view Documentation/usb/ehci.txt @ 897:329ea0ccb344

balloon: try harder to balloon up under memory pressure.

Currently if the balloon driver is unable to increase the guest's
reservation it assumes the failure was due to reaching its full
allocation, gives up on the ballooning operation and records the limit
it reached as the "hard limit". The driver will not try again until
the target is set again (even to the same value).

However it is possible that ballooning has in fact failed due to
memory pressure in the host and therefore it is desirable to keep
attempting to reach the target in case memory becomes available. The
most likely scenario is that some guests are ballooning down while
others are ballooning up and therefore there is temporary memory
pressure while things stabilise. You would not expect a well behaved
toolstack to ask a domain to balloon to more than its allocation nor
would you expect it to deliberately over-commit memory by setting
balloon targets which exceed the total host memory.

This patch drops the concept of a hard limit and causes the balloon
driver to retry increasing the reservation on a timer in the same
manner as when decreasing the reservation.

Also if we partially succeed in increasing the reservation
(i.e. receive less pages than we asked for) then we may as well keep
those pages rather than returning them to Xen.

Signed-off-by: Ian Campbell <ian.campbell@citrix.com>
author Keir Fraser <keir.fraser@citrix.com>
date Fri Jun 05 14:01:20 2009 +0100 (2009-06-05)
parents 831230e53067
line source
1 27-Dec-2002
3 The EHCI driver is used to talk to high speed USB 2.0 devices using
4 USB 2.0-capable host controller hardware. The USB 2.0 standard is
5 compatible with the USB 1.1 standard. It defines three transfer speeds:
7 - "High Speed" 480 Mbit/sec (60 MByte/sec)
8 - "Full Speed" 12 Mbit/sec (1.5 MByte/sec)
9 - "Low Speed" 1.5 Mbit/sec
11 USB 1.1 only addressed full speed and low speed. High speed devices
12 can be used on USB 1.1 systems, but they slow down to USB 1.1 speeds.
14 USB 1.1 devices may also be used on USB 2.0 systems. When plugged
15 into an EHCI controller, they are given to a USB 1.1 "companion"
16 controller, which is a OHCI or UHCI controller as normally used with
17 such devices. When USB 1.1 devices plug into USB 2.0 hubs, they
18 interact with the EHCI controller through a "Transaction Translator"
19 (TT) in the hub, which turns low or full speed transactions into
20 high speed "split transactions" that don't waste transfer bandwidth.
22 At this writing, this driver has been seen to work with implementations
23 of EHCI from (in alphabetical order): Intel, NEC, Philips, and VIA.
24 Other EHCI implementations are becoming available from other vendors;
25 you should expect this driver to work with them too.
27 While usb-storage devices have been available since mid-2001 (working
28 quite speedily on the 2.4 version of this driver), hubs have only
29 been available since late 2001, and other kinds of high speed devices
30 appear to be on hold until more systems come with USB 2.0 built-in.
31 Such new systems have been available since early 2002, and became much
32 more typical in the second half of 2002.
34 Note that USB 2.0 support involves more than just EHCI. It requires
35 other changes to the Linux-USB core APIs, including the hub driver,
36 but those changes haven't needed to really change the basic "usbcore"
37 APIs exposed to USB device drivers.
39 - David Brownell
40 <dbrownell@users.sourceforge.net>
45 This driver is regularly tested on x86 hardware, and has also been
46 used on PPC hardware so big/little endianness issues should be gone.
47 It's believed to do all the right PCI magic so that I/O works even on
48 systems with interesting DMA mapping issues.
50 Transfer Types
52 At this writing the driver should comfortably handle all control, bulk,
53 and interrupt transfers, including requests to USB 1.1 devices through
54 transaction translators (TTs) in USB 2.0 hubs. But you may find bugs.
56 High Speed Isochronous (ISO) transfer support is also functional, but
57 at this writing no Linux drivers have been using that support.
59 Full Speed Isochronous transfer support, through transaction translators,
60 is not yet available. Note that split transaction support for ISO
61 transfers can't share much code with the code for high speed ISO transfers,
62 since EHCI represents these with a different data structure. So for now,
63 most USB audio and video devices can't be connected to high speed buses.
65 Driver Behavior
67 Transfers of all types can be queued. This means that control transfers
68 from a driver on one interface (or through usbfs) won't interfere with
69 ones from another driver, and that interrupt transfers can use periods
70 of one frame without risking data loss due to interrupt processing costs.
72 The EHCI root hub code hands off USB 1.1 devices to its companion
73 controller. This driver doesn't need to know anything about those
74 drivers; a OHCI or UHCI driver that works already doesn't need to change
75 just because the EHCI driver is also present.
77 There are some issues with power management; suspend/resume doesn't
78 behave quite right at the moment.
80 Also, some shortcuts have been taken with the scheduling periodic
81 transactions (interrupt and isochronous transfers). These place some
82 limits on the number of periodic transactions that can be scheduled,
83 and prevent use of polling intervals of less than one frame.
88 Assuming you have an EHCI controller (on a PCI card or motherboard)
89 and have compiled this driver as a module, load this like:
91 # modprobe ehci-hcd
93 and remove it by:
95 # rmmod ehci-hcd
97 You should also have a driver for a "companion controller", such as
98 "ohci-hcd" or "uhci-hcd". In case of any trouble with the EHCI driver,
