattr operations and this changeset adds the calls to cifs specific protocol operations for xattrs (in order to reenable cifs support for xattrs which was temporarily disabled by the previous changeset. We do not have SMB2/SMB3 worker function for setting xattrs yet so this only enables it for cifs. CCing stable since without these two small changsets (its small coreq 666753c3 is also needed) calling getfattr/setfattr on smb2/smb3 mounts causes problems. Signed-off-by:
-
Steve French authored
commit 666753c3 upstream. When mounting with smb2 (or smb2.1 or smb3) we need to check to make sure that attempts to query or set extended attributes do not attempt to send the request with the older cifs protocol instead (eventually we also need to add the support in SMB2 to query/set extended attributes but this patch prevents us from using the wrong protocol for extended attribute operations). Signed-off-by:
-
Naoya Horiguchi authored
commit 8d547ff4 upstream. mce-test detected a test failure when injecting error to a thp tail page. This is because we take page refcount of the tail page in madvise_hwpoison() while the fix in commit a3e0f9e4 ("mm/memory-failure.c: transfer page count from head page to tail page after split thp") assumes that we always take refcount on the head page. When a real memory error happens we take refcount on the head page where memory_failure() is called without MF_COUNT_INCREASED set, so it seems to me that testing memory error on thp tail page using madvise makes little sense. This patch cancels moving refcount in !MF_COUNT_INCREASED for valid testing. [akpm@linux-foundation.org: s/&&/&/] Signed-off-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com> Cc: Chen Gong <gong.chen@linux.intel.com> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by:
-
Rafael Aquini authored
commit a0b54add upstream. Changes in commit a0b8cab3 ("mm: remove lru parameter from __pagevec_lru_add and remove parts of pagevec API") have introduced a call to add_to_page_cache_lru() which causes a leak in nfs_symlink() as now the page gets an extra refcount that is not dropped. Jan Stancek observed and reported the leak effect while running test8 from Connectathon Testsuite. After several iterations over the test case, which creates several symlinks on a NFS mountpoint, the test system was quickly getting into an out-of-memory scenario. This patch fixes the page leak by dropping that extra refcount add_to_page_cache_lru() is grabbing. Signed-off-by:
Jan Stancek <jstancek@redhat.com> Signed-off-by:
Rafael Aquini <aquini@redhat.com> Acked-by:
Mel Gorman <mgorman@suse.de> Acked-by:
Rik van Riel <riel@redhat.com> Cc: Jeff Layton <jlayton@redhat.com> Cc: Trond Myklebust <trond.myklebust@primarydata.com> Signed-off-by:
-
Eric W. Biederman authored
commit 96c7a2ff upstream. Recently due to a spike in connections per second memcached on 3 separate boxes triggered the OOM killer from accept. At the time the OOM killer was triggered there was 4GB out of 36GB free in zone 1. The problem was that alloc_fdtable was allocating an order 3 page (32KiB) to hold a bitmap, and there was sufficient fragmentation that the largest page available was 8KiB. I find the logic that PAGE_ALLOC_COSTLY_ORDER can't fail pretty dubious but I do agree that order 3 allocations are very likely to succeed. There are always pathologies where order > 0 allocations can fail when there are copious amounts of free memory available. Using the pigeon hole principle it is easy to show that it requires 1 page more than 50% of the pages being free to guarantee an order 1 (8KiB) allocation will succeed, 1 page more than 75% of the pages being free to guarantee an order 2 (16KiB) allocation will succeed and 1 page more than 87.5% of the pages being free to guarantee an order 3 allocate will succeed. A server churning memory with a lot of small requests and replies like memcached is a common case that if anything can will skew the odds against large pages being available. Therefore let's not give external applications a practical way to kill linux server applications, and specify __GFP_NORETRY to the kmalloc in alloc_fdmem. Unless I am misreading the code and by the time the code reaches should_alloc_retry in __alloc_pages_slowpath (where __GFP_NORETRY becomes signification). We have already tried everything reasonable to allocate a page and the only thing left to do is wait. So not waiting and falling back to vmalloc immediately seems like the reasonable thing to do even if there wasn't a chance of triggering the OOM killer. Signed-off-by:
"Eric W. Biederman" <ebiederm@xmission.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Acked-by:
David Rientjes <rientjes@google.com> Cc: Cong Wang <cwang@twopensource.com> Signed-off-by:
-