99 remove its module and then the driver for that companion controller will
100 take over (at lower speed) all the devices that were previously handled
101 by the EHCI driver.
103 Module parameters (pass to "modprobe") include:
105 log2_irq_thresh (default 0):
106 Log2 of default interrupt delay, in microframes. The default
107 value is 0, indicating 1 microframe (125 usec). Maximum value
108 is 6, indicating 2^6 = 64 microframes. This controls how often
109 the EHCI controller can issue interrupts.
111 If you're using this driver on a 2.5 kernel, and you've enabled USB
112 debugging support, you'll see three files in the "sysfs" directory for
113 any EHCI controller:
115 "async" dumps the asynchronous schedule, used for control
116 and bulk transfers. Shows each active qh and the qtds
117 pending, usually one qtd per urb. (Look at it with
118 usb-storage doing disk I/O; watch the request queues!)
119 "periodic" dumps the periodic schedule, used for interrupt
120 and isochronous transfers. Doesn't show qtds.
121 "registers" show controller register state, and
123 The contents of those files can help identify driver problems.
126 Device drivers shouldn't care whether they're running over EHCI or not,
127 but they may want to check for "usb_device->speed == USB_SPEED_HIGH".
128 High speed devices can do things that full speed (or low speed) ones
129 can't, such as "high bandwidth" periodic (interrupt or ISO) transfers.
130 Also, some values in device descriptors (such as polling intervals for
131 periodic transfers) use different encodings when operating at high speed.
133 However, do make a point of testing device drivers through USB 2.0 hubs.
134 Those hubs report some failures, such as disconnections, differently when
135 transaction translators are in use; some drivers have been seen to behave
136 badly when they see different faults than OHCI or UHCI report.
141 USB 2.0 throughput is gated by two main factors: how fast the host
142 controller can process requests, and how fast devices can respond to
143 them. The 480 Mbit/sec "raw transfer rate" is obeyed by all devices,
144 but aggregate throughput is also affected by issues like delays between
145 individual high speed packets, driver intelligence, and of course the
146 overall system load. Latency is also a performance concern.
148 Bulk transfers are most often used where throughput is an issue. It's
149 good to keep in mind that bulk transfers are always in 512 byte packets,
150 and at most 13 of those fit into one USB 2.0 microframe. Eight USB 2.0
151 microframes fit in a USB 1.1 frame; a microframe is 1 msec/8 = 125 usec.
153 So more than 50 MByte/sec is available for bulk transfers, when both
154 hardware and device driver software allow it. Periodic transfer modes
155 (isochronous and interrupt) allow the larger packet sizes which let you
156 approach the quoted 480 MBit/sec transfer rate.
158 Hardware Performance
160 At this writing, individual USB 2.0 devices tend to max out at around
161 20 MByte/sec transfer rates. This is of course subject to change;
162 and some devices now go faster, while others go slower.
164 The first NEC implementation of EHCI seems to have a hardware bottleneck
165 at around 28 MByte/sec aggregate transfer rate. While this is clearly
166 enough for a single device at 20 MByte/sec, putting three such devices
167 onto one bus does not get you 60 MByte/sec. The issue appears to be
168 that the controller hardware won't do concurrent USB and PCI access,
169 so that it's only trying six (or maybe seven) USB transactions each
170 microframe rather than thirteen. (Seems like a reasonable trade off
171 for a product that beat all the others to market by over a year!)
173 It's expected that newer implementations will better this, throwing
174 more silicon real estate at the problem so that new motherboard chip
175 sets will get closer to that 60 MByte/sec target. That includes an
176 updated implementation from NEC, as well as other vendors' silicon.
178 There's a minimum latency of one microframe (125 usec) for the host
179 to receive interrupts from the EHCI controller indicating completion
180 of requests. That latency is tunable; there's a module option. By
181 default ehci-hcd driver uses the minimum latency, which means that if
182 you issue a control or bulk request you can often expect to learn that
183 it completed in less than 250 usec (depending on transfer size).
185 Software Performance
187 To get even 20 MByte/sec transfer rates, Linux-USB device drivers will
188 need to keep the EHCI queue full. That means issuing large requests,
189 or using bulk queuing if a series of small requests needs to be issued.
190 When drivers don't do that, their performance results will show it.
192 In typical situations, a usb_bulk_msg() loop writing out 4 KB chunks is
193 going to waste more than half the USB 2.0 bandwidth. Delays between the
194 I/O completion and the driver issuing the next request will take longer
195 than the I/O. If that same loop used 16 KB chunks, it'd be better; a
196 sequence of 128 KB chunks would waste a lot less.
198 But rather than depending on such large I/O buffers to make synchronous
199 I/O be efficient, it's better to just queue up several (bulk) requests
200 to the HC, and wait for them all to complete (or be canceled on error).
201 Such URB queuing should work with all the USB 1.1 HC drivers too.
203 In the Linux 2.5 kernels, new usb_sg_*() api calls have been defined; they
204 queue all the buffers from a scatterlist. They also use scatterlist DMA
205 mapping (which might apply an IOMMU) and IRQ reduction, all of which will
206 help make high speed transfers run as fast as they can.
209 TBD: Interrupt and ISO transfer performance issues. Those periodic
210 transfers are fully scheduled, so the main issue is likely to be how
211 to trigger "high bandwidth" modes.